id
stringlengths 10
10
| title
stringlengths 28
136
| text
stringlengths 5.62k
98.6k
| num_sections
int64 4
44
|
|---|---|---|---|
1601.02403
|
Argumentation Mining in User-Generated Web Discourse
|
# Argumentation Mining in User-Generated Web Discourse
## Abstract
The goal of argumentation mining, an evolving research field in computational linguistics, is to design methods capable of analyzing people's argumentation. In this article, we go beyond the state of the art in several ways. (i) We deal with actual Web data and take up the challenges given by the variety of registers, multiple domains, and unrestricted noisy user-generated Web discourse. (ii) We bridge the gap between normative argumentation theories and argumentation phenomena encountered in actual data by adapting an argumentation model tested in an extensive annotation study. (iii) We create a new gold standard corpus (90k tokens in 340 documents) and experiment with several machine learning methods to identify argument components. We offer the data, source codes, and annotation guidelines to the community under free licenses. Our findings show that argumentation mining in user-generated Web discourse is a feasible but challenging task.
## Introduction
The art of argumentation has been studied since the early work of Aristotle, dating back to the 4th century BC BIBREF0 . It has been exhaustively examined from different perspectives, such as philosophy, psychology, communication studies, cognitive science, formal and informal logic, linguistics, computer science, educational research, and many others. In a recent and critically well-acclaimed study, Mercier.Sperber.2011 even claim that argumentation is what drives humans to perform reasoning. From the pragmatic perspective, argumentation can be seen as a verbal activity oriented towards the realization of a goal BIBREF1 or more in detail as a verbal, social, and rational activity aimed at convincing a reasonable critic of the acceptability of a standpoint by putting forward a constellation of one or more propositions to justify this standpoint BIBREF2 .
Analyzing argumentation from the computational linguistics point of view has very recently led to a new field called argumentation mining BIBREF3 . Despite the lack of an exact definition, researchers within this field usually focus on analyzing discourse on the pragmatics level and applying a certain argumentation theory to model and analyze textual data at hand.
Our motivation for argumentation mining stems from a practical information seeking perspective from the user-generated content on the Web. For example, when users search for information in user-generated Web content to facilitate their personal decision making related to controversial topics, they lack tools to overcome the current information overload. One particular use-case example dealing with a forum post discussing private versus public schools is shown in Figure FIGREF4 . Here, the lengthy text on the left-hand side is transformed into an argument gist on the right-hand side by (i) analyzing argument components and (ii) summarizing their content. Figure FIGREF5 shows another use-case example, in which users search for reasons that underpin certain standpoint in a given controversy (which is homeschooling in this case). In general, the output of automatic argument analysis performed on the large scale in Web data can provide users with analyzed arguments to a given topic of interest, find the evidence for the given controversial standpoint, or help to reveal flaws in argumentation of others.
Satisfying the above-mentioned information needs cannot be directly tackled by current methods for, e.g., opinion mining, questions answering, or summarization and requires novel approaches within the argumentation mining field. Although user-generated Web content has already been considered in argumentation mining, many limitations and research gaps can be identified in the existing works. First, the scope of the current approaches is restricted to a particular domain or register, e.g., hotel reviews BIBREF5 , Tweets related to local riot events BIBREF6 , student essays BIBREF7 , airline passenger rights and consumer protection BIBREF8 , or renewable energy sources BIBREF9 . Second, not all the related works are tightly connected to argumentation theories, resulting into a gap between the substantial research in argumentation itself and its adaptation in NLP applications. Third, as an emerging research area, argumentation mining still suffers from a lack of labeled corpora, which is crucial for designing, training, and evaluating the algorithms. Although some works have dealt with creating new data sets, the reliability (in terms of inter-annotator agreement) of the annotated resources is often unknown BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 .
Annotating and automatically analyzing arguments in unconstrained user-generated Web discourse represent challenging tasks. So far, the research in argumentation mining “has been conducted on domains like news articles, parliamentary records and legal documents, where the documents contain well-formed explicit arguments, i.e., propositions with supporting reasons and evidence present in the text” BIBREF8 . [p. 50]Boltuzic.Snajder.2014 point out that “unlike in debates or other more formal argumentation sources, the arguments provided by the users, if any, are less formal, ambiguous, vague, implicit, or often simply poorly worded.” Another challenge stems from the different nature of argumentation theories and computational linguistics. Whereas computational linguistics is mainly descriptive, the empirical research that is carried out in argumentation theories does not constitute a test of the theoretical model that is favored, because the model of argumentation is a normative instrument for assessing the argumentation BIBREF15 . So far, no fully fledged descriptive argumentation theory based on empirical research has been developed, thus feasibility of adapting argumentation models to the Web discourse represents an open issue.
These challenges can be formulated into the following research questions:
In this article, we push the boundaries of the argumentation mining field by focusing on several novel aspects. We tackle the above-mentioned research questions as well as the previously discussed challenges and issues. First, we target user-generated Web discourse from several domains across various registers, to examine how argumentation is communicated in different contexts. Second, we bridge the gap between argumentation theories and argumentation mining through selecting the argumenation model based on research into argumentation theories and related fields in communication studies or psychology. In particular, we adapt normative models from argumentation theory to perform empirical research in NLP and support our application of argumentation theories with an in-depth reliability study. Finally, we use state-of-the-art NLP techniques in order to build robust computational models for analyzing arguments that are capable of dealing with a variety of genres on the Web.
## Our contributions
We create a new corpus which is, to the best of our knowledge, the largest corpus that has been annotated within the argumentation mining field to date. We choose several target domains from educational controversies, such as homeschooling, single-sex education, or mainstreaming. A novel aspect of the corpus is its coverage of different registers of user-generated Web content, such as comments to articles, discussion forum posts, blog posts, as well as professional newswire articles.
Since the data come from a variety of sources and no assumptions about its actual content with respect to argumentation can be drawn, we conduct two extensive annotation studies. In the first study, we tackle the problem of relatively high “noise” in the retrieved data. In particular, not all of the documents are related to the given topics in a way that makes them candidates for further deep analysis of argumentation (this study results into 990 annotated documents). In the second study, we discuss the selection of an appropriate argumentation model based on evidence in argumentation research and propose a model that is suitable for analyzing micro-level argumention in user-generated Web content. Using this model, we annotate 340 documents (approx. 90,000 tokens), reaching a substantial inter-annotator agreement. We provide a hand-analysis of all the phenomena typical to argumentation that are prevalent in our data. These findings may also serve as empirical evidence to issues that are on the spot of current argumentation research.
From the computational perspective, we experiment on the annotated data using various machine learning methods in order to extract argument structure from documents. We propose several novel feature sets and identify configurations that run best in in-domain and cross-domain scenarios. To foster research in the community, we provide the annotated data as well as all the experimental software under free license.
The rest of the article is structured as follows. First, we provide an essential background in argumentation theory in section SECREF2 . Section SECREF3 surveys related work in several areas. Then we introduce the dataset and two annotation studies in section SECREF4 . Section SECREF5 presents our experimental work and discusses the results and errors and section SECREF6 concludes this article.
## Theoretical background
Let us first present some definitions of the term argumentation itself. [p. 3]Ketcham.1917 defines argumentation as “the art of persuading others to think or act in a definite way. It includes all writing and speaking which is persuasive in form.” According to MacEwan.1898, “argumentation is the process of proving or disproving a proposition. Its purpose is to induce a new belief, to establish truth or combat error in the mind of another.” [p. 2]Freeley.Steinberg.2008 narrow the scope of argumentation to “reason giving in communicative situations by people whose purpose is the justification of acts, beliefs, attitudes, and values.” Although these definitions vary, the purpose of argumentation remains the same – to persuade others.
We would like to stress that our perception of argumentation goes beyond somehow limited giving reasons BIBREF17 , BIBREF18 . Rather, we see the goal of argumentation as to persuade BIBREF19 , BIBREF20 , BIBREF21 . Persuasion can be defined as a successful intentional effort at influencing another's mental state through communication in a circumstance in which the persuadee has some measure of freedom BIBREF22 , although, as OKeefe2011 points out, there is no correct or universally-endorsed definition of either `persuasion' or `argumentation'. However, broader understanding of argumentation as a means of persuasion allows us to take into account not only reasoned discourse, but also non-reasoned mechanisms of influence, such as emotional appeals BIBREF23 .
Having an argument as a product within the argumentation process, we should now define it. One typical definition is that an argument is a claim supported by reasons BIBREF24 . The term claim has been used since 1950's, introduced by Toulmin.1958, and in argumentation theory it is a synonym for standpoint or point of view. It refers to what is an issue in the sense what is being argued about. The presence of a standpoint is thus crucial for argumentation analysis. However, the claim as well as other parts of the argument might be implicit; this is known as enthymematic argumentation, which is rather usual in ordinary argumentative discourse BIBREF25 .
One fundamental problem with the definition and formal description of arguments and argumentation is that there is no agreement even among argumentation theorists. As [p. 29]vanEmeren.et.al.2014 admit in their very recent and exhaustive survey of the field, ”as yet, there is no unitary theory of argumentation that encompasses the logical, dialectical, and rhetorical dimensions of argumentation and is universally accepted. The current state of the art in argumentation theory is characterized by the coexistence of a variety of theoretical perspectives and approaches, which differ considerably from each other in conceptualization, scope, and theoretical refinement.”
## Argumentation models
Despite the missing consensus on the ultimate argumentation theory, various argumentation models have been proposed that capture argumentation on different levels. Argumentation models abstract from the language level to a concept level that stresses the links between the different components of an argument or how arguments relate to each other BIBREF26 . Bentahar.et.al.2010 propose a taxonomy of argumentation models, that is horizontally divided into three categories – micro-level models, macro-level models, and rhetorical models.
In this article, we deal with argumentation on the micro-level (also called argumentation as a product or monological models). Micro-level argumentation focuses on the structure of a single argument. By contrast, macro-level models (also called dialogical models) and rhetorical models highlight the process of argumentation in a dialogue BIBREF27 . In other words, we examine the structure of a single argument produced by a single author in term of its components, not the relations that can exist among arguments and their authors in time. A detailed discussion of these different perspectives can be found, e.g., in BIBREF28 , BIBREF29 , BIBREF30 , BIBREF1 , BIBREF31 , BIBREF32 .
## Dimensions of argument
The above-mentioned models focus basically only on one dimension of the argument, namely the logos dimension. According to the classical Aristotle's theory BIBREF0 , argument can exist in three dimensions, which are logos, pathos, and ethos. Logos dimension represents a proof by reason, an attempt to persuade by establishing a logical argument. For example, syllogism belongs to this argumentation dimension BIBREF34 , BIBREF25 . Pathos dimension makes use of appealing to emotions of the receiver and impacts its cognition BIBREF35 . Ethos dimension of argument relies on the credibility of the arguer. This distinction will have practical impact later in section SECREF51 which deals with argumentation on the Web.
## Original Toulmin's model
We conclude the theoretical section by presenting one (micro-level) argumentation model in detail – a widely used conceptual model of argumentation introduced by Toulmin.1958, which we will henceforth denote as the Toulmin's original model. This model will play an important role later in the annotation studies (section SECREF51 ) and experimental work (section SECREF108 ). The model consists of six parts, referred as argument components, where each component plays a distinct role.
is an assertion put forward publicly for general acceptance BIBREF38 or the conclusion we seek to establish by our arguments BIBREF17 .
It is the evidence to establish the foundation of the claim BIBREF24 or, as simply put by Toulmin, “the data represent what we have to go on.” BIBREF37 . The name of this concept was later changed to grounds in BIBREF38 .
The role of warrant is to justify a logical inference from the grounds to the claim.
is a set of information that stands behind the warrant, it assures its trustworthiness.
limits the degree of certainty under which the argument should be accepted. It is the degree of force which the grounds confer on the claim in virtue of the warrant BIBREF37 .
presents a situation in which the claim might be defeated.
A schema of the Toulmin's original model is shown in Figure FIGREF29 . The lines and arrows symbolize implicit relations between the components. An example of an argument rendered using the Toulmin's scheme can be seen in Figure FIGREF30 .
We believe that this theoretical overview should provide sufficient background for the argumentation mining research covered in this article; for further references, we recommend for example BIBREF15 .
## Related work in computational linguistics
We structure the related work into three sub-categories, namely argumentation mining, stance detection, and persuasion and on-line dialogs, as these areas are closest to this article's focus. For a recent overview of general discourse analysis see BIBREF39 . Apart from these, research on computer-supported argumentation has been also very active; see, e.g., BIBREF40 for a survey of various models and argumentation formalisms from the educational perspective or BIBREF41 which examines argumentation in the Semantic Web.
## Argumentation Mining
The argumentation mining field has been evolving very rapidly in the recent years, resulting into several workshops co-located with major NLP conferences. We first present related works with a focus on annotations and then review experiments with classifying argument components, schemes, or relations.
One of the first papers dealing with annotating argumentative discourse was Argumentative Zoning for scientific publications BIBREF42 . Later, Teufel.et.al.2009 extended the original 7 categories to 15 and annotated 39 articles from two domains, where each sentence is assigned a category. The obtained Fleiss' INLINEFORM0 was 0.71 and 0.65. In their approach, they tried to deliberately ignore the domain knowledge and rely only on general, rhetorical and logical aspect of the annotated texts. By contrast to our work, argumentative zoning is specific to scientific publications and has been developed solely for that task.
Reed.Rowe.2004 presented Araucaria, a tool for argumentation diagramming which supports both convergent and linked arguments, missing premises (enthymemes), and refutations. They also released the AracuariaDB corpus which has later been used for experiments in the argumentation mining field. However, the creation of the dataset in terms of annotation guidelines and reliability is not reported – these limitations as well as its rather small size have been identified BIBREF10 .
Biran.Rambow.2011 identified justifications for subjective claims in blog threads and Wikipedia talk pages. The data were annotated with claims and their justifications reaching INLINEFORM0 0.69, but a detailed description of the annotation approach was missing.
[p. 1078]Schneider.et.al.2013b annotated Wikipedia talk pages about deletion using 17 Walton's schemes BIBREF43 , reaching a moderate agreement (Cohen's INLINEFORM0 0.48) and concluded that their analysis technique can be reused, although “it is intensive and difficult to apply.”
Stab.Gurevych.2014 annotated 90 argumentative essays (about 30k tokens), annotating claims, major claims, and premises and their relations (support, attack). They reached Krippendorff's INLINEFORM0 0.72 for argument components and Krippendorff's INLINEFORM1 0.81 for relations between components.
Rosenthal2012 annotated sentences that are opinionated claims, in which the author expresses a belief that should be adopted by others. Two annotators labeled sentences as claims without any context and achieved Cohen's INLINEFORM0 0.50 (2,000 sentences from LiveJournal) and 0.56 (2,000 sentences from Wikipedia).
Aharoni.et.al.2014 performed an annotation study in order to find context-dependent claims and three types of context-dependent evidence in Wikipedia, that were related to 33 controversial topics. The claim and evidence were annotated in 104 articles. The average Cohen's INLINEFORM0 between a group of 20 expert annotators was 0.40. Compared to our work, the linguistic properties of Wikipedia are qualitatively different from other user-generated content, such as blogs or user comments BIBREF44 .
Wacholder.et.al.2014 annotated “argument discourse units” in blog posts and criticized the Krippendorff's INLINEFORM0 measure. They proposed a new inter-annotator metric by taking the most overlapping part of one annotation as the “core” and all annotations as a “cluster”. The data were extended by Ghosh2014, who annotated “targets” and “callouts” on the top of the units.
Park.Cardie.2014 annotated about 10k sentences from 1,047 documents into four types of argument propositions with Cohen's INLINEFORM0 0.73 on 30% of the dataset. Only 7% of the sentences were found to be non-argumentative.
Faulkner2014 used Amazon Mechanical Turk to annotate 8,179 sentences from student essays. Three annotators decided whether the given sentence offered reasons for or against the main prompt of the essay (or no reason at all; 66% of the sentences were found to be neutral and easy to identify). The achieved Cohen's INLINEFORM0 was 0.70.
The research has also been active on non-English datasets. Goudas.et.al.2014 focused on user-generated Greek texts. They selected 204 documents and manually annotated sentences that contained an argument (760 out of 16,000). They distinguished claims and premises, but the claims were always implicit. However, the annotation agreement was not reported, neither was the number of annotators or the guidelines. A study on annotation of arguments was conducted by Peldszus.Stede.2013, who evaluate agreement among 26 “naive" annotators (annotators with very little training). They manually constructed 23 German short texts, each of them contains exactly one central claim, two premises, and one objection (rebuttal or undercut) and analyzed annotator agreement on this artificial data set. Peldszus.2014 later achieved higher inter-rater agreement with expert annotators on an extended version of the same data. Kluge.2014 built a corpus of argumentative German Web documents, containing 79 documents from 7 educational topics, which were annotated by 3 annotators according to the claim-premise argumentation model. The corpus comprises 70,000 tokens and the inter-annotator agreement was 0.40 (Krippendorff's INLINEFORM0 ). Houy.et.al.2013 targeted argumentation mining of German legal cases.
Table TABREF33 gives an overview of annotation studies with their respective argumentation model, domain, size, and agreement. It also contains other studies outside of computational linguistics and few proposals and position papers.
Arguments in the legal domain were targeted in BIBREF11 . Using argumentation formalism inspired by Walton.2012, they employed multinomial Naive Bayes classifier and maximum entropy model for classifying argumentative sentences on the AraucariaDB corpus BIBREF45 . The same test dataset was used by Feng.Hirst.2011, who utilized the C4.5 decision classifier. Rooney.et.al.2012 investigated the use of convolution kernel methods for classifying whether a sentence belongs to an argumentative element or not using the same corpus.
Stab.Gurevych.2014b classified sentences to four categories (none, major claim, claim, premise) using their previously annotated corpus BIBREF7 and reached 0.72 macro- INLINEFORM0 score. In contrast to our work, their documents are expected to comply with a certain structure of argumentative essays and are assumed to always contain argumentation.
Biran.Rambow.2011 identified justifications on the sentence level using a naive Bayes classifier over a feature set based on statistics from the RST Treebank, namely n-grams which were manually processed by deleting n-grams that “seemed irrelevant, ambiguous or domain-specific.”
Llewellyn2014 experimented with classifying tweets into several argumentative categories, namely claims and counter-claims (with and without evidence) and verification inquiries previously annotated by Procter.et.al.2013. They used unigrams, punctuations, and POS as features in three classifiers.
Park.Cardie.2014 classified propositions into three classes (unverifiable, verifiable non-experimental, and verifiable experimental) and ignored non-argumentative texts. Using multi-class SVM and a wide range of features (n-grams, POS, sentiment clue words, tense, person) they achieved Macro INLINEFORM0 0.69.
Peldszus.2014 experimented with a rather complex labeling schema of argument segments, but their data were artificially created for their task and manually cleaned, such as removing segments that did not meet the criteria or non-argumentative segments.
In the first step of their two-phase approach, Goudas.et.al.2014 sampled the dataset to be balanced and identified argumentative sentences with INLINEFORM0 0.77 using the maximum entropy classifier. For identifying premises, they used BIO encoding of tokens and achieved INLINEFORM1 score 0.42 using CRFs.
Saint-Dizier.2012 developed a Prolog engine using a lexicon of 1300 words and a set of 78 hand-crafted rules with the focus on a particular argument structure “reasons supporting conclusions” in French.
Taking the dialogical perspective, Cabrio.Villata.2012 built upon an argumentation framework proposed by Dung.1995 which models arguments within a graph structure and provides a reasoning mechanism for resolving accepted arguments. For identifying support and attack, they relied on existing research on textual entailment BIBREF46 , namely using the off-the-shelf EDITS system. The test data were taken from a debate portal Debatepedia and covered 19 topics. Evaluation was performed in terms of measuring the acceptance of the “main argument" using the automatically recognized entailments, yielding INLINEFORM0 score about 0.75. By contrast to our work which deals with micro-level argumentation, the Dung's model is an abstract framework intended to model dialogical argumentation.
Finding a bridge between existing discourse research and argumentation has been targeted by several researchers. Peldszus2013a surveyed literature on argumentation and proposed utilization of Rhetorical Structure Theory (RST) BIBREF47 . They claimed that RST is by its design well-suited for studying argumentative texts, but an empirical evidence has not yet been provided. Penn Discourse Tree Bank (PDTB) BIBREF48 relations have been under examination by argumentation mining researchers too. Cabrio2013b examined a connection between five Walton's schemes and discourse markers in PDTB, however an empirical evaluation is missing.
## Stance detection
Research related to argumentation mining also involves stance detection. In this case, the whole document (discussion post, article) is assumed to represent the writer's standpoint to the discussed topic. Since the topic is stated as a controversial question, the author is either for or against it.
Somasundaran.Wiebe.2009 built a computational model for recognizing stances in dual-topic debates about named entities in the electronic products domain by combining preferences learned from the Web data and discourse markers from PDTB BIBREF48 . Hasan.Ng.2013 determined stance in on-line ideological debates on four topics using data from createdebate.com, employing supervised machine learning and features ranging from n-grams to semantic frames. Predicting stance of posts in Debatepedia as well as external articles using a probabilistic graphical model was presented in BIBREF49 . This approach also employed sentiment lexicons and Named Entity Recognition as a preprocessing step and achieved accuracy about 0.80 in binary prediction of stances in debate posts.
Recent research has involved joint modeling, taking into account information about the users, the dialog sequences, and others. Hasan.Ng.2012 proposed machine learning approach to debate stance classification by leveraging contextual information and author's stances towards the topic. Qiu.et.al.2013 introduced a computational debate side model to cluster posts or users by sides for general threaded discussions using a generative graphical model employing words from various subjectivity lexicons as well as all adjectives and adverbs in the posts. Qiu.Jiang.2013 proposed a graphical model for viewpoint discovery in discussion threads. Burfoot.et.al.2011 exploited the informal citation structure in U.S. Congressional floor-debate transcripts and use a collective classification which outperforms methods that consider documents in isolation.
Some works also utilize argumentation-motivated features. Park.et.al.2011 dealt with contentious issues in Korean newswire discourse. Although they annotate the documents with “argument frames”, the formalism remains unexplained and does not refer to any existing research in argumentation. Walker.et.al.2012b incorporated features with some limited aspects of the argument structure, such as cue words signaling rhetorical relations between posts, POS generalized dependencies, and a representation of the parent post (context) to improve stance classification over 14 topics from convinceme.net.
## Online persuasion
Another stream of research has been devoted to persuasion in online media, which we consider as a more general research topic than argumentation.
Schlosser.2011 investigated persuasiveness of online reviews and concluded that presenting two sides is not always more helpful and can even be less persuasive than presenting one side. Mohammadi.et.al.2013 explored persuasiveness of speakers in YouTube videos and concluded that people are perceived more persuasive in video than in audio and text. Miceli.et.al.2006 proposed a computational model that attempts to integrate emotional and non-emotional persuasion. In the study of Murphy.2001, persuasiveness was assigned to 21 articles (out of 100 manually preselected) and four of them are later analyzed in detail for comparing the perception of persuasion between expert and students. Bernard.et.al.2012 experimented with children's perception of discourse connectives (namely with “because”) to link statements in arguments and found out that 4- and 5-years-old and adults are sensitive to the connectives. Le.2004 presented a study of persuasive texts and argumentation in newspaper editorials in French.
A coarse-grained view on dialogs in social media was examined by Bracewell.et.al.2013, who proposed a set of 15 social acts (such as agreement, disagreement, or supportive behavior) to infer the social goals of dialog participants and presented a semi-supervised model for their classification. Their social act types were inspired by research in psychology and organizational behavior and were motivated by work in dialog understanding. They annotated a corpus in three languages using in-house annotators and achieved INLINEFORM0 in the range from 0.13 to 0.53.
Georgila.et.al.2011 focused on cross-cultural aspects of persuasion or argumentation dialogs. They developed a novel annotation scheme stemming from different literature sources on negotiation and argumentation as well as from their original analysis of the phenomena. The annotation scheme is claimed to cover three dimensions of an utterance, namely speech act, topic, and response or reference to a previous utterance. They annotated 21 dialogs and reached Krippendorff's INLINEFORM0 between 0.38 and 0.57.
Given the broad landscape of various approaches to argument analysis and persuasion studies presented in this section, we would like to stress some novel aspects of the current article. First, we aim at adapting a model of argument based on research by argumentation scholars, both theoretical and empirical. We pose several pragmatical constraints, such as register independence (generalization over several registers). Second, our emphasis is put on reliable annotations and sufficient data size (about 90k tokens). Third, we deal with fairly unrestricted Web-based sources, so additional steps of distinguishing whether the texts are argumentative are required. Argumentation mining has been a rapidly evolving field with several major venues in 2015. We encourage readers to consult an upcoming survey article by Lippi.Torroni.2016 or the proceedings of the 2nd Argumentation Mining workshop BIBREF50 to keep up with recent developments. However, to the best of our knowledge, the main findings of this article have not yet been made obsolete by any related work.
## Annotation studies and corpus creation
This section describes the process of data selection, annotation, curation, and evaluation with the goal of creating a new corpus suitable for argumentation mining research in the area of computational linguistics. As argumentation mining is an evolving discipline without established and widely-accepted annotation schemes, procedures, and evaluation, we want to keep this overview detailed to ensure full reproducibility of our approach. Given the wide range of perspectives on argumentation itself BIBREF15 , variety of argumentation models BIBREF27 , and high costs of discourse or pragmatic annotations BIBREF48 , creating a new, reliable corpus for argumentation mining represents a substantial effort.
A motivation for creating a new corpus stems from the various use-cases discussed in the introduction, as well as some research gaps pointed in section SECREF1 and further discussed in the survey in section SECREF31 (e.g., domain restrictions, missing connection to argumentation theories, non-reported reliability or detailed schemes).
## Topics and registers
As a main field of interest in the current study, we chose controversies in education. One distinguishing feature of educational topics is their breadth of sub-topics and points of view, as they attract researchers, practitioners, parents, students, or policy-makers. We assume that this diversity leads to the linguistic variability of the education topics and thus represents a challenge for NLP. In a cooperation with researchers from the German Institute for International Educational Research we identified the following current controversial topics in education in English-speaking countries: (1) homeschooling, (2) public versus private schools, (3) redshirting — intentionally delaying the entry of an age-eligible child into kindergarten, allowing their child more time to mature emotionally and physically BIBREF51 , (4) prayer in schools — whether prayer in schools should be allowed and taken as a part of education or banned completely, (5) single-sex education — single-sex classes (males and females separate) versus mixed-sex classes (“co-ed”), and (6) mainstreaming — including children with special needs into regular classes.
Since we were also interested in whether argumentation differs across registers, we included four different registers — namely (1) user comments to newswire articles or to blog posts, (2) posts in discussion forums (forum posts), (3) blog posts, and (4) newswire articles. Throughout this work, we will refer to each article, blog post, comment, or forum posts as a document. This variety of sources covers mainly user-generated content except newswire articles which are written by professionals and undergo an editing procedure by the publisher. Since many publishers also host blog-like sections on their portals, we consider as blog posts all content that is hosted on personal blogs or clearly belong to a blog category within a newswire portal.
## Raw corpus statistics
Given the six controversial topics and four different registers, we compiled a collection of plain-text documents, which we call the raw corpus. It contains 694,110 tokens in 5,444 documents. As a coarse-grained analysis of the data, we examined the lengths and the number of paragraphs (see Figure FIGREF43 ). Comments and forum posts follow a similar distribution, being shorter than 300 tokens on average. By contrast, articles and blogs are longer than 400 tokens and have 9.2 paragraphs on average. The process of compiling the raw corpus and its further statistics are described in detail in Appendix UID158 .
## Annotation study 1: Identifying persuasive documents in forums and comments
The goal of this study was to select documents suitable for a fine-grained analysis of arguments. In a preliminary study on annotating argumentation using a small sample (50 random documents) of forum posts and comments from the raw corpus, we found that many documents convey no argumentation at all, even in discussions about controversies. We observed that such contributions do not intend to persuade; these documents typically contain story-sharing, personal worries, user interaction (asking questions, expressing agreement), off-topic comments, and others. Such characteristics are typical to on-line discussions in general, but they have not been examined with respect to argumentation or persuasion. Indeed, we observed that there are (1) documents that are completely unrelated and (2) documents that are related to the topic, but do not contain any argumentation. This issue has been identified among argumentation theorist; for example as external relevance by Paglieri.Castelfranchia.2014. Similar findings were also confirmed in related literature in argumentation mining, however never tackled empirically BIBREF53 , BIBREF8 These documents are thus not suitable for analyzing argumentation.
In order to filter documents that are suitable for argumentation annotation, we defined a binary document-level classification task. The distinction is made between either persuasive documents or non-persuasive (which includes all other sorts of texts, such as off-topic, story sharing, unrelated dialog acts, etc.).
The two annotated categories were on-topic persuasive and non-persuasive. Three annotators with near-native English proficiency annotated a set of 990 documents (a random subset of comments and forum posts) reaching 0.59 Fleiss' INLINEFORM0 . The final label was selected by majority voting. The annotation study took on average of 15 hours per annotator with approximately 55 annotated documents per hour. The resulting labels were derived by majority voting. Out of 990 documents, 524 (53%) were labeled as on-topic persuasive. We will refer to this corpus as gold data persuasive.
We examined all disagreements between annotators and discovered some typical problems, such as implicitness or topic relevance. First, the authors often express their stance towards the topic implicitly, so it must be inferred by the reader. To do so, certain common-ground knowledge is required. However, such knowledge heavily depends on many aspects, such as the reader's familiarity with the topic or her cultural background, as well as the context of the source website or the discussion forum thread. This also applies for sarcasm and irony. Second, the decision whether a particular topic is persuasive was always made with respect to the controversial topic under examination. Some authors shift the focus to a particular aspect of the given controversy or a related issue, making the document less relevant.
We achieved moderate agreement between the annotators, although the definition of persuasiveness annotation might seem a bit fuzzy. We found different amounts of persuasion in the specific topics. For instance, prayer in schools or private vs. public schools attract persuasive discourse, while other discussed controversies often contain non-persuasive discussions, represented by redshirting and mainstreaming. Although these two topics are also highly controversial, the participants of on-line discussions seem to not attempt to persuade but they rather exchange information, support others in their decisions, etc. This was also confirmed by socio-psychological researchers. Ammari.et.al.2014 show that parents of children with special needs rely on discussion sites for accessing information and social support and that, in particular, posts containing humor, achievement, or treatment suggestions are perceived to be more socially appropriate than posts containing judgment, violence, or social comparisons. According to Nicholson.Leask.2012, in the online forum, parents of autistic children were seen to understand the issue because they had lived it. Assuming that participants in discussions related to young kids (e.g., redshirting, or mainstreaming) are usually females (mothers), the gender can also play a role. In a study of online persuasion, Guadagno.Cialdini.2002 conclude that women chose to bond rather than compete (women feel more comfortable cooperating, even in a competitive environment), whereas men are motivated to compete if necessary to achieve independence.
## Annotation study 2: Annotating micro-structure of arguments
The goal of this study was to annotate documents on a detailed level with respect to an argumentation model. First, we will present the annotation scheme. Second, we will describe the annotation process. Finally, we will evaluate the agreement and draw some conclusions.
Given the theoretical background briefly introduced in section SECREF2 , we motivate our selection of the argumentation model by the following requirements. First, the scope of this work is to capture argumentation within a single document, thus focusing on micro-level models. Second, there should exist empirical evidence that such a model has been used for analyzing argumentation in previous works, so it is likely to be suitable for our purposes of argumentative discourse analysis in user-generated content. Regarding the first requirement, two typical examples of micro-level models are the Toulmin's model BIBREF36 and Walton's schemes BIBREF55 . Let us now elaborate on the second requirement.
Walton's argumentation schemes are claimed to be general and domain independent. Nevertheless, evidence from the computational linguistics field shows that the schemes lack coverage for analyzing real argumentation in natural language texts. In examining real-world political argumentation from BIBREF56 , Walton.2012 found out that 37.1% of the arguments collected did not fit any of the fourteen schemes they chose so they created new schemes ad-hoc. Cabrio2013b selected five argumentation schemes from Walton and map these patterns to discourse relation categories in the Penn Discourse TreeBank (PDTB) BIBREF48 , but later they had to define two new argumentation schemes that they discovered in PDTB. Similarly, Song.et.al.2014 admitted that the schemes are ambiguous and hard to directly apply for annotation, therefore they modified the schemes and created new ones that matched the data.
Although Macagno.Konstantinidou.2012 show several examples of two argumentation schemes applied to few selected arguments in classroom experiments, empirical evidence presented by Anthony.Kim.2014 reveals many practical and theoretical difficulties of annotating dialogues with schemes in classroom deliberation, providing many details on the arbitrary selection of the sub-set of the schemes, the ambiguity of the scheme definitions, concluding that the presence of the authors during the experiment was essential for inferring and identifying the argument schemes BIBREF57 .
Although this model (refer to section SECREF21 ) was designed to be applicable to real-life argumentation, there are numerous studies criticizing both the clarity of the model definition and the differentiation between elements of the model. Ball1994 claims that the model can be used only for the most simple arguments and fails on the complex ones. Also Freeman1991 and other argumentation theorists criticize the usefulness of Toulmin's framework for the description of real-life argumentative texts. However, others have advocated the model and claimed that it can be applied to the people's ordinary argumenation BIBREF58 , BIBREF59 .
A number of studies (outside the field of computational linguistics) used Toulmin's model as their backbone argumentation framework. Chambliss1995 experimented with analyzing 20 written documents in a classroom setting in order to find the argument patterns and parts. Simosi2003 examined employees' argumentation to resolve conflicts. Voss2006 analyzed experts' protocols dealing with problem-solving.
The model has also been used in research on computer-supported collaborative learning. Erduran2004 adapt Toulmin's model for coding classroom argumentative discourse among teachers and students. Stegmann2011 builds on a simplified Toulmin's model for scripted construction of argument in computer-supported collaborative learning. Garcia-Mila2013 coded utterances into categories from Toulmin's model in persuasion and consensus-reaching among students. Weinberger.Fischer.2006 analyze asynchronous discussion boards in which learners engage in an argumentative discourse with the goal to acquire knowledge. For coding the argument dimension, they created a set of argumentative moves based on Toulmin's model. Given this empirical evidence, we decided to build upon the Toulmin's model.
In this annotation task, a sequence of tokens (e.g. a phrase, a sentence, or any arbitrary text span) is labeled with a corresponding argument component (such as the claim, the grounds, and others). There are no explicit relations between these annotation spans as the relations are implicitly encoded in the pragmatic function of the components in the Toulmin's model.
In order to prove the suitability of the Toulmin's model, we analyzed 40 random documents from the gold data persuasive dataset using the original Toulmin's model as presented in section SECREF21 . We took into account sever criteria for assessment, such as frequency of occurrence of the components or their importance for the task. We proposed some modifications of the model based on the following observations.
Authors do not state the degree of cogency (the probability of their claim, as proposed by Toulmin). Thus we omitted qualifier from the model due to its absence in the data.
The warrant as a logical explanation why one should accept the claim given the evidence is almost never stated. As pointed out by BIBREF37 , “data are appealed to explicitly, warrants implicitly.” This observation has also been made by Voss2006. Also, according to [p. 205]Eemeren.et.al.1987, the distinction of warrant is perfectly clear only in Toulmin’s examples, but the definitions fail in practice. We omitted warrant from the model.
Rebuttal is a statement that attacks the claim, thus playing a role of an opposing view. In reality, the authors often attack the presented rebuttals by another counter-rebuttal in order to keep the whole argument's position consistent. Thus we introduced a new component – refutation – which is used for attacking the rebuttal. Annotation of refutation was conditioned of explicit presence of rebuttal and enforced by the annotation guidelines. The chain rebuttal–refutation is also known as the procatalepsis figure in rhetoric, in which the speaker raises an objection to his own argument and then immediately answers it. By doing so, the speaker hopes to strengthen the argument by dealing with possible counter-arguments before the audience can raise them BIBREF43 .
The claim of the argument should always reflect the main standpoint with respect to the discussed controversy. We observed that this standpoint is not always explicitly expressed, but remains implicit and must be inferred by the reader. Therefore, we allow the claim to be implicit. In such a case, the annotators must explicitly write down the (inferred) stance of the author.
By definition, the Toulmin's model is intended to model single argument, with the claim in its center. However, we observed in our data, that some authors elaborate on both sides of the controversy equally and put forward an argument for each side (by argument here we mean the claim and its premises, backings, etc.). Therefore we allow multiple arguments to be annotated in one document. At the same time, we restrained the annotators from creating complex argument hierarchies.
Toulmin's grounds have an equivalent role to a premise in the classical view on an argument BIBREF15 , BIBREF60 in terms that they offer the reasons why one should accept the standpoint expressed by the claim. As this terminology has been used in several related works in the argumentation mining field BIBREF7 , BIBREF61 , BIBREF62 , BIBREF11 , we will keep this convention and denote the grounds as premises.
One of the main critiques of the original Toulmin's model was the vague distinction between grounds, warrant, and backing BIBREF63 , BIBREF64 , BIBREF65 . The role of backing is to give additional support to the warrant, but there is no warrant in our model anymore. However, what we observed during the analysis, was a presence of some additional evidence. Such evidence does not play the role of the grounds (premises) as it is not meant as a reason supporting the claim, but it also does not explain the reasoning, thus is not a warrant either. It usually supports the whole argument and is stated by the author as a certain fact. Therefore, we extended the scope of backing as an additional support to the whole argument.
The annotators were instructed to distinguish between premises and backing, so that premises should cover generally applicable reasons for the claim, whereas backing is a single personal experience or statements that give credibility or attribute certain expertise to the author. As a sanity check, the argument should still make sense after removing backing (would be only considered “weaker”).
We call the model as a modified Toulmin's model. It contains five argument components, namely claim, premise, backing, rebuttal, and refutation. When annotating a document, any arbitrary token span can be labeled with an argument component; the components do not overlap. The spans are not known in advance and the annotator thus chooses the span and the component type at the same time. All components are optional (they do not have to be present in the argument) except the claim, which is either explicit or implicit (see above). If a token span is not labeled by any argument component, it is not considered as a part of the argument and is later denoted as none (this category is not assigned by the annotators).
An example analysis of a forum post is shown in Figure FIGREF65 . Figure FIGREF66 then shows a diagram of the analysis from that example (the content of the argument components was shortened or rephrased).
The annotation experiment was split into three phases. All documents were annotated by three independent annotators, who participated in two training sessions. During the first phase, 50 random comments and forum posts were annotated. Problematic cases were resolved after discussion and the guidelines were refined. In the second phase, we wanted to extend the range of annotated registers, so we selected 148 comments and forum posts as well as 41 blog posts. After the second phase, the annotation guidelines were final.
In the final phase, we extended the range of annotated registers and added newswire articles from the raw corpus in order to test whether the annotation guidelines (and inherently the model) is general enough. Therefore we selected 96 comments/forum posts, 8 blog posts, and 8 articles for this phase. A detailed inter-annotator agreement study on documents from this final phase will be reported in section UID75 .
The annotations were very time-consuming. In total, each annotator spent 35 hours by annotating in the course of five weeks. Discussions and consolidation of the gold data took another 6 hours. Comments and forum posts required on average of 4 minutes per document to annotate, while blog posts and articles on average of 14 minutes per document. Examples of annotated documents from the gold data are listed in Appendix UID158 .
We discarded 11 documents out of the total 351 annotated documents. Five forum posts, although annotated as persuasive in the first annotation study, were at a deeper look a mixture of two or more posts with missing quotations, therefore unsuitable for analyzing argumentation. Three blog posts and two articles were found not to be argumentative (the authors took no stance to the discussed controversy) and one article was an interview, which the current model cannot capture (a dialogical argumentation model would be required).
For each of the 340 documents, the gold standard annotations were obtained using the majority vote. If simple majority voting was not possible (different boundaries of the argument component together with a different component label), the gold standard was set after discussion among the annotators. We will refer to this corpus as the gold standard Toulmin corpus. The distribution of topics and registers in this corpus in shown in Table TABREF71 , and Table TABREF72 presents some lexical statistics.
Based on pre-studies, we set the minimal unit for annotation as token. The documents were pre-segmented using the Stanford Core NLP sentence splitter BIBREF69 embedded in the DKPro Core framework BIBREF70 . Annotators were asked to stick to the sentence level by default and label entire pre-segmented sentences. They should switch to annotations on the token level only if (a) a particular sentence contained more than one argument component, or (b) if the automatic sentence segmentation was wrong. Given the “noise” in user-generated Web data (wrong or missing punctuation, casing, etc.), this was often the case.
Annotators were also asked to rephrase (summarize) each annotated argument component into a simple statement when applicable, as shown in Figure FIGREF66 . This was used as a first sanity checking step, as each argument component is expected to be a coherent discourse unit. For example, if a particular occurrence of a premise cannot be summarized/rephrased into one statement, this may require further splitting into two or more premises.
For the actual annotations, we developed a custom-made web-based application that allowed users to switch between different granularity of argument components (tokens or sentences), to annotate the same document in different argument “dimensions” (logos and pathos), and to write summary for each annotated argument component.
As a measure of annotation reliability, we rely on Krippendorff's unitized alpha ( INLINEFORM0 ) BIBREF71 . To the best of our knowledge, this is the only agreement measure that is applicable when both labels and boundaries of segments are to be annotated.
Although the measure has been used in related annotation works BIBREF61 , BIBREF7 , BIBREF72 , there is one important detail that has not been properly communicated. The INLINEFORM0 is computed over a continuum of the smallest units, such as tokens. This continuum corresponds to a single document in the original Krippendorff's work. However, there are two possible extensions to multiple documents (a corpus), namely (a) to compute INLINEFORM1 for each document first and then report an average value, or (b) to concatenate all documents into one large continuum and compute INLINEFORM2 over it. The first approach with averaging yielded extremely high the standard deviation of INLINEFORM3 (i.e., avg. = 0.253; std. dev. = 0.886; median = 0.476 for the claim). This says that some documents are easy to annotate while others are harder, but interpretation of such averaged value has no evidence either in BIBREF71 or other papers based upon it. Thus we use the other methodology and treat the whole corpus as a single long continuum (which yields in the example of claim 0.541 INLINEFORM4 ).
Table TABREF77 shows the inter-annotator agreement as measured on documents from the last annotation phase (see section UID67 ). The overall INLINEFORM0 for all register types, topics, and argument components is 0.48 in the logos dimension (annotated with the modified Toulmin's model). Such agreement can be considered as moderate by the measures proposed by Landis.Koch.1977, however, direct interpretation of the agreement value lacks consensus BIBREF54 . Similar inter-annotator agreement numbers were achieved in the relevant works in argumentation mining (refer to Table TABREF33 in section SECREF31 ; although most of the numbers are not directly comparable, as different inter-annotator metrics were used on different tasks).
There is a huge difference in INLINEFORM0 regarding the registers between comments + forums posts ( INLINEFORM1 0.60, Table TABREF77 a) and articles + blog posts ( INLINEFORM2 0.09, Table TABREF77 b) in the logos dimension. If we break down the value with respect to the individual argument components, the agreement on claim and premise is substantial in the case of comments and forum posts (0.59 and 0.69, respectively). By contrast, these argument components were annotated only with a fair agreement in articles and blog posts (0.22 and 0.24, respectively).
As can be also observed from Table TABREF77 , the annotation agreement in the logos dimension varies regarding the document topic. While it is substantial/moderate for prayer in schools (0.68) or private vs. public schools (0.44), for some topics it remains rather slight, such as in the case of redshirting (0.14) or mainstreaming (0.08).
First, we examine the disagreement in annotations by posing the following research question: are there any measurable properties of the annotated documents that might systematically cause low inter-annotator agreement? We use Pearson's correlation coefficient between INLINEFORM0 on each document and the particular property under investigation. We investigated the following set of measures.
Full sentence coverage ratio represents a ratio of argument component boundaries that are aligned to sentence boundaries. The value is 1.0 if all annotations in the particular document are aligned to sentences and 0.0 if no annotations match the sentence boundaries. Our hypothesis was that automatic segmentation to sentences was often incorrect, therefore annotators had to switch to the token level annotations and this might have increased disagreement on boundaries of the argument components.
Document length, paragraph length and average sentence length. Our hypotheses was that the length of documents, paragraphs, or sentences negatively affects the agreement.
Readability measures. We tested four standard readability measures, namely Ari BIBREF73 , Coleman-Liau BIBREF74 , Flesch BIBREF75 , and Lix BIBREF76 to find out whether readability of the documents plays any role in annotation agreement.
Correlation results are listed in Table TABREF82 . We observed the following statistically significant ( INLINEFORM0 ) correlations. First, document length negatively correlates with agreement in comments. The longer the comment was the lower the agreement was. Second, average paragraph length negatively correlates with agreement in blog posts. The longer the paragraphs in blogs were, the lower agreement was reached. Third, all readability scores negatively correlate with agreement in the public vs. private school domain, meaning that the more complicated the text in terms of readability is, the lower agreement was reached. We observed no significant correlation in sentence coverage and average sentence length measures. We cannot draw any general conclusion from these results, but we can state that some registers and topics, given their properties, are more challenging to annotate than others.
Another qualitative analysis of disagreements between annotators was performed by constructing a probabilistic confusion matrix BIBREF77 on the token level. The biggest disagreements, as can be seen in Table TABREF85 , is caused by rebuttal and refutation confused with none (0.27 and 0.40, respectively). This is another sign that these two argument components were very hard to annotate. As shown in Table TABREF77 , the INLINEFORM5 was also low – 0.08 for rebuttal and 0.17 for refutation.
We analyzed the annotations and found the following phenomena that usually caused disagreements between annotators.
Each argument component (e.g., premise or backing) should express one consistent and coherent piece of information, for example a single reason in case of the premise (see Section UID73 ). However, the decision whether a longer text should be kept as a single argument component or segmented into multiple components is subjective and highly text-specific.
While rhetorical questions have been researched extensively in linguistics BIBREF78 , BIBREF79 , BIBREF80 , BIBREF81 , their role in argumentation represents a substantial research question BIBREF82 , BIBREF83 , BIBREF84 , BIBREF85 , BIBREF86 . Teninbaum.2011 provides a brief history of rhetorical questions in persuasion. In short, rhetorical questions should provoke the reader. From the perspective of our argumentation model, rhetorical questions might fall both into the logos dimension (and thus be labeled as, e.g., claim, premise, etc.) or into the pathos dimension (refer to Section SECREF20 ). Again, the decision is usually not clear-cut.
As introduced in section UID55 , rebuttal attacks the claim by presenting an opponent's view. In most cases, the rebuttal is again attacked by the author using refutation. From the pragmatic perspective, refutation thus supports the author's stance expressed by the claim. Therefore, it can be easily confused with premises, as the function of both is to provide support for the claim. Refutation thus only takes place if it is meant as a reaction to the rebuttal. It follows the discussed matter and contradicts it. Such a discourse is usually expressed as:
[claim: My claim.] [rebuttal: On the other hand, some people claim XXX which makes my claim wrong.] [refutation: But this is not true, because of YYY.]
However, the author might also take the following defensible approach to formulate the argument:
[rebuttal: Some people claim XXX-1 which makes my claim wrong.] [refutation: But this is not true, because of YYY-1.] [rebuttal: Some people claim XXX-2 which makes my claim wrong.] [refutation: But this is not true, because of YYY-2.] [claim: Therefore my claim.]
If this argument is formulated without stating the rebuttals, it would be equivalent to the following:
[premise: YYY-1.] [premise: YYY-2.] [claim: Therefore my claim.]
This example shows that rebuttal and refutation represent a rhetorical device to produce arguments, but the distinction between refutation and premise is context-dependent and on the functional level both premise and refutation have very similar role – to support the author's standpoint. Although introducing dialogical moves into monological model and its practical consequences, as described above, can be seen as a shortcoming of our model, this rhetoric figure has been identified by argumentation researchers as procatalepsis BIBREF43 . A broader view on incorporating opposing views (or lack thereof) is discussed under the term confirmation bias by BIBREF21 who claim that “[...] people are trying to convince others. They are typically looking for arguments and evidence to confirm their own claim, and ignoring negative arguments and evidence unless they anticipate having to rebut them.” The dialectical attack of possible counter-arguments may thus strengthen one's own argument.
One possible solution would be to refrain from capturing this phenomena completely and to simplify the model to claims and premises, for instance. However, the following example would then miss an important piece of information, as the last two clauses would be left un-annotated. At the same time, annotating the last clause as premise would be misleading, because it does not support the claim (in fact, it supports it only indirectly by attacking the rebuttal; this can be seen as a support is considered as an admissible extension of abstract argument graph by BIBREF87 ).
Doc#422 (forumpost, homeschooling) [claim: I try not to be anti-homeschooling, but... it's just hard for me.] [premise: I really haven't met any homeschoolers who turned out quite right, including myself.] I apologize if what I'm saying offends any of you - that's not my intention, [rebuttal: I know that there are many homeschooled children who do just fine,] but [refutation: that hasn't been my experience.]
To the best of our knowledge, these context-dependent dialogical properties of argument components using Toulmin's model have not been solved in the literature on argumentation theory and we suggest that these observations should be taken into account in the future research in monological argumentation.
Appeal to emotion, sarcasm, irony, or jokes are common in argumentation in user-generated Web content. We also observed documents in our data that were purely sarcastic (the pathos dimension), therefore logical analysis of the argument (the logos dimension) would make no sense. However, given the structure of such documents, some claims or premises might be also identified. Such an argument is a typical example of fallacious argumentation, which intentionally pretends to present a valid argument, but its persuasion is conveyed purely for example by appealing to emotions of the reader BIBREF88 .
We present some statistics of the annotated data that are important from the argumentation research perspective. Regardless of the register, 48% of claims are implicit. This means that the authors assume that their standpoint towards the discussed controversy can be inferred by the reader and give only reasons for that standpoint. Also, explicit claims are mainly written just once, only in 3% of the documents the claim was rephrased and occurred multiple times.
In 6% of the documents, the reasons for an implicit claim are given only in the pathos dimension, making the argument purely persuasive without logical argumentation.
The “myside bias”, defined as a bias against information supporting another side of an argument BIBREF89 , BIBREF90 , can be observed by the presence of rebuttals to the author's claim or by formulating arguments for both sides when the overall stance is neutral. While 85% of the documents do not consider any opposing side, only 8% documents present a rebuttal, which is then attacked by refutation in 4% of the documents. Multiple rebuttals and refutations were found in 3% of the documents. Only 4% of the documents were overall neutral and presented arguments for both sides, mainly in blog posts.
We were also interested whether mitigating linguistic devices are employed in the annotated arguments, namely in their main stance-taking components, the claims. Such devices typically include parenthetical verbs, syntactic constructions, token agreements, hedges, challenge questions, discourse markers, and tag questions, among others BIBREF91 . In particular, [p. 1]Kaltenbock.et.al.2010 define hedging as a discourse strategy that reduces the force or truth of an utterance and thus reduces the risk a speaker runs when uttering a strong or firm assertion or other speech act. We manually examined the use of hedging in the annotated claims.
Our main observation is that hedging is used differently across topics. For instance, about 30-35% of claims in homeschooling and mainstreaming signal the lack of a full commitment to the expressed stance, in contrast to prayer in schools (15%) or public vs. private schools (about 10%). Typical hedging cues include speculations and modality (“If I have kids, I will probably homeschool them.”), statements as neutral observations (“It's not wrong to hold the opinion that in general it's better for kids to go to school than to be homeschooled.”), or weasel phrases BIBREF92 (“In some cases, inclusion can work fantastically well.”, “For the majority of the children in the school, mainstream would not have been a suitable placement.”).
On the other hand, most claims that are used for instance in the prayer in schools arguments are very direct, without trying to diminish its commitment to the conveyed belief (for example, “NO PRAYER IN SCHOOLS!... period.”, “Get it out of public schools”, “Pray at home.”, or “No organized prayers or services anywhere on public school board property - FOR ANYONE.”). Moreover, some claims are clearly offensive, persuading by direct imperative clauses towards the opponents/audience (“TAKE YOUR KIDS PRIVATE IF YOU CARE AS I DID”, “Run, don't walk, to the nearest private school.”) or even accuse the opponents for taking a certain stance (“You are a bad person if you send your children to private school.”).
These observations are consistent with the findings from the first annotation study on persuasion (see section UID48 ), namely that some topics attract heated argumentation where participant take very clear and reserved standpoints (such as prayer in schools or private vs. public schools), while discussions about other topics are rather milder. It has been shown that the choices a speaker makes to express a position are informed by their social and cultural background, as well as their ability to speak the language BIBREF93 , BIBREF94 , BIBREF91 . However, given the uncontrolled settings of the user-generated Web content, we cannot infer any similar conclusions in this respect.
We investigated premises across all topics in order to find the type of support used in the argument. We followed the approach of Park.Cardie.2014, who distinguished three types of propositions in their study, namely unverifiable, verifiable non-experiential, and verifiable experiential.
Verifiable non-experiential and verifiable experiential propositions, unlike unverifiable propositions, contain an objective assertion, where objective means “expressing or dealing with facts or conditions as perceived without distortion by personal feelings, prejudices, or interpretations.” Such assertions have truth values that can be proved or disproved with objective evidence; the correctness of the assertion or the availability of the objective evidence does not matter BIBREF8 . A verifiable proposition can further be distinguished as experiential or not, depending on whether the proposition is about the writer's personal state or experience or something non-experiential. Verifiable experiential propositions are sometimes referred to as anectotal evidence, provide the novel knowledge that readers are seeking BIBREF8 .
Table TABREF97 shows the distribution of the premise types with examples for each topic from the annotated corpus. As can be seen in the first row, arguments in prayer in schools contain majority (73%) of unverifiable premises. Closer examination reveals that their content vary from general vague propositions to obvious fallacies, such as a hasty generalization, straw men, or slippery slope. As Nieminen.Mustonen.2014 found out, fallacies are very common in argumentation about religion-related issues. On the other side of the spectrum, arguments about redshirting rely mostly on anecdotal evidence (61% of verifiable experiential propositions). We will discuss the phenomena of narratives in argumentation in more detail later in section UID98 . All the topics except private vs. public schools exhibit similar amount of verifiable non-experiential premises (9%–22%), usually referring to expert studies or facts. However, this type of premises has usually the lowest frequency.
Manually analyzing argumentative discourse and reconstructing (annotating) the underlying argument structure and its components is difficult. As [p. 267]Reed2006 point out, “the analysis of arguments is often hard, not only for students, but for experts too.” According to [p. 81]Harrell.2011b, argumentation is a skill and “even for simple arguments, untrained college students can identify the conclusion but without prompting are poor at both identifying the premises and how the premises support the conclusion.” [p. 81]Harrell.2011 further claims that “a wide literature supports the contention that the particular skills of understanding, evaluating, and producing arguments are generally poor in the population of people who have not had specific training and that specific training is what improves these skills.” Some studies, for example, show that students perform significantly better on reasoning tasks when they have learned to identify premises and conclusions BIBREF95 or have learned some standard argumentation norms BIBREF96 .
One particular extra challenge in analyzing argumentation in Web user-generated discourse is that the authors produce their texts probably without any existing argumentation theory or model in mind. We assume that argumentation or persuasion is inherent when users discuss controversial topics, but the true reasons why people participate in on-line communities and what drives their behavior is another research question BIBREF97 , BIBREF98 , BIBREF99 , BIBREF100 . When the analyzed texts have a clear intention to produce argumentative discourse, such as in argumentative essays BIBREF7 , the argumentation is much more explicit and a substantially higher inter-annotator agreement can be achieved.
The model seems to be suitable for short persuasive documents, such as comments and forum posts. Its applicability to longer documents, such as articles or blog posts, is problematic for several reasons.
The argument components of the (modified) Toulmin's model and their roles are not expressive enough to capture argumentation that not only conveys the logical structure (in terms of reasons put forward to support the claim), but also relies heavily on the rhetorical power. This involves various stylistic devices, pervading narratives, direct and indirect speech, or interviews. While in some cases the argument components are easily recognizable, the vast majority of the discourse in articles and blog posts does not correspond to any distinguishable argumentative function in the logos dimension. As the purpose of such discourse relates more to rhetoric than to argumentation, unambiguous analysis of such phenomena goes beyond capabilities of the current argumentation model. For a discussion about metaphors in Toulmin's model of argumentation see, e.g., BIBREF102 , BIBREF103 .
Articles without a clear standpoint towards the discussed controversy cannot be easily annotated with the model either. Although the matter is viewed from both sides and there might be reasons presented for either of them, the overall persuasive intention is missing and fitting such data to the argumentation framework causes disagreements. One solution might be to break the document down to paragraphs and annotate each paragraph separately, examining argumentation on a different level of granularity.
As introduced in section SECREF20 , there are several dimensions of an argument. The Toulmin's model focuses solely on the logos dimension. We decided to ignore the ethos dimension, because dealing with the author's credibility remains unclear, given the variety of the source web data. However, exploiting the pathos dimension of an argument is prevalent in the web data, for example as an appeal to emotions. Therefore we experimented with annotating appeal to emotions as a separate category independent of components in the logos dimension. We defined some features for the annotators how to distinguish appeal to emotions. Figurative language such as hyperbole, sarcasm, or obvious exaggerating to “spice up” the argument are the typical signs of pathos. In an extreme case, the whole argument might be purely emotional, as in the following example.
Doc#1698 (comment, prayer in schools) [app-to-emot: Prayer being removed from school is just the leading indicator of a nation that is ‘Falling Away’ from Jehovah. [...] And the disasters we see today are simply God’s finger writing on the wall: Mene, mene, Tekel, Upharsin; that is, God has weighed America in the balances, and we’ve been found wanting. No wonder 50 million babies have been aborted since 1973. [...]]
We kept annotations on the pathos dimension as simple as possible (with only one appeal to emotions label), but the resulting agreement was unsatisfying ( INLINEFORM0 0.30) even after several annotation iterations. Appeal to emotions is considered as a type of fallacy BIBREF104 , BIBREF18 . Given the results, we assume that more carefully designed approach to fallacy annotation should be applied. To the best of our knowledge, there have been very few research works on modeling fallacies similarly to arguments on the discourse level BIBREF105 . Therefore the question, in which detail and structure fallacies should be annotated, remains open. For the rest of the paper, we thus focus on the logos dimension solely.
Some of the educational topics under examination relate to young children (e.g., redshirting or mainstreaming); therefore we assume that the majority of participants in discussions are their parents. We observed that many documents related to these topics contain narratives. Sometimes the story telling is meant as a support for the argument, but there are documents where the narrative has no intention to persuade and is simply a story sharing.
There is no widely accepted theory of the role of narratives among argumentation scholars. According to Fisher.1987, humans are storytellers by nature, and the “reason” in argumentation is therefore better understood in and through the narratives. He found that good reasons often take the form of narratives. Hoeken.Fikkers.2014 investigated how integration of explicit argumentative content into narratives influences issue-relevant thinking and concluded that identifying with the character being in favor of the issue yielded a more positive attitude toward the issue. In a recent research, Bex.2011 proposes an argumentative-narrative model of reasoning with evidence, further elaborated in BIBREF106 ; also Niehaus.et.al.2012 proposes a computational model of narrative persuasion.
Stemming from another research field, LeytonEscobar2014 found that online community members who use and share narratives have higher participation levels and that narratives are useful tools to build cohesive cultures and increase participation. Betsch.et.al.2010 examined influencing vaccine intentions among parents and found that narratives carry more weight than statistics.
## Summary of annotation studies
This section described two annotation studies that deal with argumentation in user-generated Web content on different levels of detail. In section SECREF44 , we argued for a need of document-level distinction of persuasiveness. We annotated 990 comments and forum posts, reaching moderate inter-annotator agreement (Fleiss' INLINEFORM0 0.59). Section SECREF51 motivated the selection of a model for micro-level argument annotation, proposed its extension based on pre-study observations, and outlined the annotation set-up. This annotation study resulted into 340 documents annotated with the modified Toulmin's model and reached moderate inter-annotator agreement in the logos dimension (Krippendorff's INLINEFORM1 0.48). These results make the annotated corpora suitable for training and evaluation computational models and each of these two annotation studies will have their experimental counterparts in the following section.
## Experiments
This section presents experiments conducted on the annotated corpora introduced in section SECREF4 . We put the main focus on identifying argument components in the discourse. To comply with the machine learning terminology, in this section we will use the term domain as an equivalent to a topic (remember that our dataset includes six different topics; see section SECREF38 ).
We evaluate three different scenarios. First, we report ten-fold cross validation over a random ordering of the entire data set. Second, we deal with in-domain ten-fold cross validation for each of the six domains. Third, in order to evaluate the domain portability of our approach, we train the system on five domains and test on the remaining one for all six domains (which we report as cross-domain validation).
## Identification of argument components
In the following experiment, we focus on automatic identification of arguments in the discourse. Our approach is based on supervised and semi-supervised machine learning methods on the gold data Toulmin dataset introduced in section SECREF51 .
An argument consists of different components (such as premises, backing, etc.) which are implicitly linked to the claim. In principle one document can contain multiple independent arguments. However, only 4% of the documents in our dataset contain arguments for both sides of the issue. Thus we simplify the task and assume there is only one argument per document.
Given the low inter-annotator agreement on the pathos dimension (Table TABREF77 ), we focus solely on recognizing the logical dimension of argument. The pathos dimension of argument remains an open problem for a proper modeling as well as its later recognition.
Since the smallest annotation unit is a token and the argument components do not overlap, we approach identification of argument components as a sequence labeling problem. We use the BIO encoding, so each token belongs to one of the following 11 classes: O (not a part of any argument component), Backing-B, Backing-I, Claim-B, Claim-I, Premise-B, Premise-I, Rebuttal-B, Rebuttal-I, Refutation-B, Refutation-I. This is the minimal encoding that is able to distinguish two adjacent argument components of the same type. In our data, 48% of all adjacent argument components of the same type are direct neighbors (there are no "O" tokens in between).
We report Macro- INLINEFORM0 score and INLINEFORM1 scores for each of the 11 classes as the main evaluation metric. This evaluation is performed on the token level, and for each token the predicted label must exactly match the gold data label (classification of tokens into 11 classes).
As instances for the sequence labeling model, we chose sentences rather than tokens. During our initial experiments, we observed that building a sequence labeling model for recognizing argument components as sequences of tokens is too fine-grained, as a single token does not convey enough information that could be encoded as features for a machine learner. However, as discussed in section UID73 , the annotations were performed on data pre-segmented to sentences and annotating tokens was necessary only when the sentence segmentation was wrong or one sentence contained multiple argument components. Our corpus consists of 3899 sentences, from which 2214 sentences (57%) contain no argument component. From the remaining ones, only 50 sentences (1%) have more than one argument component. Although in 19 cases (0.5%) the sentence contains a Claim-Premise pair which is an important distinction from the argumentation perspective, given the overall small number of such occurrences, we simplify the task by treating each sentence as if it has either one argument component or none.
The approximation with sentence-level units is explained in the example in Figure FIGREF112 . In order to evaluate the expected performance loss using this approximation, we used an oracle that always predicts the correct label for the unit (sentence) and evaluated it against the true labels (recall that the evaluation against the true gold labels is done always on token level). We lose only about 10% of macro INLINEFORM0 score (0.906) and only about 2% of accuracy (0.984). This performance is still acceptable, while allowing to model sequences where the minimal unit is a sentence.
Table TABREF114 shows the distribution of the classes in the gold data Toulmin, where the labeling was already mapped to the sentences. The little presence of rebuttal and refutation (4 classes account only for 3.4% of the data) makes this dataset very unbalanced.
We chose SVMhmm BIBREF111 implementation of Structural Support Vector Machines for sequence labeling. Each sentence ( INLINEFORM0 ) is represented as a vector of real-valued features.
We defined the following feature sets:
FS0: Baseline lexical features
word uni-, bi-, and tri-grams (binary)
FS1: Structural, morphological, and syntactic features
First and last 3 tokens. Motivation: these tokens may contain discourse markers or other indicators for argument components, such as “therefore” and “since” for premises or “think” and “believe” for claims.
Relative position in paragraph and relative position in document. Motivation: We expect that claims are more likely to appear at the beginning or at the end of the document.
Number of POS 1-3 grams, dependency tree depth, constituency tree production rules, and number of sub-clauses. Based on BIBREF113 .
FS2: Topic and sentiment features
30 features taken from a vector representation of the sentence obtained by using Gibbs sampling on LDA model BIBREF114 , BIBREF115 with 30 topics trained on unlabeled data from the raw corpus. Motivation: Topic representation of a sentence might be valuable for detecting off-topic sentences, namely non-argument components.
Scores for five sentiment categories (from very negative to very positive) obtained from Stanford sentiment analyzer BIBREF116 . Motivation: Claims usually express opinions and carry sentiment.
FS3: Semantic, coreference, and discourse features
Binary features from Clear NLP Semantic Role Labeler BIBREF117 . Namely, we extract agent, predicate + agent, predicate + agent + patient + (optional) negation, argument type + argument value, and discourse marker which are based on PropBank semantic role labels. Motivation: Exploit the semantics of Capturing the semantics of the sentences.
Binary features from Stanford Coreference Chain Resolver BIBREF118 , e.g., presence of the sentence in a chain, transition type (i.e., nominal–pronominal), distance to previous/next sentences in the chain, or number of inter-sentence coreference links. Motivation: Presence of coreference chains indicates links outside the sentence and thus may be informative, for example, for classifying whether the sentence is a part of a larger argument component.
Results of a PTDB-style discourse parser BIBREF119 , namely the type of discourse relation (explicit, implicit), presence of discourse connectives, and attributions. Motivation: It has been claimed that discourse relations play a role in argumentation mining BIBREF120 .
FS4: Embedding features
300 features from word embedding vectors using word embeddings trained on part of the Google News dataset BIBREF121 . In particular, we sum up embedding vectors (dimensionality 300) of each word, resulting into a single vector for the entire sentence. This vector is then directly used as a feature vector. Motivation: Embeddings helped to achieve state-of-the-art results in various NLP tasks BIBREF116 , BIBREF122 .
Except the baseline lexical features, all feature types are extracted not only for the current sentence INLINEFORM0 , but also for INLINEFORM1 preceding and subsequent sentences, namely INLINEFORM2 , INLINEFORM3 , INLINEFORM4 INLINEFORM5 , INLINEFORM6 , where INLINEFORM7 was empirically set to 4. Each feature is then represented with a prefix to determine its relative position to the current sequence unit.
Let us first discuss the upper bounds of the system. Performance of the three human annotators is shown in the first column of Table TABREF139 (results are obtained from a cumulative confusion matrix). The overall Macro- INLINEFORM0 score is 0.602 (accuracy 0.754). If we look closer at the different argument components, we observe that humans are good at predicting claims, premises, backing and non-argumentative text (about 0.60-0.80 INLINEFORM1 ), but on rebuttal and refutation they achieve rather low scores. Without these two components, the overall human Macro- INLINEFORM2 would be 0.707. This trend follows the inter-annotator agreement scores, as discussed in section UID75 .
In our experiments, the feature sets were combined in the bottom-up manner, starting with the simple lexical features (FS0), adding structural and syntactic features (FS1), then adding topic and sentiment features (FS2), then features reflecting the discourse structure (FS3), and finally enriched with completely unsupervised latent vector space representation (FS4). In addition, we were gradually removing the simple features (e.g., without lexical features, without syntactic features, etc.) to test the system with more “abstract” feature sets (feature ablation). The results are shown in Table TABREF139 .
The overall best performance (Macro- INLINEFORM0 0.251) was achieved using the rich feature sets (01234 and 234) and significantly outperformed the baseline as well as other feature sets. Classification of non-argumentative text (the "O" class) yields about 0.7 INLINEFORM1 score even in the baseline setting. The boundaries of claims (Cla-B), premises (Pre-B), and backing (Bac-B) reach in average lower scores then their respective inside tags (Cla-I, Pre-I, Bac-I). It can be interpreted such that the system is able to classify that a certain sentence belongs to a certain argument component, but the distinction whether it is a beginning of the argument component is harder. The very low numbers for rebuttal and refutation have two reasons. First, these two argument components caused many disagreements in the annotations, as discussed in section UID86 , and were hard to recognize for the humans too. Second, these four classes have very few instances in the corpus (about 3.4%, see Table TABREF114 ), so the classifier suffers from the lack of training data.
The results for the in-domain cross validation scenario are shown in Table TABREF140 . Similarly to the cross-validation scenario, the overall best results were achieved using the largest feature set (01234). For mainstreaming and red-shirting, the best results were achieved using only the feature set 4 (embeddings). These two domains contain also fewer documents, compared to other domains (refer to Table TABREF71 ). We suspect that embeddings-based features convey important information when not enough in-domain data are available. This observation will become apparent in the next experiment.
The cross-domain experiments yield rather poor results for most of the feature combinations (Table TABREF141 ). However, using only feature set 4 (embeddings), the system performance increases rapidly, so it is even comparable to numbers achieved in the in-domain scenario. These results indicate that embedding features generalize well across domains in our task of argument component identification. We leave investigating better performing vector representations, such as paragraph vectors BIBREF123 , for future work.
Error analysis based on the probabilistic confusion matrix BIBREF124 shown in Table TABREF142 reveals further details. About a half of the instances for each class are misclassified as non-argumentative (the "O" prediction).
Backing-B is often confused with Premise-B (12%) and Backing-I with Premise-I (23%). Similarly, Premise-I is misclassified as Backing-I in 9%. This shows that distinguishing between backing and premises is not easy because these two components are similar such that they support the claim, as discussed in section UID86 . We can also see that the misclassification is consistent among *-B and *-I tags.
Rebuttal is often misclassified as Premise (28% for Rebuttal-I and 18% for Rebuttal-B; notice again the consistency in *-B and *-I tags). This is rather surprising, as one would expect that rebuttal would be confused with a claim, because its role is to provide an opposing view.
Refutation-B and Refutation-I is misclassified as Premise-I in 19% and 27%, respectively. This finding confirms the discussion in section UID86 , because the role of refutation is highly context-dependent. In a pragmatic perspective, it is put forward to indirectly support the claim by attacking the rebuttal, thus having a similar function to the premise.
We manually examined miss-classified examples produced the best-performing system to find out which phenomena pose biggest challenges. Properly detecting boundaries of argument components caused problems, as shown in Figure FIGREF146 (a). This goes in line with the granularity annotation difficulties discussed in section UID86 . The next example in Figure FIGREF146 (b) shows that even if boundaries of components were detected precisely, the distinction between premise and backing fails. The example also shows that in some cases, labeling on clause level is required (left-hand side claim and premise) but the approximation in the system cannot cope with this level of detail (as explained in section UID111 ). Confusing non-argumentative text and argument components by the system is sometimes plausible, as is the case of the last rhetorical question in Figure FIGREF146 (c). On the other hand, the last example in Figure FIGREF146 (d) shows that some claims using figurative language were hard to be identified. The complete predictions along with the gold data are publicly available.
SVMhmm offers many hyper-parameters with suggested default values, from which three are of importance. Parameter INLINEFORM0 sets the order of dependencies of transitions in HMM, parameter INLINEFORM1 sets the order of dependencies of emissions in HMM, and parameter INLINEFORM2 represents a trading-off slack versus magnitude of the weight-vector. For all experiments, we set all the hyper-parameters to their default values ( INLINEFORM3 , INLINEFORM4 , INLINEFORM5 ). Using the best performing feature set from Table TABREF139 , we experimented with a grid search over different values ( INLINEFORM6 , INLINEFORM7 , INLINEFORM8 ) but the results did not outperform the system trained with default parameter values.
The INLINEFORM0 scores might seem very low at the first glance. One obvious reason is the actual performance of the system, which gives a plenty of room for improvement in the future. But the main cause of low INLINEFORM2 numbers is the evaluation measure — using 11 classes on the token level is very strict, as it penalizes a mismatch in argument component boundaries the same way as a wrongly predicted argument component type. Therefore we also report two another evaluation metrics that help to put our results into a context.
Krippendorff's INLINEFORM0 — It was also used for evaluating inter-annotator agreement (see section UID75 ).
Boundary similarity BIBREF125 — Using this metric, the problem is treated solely as a segmentation task without recognizing the argument component types.
As shown in Table TABREF157 (the Macro- INLINEFORM0 scores are repeated from Table TABREF139 ), the best-performing system achieves 0.30 score using Krippendorf's INLINEFORM1 , which is in the middle between the baseline and the human performance (0.48) but is considered as poor from the inter-annotator agreement point of view BIBREF54 . The boundary similarity metrics is not directly suitable for evaluating argument component classification, but reveals a sub-task of finding the component boundaries. The best system achieved 0.32 on this measure. Vovk2013MT used this measure to annotate argument spans and his annotators achieved 0.36 boundary similarity score. Human annotators in BIBREF125 reached 0.53 boundary similarity score.
The overall performance of the system is also affected by the accuracy of individual NPL tools used for extracting features. One particular problem is that the preprocessing models we rely on (POS, syntax, semantic roles, coreference, discourse; see section UID115 ) were trained on newswire corpora, so one has to expect performance drop when applied on user-generated content. This is however a well-known issue in NLP BIBREF126 , BIBREF127 , BIBREF128 .
To get an impression of the actual performance of the system on the data, we also provide the complete output of our best performing system in one PDF document together with the gold annotations in the logos dimension side by side in the accompanying software package. We believe this will help the community to see the strengths of our model as well as possible limitations of our current approaches.
## Conclusions
Let us begin with summarizing answers to the research questions stated in the introduction. First, as we showed in section UID55 , existing argumentation theories do offer models for capturing argumentation in user-generated content on the Web. We built upon the Toulmin's model and proposed some extensions.
Second, as compared to the negative experiences with annotating using Walton's schemes (see sections UID52 and SECREF31 ), our modified Toulmin's model offers a trade-off between its expressiveness and annotation reliability. However, we found that the capabilities of the model to capture argumentation depend on the register and topic, the length of the document, and inherently on the literary devices and structures used for expressing argumentation as these properties influenced the agreement among annotators.
Third, there are aspects of online argumentation that lack their established theoretical counterparts, such as rhetorical questions, figurative language, narratives, and fallacies in general. We tried to model some of them in the pathos dimension of argument (section UID103 ), but no satisfying agreement was reached. Furthermore, we dealt with a step that precedes argument analysis by filtering documents given their persuasiveness with respect to the controversy. Finally, we proposed a computational model based on machine learning for identifying argument components (section SECREF108 ). In this identification task, we experimented with a wide range of linguistically motivated features and found that (1) the largest feature set (including n-grams, structural features, syntactic features, topic distribution, sentiment distribution, semantic features, coreference feaures, discourse features, and features based on word embeddings) performs best in both in-domain and all-data cross validation, while (2) features based only on word embeddings yield best results in cross-domain evaluation.
Since there is no one-size-fits-all argumentation theory to be applied to actual data on the Web, the argumentation model and an annotation scheme for argumentation mining is a function of the task requirements and the corpus properties. Its selection should be based on the data at hand and the desired application. Given the proposed use-case scenarios (section SECREF1 ) and the results of our annotation study (section SECREF51 ), we recommend a scheme based on Toulmin's model for short documents, such as comments or forum posts.
| 19
|
1603.00968
|
MGNC-CNN: A Simple Approach to Exploiting Multiple Word Embeddings for Sentence Classification
|
# MGNC-CNN: A Simple Approach to Exploiting Multiple Word Embeddings for Sentence Classification
## Abstract
We introduce a novel, simple convolution neural network (CNN) architecture - multi-group norm constraint CNN (MGNC-CNN) that capitalizes on multiple sets of word embeddings for sentence classification. MGNC-CNN extracts features from input embedding sets independently and then joins these at the penultimate layer in the network to form a final feature vector. We then adopt a group regularization strategy that differentially penalizes weights associated with the subcomponents generated from the respective embedding sets. This model is much simpler than comparable alternative architectures and requires substantially less training time. Furthermore, it is flexible in that it does not require input word embeddings to be of the same dimensionality. We show that MGNC-CNN consistently outperforms baseline models.
## Introduction
Neural models have recently gained popularity for Natural Language Processing (NLP) tasks BIBREF0 , BIBREF1 , BIBREF2 . For sentence classification, in particular, Convolution Neural Networks (CNN) have realized impressive performance BIBREF3 , BIBREF4 . These models operate over word embeddings, i.e., dense, low dimensional vector representations of words that aim to capture salient semantic and syntactic properties BIBREF1 .
An important consideration for such models is the specification of the word embeddings. Several options exist. For example, Kalchbrenner et al. kalchbrenner2014convolutional initialize word vectors to random low-dimensional vectors to be fit during training, while Johnson and Zhang johnson2014effective use fixed, one-hot encodings for each word. By contrast, Kim kim2014convolutional initializes word vectors to those estimated via the word2vec model trained on 100 billion words of Google News BIBREF5 ; these are then updated during training. Initializing embeddings to pre-trained word vectors is intuitively appealing because it allows transfer of learned distributional semantics. This has allowed a relatively simple CNN architecture to achieve remarkably strong results.
Many pre-trained word embeddings are now readily available on the web, induced using different models, corpora, and processing steps. Different embeddings may encode different aspects of language BIBREF6 , BIBREF7 , BIBREF8 : those based on bag-of-words (BoW) statistics tend to capture associations (doctor and hospital), while embeddings based on dependency-parses encode similarity in terms of use (doctor and surgeon). It is natural to consider how these embeddings might be combined to improve NLP models in general and CNNs in particular.
Contributions. We propose MGNC-CNN, a novel, simple, scalable CNN architecture that can accommodate multiple off-the-shelf embeddings of variable sizes. Our model treats different word embeddings as distinct groups, and applies CNNs independently to each, thus generating corresponding feature vectors (one per embedding) which are then concatenated at the classification layer. Inspired by prior work exploiting regularization to encode structure for NLP tasks BIBREF9 , BIBREF10 , we impose different regularization penalties on weights for features generated from the respective word embedding sets.
Our approach enjoys the following advantages compared to the only existing comparable model BIBREF11 : (i) It can leverage diverse, readily available word embeddings with different dimensions, thus providing flexibility. (ii) It is comparatively simple, and does not, for example, require mutual learning or pre-training. (iii) It is an order of magnitude more efficient in terms of training time.
## Related Work
Prior work has considered combining latent representations of words that capture syntactic and semantic properties BIBREF12 , and inducing multi-modal embeddings BIBREF13 for general NLP tasks. And recently, Luo et al. luo2014pre proposed a framework that combines multiple word embeddings to measure text similarity, however their focus was not on classification.
More similar to our work, Yin and Schütze yin-schutze:2015:CoNLL proposed MVCNN for sentence classification. This CNN-based architecture accepts multiple word embeddings as inputs. These are then treated as separate `channels', analogous to RGB channels in images. Filters consider all channels simultaneously. MVCNN achieved state-of-the-art performance on multiple sentence classification tasks. However, this model has practical drawbacks. (i) MVCNN requires that input word embeddings have the same dimensionality. Thus to incorporate a second set of word vectors trained on a corpus (or using a model) of interest, one needs to either find embeddings that happen to have a set number of dimensions or to estimate embeddings from scratch. (ii) The model is complex, both in terms of implementation and run-time. Indeed, this model requires pre-training and mutual-learning and requires days of training time, whereas the simple architecture we propose requires on the order of an hour (and is easy to implement).
## Model Description
We first review standard one-layer CNN (which exploits a single set of embeddings) for sentence classification BIBREF3 , and then propose our augmentations, which exploit multiple embedding sets.
Basic CNN. In this model we first replace each word in a sentence with its vector representation, resulting in a sentence matrix INLINEFORM0 , where INLINEFORM1 is the (zero-padded) sentence length, and INLINEFORM2 is the dimensionality of the embeddings. We apply a convolution operation between linear filters with parameters INLINEFORM3 and the sentence matrix. For each INLINEFORM4 , where INLINEFORM5 denotes `height', we slide filter INLINEFORM6 across INLINEFORM7 , considering `local regions' of INLINEFORM8 adjacent rows at a time. At each local region, we perform element-wise multiplication and then take the element-wise sum between the filter and the (flattened) sub-matrix of INLINEFORM9 , producing a scalar. We do this for each sub-region of INLINEFORM10 that the filter spans, resulting in a feature map vector INLINEFORM11 . We can use multiple filter sizes with different heights, and for each filter size we can have multiple filters. Thus the model comprises INLINEFORM12 weight vectors INLINEFORM13 , each of which is associated with an instantiation of a specific filter size. These in turn generate corresponding feature maps INLINEFORM14 , with dimensions varying with filter size. A 1-max pooling operation is applied to each feature map, extracting the largest number INLINEFORM15 from each feature map INLINEFORM16 . Finally, we combine all INLINEFORM17 together to form a feature vector INLINEFORM18 to be fed through a softmax function for classification. We regularize weights at this level in two ways. (1) Dropout, in which we randomly set elements in INLINEFORM19 to zero during the training phase with probability INLINEFORM20 , and multiply INLINEFORM21 with the parameters trained in INLINEFORM22 at test time. (2) An l2 norm penalty, for which we set a threshold INLINEFORM23 for the l2 norm of INLINEFORM24 during training; if this is exceeded, we rescale the vector accordingly. For more details, see BIBREF4 .
MG-CNN. Assuming we have INLINEFORM0 word embeddings with corresponding dimensions INLINEFORM1 , we can simply treat each word embedding independently. In this case, the input to the CNN comprises multiple sentence matrices INLINEFORM2 , where each INLINEFORM3 may have its own width INLINEFORM4 . We then apply different groups of filters INLINEFORM5 independently to each INLINEFORM6 , where INLINEFORM7 denotes the set of filters for INLINEFORM8 . As in basic CNN, INLINEFORM9 may have multiple filter sizes, and multiple filters of each size may be introduced. At the classification layer we then obtain a feature vector INLINEFORM10 for each embedding set, and we can simply concatenate these together to form the final feature vector INLINEFORM11 to feed into the softmax function, where INLINEFORM12 . This representation contains feature vectors generated from all sets of embeddings under consideration. We call this method multiple group CNN (MG-CNN). Here groups refer to the features generated from different embeddings. Note that this differs from `multi-channel' models because at the convolution layer we use different filters on each word embedding matrix independently, whereas in a standard multi-channel approach each filter would consider all channels simultaneously and generate a scalar from all channels at each local region. As above, we impose a max l2 norm constraint on the final feature vector INLINEFORM13 for regularization. Figure FIGREF1 illustrates this approach.
MGNC-CNN. We propose an augmentation of MG-CNN, Multi-Group Norm Constraint CNN (MGNC-CNN), which differs in its regularization strategy. Specifically, in this variant we impose grouped regularization constraints, independently regularizing subcomponents INLINEFORM0 derived from the respective embeddings, i.e., we impose separate max norm constraints INLINEFORM1 for each INLINEFORM2 (where INLINEFORM3 again indexes embedding sets); these INLINEFORM4 hyper-parameters are to be tuned on a validation set. Intuitively, this method aims to better capitalize on features derived from word embeddings that capture discriminative properties of text for the task at hand by penalizing larger weight estimates for features derived from less discriminative embeddings.
## Datasets
Stanford Sentiment Treebank Stanford Sentiment Treebank (SST) BIBREF14 . This concerns predicting movie review sentiment. Two datasets are derived from this corpus: (1) SST-1, containing five classes: very negative, negative, neutral, positive, and very positive. (2) SST-2, which has only two classes: negative and positive. For both, we remove phrases of length less than 4 from the training set.
Subj BIBREF15 . The aim here is to classify sentences as either subjective or objective. This comprises 5000 instances of each.
TREC BIBREF16 . A question classification dataset containing six classes: abbreviation, entity, description, human, location and numeric. There are 5500 training and 500 test instances.
Irony BIBREF17 . This dataset contains 16,006 sentences from reddit labeled as ironic (or not). The dataset is imbalanced (relatively few sentences are ironic). Thus before training, we under-sampled negative instances to make classes sizes equal. Note that for this dataset we report the Area Under Curve (AUC), rather than accuracy, because it is imbalanced.
## Pre-trained Word Embeddings
We consider three sets of word embeddings for our experiments: (i) word2vec is trained on 100 billion tokens of Google News dataset; (ii) GloVe BIBREF18 is trained on aggregated global word-word co-occurrence statistics from Common Crawl (840B tokens); and (iii) syntactic word embedding trained on dependency-parsed corpora. These three embedding sets happen to all be 300-dimensional, but our model could accommodate arbitrary and variable sizes. We pre-trained our own syntactic embeddings following BIBREF8 . We parsed the ukWaC corpus BIBREF19 using the Stanford Dependency Parser v3.5.2 with Stanford Dependencies BIBREF20 and extracted (word, relation+context) pairs from parse trees. We “collapsed" nodes with prepositions and notated inverse relations separately, e.g., “dog barks" emits two tuples: (barks, nsubj_dog) and (dog, nsubj INLINEFORM0 _barks). We filter words and contexts that appear fewer than 100 times, resulting in INLINEFORM1 173k words and 1M contexts. We trained 300d vectors using word2vecf with default parameters.
## Setup
We compared our proposed approaches to a standard CNN that exploits a single set of word embeddings BIBREF3 . We also compared to a baseline of simply concatenating embeddings for each word to form long vector inputs. We refer to this as Concatenation-CNN C-CNN. For all multiple embedding approaches (C-CNN, MG-CNN and MGNC-CNN), we explored two combined sets of embedding: word2vec+Glove, and word2vec+syntactic, and one three sets of embedding: word2vec+Glove+syntactic. For all models, we tuned the l2 norm constraint INLINEFORM0 over the range INLINEFORM1 on a validation set. For instantiations of MGNC-CNN in which we exploited two embeddings, we tuned both INLINEFORM2 , and INLINEFORM3 ; where we used three embedding sets, we tuned INLINEFORM4 and INLINEFORM5 .
We used standard train/test splits for those datasets that had them. Otherwise, we performed 10-fold cross validation, creating nested development sets with which to tune hyperparameters. For all experiments we used filters sizes of 3, 4 and 5 and we created 100 feature maps for each filter size. We applied 1 max-pooling and dropout (rate: 0.5) at the classification layer. For training we used back-propagation in mini-batches and used AdaDelta as the stochastic gradient descent (SGD) update rule, and set mini-batch size as 50. In this work, we treat word embeddings as part of the parameters of the model, and update them as well during training. In all our experiments, we only tuned the max norm constraint(s), fixing all other hyperparameters.
## Results and Discussion
We repeated each experiment 10 times and report the mean and ranges across these. This replication is important because training is stochastic and thus introduces variance in performance BIBREF4 . Results are shown in Table TABREF2 , and the corresponding best norm constraint value is shown in Table TABREF2 . We also show results on Subj, SST-1 and SST-2 achieved by the more complex model of BIBREF11 for comparison; this represents the state-of-the-art on the three datasets other than TREC.
We can see that MGNC-CNN and MG-CNN always outperform baseline methods (including C-CNN), and MGNC-CNN is usually better than MG-CNN. And on the Subj dataset, MG-CNN actually achieves slightly better results than BIBREF11 , with far less complexity and required training time (MGNC-CNN performs comparably, although no better, here). On the TREC dataset, the best-ever accuracy we are aware of is 96.0% BIBREF21 , which falls within the range of the result of our MGNC-CNN model with three word embeddings. On the irony dataset, our model with three embeddings achieves 4% improvement (in terms of AUC) compared to the baseline model. On SST-1 and SST-2, our model performs slightly worse than BIBREF11 . However, we again note that their performance is achieved using a much more complex model which involves pre-training and mutual-learning steps. This model takes days to train, whereas our model requires on the order of an hour.
We note that the method proposed by Astudillo et al. astudillo2015learning is able to accommodate multiple embedding sets with different dimensions by projecting the original word embeddings into a lower-dimensional space. However, this work requires training the optimal projection matrix on laebled data first, which again incurs large overhead.
Of course, our model also has its own limitations: in MGNC-CNN, we need to tune the norm constraint hyperparameter for all the word embeddings. As the number of word embedding increases, this will increase the running time. However, this tuning procedure is embarrassingly parallel.
## Conclusions
We have proposed MGNC-CNN: a simple, flexible CNN architecture for sentence classification that can exploit multiple, variable sized word embeddings. We demonstrated that this consistently achieves better results than a baseline architecture that exploits only a single set of word embeddings, and also a naive concatenation approach to capitalizing on multiple embeddings. Furthermore, our results are comparable to those achieved with a recently proposed model BIBREF11 that is much more complex. However, our simple model is easy to implement and requires an order of magnitude less training time. Furthermore, our model is much more flexible than previous approaches, because it can accommodate variable-size word embeddings.
## Acknowledgments
This work was supported in part by the Army Research Office (grant W911NF-14-1-0442) and by The Foundation for Science and Technology, Portugal (grant UTAP-EXPL/EEIESS/0031/2014). This work was also made possible by the support of the Texas Advanced Computer Center (TACC) at UT Austin.
| 9
|
1603.01417
|
Dynamic Memory Networks for Visual and Textual Question Answering
|
# Dynamic Memory Networks for Visual and Textual Question Answering
## Abstract
Neural network architectures with memory and attention mechanisms exhibit certain reasoning capabilities required for question answering. One such architecture, the dynamic memory network (DMN), obtained high accuracy on a variety of language tasks. However, it was not shown whether the architecture achieves strong results for question answering when supporting facts are not marked during training or whether it could be applied to other modalities such as images. Based on an analysis of the DMN, we propose several improvements to its memory and input modules. Together with these changes we introduce a novel input module for images in order to be able to answer visual questions. Our new DMN+ model improves the state of the art on both the Visual Question Answering dataset and the \babi-10k text question-answering dataset without supporting fact supervision.
## Introduction
Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 .
The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging.
We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set.
In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 .
## Dynamic Memory Networks
We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 .
The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows:
Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words.
As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by
$$u_i &=& \sigma \left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \right)\\
r_i &=& \sigma \left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \right)\\
\tilde{h}_i &=& \tanh \left(Wx_{i} + r_i \circ U h_{i-1} + b^{(h)}\right)\\
h_i &=& u_i\circ \tilde{h}_i + (1-u_i) \circ h_{i-1}$$ (Eq. 2)
where $\sigma $ is the sigmoid activation function, $\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \in \mathbb {R}^{n_H \times n_I}$ , $U^{(z)}, U^{(r)}, U \in \mathbb {R}^{n_H \times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size.
Question Module: This module computes a vector representation $q$ of the question, where $q \in \mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question.
Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \in \mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ .
The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \in \mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \in \mathbb {R}^{n_H}$0 .
Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network.
## Improved Dynamic Memory Networks: DMN+
We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset.
## Input Module for Text QA
In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. "Model Analysis" ).
We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU.
Input Fusion Layer
For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction.
Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence.
The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used.
For the positional encoding scheme, the sentence representation is produced by $f_i = \sum ^{j=1}_M l_j \circ w^i_j$ , where $\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding.
The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU.
$$\overrightarrow{f_i} = GRU_{fwd}(f_i, \overrightarrow{f_{i-1}}) \\
\overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \overleftarrow{f_{i+1}}) \\
\overleftrightarrow{f_i} = \overleftarrow{f_i} + \overrightarrow{f_i}$$ (Eq. 5)
where $f_i$ is the input fact at timestep $i$ , $ \overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\overleftrightarrow{f_i}$ .
We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used.
## Input Module for VQA
To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. "Input Module for Text QA" .
Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \times 14 \times 14$ . The pooling layer divides the image into a grid of $14 \times 14$ , resulting in 196 local regional vectors of $d = 512$ .
Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ .
Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems.
To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. "Input Module for Text QA" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information.
## The Episodic Memory Module
The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\overleftrightarrow{F} = [\overleftrightarrow{f_1}, \hdots , \overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state.
$$z^t_i &=& [\overleftrightarrow{f_i} \circ q; \overleftrightarrow{f_i} \circ m^{t-1}; \vert \overleftrightarrow{f_i} - q \vert ; \vert \overleftrightarrow{f_i} - m^{t-1} \vert ] \\
Z^t_i &=& W^{(2)} \tanh \left(W^{(1)}z^t_i + b^{(1)} \right)+ b^{(2)} \\
g^t_i &=& \frac{\exp (Z^t_i)}{\sum _{k=1}^{M_i} \exp (Z^t_k)} $$ (Eq. 10)
where $\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\circ $ is the element-wise product, $|\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors.
The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required.
Attention Mechanism
Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+.
Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \sum _{i=1}^N g^t_i \overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient.
Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation .
We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory.
By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 .
$$h_i &=& g^t_i \circ \tilde{h}_i + (1-g^t_i) \circ h_{i-1}$$ (Eq. 12)
An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \in \mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \in \mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU.
Episode Memory Updates
After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by
$$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13)
The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the “Mem Weights” row in Table 1 .
Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by
$$m^t = ReLU\left(W^t [m^{t-1} ; c^t ; q] + b\right)$$ (Eq. 14)
where $;$ is the concatenation operator, $W^t \in \mathbb {R}^{n_H \times n_H}$ , $b \in \mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN.
## Related Work
The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities.
Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps.
Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 .
Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 .
Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset.
Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 .
## Datasets
To analyze our proposed model changes and compare our performance with other architectures, we use three datasets.
## bAbI-10k
For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k.
## DAQUAR-ALL visual dataset
The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. "Model Analysis" ).
## Visual Question Answering
The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 .
Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task.
## Model Analysis
To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets.
The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. "Input Module for Text QA" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. "The Episodic Memory Module" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 .
A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required.
The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases.
From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets.
## Comparison to state of the art using bAbI-10k
We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\sqrt{3}, \sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 .
On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss.
Text QA Results
We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training.
The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component.
The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time.
In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset.
One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks.
The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting.
## Comparison to state of the art using VQA
For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \frac{1}{N}\sum _{i=1}^Nmin(\frac{\sum _{t\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\%$ accurate if at least 3 people provide that exact answer.
Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ .
Results and Analysis
The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities.
The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach.
Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image.
However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries.
## Conclusion
We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains.
| 15
|
1603.01514
|
A Bayesian Model of Multilingual Unsupervised Semantic Role Induction
|
# A Bayesian Model of Multilingual Unsupervised Semantic Role Induction
## Abstract
We propose a Bayesian model of unsupervised semantic role induction in multiple languages, and use it to explore the usefulness of parallel corpora for this task. Our joint Bayesian model consists of individual models for each language plus additional latent variables that capture alignments between roles across languages. Because it is a generative Bayesian model, we can do evaluations in a variety of scenarios just by varying the inference procedure, without changing the model, thereby comparing the scenarios directly. We compare using only monolingual data, using a parallel corpus, using a parallel corpus with annotations in the other language, and using small amounts of annotation in the target language. We find that the biggest impact of adding a parallel corpus to training is actually the increase in mono-lingual data, with the alignments to another language resulting in small improvements, even with labeled data for the other language.
## Introduction
Semantic Role Labeling (SRL) has emerged as an important task in Natural Language Processing (NLP) due to its applicability in information extraction, question answering, and other NLP tasks. SRL is the problem of finding predicate-argument structure in a sentence, as illustrated below:
INLINEFORM0
Here, the predicate WRITE has two arguments: `Mike' as A0 or the writer, and `a book' as A1 or the thing written. The labels A0 and A1 correspond to the PropBank annotations BIBREF0 .
As the need for SRL arises in different domains and languages, the existing manually annotated corpora become insufficient to build supervised systems. This has motivated work on unsupervised SRL BIBREF1 , BIBREF2 , BIBREF3 . Previous work has indicated that unsupervised systems could benefit from the word alignment information in parallel text in two or more languages BIBREF4 , BIBREF5 , BIBREF6 . For example, consider the German translation of sentence INLINEFORM0 :
INLINEFORM0
If sentences INLINEFORM0 and INLINEFORM1 have the word alignments: Mike-Mike, written-geschrieben, and book-Buch, the system might be able to predict A1 for Buch, even if there is insufficient information in the monolingual German data to learn this assignment. Thus, in languages where the resources are sparse or not good enough, or the distributions are not informative, SRL systems could be made more accurate by using parallel data with resource rich or more amenable languages.
In this paper, we propose a joint Bayesian model for unsupervised semantic role induction in multiple languages. The model consists of individual Bayesian models for each language BIBREF3 , and crosslingual latent variables to incorporate soft role agreement between aligned constituents. This latent variable approach has been demonstrated to increase the performance in a multilingual unsupervised part-of-speech tagging model based on HMMs BIBREF4 . We investigate the application of this approach to unsupervised SRL, presenting the performance improvements obtained in different settings involving labeled and unlabeled data, and analyzing the annotation effort required to obtain similar gains using labeled data.
We begin by briefly describing the unsupervised SRL pipeline and the monolingual semantic role induction model we use, and then describe our multilingual model.
## Unsupervised SRL Pipeline
As established in previous work BIBREF7 , BIBREF8 , we use a standard unsupervised SRL setup, consisting of the following steps:
The task we model, unsupervised semantic role induction, is the step 4 of this pipeline.
## Monolingual Model
We use the Bayesian model of garg2012unsupervised as our base monolingual model. The semantic roles are predicate-specific. To model the role ordering and repetition preferences, the role inventory for each predicate is divided into Primary and Secondary roles as follows:
For example, the complete role sequence in a frame could be: INLINEFORM0 INLINEFORM1 , INLINEFORM2 , INLINEFORM3 , INLINEFORM4 , INLINEFORM5 , INLINEFORM6 , INLINEFORM7 , INLINEFORM8 INLINEFORM9 . The ordering is defined as the sequence of PRs, INLINEFORM10 INLINEFORM11 , INLINEFORM12 , INLINEFORM13 , INLINEFORM14 , INLINEFORM15 INLINEFORM16 . Each pair of consecutive PRs in an ordering is called an interval. Thus, INLINEFORM17 is an interval that contains two SRs, INLINEFORM18 and INLINEFORM19 . An interval could also be empty, for instance INLINEFORM20 contains no SRs. When we evaluate, these roles get mapped to gold roles. For instance, the PR INLINEFORM21 could get mapped to a core role like INLINEFORM22 , INLINEFORM23 , etc. or to a modifier role like INLINEFORM24 , INLINEFORM25 , etc. garg2012unsupervised reported that, in practice, PRs mostly get mapped to core roles and SRs to modifier roles, which conforms to the linguistic motivations for this distinction.
Figure FIGREF16 illustrates two copies of the monolingual model, on either side of the crosslingual latent variables. The generative process is as follows:
All the multinomial and binomial distributions have symmetric Dirichlet and beta priors respectively. Figure FIGREF7 gives the probability equations for the monolingual model. This formulation models the global role ordering and repetition preferences using PRs, and limited context for SRs using intervals. Ordering and repetition information was found to be helpful in supervised SRL as well BIBREF9 , BIBREF8 , BIBREF10 . More details, including the motivations behind this model, are in BIBREF3 .
## Multilingual Model
The multilingual model uses word alignments between sentences in a parallel corpus to exploit role correspondences across languages. We make copies of the monolingual model for each language and add additional crosslingual latent variables (CLVs) to couple the monolingual models, capturing crosslingual semantic role patterns. Concretely, when training on parallel sentences, whenever the head words of the arguments are aligned, we add a CLV as a parent of the two corresponding role variables. Figure FIGREF16 illustrates this model. The generative process, as explained below, remains the same as the monolingual model for the most part, with the exception of aligned roles which are now generated by both the monolingual process as well as the CLV.
Every predicate-tuple has its own inventory of CLVs specific to that tuple. Each CLV INLINEFORM0 is a multi-valued variable where each value defines a distribution over role labels for each language (denoted by INLINEFORM1 above). These distributions over labels are trained to be peaky, so that each value INLINEFORM2 for a CLV represents a correlation between the labels that INLINEFORM3 predicts in the two languages. For example, a value INLINEFORM4 for the CLV INLINEFORM5 might give high probabilities to INLINEFORM6 and INLINEFORM7 in language 1, and to INLINEFORM8 in language 2. If INLINEFORM9 is the only value for INLINEFORM10 that gives high probability to INLINEFORM11 in language 1, and the monolingual model in language 1 decides to assign INLINEFORM12 to the role for INLINEFORM13 , then INLINEFORM14 will predict INLINEFORM15 in language 2, with high probability. We generate the CLVs via a Chinese Restaurant Process BIBREF11 , a non-parametric Bayesian model, which allows us to induce the number of CLVs for every predicate-tuple from the data. We continue to train on the non-parallel sentences using the respective monolingual models.
The multilingual model is deficient, since the aligned roles are being generated twice. Ideally, we would like to add the CLV as additional conditioning variables in the monolingual models. The new joint probability can be written as equation UID11 (Figure FIGREF7 ), which can be further decomposed following the decomposition of the monolingual model in Figure FIGREF7 . However, having this additional conditioning variable breaks the Dirichlet-multinomial conjugacy, which makes it intractable to marginalize out the parameters during inference. Hence, we use an approximation where we treat each of the aligned roles as being generated twice, once by the monolingual model and once by the corresponding CLV (equation ).
This is the first work to incorporate the coupling of aligned arguments directly in a Bayesian SRL model. This makes it easier to see how to extend this model in a principled way to incorporate additional sources of information. First, the model scales gracefully to more than two languages. If there are a total of INLINEFORM0 languages, and there is an aligned argument in INLINEFORM1 of them, the multilingual latent variable is connected to only those INLINEFORM2 aligned arguments.
Second, having one joint Bayesian model allows us to use the same model in various semi-supervised learning settings, just by fixing the annotated variables during training. Section SECREF29 evaluates a setting where we have some labeled data in one language (called source), while no labeled data in the second language (called target). Note that this is different from a classic annotation projection setting (e.g. BIBREF12 ), where the role labels are mapped from source constituents to aligned target constituents.
## Inference and Training
The inference problem consists of predicting the role labels and CLVs (the hidden variables) given the predicate, its voice, and syntactic features of all the identified arguments (the visible variables). We use a collapsed Gibbs-sampling based approach to generate samples for the hidden variables (model parameters are integrated out). The sample counts and the priors are then used to calculate the MAP estimate of the model parameters.
For the monolingual model, the role at a given position is sampled as:
DISPLAYFORM0
where the subscript INLINEFORM0 refers to all the variables except at position INLINEFORM1 , INLINEFORM2 refers to the variables in all the training instances except the current one, and INLINEFORM3 refers to all the model parameters. The above integral has a closed form solution due to Dirichlet-multinomial conjugacy.
For sampling roles in the multilingual model, we also need to consider the probabilities of roles being generated by the CLVs:
DISPLAYFORM0
For sampling CLVs, we need to consider three factors: two corresponding to probabilities of generating the aligned roles, and the third one corresponding to selecting the CLV according to CRP.
DISPLAYFORM0
where the aligned roles INLINEFORM0 and INLINEFORM1 are connected to INLINEFORM2 , and INLINEFORM3 refers to all the variables except INLINEFORM4 , INLINEFORM5 , and INLINEFORM6 .
We use the trained parameters to parse the monolingual data using the monolingual model. The crosslingual parameters are ignored even if they were used during training. Thus, the information coming from the CLVs acts as a regularizer for the monolingual models.
## Evaluation
Following the setting of titovcrosslingual, we evaluate only on the arguments that were correctly identified, as the incorrectly identified arguments do not have any gold semantic labels. Evaluation is done using the metric proposed by lang2011unsupervised, which has 3 components: (i) Purity (PU) measures how well an induced cluster corresponds to a single gold role, (ii) Collocation (CO) measures how well a gold role corresponds to a single induced cluster, and (iii) F1 is the harmonic mean of PU and CO. For each predicate, let INLINEFORM0 denote the total number of argument instances, INLINEFORM1 the instances in the induced cluster INLINEFORM2 , and INLINEFORM3 the instances having label INLINEFORM4 in gold annotations. INLINEFORM5 , INLINEFORM6 , and INLINEFORM7 . The score for each predicate is weighted by the number of its argument instances, and a weighted average is computed over all the predicates.
## Baseline
We use the same baseline as used by lang2011unsupervised which has been shown to be difficult to outperform. This baseline assigns a semantic role to a constituent based on its syntactic function, i.e. the dependency relation to its head. If there is a total of INLINEFORM0 clusters, INLINEFORM1 most frequent syntactic functions get a cluster each, and the rest are assigned to the INLINEFORM2 th cluster.
## Closest Previous Work
This work is closely related to the cross-lingual unsupervised SRL work of titovcrosslingual. Their model has separate monolingual models for each language and an extra penalty term which tries to maximize INLINEFORM0 and INLINEFORM1 i.e. for all the aligned arguments with role label INLINEFORM2 in language 1, it tries to find a role label INLINEFORM3 in language 2 such that the given proportion is maximized and vice verse. However, there is no efficient way to optimize the objective with this penalty term and the authors used an inference method similar to annotation projection. Further, the method does not scale naturally to more than two languages. Their algorithm first does monolingual inference in one language ignoring the penalty and then does the inference in the second language taking into account the penalty term. In contrast, our model adds the latent variables as a part of the model itself, and not an external penalty, which enables us to use the standard Bayesian learning methods such as sampling.
The monolingual model we use BIBREF3 also has two main advantages over titovcrosslingual. First, the former incorporates a global role ordering probability that is missing in the latter. Secondly, the latter defines argument-keys as a tuple of four syntactic features and all the arguments having the same argument-keys are assigned the same role. This kind of hard clustering is avoided in the former model where two constituents having the same set of features might get assigned different roles if they appear in different contexts.
## Data
Following titovcrosslingual, we run our experiments on the English (EN) and German (DE) sections of the CoNLL 2009 corpus BIBREF13 , and EN-DE section of the Europarl corpus BIBREF14 . We get about 40k EN and 36k DE sentences from the CoNLL 2009 training set, and about 1.5M parallel EN-DE sentences from Europarl. For appropriate comparison, we keep the same setting as in BIBREF6 for automatic parses and argument identification, which we briefly describe here. The EN sentences are parsed syntactically using MaltParser BIBREF15 and DE using LTH parser BIBREF16 . All the non-auxiliary verbs are selected as predicates. In CoNLL data, this gives us about 3k EN and 500 DE predicates. The total number of predicate instances are 3.4M in EN (89k CoNLL + 3.3M Europarl) and 2.62M in DE (17k CoNLL + 2.6M Europarl). The arguments for EN are identified using the heuristics proposed by lang2011unsupervised. However, we get an F1 score of 85.1% for argument identification on CoNLL 2009 EN data as opposed to 80.7% reported by titovcrosslingual. This could be due to implementation differences, which unfortunately makes our EN results incomparable. For DE, the arguments are identified using the LTH system BIBREF16 , which gives an F1 score of 86.5% on the CoNLL 2009 DE data. The word alignments for the EN-DE parallel Europarl corpus are computed using GIZA++ BIBREF17 . For high-precision, only the intersecting alignments in the two directions are kept. We define two semantic arguments as aligned if their head-words are aligned. In total we get 9.3M arguments for EN (240k CoNLL + 9.1M Europarl) and 4.43M for DE (32k CoNLL + 4.4M Europarl). Out of these, 0.76M arguments are aligned.
## Main Results
Since the CoNLL annotations have 21 semantic roles in total, we use 21 roles in our model as well as the baseline. Following garg2012unsupervised, we set the number of PRs to 2 (excluding INLINEFORM0 , INLINEFORM1 and INLINEFORM2 ), and SRs to 21-2=19. Table TABREF27 shows the results.
In the first setting (Line 1), we train and test the monolingual model on the CoNLL data. We observe significant improvements in F1 score over the Baseline (Line 0) in both languages. Using the CoNLL 2009 dataset alone, titovcrosslingual report an F1 score of 80.9% (PU=86.8%, CO=75.7%) for German. Thus, our monolingual model outperforms their monolingual model in German. For English, they report an F1 score of 83.6% (PU=87.5%, CO=80.1%), but note that our English results are not directly comparable to theirs due to differences argument identification, as discussed in section SECREF25 . As their argument identification score is lower, perhaps their system is discarding “difficult” arguments which leads to a higher clustering score.
In the second setting (Line 2), we use the additional monolingual Europarl (EP) data for training. We get equivalent results in English and a significant improvement in German compared to our previous setting (Line 1). The German dataset in CoNLL is quite small and benefits from the additional EP training data. In contrast, the English model is already quite good due to a relatively big dataset from CoNLL, and good accuracy syntactic parsers. Unfortunately, titovcrosslingual do not report results with this setting.
The third setting (Line 3) gives the results of our multilingual model, which adds the word alignments in the EP data. Comparing with Line 2, we get non-significant improvements in both languages. titovcrosslingual obtain an F1 score of 82.7% (PU=85.0%, CO=80.6%) for German, and 83.7% (PU=86.8%, CO=80.7%) for English. Thus, for German, our multilingual Bayesian model is able to capture the cross-lingual patterns at least as well as the external penalty term in BIBREF6 . We cannot compare the English results unfortunately due to differences in argument identification.
We also compared monolingual and bilingual training data using a setting that emulates the standard supervised setup of separate training and test data sets. We train only on the EP dataset and test on the CoNLL dataset. Lines 4 and 5 of Table TABREF27 give the results. The multilingual model obtains small improvements in both languages, which confirms the results from the standard unsupervised setup, comparing lines 2 to 3.
These results indicate that little information can be learned about semantic roles from this parallel data setup. One possible explanation for this result is that the setup itself is inadequate. Given the definition of aligned arguments, only 8% of English arguments and 17% of German arguments are aligned. This plus our experiments suggest that improving the alignment model is a necessary step to making effective use of parallel data in multilingual SRI, for example by joint modeling with SRI. We leave this exploration to future work.
## Multilingual Training with Labeled Data for One Language
Another motivation for jointly modeling SRL in multiple languages is the transfer of information from a resource rich language to a resource poor language. We evaluated our model in a very general annotation transfer scenario, where we have a small labeled dataset for one language (source), and a large parallel unlabeled dataset for the source and another (target) language. We investigate whether this setting improves the parameter estimates for the target language. To this end, we clamp the role annotations of the source language in the CoNLL dataset using a predefined mapping, and do not sample them during training. This data gives us good parameters for the source language, which are used to sample the roles of the source language in the unlabeled Europarl data. The CLVs aim to capture this improvement and thereby improve sampling and parameter estimates for the target language. Table TABREF28 shows the results of this experiment. We obtain small improvements in the target languages. As in the unsupervised setting, the small percentage of aligned roles probably limits the impact of the cross-lingual information.
## Labeled Data in Monolingual Model
We explored the improvement in the monolingual model in a semi-supervised setting. To this end, we randomly selected INLINEFORM0 of the sentences in the CoNLL dataset as “supervised sentences” and the rest INLINEFORM1 were kept unsupervised. Next, we clamped the role labels of the supervised sentences using the predefined mapping from Section SECREF29 . Sampling was done on the unsupervised sentences as usual. We then measured the clustering performance using the trained parameters.
To access the contribution of partial supervision better, we constructed a “supervised baseline” as follows. For predicates seen in the supervised sentences, a MAP estimate of the parameters was calculated using the predefined mapping. For the unseen predicates, the standard baseline was used.
Figures FIGREF33 and FIGREF33 show the performance variation with INLINEFORM0 . We make the following observations:
[leftmargin=*]
In both languages, at around INLINEFORM0 , the supervised baseline starts outperforming the semi-supervised model, which suggests that manually labeling about 10% of the sentences is a good enough alternative to our training procedure. Note that 10% amounts to about 3.6k sentences in German and 4k in English. We noticed that the proportion of seen predicates increases dramatically as we increase the proportion of supervised sentences. At 10% supervised sentences, the model has already seen 63% of predicates in German and 44% in English. This explains to some extent why only 10% labeled sentences are enough.
For German, it takes about 3.5% or 1260 supervised sentences to have the same performance increase as 1.5M unlabeled sentences (Line 1 to Line 2 in Table TABREF27 ). Adding about 180 more supervised sentences also covers the benefit obtained by alignments in the multilingual model (Line 2 to Line 3 in Table TABREF27 ). There is no noticeable performance difference in English.
We also evaluated the performance variation on a completely unseen CoNLL test set. Since the test set is very small compared to the training set, the clustering evaluation is not as reliable. Nonetheless, we broadly obtained the same pattern.
## Related Work
As discussed in section SECREF24 , our work is closely related to the crosslingual unsupervised SRL work of titovcrosslingual. The idea of using superlingual latent variables to capture cross-lingual information was proposed for POS tagging by naseem2009multilingual, which we use here for SRL. In a semi-supervised setting, pado2009cross used a graph based approach to transfer semantic role annotations from English to German. furstenau2009graph used a graph alignment method to measure the semantic and syntactic similarity between dependency tree arguments of known and unknown verbs.
For monolingual unsupervised SRL, swier2004unsupervised presented the first work on a domain-general corpus, the British National Corpus, using 54 verbs taken from VerbNet. garg2012unsupervised proposed a Bayesian model for this problem that we use here. titov2012bayesian also proposed a closely related Bayesian model. grenager2006unsupervised proposed a generative model but their parameter space consisted of all possible linkings of syntactic constituents and semantic roles, which made unsupervised learning difficult and a separate language-specific rule based method had to be used to constrain this space. Other proposed models include an iterative split-merge algorithm BIBREF18 and a graph-partitioning based approach BIBREF1 . marquez2008semantic provide a good overview of the supervised SRL systems.
## Conclusions
We propose a Bayesian model of semantic role induction (SRI) that uses crosslingual latent variables to capture role alignments in parallel corpora. The crosslingual latent variables capture correlations between roles in different languages, and regularize the parameter estimates of the monolingual models. Because this is a joint Bayesian model of multilingual SRI, we can apply the same model to a variety of training scenarios just by changing the inference procedure appropriately. We evaluate monolingual SRI with a large unlabeled dataset, bilingual SRI with a parallel corpus, bilingual SRI with annotations available for the source language, and monolingual SRI with a small labeled dataset. Increasing the amount of monolingual unlabeled data significantly improves SRI in German but not in English. Adding word alignments in parallel sentences results in small, non significant improvements, even if there is some labeled data available in the source language. This difficulty in showing the usefulness of parallel corpora for SRI may be due to the current assumptions about role alignments, which mean that only a small percentage of roles are aligned. Further analyses reveals that annotating small amounts of data can easily outperform the performance gains obtained by adding large unlabeled dataset as well as adding parallel corpora.
Future work includes training on different language pairs, on more than two languages, and with more inclusive models of role alignment.
## Acknowledgments
This work was funded by the Swiss NSF grant 200021_125137 and EC FP7 grant PARLANCE.
| 15
|
1603.04513
|
Multichannel Variable-Size Convolution for Sentence Classification
|
# Multichannel Variable-Size Convolution for Sentence Classification
## Abstract
We propose MVCNN, a convolution neural network (CNN) architecture for sentence classification. It (i) combines diverse versions of pretrained word embeddings and (ii) extracts features of multigranular phrases with variable-size convolution filters. We also show that pretraining MVCNN is critical for good performance. MVCNN achieves state-of-the-art performance on four tasks: on small-scale binary, small-scale multi-class and largescale Twitter sentiment prediction and on subjectivity classification.
## Introduction
Different sentence classification tasks are crucial for many Natural Language Processing (NLP) applications. Natural language sentences have complicated structures, both sequential and hierarchical, that are essential for understanding them. In addition, how to decode and compose the features of component units, including single words and variable-size phrases, is central to the sentence classification problem.
In recent years, deep learning models have achieved remarkable results in computer vision BIBREF0 , speech recognition BIBREF1 and NLP BIBREF2 . A problem largely specific to NLP is how to detect features of linguistic units, how to conduct composition over variable-size sequences and how to use them for NLP tasks BIBREF3 , BIBREF4 , BIBREF5 . socher2011dynamic proposed recursive neural networks to form phrases based on parsing trees. This approach depends on the availability of a well performing parser; for many languages and domains, especially noisy domains, reliable parsing is difficult. Hence, convolution neural networks (CNN) are getting increasing attention, for they are able to model long-range dependencies in sentences via hierarchical structures BIBREF6 , BIBREF5 , BIBREF7 . Current CNN systems usually implement a convolution layer with fixed-size filters (i.e., feature detectors), in which the concrete filter size is a hyperparameter. They essentially split a sentence into multiple sub-sentences by a sliding window, then determine the sentence label by using the dominant label across all sub-sentences. The underlying assumption is that the sub-sentence with that granularity is potentially good enough to represent the whole sentence. However, it is hard to find the granularity of a “good sub-sentence” that works well across sentences. This motivates us to implement variable-size filters in a convolution layer in order to extract features of multigranular phrases.
Breakthroughs of deep learning in NLP are also based on learning distributed word representations – also called “word embeddings” – by neural language models BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 . Word embeddings are derived by projecting words from a sparse, 1-of- $V$ encoding ( $V$ : vocabulary size) onto a lower dimensional and dense vector space via hidden layers and can be interpreted as feature extractors that encode semantic and syntactic features of words.
Many papers study the comparative performance of different versions of word embeddings, usually learned by different neural network (NN) architectures. For example, chen2013expressive compared HLBL BIBREF9 , SENNA BIBREF2 , Turian BIBREF13 and Huang BIBREF14 , showing great variance in quality and characteristics of the semantics captured by the tested embedding versions. hill2014not showed that embeddings learned by neural machine translation models outperform three representative monolingual embedding versions: skip-gram BIBREF15 , GloVe BIBREF16 and C&W BIBREF3 in some cases. These prior studies motivate us to explore combining multiple versions of word embeddings, treating each of them as a distinct description of words. Our expectation is that the combination of these embedding versions, trained by different NNs on different corpora, should contain more information than each version individually. We want to leverage this diversity of different embedding versions to extract higher quality sentence features and thereby improve sentence classification performance.
The letters “M” and “V” in the name “MVCNN” of our architecture denote the multichannel and variable-size convolution filters, respectively. “Multichannel” employs language from computer vision where a color image has red, green and blue channels. Here, a channel is a description by an embedding version.
For many sentence classification tasks, only relatively small training sets are available. MVCNN has a large number of parameters, so that overfitting is a danger when they are trained on small training sets. We address this problem by pretraining MVCNN on unlabeled data. These pretrained weights can then be fine-tuned for the specific classification task.
In sum, we attribute the success of MVCNN to: (i) designing variable-size convolution filters to extract variable-range features of sentences and (ii) exploring the combination of multiple public embedding versions to initialize words in sentences. We also employ two “tricks” to further enhance system performance: mutual learning and pretraining.
In remaining parts, Section "Related Work" presents related work. Section "Model Description" gives details of our classification model. Section "Model Enhancements" introduces two tricks that enhance system performance: mutual-learning and pretraining. Section "Experiments" reports experimental results. Section "Conclusion" concludes this work.
## Related Work
Much prior work has exploited deep neural networks to model sentences.
blacoe2012comparison represented a sentence by element-wise addition, multiplication, or recursive autoencoder over embeddings of component single words. yin2014exploration extended this approach by composing on words and phrases instead of only single words.
collobert2008unified and yu2014deep used one layer of convolution over phrases detected by a sliding window on a target sentence, then used max- or average-pooling to form a sentence representation.
blunsom2014convolutional stacked multiple layers of one-dimensional convolution by dynamic k-max pooling to model sentences. We also adopt dynamic k-max pooling while our convolution layer has variable-size filters.
kimEMNLP2014 also studied multichannel representation and variable-size filters. Differently, their multichannel relies on a single version of pretrained embeddings (i.e., pretrained Word2Vec embeddings) with two copies: one is kept stable and the other one is fine-tuned by backpropagation. We develop this insight by incorporating diverse embedding versions. Additionally, their idea of variable-size filters is further developed.
le2014distributed initialized the representation of a sentence as a parameter vector, treating it as a global feature and combining this vector with the representations of context words to do word prediction. Finally, this fine-tuned vector is used as representation of this sentence. Apparently, this method can only produce generic sentence representations which encode no task-specific features.
Our work is also inspired by studies that compared the performance of different word embedding versions or investigated the combination of them. For example, turian2010word compared Brown clusters, C&W embeddings and HLBL embeddings in NER and chunking tasks. They found that Brown clusters and word embeddings both can improve the accuracy of supervised NLP systems; and demonstrated empirically that combining different word representations is beneficial. luo2014pre adapted CBOW BIBREF12 to train word embeddings on different datasets: free text documents from Wikipedia, search click-through data and user query data, showing that combining them gets stronger results than using individual word embeddings in web search ranking and word similarity task. However, these two papers either learned word representations on the same corpus BIBREF13 or enhanced the embedding quality by extending training corpora, not learning algorithms BIBREF17 . In our work, there is no limit to the type of embedding versions we can use and they leverage not only the diversity of corpora, but also the different principles of learning algorithms.
## Model Description
We now describe the architecture of our model MVCNN, illustrated in Figure 1 .
Multichannel Input. The input of MVCNN includes multichannel feature maps of a considered sentence, each is a matrix initialized by a different embedding version. Let $s$ be sentence length, $d$ dimension of word embeddings and $c$ the total number of different embedding versions (i.e., channels). Hence, the whole initialized input is a three-dimensional array of size $c\times d\times s$ . Figure 1 depicts a sentence with $s=12$ words. Each word is initialized by $c=5$ embeddings, each coming from a different channel. In implementation, sentences in a mini-batch will be padded to the same length, and unknown words for corresponding channel are randomly initialized or can acquire good initialization from the mutual-learning phase described in next section.
Multichannel initialization brings two advantages: 1) a frequent word can have $c$ representations in the beginning (instead of only one), which means it has more available information to leverage; 2) a rare word missed in some embedding versions can be “made up” by others (we call it “partially known word”). Therefore, this kind of initialization is able to make use of information about partially known words, without having to employ full random initialization or removal of unknown words. The vocabulary of the binary sentiment prediction task described in experimental part contains 5232 words unknown in HLBL embeddings, 4273 in Huang embeddings, 3299 in GloVe embeddings, 4136 in SENNA embeddings and 2257 in Word2Vec embeddings. But only 1824 words find no embedding from any channel! Hence, multichannel initialization can considerably reduce the number of unknown words.
Convolution Layer (Conv). For convenience, we first introduce how this work uses a convolution layer on one input feature map to generate one higher-level feature map. Given a sentence of length $s$ : $w_1, w_2, \ldots , w_s$ ; $\mathbf {w}_i\in \mathbb {R}^{d}$ denotes the embedding of word $w_i$ ; a convolution layer uses sliding filters to extract local features of that sentence. The filter width $l$ is a parameter. We first concatenate the initialized embeddings of $l$ consecutive words ( $\mathbf {w}_{i-l+1},
\ldots , \mathbf {w}_i$ ) as $\mathbf {c}_i\in \mathbb {R}^{ld}$ $(1\le i <s+l)$ , then generate the feature value of this phrase as $\textbf {p}_i$ (the whole vector $w_1, w_2, \ldots , w_s$0 contains all the local features) using a tanh activation function and a linear projection vector $w_1, w_2, \ldots , w_s$1 as:
$$\mathbf {p}_i=\mathrm {tanh}(\mathbf {v}^\mathrm {T}\mathbf {c}_i+b)$$ (Eq. 2)
More generally, convolution operation can deal with multiple input feature maps and can be stacked to yield feature maps of increasing layers. In each layer, there are usually multiple filters of the same size, but with different weights BIBREF4 . We refer to a filter with a specific set of weights as a kernel. The goal is often to train a model in which different kernels detect different kinds of features of a local region. However, this traditional way can not detect the features of regions of different granularity. Hence we keep the property of multi-kernel while extending it to variable-size in the same layer.
As in CNN for object recognition, to increase the number of kernels of a certain layer, multiple feature maps may be computed in parallel at the same layer. Further, to increase the size diversity of kernels in the same layer, more feature maps containing various-range dependency features can be learned. We denote a feature map of the $i^{\mathrm {th}}$ layer by $\mathbf {F}_i$ , and assume totally $n$ feature maps exist in layer $i-1$ : $\mathbf {F}_{i-1}^1, \ldots ,
\mathbf {F}_{i-1}^n$ . Considering a specific filter size $l$ in layer $i$ , each feature map $\mathbf {F}_{i,l}^j$ is computed by convolving a distinct set of filters of size $l$ , arranged in a matrix $\mathbf {V}_{i,l}^{j,k}$ , with each feature map $\mathbf {F}_i$0 and summing the results:
$$\mathbf {F}_{i,l}^j=\sum ^n_{k=1}\mathbf {V}_{i,l}^{j,k}*\mathbf {F}^k_{i-1}$$ (Eq. 3)
where $*$ indicates the convolution operation and $j$ is the index of a feature map in layer $i$ . The weights in $\mathbf {V}$ form a rank 4 tensor.
Note that we use wide convolution in this work: it means word representations $\mathbf {w}_g$ for $g\le 0$ or $g\ge s+1$ are actually zero embeddings. Wide convolution enables that each word can be detected by all filter weights in $\mathbf {V}$ .
In Figure 1 , the first convolution layer deals with an input with $n=5$ feature maps. Its filters have sizes 3 and 5 respectively (i.e., $l=3, 5$ ), and each filter has $j=3$ kernels. This means this convolution layer can detect three kinds of features of phrases with length 3 and 5, respectively.
DCNN in BIBREF4 used one-dimensional convolution: each higher-order feature is produced from values of a single dimension in the lower-layer feature map. Even though that work proposed folding operation to model the dependencies between adjacent dimensions, this type of dependency modeling is still limited. Differently, convolution in present work is able to model dependency across dimensions as well as adjacent words, which obviates the need for a folding step. This change also means our model has substantially fewer parameters than the DCNN since the output of each convolution layer is smaller by a factor of $d$ .
Dynamic k-max Pooling. blunsom2014convolutional pool the $k$ most active features compared with simple max (1-max) pooling BIBREF2 . This property enables it to connect multiple convolution layers to form a deep architecture to extract high-level abstract features. In this work, we directly use it to extract features for variable-size feature maps. For a given feature map in layer $i$ , dynamic k-max pooling extracts $k_{i}$ top values from each dimension and $k_{top}$ top values in the top layer. We set
$$\nonumber k_{i}=\mathrm {max}(k_{top}, \lceil \frac{L-i}{L}s\rceil )$$ (Eq. 5)
where $i\in \lbrace 1,2,\ldots \, L\rbrace $ is the order of convolution layer from bottom to top in Figure 1 ; $L$ is the total numbers of convolution layers; $k_{top}$ is a constant determined empirically, we set it to 4 as BIBREF4 .
As a result, the second convolution layer in Figure 1 has an input with two same-size feature maps, one results from filter size 3, one from filter size 5. The values in the two feature maps are for phrases with different granularity. The motivation of this convolution layer lies in that a feature reflected by a short phrase may be not trustworthy while the longer phrase containing the short one is trustworthy, or the long phrase has no trustworthy feature while its component short phrase is more reliable. This and even higher-order convolution layers therefore can make a trade-off between the features of different granularity.
Hidden Layer. On the top of the final k-max pooling, we stack a fully connected layer to learn sentence representation with given dimension (e.g., $d$ ).
Logistic Regression Layer. Finally, sentence representation is forwarded into logistic regression layer for classification.
In brief, our MVCNN model learns from BIBREF4 to use dynamic k-max pooling to stack multiple convolution layers, and gets insight from BIBREF5 to investigate variable-size filters in a convolution layer. Compared to BIBREF4 , MVCNN has rich feature maps as input and as output of each convolution layer. Its convolution operation is not only more flexible to extract features of variable-range phrases, but also able to model dependency among all dimensions of representations. MVCNN extends the network in BIBREF5 by hierarchical convolution architecture and further exploration of multichannel and variable-size feature detectors.
## Model Enhancements
This part introduces two training tricks that enhance the performance of MVCNN in practice.
Mutual-Learning of Embedding Versions. One observation in using multiple embedding versions is that they have different vocabulary coverage. An unknown word in an embedding version may be a known word in another version. Thus, there exists a proportion of words that can only be partially initialized by certain versions of word embeddings, which means these words lack the description from other versions.
To alleviate this problem, we design a mutual-learning regime to predict representations of unknown words for each embedding version by learning projections between versions. As a result, all embedding versions have the same vocabulary. This processing ensures that more words in each embedding version receive a good representation, and is expected to give most words occurring in a classification dataset more comprehensive initialization (as opposed to just being randomly initialized).
Let $c$ be the number of embedding versions in consideration, $V_1, V_2, \ldots , V_i, \ldots , V_c$ their vocabularies, $V^*=\cup ^c_{i=1} V_i$ their union, and $V_i^-=V^*\backslash V_i$ ( $i=1, \ldots , c$ ) the vocabulary of unknown words for embedding version $i$ . Our goal is to learn embeddings for the words in $V_i^-$ by knowledge from the other $c-1$ embedding versions.
We use the overlapping vocabulary between $V_i$ and $V_j$ , denoted as $V_{ij}$ , as training set, formalizing a projection $f_{ij}$ from space $V_i$ to space $V_j$ ( $i\ne j; i,
j\in \lbrace 1,2,\ldots ,c\rbrace $ ) as follows:
$$\mathbf {\hat{w}}_j=\mathbf {M}_{ij}\mathbf {w}_i$$ (Eq. 6)
where $\mathbf {M}_{ij}\in \mathbb {R}^{d\times d}$ , $\mathbf {w}_i\in \mathbb {R}^d$ denotes the representation of word $w$ in space $V_i$ and $\mathbf {\hat{w}}_j$ is the projected (or learned) representation of word $w$ in space $V_j$ . Squared error between $\mathbf {w}_j$ and $\mathbf {\hat{w}}_j$ is the training loss to minimize. We use $\hat{\mathbf {}{w}_j=f_{ij}(\mathbf {w}_i) to reformat
Equation \ref {equ:proj}.
Totally c(c-1)/2 projections f_{ij} are trained, each on the
vocabulary intersection V_{ij}.
}Let $ w $\mathbf {w}_i\in \mathbb {R}^d$0 Vi $\mathbf {w}_i\in \mathbb {R}^d$1 V1, V2, ..., Vk $\mathbf {w}_i\in \mathbb {R}^d$2 w $\mathbf {w}_i\in \mathbb {R}^d$3 Vi $\mathbf {w}_i\in \mathbb {R}^d$4 k $\mathbf {w}_i\in \mathbb {R}^d$5 f1i(w1) $\mathbf {w}_i\in \mathbb {R}^d$6 f2i(w2) $\mathbf {w}_i\in \mathbb {R}^d$7 ... $\mathbf {w}_i\in \mathbb {R}^d$8 fki(wk) $\mathbf {w}_i\in \mathbb {R}^d$9 V1, V2, ..., Vk $w$0 Vi $w$1 f1i(w1) $w$2 f2i(w2) $w$3 ... $w$4 fki(wk) $w$5 w $w$6 Vi $w$7 w $w$8 Vi $w$9
As discussed in Section "Model Description" , we found that for the binary sentiment classification dataset, many words were unknown in at least one embedding version. But of these words, a total of 5022 words did have coverage in another embedding version and so will benefit from mutual-learning. In the experiments, we will show that this is a very effective method to learn representations for unknown words that increases system performance if learned representations are used for initialization.
Pretraining. Sentence classification systems are usually implemented as supervised training regimes where training loss is between true label distribution and predicted label distribution. In this work, we use pretraining on the unlabeled data of each task and show that it can increase the performance of classification systems.
Figure 1 shows our pretraining setup. The “sentence representation” – the output of “Fully connected” hidden layer – is used to predict the component words (“on” in the figure) in the sentence (instead of predicting the sentence label Y/N as in supervised learning). Concretely, the sentence representation is averaged with representations of some surrounding words (“the”, “cat”, “sat”, “the”, “mat”, “,” in the figure) to predict the middle word (“on”).
Given sentence representation $\mathbf {s}\in \mathbb {R}^d$ and initialized representations of $2t$ context words ( $t$ left words and $t$ right words): $\mathbf {w}_{i-t}$ , $\ldots $ , $\mathbf {w}_{i-1}$ , $\mathbf {w}_{i+1}$ , $\ldots $ , $\mathbf {w}_{i+t}$ ; $2t$0 , we average the total $2t$1 vectors element-wise, depicted as “Average” operation in Figure 1 . Then, this resulting vector is treated as a predicted representation of the middle word and is used to find the true middle word by means of noise-contrastive estimation (NCE) BIBREF18 . For each true example, 10 noise words are sampled.
Note that in pretraining, there are three places where each word needs initialization. (i) Each word in the sentence is initialized in the “Multichannel input” layer to the whole network. (ii) Each context word is initialized as input to the average layer (“Average” in the figure). (iii) Each target word is initialized as the output of the “NCE” layer (“on” in the figure). In this work, we use multichannel initialization for case (i) and random initialization for cases (ii) and (iii). Only fine-tuned multichannel representations (case (i)) are kept for subsequent supervised training.
The rationale for this pretraining is similar to auto-encoder: for an object composed of smaller-granular elements, the representations of the whole object and its components can learn each other. The CNN architecture learns sentence features layer by layer, then those features are justified by all constituent words.
During pretraining, all the model parameters, including mutichannel input, convolution parameters and fully connected layer, will be updated until they are mature to extract the sentence features. Subsequently, the same sets of parameters will be fine-tuned for supervised classification tasks.
In sum, this pretraining is designed to produce good initial values for both model parameters and word embeddings. It is especially helpful for pretraining the embeddings of unknown words.
## Experiments
We test the network on four classification tasks. We begin by specifying aspects of the implementation and the training of the network. We then report the results of the experiments.
## Hyperparameters and Training
In each of the experiments, the top of the network is a logistic regression that predicts the probability distribution over classes given the input sentence. The network is trained to minimize cross-entropy of predicted and true distributions; the objective includes an $L_2$ regularization term over the parameters. The set of parameters comprises the word embeddings, all filter weights and the weights in fully connected layers. A dropout operation BIBREF19 is put before the logistic regression layer. The network is trained by back-propagation in mini-batches and the gradient-based optimization is performed using the AdaGrad update rule BIBREF20
In all data sets, the initial learning rate is 0.01, dropout probability is 0.8, $L_2$ weight is $5\cdot 10^{-3}$ , batch size is 50. In each convolution layer, filter sizes are {3, 5, 7, 9} and each filter has five kernels (independent of filter size).
## Datasets and Experimental Setup
Standard Sentiment Treebank BIBREF21 . This small-scale dataset includes two tasks predicting the sentiment of movie reviews. The output variable is binary in one experiment and can have five possible outcomes in the other: {negative, somewhat negative, neutral, somewhat positive, positive}. In the binary case, we use the given split of 6920 training, 872 development and 1821 test sentences. Likewise, in the fine-grained case, we use the standard 8544/1101/2210 split. socher2013recursive used the Stanford Parser BIBREF22 to parse each sentence into subphrases. The subphrases were then labeled by human annotators in the same way as the sentences were labeled. Labeled phrases that occur as subparts of the training sentences are treated as independent training instances as in BIBREF23 , BIBREF4 .
Sentiment140 BIBREF24 . This is a large-scale dataset of tweets about sentiment classification, where a tweet is automatically labeled as positive or negative depending on the emoticon that occurs in it. The training set consists of 1.6 million tweets with emoticon-based labels and the test set of about 400 hand-annotated tweets. We preprocess the tweets minimally as follows. 1) The equivalence class symbol “url” (resp. “username”) replaces all URLs (resp. all words that start with the @ symbol, e.g., @thomasss). 2) A sequence of $k>2$ repetitions of a letter $c$ (e.g., “cooooooool”) is replaced by two occurrences of $c$ (e.g., “cool”). 3) All tokens are lowercased.
Subj. Subjectivity classification dataset released by BIBREF25 has 5000 subjective sentences and 5000 objective sentences. We report the result of 10-fold cross validation as baseline systems did.
In this work, we use five embedding versions, as shown in Table 1 , to initialize words. Four of them are directly downloaded from the Internet. (i) HLBL. Hierarchical log-bilinear model presented by mnih2009scalable and released by turian2010word; size: 246,122 word embeddings; training corpus: RCV1 corpus, one year of Reuters English newswire from August 1996 to August 1997. (ii) Huang. huang2012improving incorporated global context to deal with challenges raised by words with multiple meanings; size: 100,232 word embeddings; training corpus: April 2010 snapshot of Wikipedia. (iii) GloVe. Size: 1,193,514 word embeddings; training corpus: a Twitter corpus of 2B tweets with 27B tokens. (iv) SENNA. Size: 130,000 word embeddings; training corpus: Wikipedia. Note that we use their 50-dimensional embeddings. (v) Word2Vec. It has no 50-dimensional embeddings available online. We use released code to train skip-gram on English Gigaword Corpus BIBREF26 with setup: window size 5, negative sampling, sampling rate $10^{-3}$ , threads 12. It is worth emphasizing that above embeddings sets are derived on different corpora with different algorithms. This is the very property that we want to make use of to promote the system performance.
Table 2 shows the number of unknown words in each task when using corresponding embedding version to initialize (rows “HLBL”, “Huang”, “Glove”, “SENNA”, “W2V”) and the number of words fully initialized by five embedding versions (“Full hit” row), the number of words partially initialized (“Partial hit” row) and the number of words that cannot be initialized by any of the embedding versions (“No hit” row).
About 30% of words in each task have partially initialized embeddings and our mutual-learning is able to initialize the missing embeddings through projections. Pretraining is expected to learn good representations for all words, but pretraining is especially important for words without initialization (“no hit”); a particularly clear example for this is the Senti140 task: 236,484 of 387,877 words or 61% are in the “no hit” category.
Table 3 compares results on test of MVCNN and its variants with other baselines in the four sentence classification tasks. Row 34, “MVCNN (overall)”, shows performance of the best configuration of MVCNN, optimized on dev. This version uses five versions of word embeddings, four filter sizes (3, 5, 7, 9), both mutual-learning and pretraining, three convolution layers for Senti140 task and two convolution layers for the other tasks. Overall, our system gets the best results, beating all baselines.
The table contains five blocks from top to bottom. Each block investigates one specific configurational aspect of the system. All results in the five blocks are with respect to row 34, “MVCNN (overall)”; e.g., row 19 shows what happens when HLBL is removed from row 34, row 28 shows what happens when mutual learning is removed from row 34 etc.
The block “baselines” (1–18) lists some systems representative of previous work on the corresponding datasets, including the state-of-the-art systems (marked as italic). The block “versions” (19–23) shows the results of our system when one of the embedding versions was not used during training. We want to explore to what extend different embedding versions contribute to performance. The block “filters” (24–27) gives the results when individual filter width is discarded. It also tells us how much a filter with specific size influences. The block “tricks” (28–29) shows the system performance when no mutual-learning or no pretraining is used. The block “layers” (30–33) demonstrates how the system performs when it has different numbers of convolution layers.
From the “layers” block, we can see that our system performs best with two layers of convolution in Standard Sentiment Treebank and Subjectivity Classification tasks (row 31), but with three layers of convolution in Sentiment140 (row 32). This is probably due to Sentiment140 being a much larger dataset; in such a case deeper neural networks are beneficial.
The block “tricks” demonstrates the effect of mutual-learning and pretraining. Apparently, pretraining has a bigger impact on performance than mutual-learning. We speculate that it is because pretraining can influence more words and all learned word embeddings are tuned on the dataset after pretraining.
The block “filters” indicates the contribution of each filter size. The system benefits from filters of each size. Sizes 5 and 7 are most important for high performance, especially 7 (rows 25 and 26).
In the block “versions”, we see that each embedding version is crucial for good performance: performance drops in every single case. Though it is not easy to compare fairly different embedding versions in NLP tasks, especially when those embeddings were trained on different corpora of different sizes using different algorithms, our results are potentially instructive for researchers making decision on which embeddings to use for their own tasks.
## Conclusion
This work presented MVCNN, a novel CNN architecture for sentence classification. It combines multichannel initialization – diverse versions of pretrained word embeddings are used – and variable-size filters – features of multigranular phrases are extracted with variable-size convolution filters. We demonstrated that multichannel initialization and variable-size filters enhance system performance on sentiment classification and subjectivity classification tasks.
## Future Work
As pointed out by the reviewers the success of the multichannel approach is likely due to a combination of several quite different effects.
First, there is the effect of the embedding learning algorithm. These algorithms differ in many aspects, including in sensitivity to word order (e.g., SENNA: yes, word2vec: no), in objective function and in their treatment of ambiguity (explicitly modeled only by huang2012improving.
Second, there is the effect of the corpus. We would expect the size and genre of the corpus to have a big effect even though we did not analyze this effect in this paper.
Third, complementarity of word embeddings is likely to be more useful for some tasks than for others. Sentiment is a good application for complementary word embeddings because solving this task requires drawing on heterogeneous sources of information, including syntax, semantics and genre as well as the core polarity of a word. Other tasks like part of speech (POS) tagging may benefit less from heterogeneity since the benefit of embeddings in POS often comes down to making a correct choice between two alternatives – a single embedding version may be sufficient for this.
We plan to pursue these questions in future work.
## Acknowledgments
Thanks to CIS members and anonymous reviewers for constructive comments. This work was supported by Baidu (through a Baidu scholarship awarded to Wenpeng Yin) and by Deutsche Forschungsgemeinschaft (grant DFG SCHU 2246/8-2, SPP 1335).
| 10
|
1604.00400
|
Revisiting Summarization Evaluation for Scientific Articles
|
# Revisiting Summarization Evaluation for Scientific Articles
## Abstract
Evaluation of text summarization approaches have been mostly based on metrics that measure similarities of system generated summaries with a set of human written gold-standard summaries. The most widely used metric in summarization evaluation has been the ROUGE family. ROUGE solely relies on lexical overlaps between the terms and phrases in the sentences; therefore, in cases of terminology variations and paraphrasing, ROUGE is not as effective. Scientific article summarization is one such case that is different from general domain summarization (e.g. newswire data). We provide an extensive analysis of ROUGE's effectiveness as an evaluation metric for scientific summarization; we show that, contrary to the common belief, ROUGE is not much reliable in evaluating scientific summaries. We furthermore show how different variants of ROUGE result in very different correlations with the manual Pyramid scores. Finally, we propose an alternative metric for summarization evaluation which is based on the content relevance between a system generated summary and the corresponding human written summaries. We call our metric SERA (Summarization Evaluation by Relevance Analysis). Unlike ROUGE, SERA consistently achieves high correlations with manual scores which shows its effectiveness in evaluation of scientific article summarization.
## Introduction
Automatic text summarization has been an active research area in natural language processing for several decades. To compare and evaluate the performance of different summarization systems, the most intuitive approach is assessing the quality of the summaries by human evaluators. However, manual evaluation is expensive and the obtained results are subjective and difficult to reproduce BIBREF0 . To address these problems, automatic evaluation measures for summarization have been proposed. Rouge BIBREF1 is one of the first and most widely used metrics in summarization evaluation. It facilitates evaluation of system generated summaries by comparing them to a set of human written gold-standard summaries. It is inspired by the success of a similar metric Bleu BIBREF2 which is being used in Machine Translation (MT) evaluation. The main success of Rouge is due to its high correlation with human assessment scores on standard benchmarks BIBREF1 . Rouge has been used as one of the main evaluation metrics in later summarization benchmarks such as TAC[1] BIBREF3 .
[1]Text Analysis Conference (TAC) is a series of workshops for evaluating research in Natural Language Processing
Since the establishment of Rouge, almost all research in text summarization have used this metric as the main means for evaluating the quality of the proposed approaches. The public availability of Rouge as a toolkit for summarization evaluation has contributed to its wide usage. While Rouge has originally shown good correlations with human assessments, the study of its effectiveness was only limited to a few benchmarks on news summarization data (DUC[2] 2001-2003 benchmarks). Since 2003, summarization has grown to much further domains and genres such as scientific documents, social media and question answering. While there is not enough compelling evidence about the effectiveness of Rouge on these other summarization tasks, published research is almost always evaluated by Rouge. In addition, Rouge has a large number of possible variants and the published research often (arbitrarily) reports only a few of these variants.
[2]Document Understanding Conference (DUC) was one of NIST workshops that provided infrastructure for evaluation of text summarization methodologies (http://duc.nist.gov/).
By definition, Rouge solely relies on lexical overlaps (such as n-gram and sequence overlaps) between the system generated and human written gold-standard summaries. Higher lexical overlaps between the two show that the system generated summary is of higher quality. Therefore, in cases of terminology nuances and paraphrasing, Rouge is not accurate in estimating the quality of the summary.
We study the effectiveness of Rouge for evaluating scientific summarization. Scientific summarization targets much more technical and focused domains in which the goal is providing summaries for scientific articles. Scientific articles are much different than news articles in elements such as length, complexity and structure. Thus, effective summarization approaches usually have much higher compression rate, terminology variations and paraphrasing BIBREF4 .
Scientific summarization has attracted more attention recently (examples include works by abu2011coherent, qazvinian2013generating, and cohan2015scientific). Thus, it is important to study the validity of existing methodologies applied to the evaluation of news article summarization for this task. In particular, we raise the important question of how effective is Rouge, as an evaluation metric for scientific summarization? We answer this question by comparing Rouge scores with semi-manual evaluation score (Pyramid) in TAC 2014 scientific summarization dataset[1]. Results reveal that, contrary to the common belief, correlations between Rouge and the Pyramid scores are weak, which challenges its effectiveness for scientific summarization. Furthermore, we show a large variance of correlations between different Rouge variants and the manual evaluations which further makes the reliability of Rouge for evaluating scientific summaries less clear. We then propose an evaluation metric based on relevance analysis of summaries which aims to overcome the limitation of high lexical dependence in Rouge. We call our metric Sera (Summarization Evaluation by Relevance Analysis). Results show that the proposed metric achieves higher and more consistent correlations with semi-manual assessment scores.
[1]http://www.nist.gov/tac/2014/BiomedSumm/
Our contributions are as follows:
[2]The annotations can be accessed via the following repository: https://github.com/acohan/TAC-pyramid-Annotations/
## Summarization evaluation by Rouge
Rouge has been the most widely used family of metrics in summarization evaluation. In the following, we briefly describe the different variants of Rouge:
Rouge-L, Rouge-W, Rouge-S and Rouge-SU were later extended to consider both the recall and precision. In calculating Rouge, stopword removal or stemming can also be considered, resulting in more variants.
In the summarization literature, despite the large number of variants of Rouge, only one or very few of these variants are often chosen (arbitrarily) for evaluation of the quality of the summarization approaches. When Rouge was proposed, the original variants were only recall-oriented and hence the reported correlation results BIBREF1 . The later extension of Rouge family by precision were only reflected in the later versions of the Rouge toolkit and additional evaluation of its effectiveness was not reported. Nevertheless, later published work in summarization adopted this toolkit for its ready implementation and relatively efficient performance.
The original Rouge metrics show high correlations with human judgments of the quality of summaries on the DUC 2001-2003 benchmarks. However, these benchmarks consist of newswire data and are intrinsically very different than other summarization tasks such as summarization of scientific papers. We argue that Rouge is not the best metric for all summarization tasks and we propose an alternative metric for evaluation of scientific summarization. The proposed alternative metric shows much higher and more consistent correlations with manual judgments in comparison with the well-established Rouge.
## Summarization Evaluation by Relevance Analysis (Sera)
Rouge functions based on the assumption that in order for a summary to be of high quality, it has to share many words or phrases with a human gold summary. However, different terminology may be used to refer to the same concepts and thus relying only on lexical overlaps may underrate content quality scores. To overcome this problem, we propose an approach based on the premise that concepts take meanings from the context they are in, and that related concepts co-occur frequently.
Our proposed metric is based on analysis of the content relevance between a system generated summary and the corresponding human written gold-standard summaries. On high level, we indirectly evaluate the content relevance between the candidate summary and the human summary using information retrieval. To accomplish this, we use the summaries as search queries and compare the overlaps of the retrieved results. Larger number of overlaps, suggest that the candidate summary has higher content quality with respect to the gold-standard. This method, enables us to also reward for terms that are not lexically equivalent but semantically related. Our method is based on the well established linguistic premise that semantically related words occur in similar contexts BIBREF5 . The context of the words can be considered as surrounding words, sentences in which they appear or the documents. For scientific summarization, we consider the context of the words as the scientific articles in which they appear. Thus, if two concepts appear in identical set of articles, they are semantically related. We consider the two summaries as similar if they refer to same set of articles even if the two summaries do not have high lexical overlaps. To capture if a summary relates to a article, we use information retrieval by considering the summaries as queries and the articles as documents and we rank the articles based on their relatedness to a given summary. For a given pair of system summary and the gold summary, similar rankings of the retrieved articles suggest that the summaries are semantically related, and thus the system summary is of higher quality.
Based on the domain of interest, we first construct an index from a set of articles in the same domain. Since TAC 2014 was focused on summarization in the biomedical domain, our index also comprises of biomedical articles. Given a candidate summary INLINEFORM0 and a set of gold summaries INLINEFORM1 ( INLINEFORM2 ; INLINEFORM3 is the total number of human summaries), we submit the candidate summary and gold summaries to the search engine as queries and compare their ranked results. Let INLINEFORM4 be the entire index which comprises of INLINEFORM5 total documents.
Let INLINEFORM0 be the ranked list of retrieved documents for candidate summary INLINEFORM1 , and INLINEFORM2 the ranked list of results for the gold summary INLINEFORM3 . These lists of results are based on a rank cut-off point INLINEFORM4 that is a parameter of the system. We provide evaluation results on different choices of cut-off point INLINEFORM5 in the Section SECREF5 We consider the following two scores: (i) simple intersection and (ii) discounted intersection by rankings. The simple intersection just considers the overlaps of the results in the two ranked lists and ignores the rankings. The discounted ranked scores, on the other hand, penalizes ranking differences between the two result sets. As an example consider the following list of retrieved documents (denoted by INLINEFORM6 s) for a candidate and a gold summary as queries:
Results for candidate summary: INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3
Results for gold summary: INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3
These two sets of results consist of identical documents but the ranking of the retrieved documents differ. Therefore, the simple intersection method assigns a score of 1.0 while in the discounted ranked score, the score will be less than 1.0 (due to ranking differences between the result lists).
We now define the metrics more precisely. Using the above notations, without loss of generality, we assume that INLINEFORM0 . Sera is defined as follows: INLINEFORM1
To also account for the ranked position differences, we modify this score to discount rewards based on rank differences. That is, in ideal score, we want search results from candidate summary ( INLINEFORM0 ) to be the same as results for gold-standard summaries ( INLINEFORM1 ) and the rankings of the results also be the same. If the rankings differ, we discount the reward by log of the differences of the ranks. More specifically, the discounted score (Sera-Dis) is defined as: INLINEFORM2
where, as previously defined, INLINEFORM0 , INLINEFORM1 and INLINEFORM2 are total number of human gold summaries, result list for the candidate summary and result list for the human gold summary, respectively. In addition, INLINEFORM3 shows the INLINEFORM4 th results in the ranked list INLINEFORM5 and INLINEFORM6 is the maximum attainable score used as the normalizing factor.
We use elasticsearch[1], an open-source search engine, for indexing and querying the articles. For retrieval model, we use the Language Modeling retrieval model with Dirichlet smoothing BIBREF6 . Since TAC 2014 benchmark is on summarization of biomedical articles, the appropriate index would be the one constructed from articles in the same domain. Therefore, we use the open access subset of Pubmed[2] which consists of published articles in biomedical literature.
[1]https://github.com/elastic/elasticsearch [2]PubMed is a comprehensive resource of articles and abstracts published in life sciences and biomedical literature http://www.ncbi.nlm.nih.gov/pmc/
We also experiment with different query (re)formulation approaches. Query reformulation is a method in Information Retrieval that aims to refine the query for better retrieval of results. Query reformulation methods often consist of removing ineffective terms and expressions from the query (query reduction) or adding terms to the query that help the retrieval (query expansion). Query reduction is specially important when queries are verbose. Since we use the summaries as queries, the queries are usually long and therefore we consider query reductions.
In our experiments, the query reformulation is done by 3 different ways: (i) Plain: The entire summary without stopwords and numeric values; (ii) Noun Phrases (NP): We only keep the noun phrases as informative concepts in the summary and eliminate all other terms; and (iii) Keywords (KW): We only keep the keywords and key phrases in the summary. For extracting the keywords and keyphrases (with length of up to 3 terms), we extract expressions whose idf[1] values is higher than a predefined threshold that is set as a parameter. We set this threshold to the average idf values of all terms except stopwords. idf values are calculated on the same index that is used for the retrieval.
[1]Inverted Document Frequency
We hypothesize that using only informative concepts in the summary prevents query drift and leads to retrieval of more relevant documents. Noun phrases and keywords are two heuristics for identifying the informative concepts.
## Data
To the best of our knowledge, the only scientific summarization benchmark is from TAC 2014 summarization track. For evaluating the effectiveness of Rouge variants and our metric (Sera), we use this benchmark, which consists of 20 topics each with a biomedical journal article and 4 gold human written summaries.
## Annotations
In the TAC 2014 summarization track, Rouge was suggested as the evaluation metric for summarization and no human assessment was provided for the topics. Therefore, to study the effectiveness of the evaluation metrics, we use the semi-manual Pyramid evaluation framework BIBREF7 , BIBREF8 . In the pyramid scoring, the content units in the gold human written summaries are organized in a pyramid. In this pyramid, the content units are organized in tiers and higher tiers of the pyramid indicate higher importance. The content quality of a given candidate summary is evaluated with respect to this pyramid.
To analyze the quality of the evaluation metrics, following the pyramid framework, we design an annotation scheme that is based on identification of important content units. Consider the following example:
Endogeneous small RNAs (miRNA) were genetically screened and studied to find the miRNAs which are related to tumorigenesis.
In the above example, the underlined expressions are the content units that convey the main meaning of the text. We call these small units, nuggets which are phrases or concepts that are the main contributors to the content quality of the summary.
We asked two human annotators to review the gold summaries and extract content units in these summaries. The pyramid tiers represent the occurrences of nuggets across all the human written gold-standard summaries, and therefore the nuggets are weighted based on these tiers. The intuition is that, if a nugget occurs more frequently in the human summaries, it is a more important contributor (thus belongs to higher tier in the pyramid). Thus, if a candidate summary contains this nugget, it should be rewarded more. An example of the nuggets annotations in pyramid framework is shown in Table TABREF12 . In this example, the nugget “cell mutation” belongs to the 4th tier and it suggests that the “cell mutation” nugget is a very important representative of the content of the corresponding document.
Let INLINEFORM0 define the tiers of the pyramid with INLINEFORM1 being the bottom tier and INLINEFORM2 the top tier. Let INLINEFORM3 be the number of the nuggets in the candidate summary that appear in the tier INLINEFORM4 . Then the pyramid score INLINEFORM5 of the candidate summary will be: INLINEFORM6
where INLINEFORM0 is the maximum attainable score used for normalizing the scores: INLINEFORM1
where INLINEFORM0 is the total number of nuggets in the summary and INLINEFORM1 .
We release the pyramid annotations of the TAC 2014 dataset through a public repository[2].
[2]https://github.com/acohan/TAC-pyramid-Annotations
3.1pt
## Summarization approaches
We study the effectiveness of Rouge and our proposed method (Sera) by analyzing the correlations with semi-manual human judgments. Very few teams participated in TAC 2014 summarization track and the official results and the review paper of TAC 2014 systems were never published. Therefore, to evaluate the effectiveness of Rouge, we applied 9 well-known summarization approaches on the TAC 2014 scientific summarization dataset. Obtained Rouge and Sera results of each of these approaches are then correlated with semi-manual human judgments. In the following, we briefly describe each of these summarization approaches.
LexRank BIBREF9 : LexRank finds the most important (central) sentences in a document by using random walks in a graph constructed from the document sentences. In this graph, the sentences are nodes and the similarity between the sentences determines the edges. Sentences are ranked according to their importance. Importance is measured in terms of centrality of the sentence — the total number of edges incident on the node (sentence) in the graph. The intuition behind LexRank is that a document can be summarized using the most central sentences in the document that capture its main aspects.
Latent Semantic Analysis (LSA) based summarization BIBREF10 : In this summarization method, Singular Value Decomposition (SVD) BIBREF11 is used for deriving latent semantic structure of the document. The document is divided into sentences and a term-sentence matrix INLINEFORM0 is constructed. The matrix INLINEFORM1 is then decomposed into a number of linearly-independent singular vectors which represent the latent concepts in the document. This method, intuitively, decomposes the document into several latent topics and then selects the most representative sentences for each of these topics as the summary of the document.
Maximal Marginal Relevance (MMR) BIBREF12 : Maximal Marginal Relevance (MMR) is a greedy strategy for selecting sentences for the summary. Sentences are added iteratively to the summary based on their relatedness to the document as well as their novelty with respect to the current summary.
Citation based summarization BIBREF13 : In this method, citations are used for summarizing an article. Using the LexRank algorithm on the citation network of the article, top sentences are selected for the final summary.
Using frequency of the words BIBREF14 : In this method, which is one the earliest works in text summarization, raw word frequencies are used to estimate the saliency of sentences in the document. The most salient sentences are chosen for the final summary.
SumBasic BIBREF15 : SumBasic is an approach that weights sentences based on the distribution of words that is derived from the document. Sentence selection is applied iteratively by selecting words with highest probability and then finding the highest scoring sentence that contains that word. The word weights are updated after each iteration to prevent selection of similar sentences.
Summarization using citation-context and discourse structure BIBREF16 : In this method, the set of citations to the article are used to find the article sentences that directly reflect those citations (citation-contexts). In addition, the scientific discourse of the article is utilized to capture different aspects of the article. The scientific discourse usually follows a structure in which the authors first describe their hypothesis, then the methods, experiment, results and implications. Sentence selection is based on finding the most important sentences in each of the discourse facets of the document using the MMR heuristic.
KL Divergence BIBREF17 In this method, the document unigram distribution INLINEFORM0 and the summary unigram distributation INLINEFORM1 are considered; the goal is to find a summary whose distribution is very close to the document distribution. The difference of the distributions is captured by the Kullback-Lieber (KL) divergence, denoted by INLINEFORM2 .
Summarization based on Topic Models BIBREF17 : Instead of using unigram distributions for modeling the content distribution of the document and the summary, this method models the document content using an LDA based topic model BIBREF18 . It then uses the KL divergence between the document and the summary content models for selecting sentences for the summary.
## Results and Discussion
We calculated all variants of Rouge scores, our proposed metric, Sera, and the Pyramid score on the generated summaries from the summarizers described in Section SECREF13 . We do not report the Rouge, Sera or pyramid scores of individual systems as it is not the focus of this study. Our aim is to analyze the effectiveness of the evaluation metrics, not the summarization approaches. Therefore, we consider the correlations of the automatic evaluation metrics with the manual Pyramid scores to evaluate their effectiveness; the metrics that show higher correlations with manual judgments are more effective.
Table TABREF23 shows the Pearson, Spearman and Kendall correlation of Rouge and Sera, with pyramid scores. Both Rouge and Sera are calculated with stopwords removed and with stemming. Our experiments with inclusion of stopwords and without stemming showed similar results and thus, we do not include those to avoid redundancy.
## Sera
The results of our proposed method (Sera) are shown in the bottom part of Table TABREF23 . In general, Sera shows better correlation with pyramid scores in comparison with Rouge. We observe that the Pearson correlation of Sera with cut-off point of 5 (shown by Sera-5) is 0.823 which is higher than most of the Rouge variants. Similarly, the Spearman and Kendall correlations of the Sera evaluation score is 0.941 and 0.857 respectively, which are higher than all Rouge correlation values. This shows the effectiveness of the simple variant of our proposed summarization evaluation metric.
Table TABREF23 also shows the results of other Sera variants including discounting and query reformulation methods. Some of these variants are the result of applying query reformulation in the process of document retrieval which are described in section SECREF3 As illustrated, the Noun Phrases (NP) query reformulation at cut-off point of 5 (shown as Sera-np-5) achieves the highest correlations among all the Sera variants ( INLINEFORM0 = INLINEFORM1 , INLINEFORM2 = INLINEFORM3 = INLINEFORM4 ). In the case of Keywords (KW) query reformulation, without using discounting, we can see that there is no positive gain in correlation. However, keywords when applied on the discounted variant of Sera, result in higher correlations.
Discounting has more positive effect when applied on query reformulation-based Sera than on the simple variant of Sera. In the case of discounting and NP query reformulation (Sera-dis-np), we observe higher correlations in comparison with simple Sera. Similarly, in the case of Keywords (KW), positive correlation gain is obtained in most of correlation coefficients. NP without discounting and at cut-off point of 5 (Sera-np-5) shows the highest non-parametric correlation. In addition, the discounted NP at cut-off point of 10 (Sera-np-dis-10) shows the highest parametric correlations.
In general, using NP and KW as heuristics for finding the informative concepts in the summary effectively increases the correlations with the manual scores. Selecting informative terms from long queries results in more relevant documents and prevents query drift. Therefore, the overall similarity between the two summaries (candidate and the human written gold summary) is better captured.
## Rouge
Another important observation is regarding the effectiveness of Rouge scores (top part of Table TABREF23 ). Interestingly, we observe that many variants of Rouge scores do not have high correlations with human pyramid scores. The lowest F-score correlations are for Rouge-1 and Rouge-L (with INLINEFORM0 =0.454). Weak correlation of Rouge-1 shows that matching unigrams between the candidate summary and gold summaries is not accurate in quantifying the quality of the summary. On higher order n-grams, however, we can see that Rouge correlates better with pyramid. In fact, the highest overall INLINEFORM1 is obtained by Rouge-3. Rouge-L and its weighted version Rouge-W, both have weak correlations with pyramid. Skip-bigrams (Rouge-S) and its combination with unigrams (Rouge-SU) also show sub-optimal correlations. Note that INLINEFORM2 and INLINEFORM3 correlations are more reliable in our setup due to the small sample size.
These results confirm our initial hypothesis that Rouge is not accurate estimator of the quality of the summary in scientific summarization. We attribute this to the differences of scientific summarization with general domain summaries. When humans summarize a relatively long research paper, they might use different terminology and paraphrasing. Therefore, Rouge which only relies on term matching between a candidate and a gold summary, is not accurate in quantifying the quality of the candidate summary.
## Correlation of Sera with Rouge
Table TABREF25 shows correlations of our metric Sera with Rouge-2 and Rouge-3, which are the highest correlated Rouge variants with pyramid. We can see that in general, the correlation is not strong. Keyword based reduction variants are the only variants for which the correlation with Rouge is high. Looking at the correlations of KW variants of Sera with pyramid (Table TABREF23 , bottom part), we observe that these variants are also highly correlated with manual evaluation.
## Effect of the rank cut-off point
Finally, Figure FIGREF28 shows INLINEFORM0 correlation of different variants of Sera with pyramid based on selection of different cut-off points ( INLINEFORM1 and INLINEFORM2 correlations result in very similar graphs). When the cut-off point increases, more documents are retrieved for the candidate and the gold summaries, and therefore the final Sera score is more fine-grained. A general observation is that as the search cut-off point increases, the correlation with pyramid scores decreases. This is because when the retrieved result list becomes larger, the probability of including less related documents increases which negatively affects correct estimation of the similarity of the candidate and gold summaries. The most accurate estimations are for metrics with cut-off points of 5 and 10 which are included in the reported results of all variants in Table TABREF23 .
## Related work
Rouge BIBREF1 assesses the content quality of a candidate summary with respect to a set of human gold summaries based on their lexical overlaps. Rouge consists of several variants. Since its introduction, Rouge has been one of the most widely reported metrics in the summarization literature, and its high adoption has been due to its high correlation with human assessment scores in DUC datasets BIBREF1 . However, later research has casted doubts about the accuracy of Rouge against manual evaluations. conroy2008mind analyzed DUC 2005 to 2007 data and showed that while some systems achieve high Rouge scores with respect to human summaries, the linguistic and responsiveness scores of those systems do not correspond to the high Rouge scores.
We studied the effectiveness of Rouge through correlation analysis with manual scores. Besides correlation with human assessment scores, other approaches have been explored for analyzing the effectiveness of summarization evaluation. Rankel:2011 studied the extent to which a metric can distinguish between the human and system generated summaries. They also proposed the use of paired two-sample t-tests and the Wilcoxon signed-rank test as an alternative to Rouge in evaluating several summarizers. Similarly, owczarzak2012assessment proposed the use of multiple binary significance tests between the system summaries for ranking the best summarizers.
Since introduction of Rouge, there have been other efforts for improving automatic summarization evaluation. hovy2006automated proposed an approach based on comparison of so called Basic Elements (BE) between the candidate and reference summaries. BEs were extracted based on syntactic structure of the sentence. The work by conroy2011nouveau was another attempt for improving Rouge for update summarization which combined two different Rouge variants and showed higher correlations with manual judgments for TAC 2008 update summaries.
Apart from the content, other aspects of summarization such as linguistic quality have been also studied. pitler2010automatic evaluated a set of models based on syntactic features, language models and entity coherences for assessing the linguistic quality of the summaries. Machine translation evaluation metrics such as blue have also been compared and contrasted against Rouge BIBREF19 . Despite these works, when gold-standard summaries are available, Rouge is still the most common evaluation metric that is used in the summarization published research. Apart from Rouge's initial good results on the newswire data, the availability of the software and its efficient performance have further contributed to its popularity.
## Conclusions
We provided an analysis of existing evaluation metrics for scientific summarization with evaluation of all variants of Rouge. We showed that Rouge may not be the best metric for summarization evaluation; especially in summaries with high terminology variations and paraphrasing (e.g. scientific summaries). Furthermore, we showed that different variants of Rouge result in different correlation values with human judgments, indicating that not all Rouge scores are equally effective. Among all variants of Rouge, Rouge-2 and Rouge-3 are better correlated with manual judgments in the context of scientific summarization. We furthermore proposed an alternative and more effective approach for scientific summarization evaluation (Summarization Evaluation by Relevance Analysis - Sera). Results revealed that in general, the proposed evaluation metric achieves higher correlations with semi-manual pyramid evaluation scores in comparison with Rouge.
Our analysis on the effectiveness of evaluation measures for scientific summaries was performed using correlations with manual judgments. An alternative approach to follow would be to use statistical significance testing on the ability of the metrics to distinguish between the summarizers (similar to Rankel:2011). We studied the effectiveness of existing summarization evaluation metrics in the scientific text genre and proposed an alternative superior metric. Another extension of this work would be to evaluate automatic summarization evaluation in other genres of text (such as social media). Our proposed method only evaluates the content quality of the summary. Similar to most of existing summarization evaluation metrics, other qualities such as linguistic cohesion, coherence and readability are not captured by this method. Developing metrics that also incorporate these qualities is yet another future direction to follow.
## Acknowledgments
We would like to thank all three anonymous reviewers for their feedback and comments, and Maryam Iranmanesh for helping in annotation. This work was partially supported by National Science Foundation (NSF) through grant CNS-1204347.
| 14
|
1604.05781
|
What we write about when we write about causality: Features of causal statements across large-scale social discourse
|
# What we write about when we write about causality: Features of causal statements across large-scale social discourse
## Abstract
Identifying and communicating relationships between causes and effects is important for understanding our world, but is affected by language structure, cognitive and emotional biases, and the properties of the communication medium. Despite the increasing importance of social media, much remains unknown about causal statements made online. To study real-world causal attribution, we extract a large-scale corpus of causal statements made on the Twitter social network platform as well as a comparable random control corpus. We compare causal and control statements using statistical language and sentiment analysis tools. We find that causal statements have a number of significant lexical and grammatical differences compared with controls and tend to be more negative in sentiment than controls. Causal statements made online tend to focus on news and current events, medicine and health, or interpersonal relationships, as shown by topic models. By quantifying the features and potential biases of causality communication, this study improves our understanding of the accuracy of information and opinions found online.
## Introduction
Social media and online social networks now provide vast amounts of data on human online discourse and other activities BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 . With so much communication taking place online and with social media being capable of hosting powerful misinformation campaigns BIBREF7 such as those claiming vaccines cause autism BIBREF8 , BIBREF9 , it is more important than ever to better understand the discourse of causality and the interplay between online communication and the statement of cause and effect.
Causal inference is a crucial way that humans comprehend the world, and it has been a major focus of philosophy, statistics, mathematics, psychology, and the cognitive sciences. Philosophers such as Hume and Kant have long argued whether causality is a human-centric illusion or the discovery of a priori truth BIBREF10 , BIBREF11 . Causal inference in science is incredibly important, and researchers have developed statistical measures such as Granger causality BIBREF12 , mathematical and probabilistic frameworks BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 , and text mining procedures BIBREF17 , BIBREF18 , BIBREF19 to better infer causal influence from data. In the cognitive sciences, the famous perception experiments of Michotte et al. led to a long line of research exploring the cognitive biases that humans possess when attempting to link cause and effect BIBREF20 , BIBREF21 , BIBREF22 .
How humans understand and communicate cause and effect relationships is complicated, and is influenced by language structure BIBREF23 , BIBREF24 , BIBREF25 , BIBREF26 and sentiment or valence BIBREF27 . A key finding is that the perceived emphasis or causal weight changes between the agent (the grammatical construct responsible for a cause) and the patient (the construct effected by the cause) depending on the types of verbs used to describe the cause and effect. Researchers have hypothesized BIBREF28 that this is because of the innate weighting property of the verbs in the English language that humans use to attribute causes and effects. Another finding is the role of a valence bias: the volume and intensity of causal reasoning may increase due to negative feedback or negative events BIBREF27 .
Despite these long lines of research, causal attributions made via social media or online social networks have not been well studied. The goal of this paper is to explore the language and topics of causal statements in a large corpus of social media taken from Twitter. We hypothesize that language and sentiment biases play a significant role in these statements, and that tools from natural language processing and computational linguistics can be used to study them. We do not attempt to study the factual correctness of these statements or offer any degree of verification, nor do we exhaustively identify and extract all causal statements from these data. Instead, here we focus on statements that are with high certainty causal statements, with the goal to better understand key characteristics about causal statements that differ from everyday online communication.
The rest of this paper is organized as follows: In Sec. "Materials and Methods" we discuss our materials and methods, including the dataset we studied, how we preprocessed that data and extracted a `causal' corpus and a corresponding `control' corpus, and the details of the statistical and language analysis tools we studied these corpora with. In Sec. "Results" we present results using these tools to compare the causal statements to control statements. We conclude with a discussion in Sec. "Discussion" .
## Dataset, filtering, and corpus selection
Data was collected from a 10% uniform sample of Twitter posts made during 2013, specifically the Gardenhose API. Twitter activity consists of short posts called tweets which are limited to 140 characters. Retweets, where users repost a tweet to spread its content, were not considered. (The spread of causal statements will be considered in future work.) We considered only English-language tweets for this study. To avoid cross-language effects, we kept only tweets with a user-reported language of `English' and, as a second constraint, individual tweets needed to match more English stopwords than any other language's set of stopwords. Stopwords considered for each language were determined using NLTK's database BIBREF29 . A tweet will be referred to as a `document' for the rest of this work.
All document text was processed the same way. Punctuation, XML characters, and hyperlinks were removed, as were Twitter-specific “at-mentions” and “hashtags” (see also the Appendix). There is useful information here, but it is either not natural language text, or it is Twitter-specific, or both. Documents were broken into individual words (unigrams) on whitespace. Casing information was retained, as we will use it for our Named Entity analysis, but otherwise all words were considered lowercase only (see also the Appendix). Stemming BIBREF30 and lemmatization BIBREF31 were not performed.
Causal documents were chosen to contain one occurrence only of the exact unigrams: `caused', `causing', or `causes'. The word `cause' was not included due to its use as a popular contraction for `because'. One `cause-word' per document restricted the analysis to single relationships between two relata. Documents that contain bidirectional words (`associate', `relate', `connect', `correlate', and any of their stems) were also not selected for analysis. This is because our focus is on causality, an inherently one-sided relationship between two objects. We also did not consider additional synonyms of these cause words, although that could be pursued for future work. Control documents were also selected. These documents did not contain any of `caused', `causing', or `causes', nor any bidirectional words, and are further matched temporally to obtain the same number of control documents as causal documents in each fifteen-minute period during 2013. Control documents were otherwise selected randomly; causal synonyms may be present. The end result of this procedure identified 965,560 causal and 965,560 control documents. Each of the three “cause-words”, `caused', `causes', and `causing' appeared in 38.2%, 35.0%, and 26.8% of causal documents, respectively.
## Tagging and corpus comparison
Documents were further studied by annotating their unigrams with Parts-of-Speech (POS) and Named Entities (NE) tags. POS tagging was done using NLTK v3.1 BIBREF29 which implements an averaged perceptron classifier BIBREF32 trained on the Brown Corpus BIBREF33 . (POS tagging is affected by punctuation; we show in the Appendix that our results are relatively robust to the removal of punctuation.) POS tags denote the nouns, verbs, and other grammatical constructs present in a document. Named Entity Recognition (NER) was performed using the 4-class, distributional similarity tagger provided as part of the Stanford CoreNLP v3.6.0 toolkit BIBREF34 . NER aims to identify and classify proper words in a text. The NE classifications considered were: Organization, Location, Person, and Misc. The Stanford NER tagger uses a conditional random field model BIBREF35 trained on diverse sets of manually-tagged English-language data (CoNLL-2003) BIBREF34 . Conditional random fields allow dependencies between words so that `New York' and `New York Times', for example, are classified separately as a location and organization, respectively. These taggers are commonly used and often provide reasonably accurate results, but there is always potential ambiguity in written text and improving upon these methods remains an active area of research.
Unigrams, POS, and NEs were compared between the cause and control corpora using odds ratios (ORs):
$$\operatorname{OR}(x) = \frac{p_C(x)/ (1-p_C(x))}{p_N(x) / (1-p_N(x))},$$ (Eq. 1)
where $p_C(x)$ and $p_N(x)$ are the probabilities that a unigram, POS, or NE $x$ occurs in the causal and control corpus, respectively. These probabilities were computed for each corpus separately as $p(x) = f(x) / \sum _{x^{\prime } \in V} f(x^{\prime })$ , where $f(x)$ is the total number of occurrences of $x$ in the corpus and $V$ is the relevant set of unigrams, POS, or NEs. Confidence intervals for the ORs were computed using Wald's methodology BIBREF36 .
As there are many unique unigrams in the text, when computing unigram ORs we focused on the most meaningful unigrams within each corpus by using the following filtering criteria: we considered only the $\operatorname{OR}$ s of the 1500 most frequent unigrams in that corpus that also have a term-frequency-inverse-document-frequency (tf-idf) score above the 90th percentile for that corpus BIBREF37 . The tf-idf was computed as
$$\mbox{tf-idf}(w) = \log f(w) \times \log \left(D̑{\mathit {df}(w)} \right) ,$$ (Eq. 2)
where $D$ is the total number of documents in the corpus, and $\mathit {df}(w)$ is the number of documents in the corpus containing unigram $w$ . Intuitively, unigrams with higher tf-idf scores appear frequently, but are not so frequent that they are ubiquitous through all documents. Filtering via tf-idf is standard practice in the information retrieval and data mining fields.
## Cause-trees
For a better understanding of the higher-order language structure present in text phrases, cause-trees were constructed. A cause-tree starts with a root cause word (either `caused', `causing' or `causes'), then the two most probable words following (preceding) the root are identified. Next, the root word plus one of the top probable words is combined into a bigram and the top two most probable words following (preceding) this bigram are found. Repeatedly applying this process builds a binary tree representing the $n$ -grams that begin with (terminate at) the root word. This process can continue until a certain $n$ -gram length is reached or until there are no more documents long enough to search.
## Sentiment analysis
Sentimental analysis was applied to estimate the emotional content of documents. Two levels of analysis were used: a method where individual unigrams were given crowdsourced numeric sentiment scores, and a second method involving a trained classifier that can incorporate document-level phrase information.
For the first sentiment analysis, each unigram $w$ was assigned a crowdsourced “labMT” sentiment score $s(w)$ BIBREF5 . (Unlike BIBREF5 , scores were recentered by subtracting the mean, $s(w) \leftarrow s(w)-\left<s\right>$ .) Unigrams determined by volunteer raters to have a negative emotional sentiment (`hate',`death', etc.) have $s(w) < 0$ , while unigrams determined to have a positive emotional sentiment (`love', `happy', etc.) tend to have $s(w) > 0$ . Unigrams that have labMT scores and are above the 90th percentile of tf-idf for the corpus form the set $\tilde{V}$ . (Unigrams in $\tilde{V}$ need not be among the 1500 most frequent unigrams.) The set $\tilde{V}$ captures 87.9% (91.5%) of total unigrams in the causal (control) corpus. Crucially, the tf-idf filtering ensures that the words `caused', `causes', and `causing', which have a slight negative sentiment, are not included and do not introduce a systematic bias when comparing the two corpora.
This sentiment measure works on a per-unigram basis, and is therefore best suited for large bodies of text, not short documents BIBREF5 . Instead of considering individual documents, the distributions of labMT scores over all unigrams for each corpus was used to compare the corpora. In addition, a single sentiment score for each corpus was computed as the average sentiment score over all unigrams in that corpus, weighed by unigram frequency: $\sum _{w \in \tilde{V}} {f(w) s(w)} \Big / \sum _{w^{\prime } \in \tilde{V}} f(w^{\prime })$ .
To supplement this sentiment analysis method, we applied a second method capable of estimating with reasonable accuracy the sentiment of individual documents. We used the sentiment classifier BIBREF38 included in the Stanford CoreNLP v3.6.0 toolkit to documents in each corpus. Documents were individually classified into one of five categories: very negative, negative, neutral, positive, very positive. The data used to train this classifier is taken from positive and negative reviews of movies (Stanford Sentiment Treebank v1.0) BIBREF38 .
## Topic modeling
Lastly, we applied topic modeling to the causal corpus to determine what are the topical foci most discussed in causal statements. Topics were built from the causal corpus using Latent Dirichlet Allocation (LDA) BIBREF39 . Under LDA each document is modeled as a bag-of-words or unordered collection of unigrams. Topics are considered as mixtures of unigrams by estimating conditional distributions over unigrams: $P(w|T)$ , the probability of unigram $w$ given topic $T$ and documents are considered as mixtures of topics via $P(T|d)$ , the probability of topic $T$ given document $d$ . These distributions are then found via statistical inference given the observed distributions of unigrams across documents. The total number of topics is a parameter chosen by the practitioner. For this study we used the MALLET v2.0.8RC3 topic modeling toolkit BIBREF40 for model inference. By inspecting the most probable unigrams per topic (according to $P(w|T)$ ), we found 10 topics provided meaningful and distinct topics.
## Results
We have collected approximately 1M causal statements made on Twitter over the course of 2013, and for a control we gathered the same number of statements selected at random but controlling for time of year (see Methods). We applied Parts-of-Speech (POS) and Named Entity (NE) taggers to all these texts. Some post-processed and tagged example documents, both causal and control, are shown in Fig. 1 A. We also applied sentiment analysis methods to these documents (Methods) and we have highlighted very positive and very negative words throughout Fig. 1 .
In Fig. 1 B we present odds ratios for how frequently unigrams (words), POS, or NE appear in causal documents relative to control documents. The three unigrams most strongly skewed towards causal documents were `stress', `problems', and `trouble', while the three most skewed towards control documents were `photo', `ready', and `cute'. While these are only a small number of the unigrams present, this does imply a negative sentiment bias among causal statements (we return to this point shortly).
Figure 1 B also presents odds ratios for POS tags, to help us measure the differences in grammatical structure between causal and control documents (see also the Appendix for the effects of punctuation and casing on these odds ratios). The causal corpus showed greater odds for plural nouns (Penn Treebank tag: NNS), plural proper nouns (NNPS), Wh-determiners/pronouns (WDT, WP$) such as `whichever',`whatever', `whose', or `whosever', and predeterminers (PDT) such as `all' or `both'. Predeterminers quantify noun phrases such as `all' in `after all the events that caused you tears', showing that many causal statements, despite the potential brevity of social media, can encompass or delineate classes of agents and/or patients. On the other hand, the causal corpus has lower odds than the control corpus for list items (LS), proper singular nouns (NNP), and interjections (UH).
Lastly, Fig. 1 B contains odds ratios for NE tags, allowing us to quantify the types of proper nouns that are more or less likely to appear in causal statements. Of the four tags, only the “Person” tag is less likely in the causal corpus than the control. (This matches the odds ratio for the proper singular noun discussed above.) Perhaps surprisingly, these results together imply that causal statements are less likely to involve individual persons than non-causal statements. There is considerable celebrity news and gossip on social media BIBREF4 ; discussions of celebrities may not be especially focused on attributing causes to these celebrities. All other NE tags, Organization, Location, and Miscellaneous, occur more frequently in the causal corpus than the control. All the odds ratios in Fig. 1 B were significant at the $\alpha = 0.05$ level except the List item marker (LS) POS tag.
The unigram analysis in Fig. 1 does not incorporate higher-order phrase structure present in written language. To explore these structures specifically in the causal corpus, we constructed “cause-trees”, shown in Fig. 2 . Inspired by association mining BIBREF41 , a cause-tree is a binary tree rooted at either `caused', `causes', or `causing', that illustrates the most frequently occurring $n$ -grams that either begin or end with that root cause word (see Methods for details).
The “causes” tree shows the focused writing (sentence segments) that many people use to express either the relationship between their own actions and a cause-and-effect (“even if it causes”), or the uncontrollable effect a cause may have on themselves: “causes me to have” shows a person's inability to control a causal event (“[...] i have central heterochromia which causes me to have dual colors in both eyes”). The `causing' tree reveals our ability to confine causal patterns to specific areas, and also our ability to be affected by others causal decisions. Phrases like “causing a scene in/at” and “causing a ruckus in/at” (from documents like “causing a ruckus in the hotel lobby typical [...]”) show people commonly associate bounds on where causal actions take place. The causing tree also shows people's tendency to emphasize current negativity: Phrases like “pain this is causing” coming from documents like “cant you see the pain you are causing her” supports the sentiment bias that causal attribution is more likely for negative cause-effect associations. Finally, the `caused' tree focuses heavily on negative events and indicates people are more likely to remember negative causal events. Documents with phrases from the caused tree (“[...] appalling tragedy [...] that caused the death”, “[...] live with this pain that you caused when i was so young [...]”) exemplify the negative events that are focused on are large-scale tragedies or very personal negative events in one's life.
Taken together, the popularity of negative sentiment unigrams (Fig. 1 ) and $n$ -grams (Fig. 2 ) among causal documents shows that emotional sentiment or “valence” may play a role in how people perform causal attribution BIBREF27 . The “if it bleeds, it leads” mentality among news media, where violent and negative news are more heavily reported, may appeal to this innate causal association mechanism. (On the other hand, many news media themselves use social media for reporting.) The prevalence of negative sentiment also contrasts with the “better angels of our nature” evidence of Pinker BIBREF42 , illustrating one bias that shows why many find the results of Ref. BIBREF42 surprising.
Given this apparent sentiment skew, we further studied sentiment (Fig. 3 ). We compared the sentiment between the corpora in four different ways to investigate the observation (Figs. 1 B and 2 ) that people focus more about negative concepts when they discuss causality. First, we computed the mean sentiment score of each corpus using crowdsourced “labMT” scores weighted by unigram frequency (see Methods). We also applied tf-idf filtering (Methods) to exclude very common words, including the three cause-words, from the mean sentiment score. The causal corpus text was slightly negative on average while the control corpus was slightly positive (Fig. 3 A). The difference in mean sentiment score was significant (t-test: $p < 0.01$ ).
Second, we moved from the mean score to the distribution of sentiment across all (scored) unigrams in the causal and control corpora (Fig. 3 B). The causal corpus contained a large group of negative sentiment unigrams, with labMT scores in the approximate range $-3 < s < -1/2$ ; the control corpus had significantly fewer unigrams in this score range.
Third, in Fig. 3 C we used POS tags to categorize scored unigrams into nouns, verbs, and adjectives. Studying the distributions for each, we found that nouns explain much of the overall difference observed in Fig. 3 B, with verbs showing a similar but smaller difference between the two corpora. Adjectives showed little difference. The distributions in Fig. 3 C account for 87.8% of scored text in the causal corpus and 77.2% of the control corpus. The difference in sentiment between corpora was significant for all distributions (t-test: $p < 0.01$ ).
Fourth, to further confirm that the causal documents tend toward negative sentiment, we applied a separate, independent sentiment analysis using the Stanford NLP sentiment toolkit BIBREF38 to classify the sentiment of individual documents not unigrams (see Methods). Instead of a numeric sentiment score, this classifier assigns documents to one of five categories ranging from very negative to very positive. The classifier showed that the causal corpus contains more negative and very negative documents than the control corpus, while the control corpus contains more neutral, positive, and very positive documents (Fig. 3 D).
We have found language (Figs. 1 and 2 ) and sentiment (Fig. 3 ) differences between causal statements made on social media compared with other social media statements. But what is being discussed? What are the topical foci of causal statements? To study this, for our last analysis we applied topic models to the causal statements. Topic modeling finds groups of related terms (unigrams) by considering similarities between how those terms co-occur across a set of documents.
We used the popular topic modeling method Latent Dirichlet Allocation (LDA) BIBREF39 . We ranked unigrams by how strongly associated they were with the topic. Inspecting these unigrams we found that a 10-topic model discovered meaningful topics. See Methods for full details. The top unigrams for each topic are shown in Tab. 1 .
Topics in the causal corpus tend to fall into three main categories: (i) news, covering current events, weather, etc.; (ii) medicine and health, covering cancer, obesity, stress, etc.; and (iii) relationships, covering problems, stress, crisis, drama, sorry, etc.
While the topics are quite different, they are all similar in their use of negative sentiment words. The negative/global features in the `news' topic are captured in the most representative words: damage, fire, power, etc. Similar to news, the `accident' topic balances the more frequent day-to-day minor frustrations with the less frequent but more severe impacts of car accidents. The words `traffic' and `delays' are the most probable words for this topic, and are common, low-impact occurrences. On the contrary, `crash', `car', `accident' and `death' are the next most probable words for the accident topic, and generally show a focus on less-common but higher-impact events.
The `medical' topic also focused on negative words; highly probable words for this topic included `cancer', `break', `disease', `blood', etc. Meanwhile, the `body' topic contained words like: `stress', `lose', and `weight', giving a focus on on our more personal struggles with body image. Besides body image, the `injuries' topic uses specific pronouns (`his', `him', `her') in references to a person's own injuries or the injuries of others such as athletes.
Aside from more factual information, social information is well represented in causal statements. The `problems' topic shows people attribute their problems to many others with terms like: `dont', `people', `they', `them'. The `stress' topic also uses general words such as `more', `than', or `people' to link stress to all people, and in the same vein, the `crisis' topic focuses on problems within organizations such as governments. The `drama' and `sorry' topics tend towards more specific causal statements. Drama used the words: `like', `she', and `her' while documents in the sorry topic tended to address other people.
The topics of causal documents discovered by LDA showed that both general and specific statements are made regarding news, medicine, and relationships when individuals make causal attributions online.
## Discussion
The power of online communication is the speed and ease with which information can be propagated by potentially any connected users. Yet these strengths come at a cost: rumors and misinformation also spread easily. Causal misattribution is at the heart of many rumors, conspiracy theories, and misinformation campaigns.
Given the central role of causal statements, further studies of the interplay of information propagation and online causal attributions are crucial. Are causal statements more likely to spread online and, if so, in which ways? What types of social media users are more or less likely to make causal statements? Will a user be more likely to make a causal statement if they have recently been exposed to one or more causal statements from other users?
The topics of causal statements also bring forth important questions to be addressed: how timely are causal statements? Are certain topics always being discussed in causal statements? Are there causal topics that are very popular for only brief periods and then forgotten? Temporal dynamics of causal statements are also interesting: do time-of-day or time-of-year factors play a role in how causal statements are made? Our work here focused on a limited subset of causal statements, but more generally, these results may inform new methods for automatically detecting causal statements from unstructured, natural language text BIBREF17 . Better computational tools focused on causal statements are an important step towards further understanding misinformation campaigns and other online activities. Lastly, an important but deeply challenging open question is how, if it is even possible, to validate the accuracy of causal statements. Can causal statements be ranked by some confidence metric(s)? We hope to pursue these and other questions in future research.
Parts-of-speech tagging depends on punctuation and casing, which we filtered in our data, so a study of how robust the POS algorithm is to punctuation and casing removal is important. We computed POS tags for the corpora with and without casing as well as with and without punctuation (which includes hashtags, links and at-symbols). Two tags mentioned in Fig. 1 B, NNPS and LS (which was not significant), were affected by punctuation removal. Otherwise, there is a strong correlation (Fig. 4 ) between Odds Ratios (causal vs. control) with punctuation and without punctuation, including casing and without casing ( $\rho = 0.71$ and $0.80$ , respectively), indicating the POS differences between the corpora were primarily not due to the removal of punctuation or casing.
## Acknowledgments
We thank R. Gallagher for useful comments and gratefully acknowledge the resources provided by the Vermont Advanced Computing Core. This material is based upon work supported by the National Science Foundation under Grant No. ISS-1447634.
| 9
|
1605.03481
|
Tweet2Vec: Character-Based Distributed Representations for Social Media
|
# Tweet2Vec: Character-Based Distributed Representations for Social Media
## Abstract
Text from social media provides a set of challenges that can cause traditional NLP approaches to fail. Informal language, spelling errors, abbreviations, and special characters are all commonplace in these posts, leading to a prohibitively large vocabulary size for word-level approaches. We propose a character composition model, tweet2vec, which finds vector-space representations of whole tweets by learning complex, non-local dependencies in character sequences. The proposed model outperforms a word-level baseline at predicting user-annotated hashtags associated with the posts, doing significantly better when the input contains many out-of-vocabulary words or unusual character sequences. Our tweet2vec encoder is publicly available.
## Introduction
We understand from Zipf's Law that in any natural language corpus a majority of the vocabulary word types will either be absent or occur in low frequency. Estimating the statistical properties of these rare word types is naturally a difficult task. This is analogous to the curse of dimensionality when we deal with sequences of tokens - most sequences will occur only once in the training data. Neural network architectures overcome this problem by defining non-linear compositional models over vector space representations of tokens and hence assign non-zero probability even to sequences not seen during training BIBREF0 , BIBREF1 . In this work, we explore a similar approach to learning distributed representations of social media posts by composing them from their constituent characters, with the goal of generalizing to out-of-vocabulary words as well as sequences at test time.
Traditional Neural Network Language Models (NNLMs) treat words as the basic units of language and assign independent vectors to each word type. To constrain memory requirements, the vocabulary size is fixed before-hand; therefore, rare and out-of-vocabulary words are all grouped together under a common type `UNKNOWN'. This choice is motivated by the assumption of arbitrariness in language, which means that surface forms of words have little to do with their semantic roles. Recently, BIBREF2 challenge this assumption and present a bidirectional Long Short Term Memory (LSTM) BIBREF3 for composing word vectors from their constituent characters which can memorize the arbitrary aspects of word orthography as well as generalize to rare and out-of-vocabulary words.
Encouraged by their findings, we extend their approach to a much larger unicode character set, and model long sequences of text as functions of their constituent characters (including white-space). We focus on social media posts from the website Twitter, which are an excellent testing ground for character based models due to the noisy nature of text. Heavy use of slang and abundant misspellings means that there are many orthographically and semantically similar tokens, and special characters such as emojis are also immensely popular and carry useful semantic information. In our moderately sized training dataset of 2 million tweets, there were about 0.92 million unique word types. It would be expensive to capture all these phenomena in a word based model in terms of both the memory requirement (for the increased vocabulary) and the amount of training data required for effective learning. Additional benefits of the character based approach include language independence of the methods, and no requirement of NLP preprocessing such as word-segmentation.
A crucial step in learning good text representations is to choose an appropriate objective function to optimize. Unsupervised approaches attempt to reconstruct the original text from its latent representation BIBREF4 , BIBREF0 . Social media posts however, come with their own form of supervision annotated by millions of users, in the form of hashtags which link posts about the same topic together. A natural assumption is that the posts with the same hashtags should have embeddings which are close to each other. Hence, we formulate our training objective to maximize cross-entropy loss at the task of predicting hashtags for a post from its latent representation.
We propose a Bi-directional Gated Recurrent Unit (Bi-GRU) BIBREF5 neural network for learning tweet representations. Treating white-space as a special character itself, the model does a forward and backward pass over the entire sequence, and the final GRU states are linearly combined to get the tweet embedding. Posterior probabilities over hashtags are computed by projecting this embedding to a softmax output layer. Compared to a word-level baseline this model shows improved performance at predicting hashtags for a held-out set of posts. Inspired by recent work in learning vector space text representations, we name our model tweet2vec.
## Related Work
Using neural networks to learn distributed representations of words dates back to BIBREF0 . More recently, BIBREF4 released word2vec - a collection of word vectors trained using a recurrent neural network. These word vectors are in widespread use in the NLP community, and the original work has since been extended to sentences BIBREF1 , documents and paragraphs BIBREF6 , topics BIBREF7 and queries BIBREF8 . All these methods require storing an extremely large table of vectors for all word types and cannot be easily generalized to unseen words at test time BIBREF2 . They also require preprocessing to find word boundaries which is non-trivial for a social network domain like Twitter.
In BIBREF2 , the authors present a compositional character model based on bidirectional LSTMs as a potential solution to these problems. A major benefit of this approach is that large word lookup tables can be compacted into character lookup tables and the compositional model scales to large data sets better than other state-of-the-art approaches. While BIBREF2 generate word embeddings from character representations, we propose to generate vector representations of entire tweets from characters in our tweet2vec model.
Our work adds to the growing body of work showing the applicability of character models for a variety of NLP tasks such as Named Entity Recognition BIBREF9 , POS tagging BIBREF10 , text classification BIBREF11 and language modeling BIBREF12 , BIBREF13 .
Previously, BIBREF14 dealt with the problem of estimating rare word representations by building them from their constituent morphemes. While they show improved performance over word-based models, their approach requires a morpheme parser for preprocessing which may not perform well on noisy text like Twitter. Also the space of all morphemes, though smaller than the space of all words, is still large enough that modelling all morphemes is impractical.
Hashtag prediction for social media has been addressed earlier, for example in BIBREF15 , BIBREF16 . BIBREF15 also use a neural architecture, but compose text embeddings from a lookup table of words. They also show that the learned embeddings can generalize to an unrelated task of document recommendation, justifying the use of hashtags as supervision for learning text representations.
## Tweet2Vec
Bi-GRU Encoder: Figure 1 shows our model for encoding tweets. It uses a similar structure to the C2W model in BIBREF2 , with LSTM units replaced with GRU units.
The input to the network is defined by an alphabet of characters $C$ (this may include the entire unicode character set). The input tweet is broken into a stream of characters $c_1, c_2, ... c_m$ each of which is represented by a 1-by- $|C|$ encoding. These one-hot vectors are then projected to a character space by multiplying with the matrix $P_C \in \mathbb {R}^{|C| \times d_c}$ , where $d_c$ is the dimension of the character vector space. Let $x_1, x_2, ... x_m$ be the sequence of character vectors for the input tweet after the lookup. The encoder consists of a forward-GRU and a backward-GRU. Both have the same architecture, except the backward-GRU processes the sequence in reverse order. Each of the GRU units process these vectors sequentially, and starting with the initial state $h_0$ compute the sequence $h_1, h_2, ... h_m$ as follows: $
r_t &= \sigma (W_r x_t + U_r h_{t-1} + b_r), \\
z_t &= \sigma (W_z x_t + U_z h_{t-1} + b_z), \\
\tilde{h}_t &= tanh(W_h x_t + U_h (r_t \odot h_{t-1}) + b_h), \\
h_t &= (1-z_t) \odot h_{t-1} + z_t \odot \tilde{h}_t.
$
Here $r_t$ , $z_t$ are called the reset and update gates respectively, and $\tilde{h}_t$ is the candidate output state which is converted to the actual output state $h_t$ . $W_r, W_z, W_h$ are $d_h \times d_c$ matrices and $U_r, U_z, U_h$ are $d_h \times d_h$ matrices, where $d_h$ is the hidden state dimension of the GRU. The final states $h_m^f$ from the forward-GRU, and $z_t$0 from the backward GRU are combined using a fully-connected layer to the give the final tweet embedding $z_t$1 :
$$e_t = W^f h_m^f + W^b h_0^b$$ (Eq. 3)
Here $W^f, W^b$ are $d_t \times d_h$ and $b$ is $d_t \times 1$ bias term, where $d_t$ is the dimension of the final tweet embedding. In our experiments we set $d_t=d_h$ . All parameters are learned using gradient descent.
Softmax: Finally, the tweet embedding is passed through a linear layer whose output is the same size as the number of hashtags $L$ in the data set. We use a softmax layer to compute the posterior hashtag probabilities:
$$P(y=j |e) = \frac{exp(w_j^Te + b_j)}{\sum _{i=1}^L exp(w_i^Te + b_j)}.$$ (Eq. 4)
Objective Function: We optimize the categorical cross-entropy loss between predicted and true hashtags:
$$J = \frac{1}{B} \sum _{i=1}^{B} \sum _{j=1}^{L} -t_{i,j}log(p_{i,j}) + \lambda \Vert \Theta \Vert ^2.$$ (Eq. 5)
Here $B$ is the batch size, $L$ is the number of classes, $p_{i,j}$ is the predicted probability that the $i$ -th tweet has hashtag $j$ , and $t_{i,j} \in \lbrace 0,1\rbrace $ denotes the ground truth of whether the $j$ -th hashtag is in the $i$ -th tweet. We use L2-regularization weighted by $\lambda $ .
## Word Level Baseline
Since our objective is to compare character-based and word-based approaches, we have also implemented a simple word-level encoder for tweets. The input tweet is first split into tokens along white-spaces. A more sophisticated tokenizer may be used, but for a fair comparison we wanted to keep language specific preprocessing to a minimum. The encoder is essentially the same as tweet2vec, with the input as words instead of characters. A lookup table stores word vectors for the $V$ (20K here) most common words, and the rest are grouped together under the `UNK' token.
## Data
Our dataset consists of a large collection of global posts from Twitter between the dates of June 1, 2013 to June 5, 2013. Only English language posts (as detected by the lang field in Twitter API) and posts with at least one hashtag are retained. We removed infrequent hashtags ( $<500$ posts) since they do not have enough data for good generalization. We also removed very frequent tags ( $>19K$ posts) which were almost always from automatically generated posts (ex: #androidgame) which are trivial to predict. The final dataset contains 2 million tweets for training, 10K for validation and 50K for testing, with a total of 2039 distinct hashtags. We use simple regex to preprocess the post text and remove hashtags (since these are to be predicted) and HTML tags, and replace user-names and URLs with special tokens. We also removed retweets and convert the text to lower-case.
## Implementation Details
Word vectors and character vectors are both set to size $d_L=150$ for their respective models. There were 2829 unique characters in the training set and we model each of these independently in a character look-up table. Embedding sizes were chosen such that each model had roughly the same number of parameters (Table 2 ). Training is performed using mini-batch gradient descent with Nesterov's momentum. We use a batch size $B=64$ , initial learning rate $\eta _0=0.01$ and momentum parameter $\mu _0=0.9$ . L2-regularization with $\lambda =0.001$ was applied to all models. Initial weights were drawn from 0-mean gaussians with $\sigma =0.1$ and initial biases were set to 0. The hyperparameters were tuned one at a time keeping others fixed, and values with the lowest validation cost were chosen. The resultant combination was used to train the models until performance on validation set stopped increasing. During training, the learning rate is halved everytime the validation set precision increases by less than 0.01 % from one epoch to the next. The models converge in about 20 epochs. Code for training both the models is publicly available on github.
## Results
We test the character and word-level variants by predicting hashtags for a held-out test set of posts. Since there may be more than one correct hashtag per post, we generate a ranked list of tags for each post from the output posteriors, and report average precision@1, recall@10 and mean rank of the correct hashtags. These are listed in Table 3 .
To see the performance of each model on posts containing rare words (RW) and frequent words (FW) we selected two test sets each containing 2,000 posts. We populated these sets with posts which had the maximum and minimum number of out-of-vocabulary words respectively, where vocabulary is defined by the 20K most frequent words. Overall, tweet2vec outperforms the word model, doing significantly better on RW test set and comparably on FW set. This improved performance comes at the cost of increased training time (see Table 2 ), since moving from words to characters results in longer input sequences to the GRU.
We also study the effect of model size on the performance of these models. For the word model we set vocabulary size $V$ to 8K, 15K and 20K respectively. For tweet2vec we set the GRU hidden state size to 300, 400 and 500 respectively. Figure 2 shows precision 1 of the two models as the number of parameters is increased, for each test set described above. There is not much variation in the performance, and moreover tweet2vec always outperforms the word based model for the same number of parameters.
Table 4 compares the models as complexity of the task is increased. We created 3 datasets (small, medium and large) with an increasing number of hashtags to be predicted. This was done by varying the lower threshold of the minimum number of tags per post for it to be included in the dataset. Once again we observe that tweet2vec outperforms its word-based counterpart for each of the three settings.
Finally, table 1 shows some predictions from the word level model and tweet2vec. We selected these to highlight some strengths of the character based approach - it is robust to word segmentation errors and spelling mistakes, effectively interprets emojis and other special characters to make predictions, and also performs comparably to the word-based approach for in-vocabulary tokens.
## Conclusion
We have presented tweet2vec - a character level encoder for social media posts trained using supervision from associated hashtags. Our result shows that tweet2vec outperforms the word based approach, doing significantly better when the input post contains many rare words. We have focused only on English language posts, but the character model requires no language specific preprocessing and can be extended to other languages. For future work, one natural extension would be to use a character-level decoder for predicting the hashtags. This will allow generation of hashtags not seen in the training dataset. Also, it will be interesting to see how our tweet2vec embeddings can be used in domains where there is a need for semantic understanding of social media, such as tracking infectious diseases BIBREF17 . Hence, we provide an off-the-shelf encoder trained on medium dataset described above to compute vector-space representations of tweets along with our code on github.
## Acknowledgments
We would like to thank Alex Smola, Yun Fu, Hsiao-Yu Fish Tung, Ruslan Salakhutdinov, and Barnabas Poczos for useful discussions. We would also like to thank Juergen Pfeffer for providing access to the Twitter data, and the reviewers for their comments.
| 9
|
1605.07333
|
Combining Recurrent and Convolutional Neural Networks for Relation Classification
|
# Combining Recurrent and Convolutional Neural Networks for Relation Classification
## Abstract
This paper investigates two different neural architectures for the task of relation classification: convolutional neural networks and recurrent neural networks. For both models, we demonstrate the effect of different architectural choices. We present a new context representation for convolutional neural networks for relation classification (extended middle context). Furthermore, we propose connectionist bi-directional recurrent neural networks and introduce ranking loss for their optimization. Finally, we show that combining convolutional and recurrent neural networks using a simple voting scheme is accurate enough to improve results. Our neural models achieve state-of-the-art results on the SemEval 2010 relation classification task.
## Introduction
Relation classification is the task of assigning sentences with two marked entities to a predefined set of relations. The sentence “We poured the <e1>milk</e1> into the <e2>pumpkin mixture</e2>.”, for example, expresses the relation Entity-Destination(e1,e2). While early research mostly focused on support vector machines or maximum entropy classifiers BIBREF0 , BIBREF1 , recent research showed performance improvements by applying neural networks (NNs) BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 on the benchmark data from SemEval 2010 shared task 8 BIBREF8 .
This study investigates two different types of NNs: recurrent neural networks (RNNs) and convolutional neural networks (CNNs) as well as their combination. We make the following contributions:
(1) We propose extended middle context, a new context representation for CNNs for relation classification. The extended middle context uses all parts of the sentence (the relation arguments, left of the relation arguments, between the arguments, right of the arguments) and pays special attention to the middle part.
(2) We present connectionist bi-directional RNN models which are especially suited for sentence classification tasks since they combine all intermediate hidden layers for their final decision. Furthermore, the ranking loss function is introduced for the RNN model optimization which has not been investigated in the literature for relation classification before.
(3) Finally, we combine CNNs and RNNs using a simple voting scheme and achieve new state-of-the-art results on the SemEval 2010 benchmark dataset.
## Related Work
In 2010, manually annotated data for relation classification was released in the context of a SemEval shared task BIBREF8 . Shared task participants used, i.a., support vector machines or maximum entropy classifiers BIBREF0 , BIBREF1 . Recently, their results on this data set were outperformed by applying NNs BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 .
zeng2014 built a CNN based only on the context between the relation arguments and extended it with several lexical features. kim2014 and others used convolutional filters of different sizes for CNNs. nguyen applied this to relation classification and obtained improvements over single filter sizes. deSantos2015 replaced the softmax layer of the CNN with a ranking layer. They showed improvements and published the best result so far on the SemEval dataset, to our knowledge.
socher used another NN architecture for relation classification: recursive neural networks that built recursive sentence representations based on syntactic parsing. In contrast, zhang investigated a temporal structured RNN with only words as input. They used a bi-directional model with a pooling layer on top.
## Convolutional Neural Networks (CNN)
CNNs perform a discrete convolution on an input matrix with a set of different filters. For NLP tasks, the input matrix represents a sentence: Each column of the matrix stores the word embedding of the corresponding word. By applying a filter with a width of, e.g., three columns, three neighboring words (trigram) are convolved. Afterwards, the results of the convolution are pooled. Following collobertWeston, we perform max-pooling which extracts the maximum value for each filter and, thus, the most informative n-gram for the following steps. Finally, the resulting values are concatenated and used for classifying the relation expressed in the sentence.
## Input: Extended Middle Context
One of our contributions is a new input representation especially designed for relation classification. The contexts are split into three disjoint regions based on the two relation arguments: the left context, the middle context and the right context. Since in most cases the middle context contains the most relevant information for the relation, we want to focus on it but not ignore the other regions completely. Hence, we propose to use two contexts: (1) a combination of the left context, the left entity and the middle context; and (2) a combination of the middle context, the right entity and the right context. Due to the repetition of the middle context, we force the network to pay special attention to it. The two contexts are processed by two independent convolutional and max-pooling layers. After pooling, the results are concatenated to form the sentence representation. Figure FIGREF3 depicts this procedure. It shows an examplary sentence: “He had chest pain and <e1>headaches</e1> from <e2>mold</e2> in the bedroom.” If we only considered the middle context “from”, the network might be tempted to predict a relation like Entity-Origin(e1,e2). However, by also taking the left and right context into account, the model can detect the relation Cause-Effect(e2,e1). While this could also be achieved by integrating the whole context into the model, using the whole context can have disadvantages for longer sentences: The max pooling step can easily choose a value from a part of the sentence which is far away from the mention of the relation. With splitting the context into two parts, we reduce this danger. Repeating the middle context increases the chance for the max pooling step to pick a value from the middle context.
## Convolutional Layer
Following previous work (e.g., BIBREF5 , BIBREF6 ), we use 2D filters spanning all embedding dimensions. After convolution, a max pooling operation is applied that stores only the highest activation of each filter. We apply filters with different window sizes 2-5 (multi-windows) as in BIBREF5 , i.e. spanning a different number of input words.
## Recurrent Neural Networks (RNN)
Traditional RNNs consist of an input vector, a history vector and an output vector. Based on the representation of the current input word and the previous history vector, a new history is computed. Then, an output is predicted (e.g., using a softmax layer). In contrast to most traditional RNN architectures, we use the RNN for sentence modeling, i.e., we predict an output vector only after processing the whole sentence and not after each word. Training is performed using backpropagation through time BIBREF9 which unfolds the recurrent computations of the history vector for a certain number of time steps. To avoid exploding gradients, we use gradient clipping with a threshold of 10 BIBREF10 .
## Input of the RNNs
Initial experiments showed that using trigrams as input instead of single words led to superior results. Hence, at timestep INLINEFORM0 we do not only give word INLINEFORM1 to the model but the trigram INLINEFORM2 by concatenating the corresponding word embeddings.
## Connectionist Bi-directional RNNs
Especially for relation classification, the processing of the relation arguments might be easier with knowledge of the succeeding words. Therefore in bi-directional RNNs, not only a history vector of word INLINEFORM0 is regarded but also a future vector. This leads to the following conditioned probability for the history INLINEFORM1 at time step INLINEFORM2 : DISPLAYFORM0
Thus, the network can be split into three parts: a forward pass which processes the original sentence word by word (Equation EQREF6 ); a backward pass which processes the reversed sentence word by word (Equation ); and a combination of both (Equation ). All three parts are trained jointly. This is also depicted in Figure FIGREF7 .
Combining forward and backward pass by adding their hidden layer is similar to BIBREF7 . We, however, also add a connection to the previous combined hidden layer with weight INLINEFORM0 to be able to include all intermediate hidden layers into the final decision of the network (see Equation ). We call this “connectionist bi-directional RNN”.
In our experiments, we compare this RNN with uni-directional RNNs and bi-directional RNNs without additional hidden layer connections.
## Word Representations
Words are represented by concatenated vectors: a word embedding and a position feature vector.
Pretrained word embeddings. In this study, we used the word2vec toolkit BIBREF11 to train embeddings on an English Wikipedia from May 2014. We only considered words appearing more than 100 times and added a special PADDING token for convolution. This results in an embedding training text of about 485,000 terms and INLINEFORM0 tokens. During model training, the embeddings are updated.
Position features. We incorporate randomly initialized position embeddings similar to zeng2014, nguyen and deSantos2015. In our RNN experiments, we investigate different possibilities of integrating position information: position embeddings, position embeddings with entity presence flags (flags indicating whether the current word is one of the relation arguments), and position indicators BIBREF7 .
## Objective Function: Ranking Loss
Ranking. We applied the ranking loss function proposed in deSantos2015 to train our models. It maximizes the distance between the true label INLINEFORM0 and the best competitive label INLINEFORM1 given a data point INLINEFORM2 . The objective function is DISPLAYFORM0
with INLINEFORM0 and INLINEFORM1 being the scores for the classes INLINEFORM2 and INLINEFORM3 respectively. The parameter INLINEFORM4 controls the penalization of the prediction errors and INLINEFORM5 and INLINEFORM6 are margins for the correct and incorrect classes. Following deSantos2015, we set INLINEFORM7 . We do not learn a pattern for the class Other but increase its difference to the best competitive label by using only the second summand in Equation EQREF10 during training.
## Experiments and Results
We used the relation classification dataset of the SemEval 2010 task 8 BIBREF8 . It consists of sentences which have been manually labeled with 19 relations (9 directed relations and one artificial class Other). 8,000 sentences have been distributed as training set and 2,717 sentences served as test set. For evaluation, we applied the official scoring script and report the macro F1 score which also served as the official result of the shared task.
RNN and CNN models were implemented with theano BIBREF12 , BIBREF13 . For all our models, we use L2 regularization with a weight of 0.0001. For CNN training, we use mini batches of 25 training examples while we perform stochastic gradient descent for the RNN. The initial learning rates are 0.2 for the CNN and 0.01 for the RNN. We train the models for 10 (CNN) and 50 (RNN) epochs without early stopping. As activation function, we apply tanh for the CNN and capped ReLU for the RNN. For tuning the hyperparameters, we split the training data into two parts: 6.5k (training) and 1.5k (development) sentences. We also tuned the learning rate schedule on dev.
Beside of training single models, we also report ensemble results for which we combined the presented single models with a voting process.
## Performance of CNNs
As a baseline system, we implemented a CNN similar to the one described by zeng2014. It consists of a standard convolutional layer with filters with only one window size, followed by a softmax layer. As input it uses the middle context. In contrast to zeng2014, our CNN does not have an additional fully connected hidden layer. Therefore, we increased the number of convolutional filters to 1200 to keep the number of parameters comparable. With this, we obtain a baseline result of 73.0. After including 5 dimensional position features, the performance was improved to 78.6 (comparable to 78.9 as reported by zeng2014 without linguistic features).
In the next step, we investigate how this result changes if we successively add further features to our CNN: multi-windows for convolution (window sizes: 2,3,4,5 and 300 feature maps each), ranking layer instead of softmax and our proposed extended middle context. Table TABREF12 shows the results. Note that all numbers are produced by CNNs with a comparable number of parameters. We also report F1 for increasing the word embedding dimensionality from 50 to 400. The position embedding dimensionality is 5 in combination with 50 dimensional word embeddings and 35 with 400 dimensional word embeddings. Our results show that especially the ranking layer and the embedding size have an important impact on the performance.
## Performance of RNNs
As a baseline for the RNN models, we apply a uni-directional RNN which predicts the relation after processing the whole sentence. With this model, we achieve an F1 score of 61.2 on the SemEval test set.
Afterwards, we investigate the impact of different position features on the performance of uni-directional RNNs (position embeddings, position embeddings concatenated with a flag indicating whether the current word is an entity or not, and position indicators BIBREF7 ). The results indicate that position indicators (i.e. artificial words that indicate the entity presence) perform the best on the SemEval data. We achieve an F1 score of 73.4 with them. However, the difference to using position embeddings with entity flags is not statistically significant.
Similar to our CNN experiments, we successively vary the RNN models by using bi-directionality, by adding connections between the hidden layers (“connectionist”), by applying ranking instead of softmax to predict the relation and by increasing the word embedding dimension to 400.
The results in Table TABREF14 show that all of these variations lead to statistically significant improvements. Especially the additional hidden layer connections and the integration of the ranking layer have a large impact on the performance.
## Combination of CNNs and RNNs
Finally, we combine our CNN and RNN models using a voting process. For each sentence in the test set, we apply several CNN and RNN models presented in Tables TABREF12 and TABREF14 and predict the class with the most votes. In case of a tie, we pick one of the most frequent classes randomly. The combination achieves an F1 score of 84.9 which is better than the performance of the two NN types alone. It, thus, confirms our assumption that the networks provide complementary information: while the RNN computes a weighted combination of all words in the sentence, the CNN extracts the most informative n-grams for the relation and only considers their resulting activations.
## Comparison with State of the Art
Table TABREF16 shows the results of our models ER-CNN (extended ranking CNN) and R-RNN (ranking RNN) in the context of other state-of-the-art models. Our proposed models obtain state-of-the-art results on the SemEval 2010 task 8 data set without making use of any linguistic features.
## Conclusion
In this paper, we investigated different features and architectural choices for convolutional and recurrent neural networks for relation classification without using any linguistic features. For convolutional neural networks, we presented a new context representation for relation classification. Furthermore, we introduced connectionist recurrent neural networks for sentence classification tasks and performed the first experiments with ranking recurrent neural networks. Finally, we showed that even a simple combination of convolutional and recurrent neural networks improved results. With our neural models, we achieved new state-of-the-art results on the SemEval 2010 task 8 benchmark data.
## Acknowledgments
Heike Adel is a recipient of the Google European Doctoral Fellowship in Natural Language Processing and this research is supported by this fellowship.
This research was also supported by Deutsche Forschungsgemeinschaft: grant SCHU 2246/4-2.
| 17
|
1606.05320
|
Increasing the Interpretability of Recurrent Neural Networks Using Hidden Markov Models
|
# Increasing the Interpretability of Recurrent Neural Networks Using Hidden Markov Models
## Abstract
As deep neural networks continue to revolutionize various application domains, there is increasing interest in making these powerful models more understandable and interpretable, and narrowing down the causes of good and bad predictions. We focus on recurrent neural networks (RNNs), state of the art models in speech recognition and translation. Our approach to increasing interpretability is by combining an RNN with a hidden Markov model (HMM), a simpler and more transparent model. We explore various combinations of RNNs and HMMs: an HMM trained on LSTM states; a hybrid model where an HMM is trained first, then a small LSTM is given HMM state distributions and trained to fill in gaps in the HMM's performance; and a jointly trained hybrid model. We find that the LSTM and HMM learn complementary information about the features in the text.
## Introduction
Following the recent progress in deep learning, researchers and practitioners of machine learning are recognizing the importance of understanding and interpreting what goes on inside these black box models. Recurrent neural networks have recently revolutionized speech recognition and translation, and these powerful models could be very useful in other applications involving sequential data. However, adoption has been slow in applications such as health care, where practitioners are reluctant to let an opaque expert system make crucial decisions. If we can make the inner workings of RNNs more interpretable, more applications can benefit from their power.
There are several aspects of what makes a model or algorithm understandable to humans. One aspect is model complexity or parsimony. Another aspect is the ability to trace back from a prediction or model component to particularly influential features in the data BIBREF0 BIBREF1 . This could be useful for understanding mistakes made by neural networks, which have human-level performance most of the time, but can perform very poorly on seemingly easy cases. For instance, convolutional networks can misclassify adversarial examples with very high confidence BIBREF2 , and made headlines in 2015 when the image tagging algorithm in Google Photos mislabeled African Americans as gorillas. It's reasonable to expect recurrent networks to fail in similar ways as well. It would thus be useful to have more visibility into where these sorts of errors come from, i.e. which groups of features contribute to such flawed predictions.
Several promising approaches to interpreting RNNs have been developed recently. BIBREF3 have approached this by using gradient boosting trees to predict LSTM output probabilities and explain which features played a part in the prediction. They do not model the internal structure of the LSTM, but instead approximate the entire architecture as a black box. BIBREF4 showed that in LSTM language models, around 10% of the memory state dimensions can be interpreted with the naked eye by color-coding the text data with the state values; some of them track quotes, brackets and other clearly identifiable aspects of the text. Building on these results, we take a somewhat more systematic approach to looking for interpretable hidden state dimensions, by using decision trees to predict individual hidden state dimensions (Figure 2 ). We visualize the overall dynamics of the hidden states by coloring the training data with the k-means clusters on the state vectors (Figures 3 , 3 ).
We explore several methods for building interpretable models by combining LSTMs and HMMs. The existing body of literature mostly focuses on methods that specifically train the RNN to predict HMM states BIBREF5 or posteriors BIBREF6 , referred to as hybrid or tandem methods respectively. We first investigate an approach that does not require the RNN to be modified in order to make it understandable, as the interpretation happens after the fact. Here, we model the big picture of the state changes in the LSTM, by extracting the hidden states and approximating them with a continuous emission hidden Markov model (HMM). We then take the reverse approach where the HMM state probabilities are added to the output layer of the LSTM (see Figure 1 ). The LSTM model can then make use of the information from the HMM, and fill in the gaps when the HMM is not performing well, resulting in an LSTM with a smaller number of hidden state dimensions that could be interpreted individually (Figures 3 , 3 ).
## Methods
We compare a hybrid HMM-LSTM approach with a continuous emission HMM (trained on the hidden states of a 2-layer LSTM), and a discrete emission HMM (trained directly on data).
## LSTM models
We use a character-level LSTM with 1 layer and no dropout, based on the Element-Research library. We train the LSTM for 10 epochs, starting with a learning rate of 1, where the learning rate is halved whenever $\exp (-l_t) > \exp (-l_{t-1}) + 1$ , where $l_t$ is the log likelihood score at epoch $t$ . The $L_2$ -norm of the parameter gradient vector is clipped at a threshold of 5.
## Hidden Markov models
The HMM training procedure is as follows:
Initialization of HMM hidden states:
(Discrete HMM) Random multinomial draw for each time step (i.i.d. across time steps).
(Continuous HMM) K-means clusters fit on LSTM states, to speed up convergence relative to random initialization.
At each iteration:
Sample states using Forward Filtering Backwards Sampling algorithm (FFBS, BIBREF7 ).
Sample transition parameters from a Multinomial-Dirichlet posterior. Let $n_{ij}$ be the number of transitions from state $i$ to state $j$ . Then the posterior distribution of the $i$ -th row of transition matrix $T$ (corresponding to transitions from state $i$ ) is: $T_i \sim \text{Mult}(n_{ij} | T_i) \text{Dir}(T_i | \alpha )$
where $\alpha $ is the Dirichlet hyperparameter.
(Continuous HMM) Sample multivariate normal emission parameters from Normal-Inverse-Wishart posterior for state $i$ : $ \mu _i, \Sigma _i \sim N(y|\mu _i, \Sigma _i) N(\mu _i |0, \Sigma _i) \text{IW}(\Sigma _i) $
(Discrete HMM) Sample the emission parameters from a Multinomial-Dirichlet posterior.
Evaluation:
We evaluate the methods on how well they predict the next observation in the validation set. For the HMM models, we do a forward pass on the validation set (no backward pass unlike the full FFBS), and compute the HMM state distribution vector $p_t$ for each time step $t$ . Then we compute the predictive likelihood for the next observation as follows: $ P(y_{t+1} | p_t) =\sum _{x_t=1}^n \sum _{x_{t+1}=1}^n p_{tx_t} \cdot T_{x_t, x_{t+1}} \cdot P(y_{t+1} | x_{t+1})$
where $n$ is the number of hidden states in the HMM.
## Hybrid models
Our main hybrid model is put together sequentially, as shown in Figure 1 . We first run the discrete HMM on the data, outputting the hidden state distributions obtained by the HMM's forward pass, and then add this information to the architecture in parallel with a 1-layer LSTM. The linear layer between the LSTM and the prediction layer is augmented with an extra column for each HMM state. The LSTM component of this architecture can be smaller than a standalone LSTM, since it only needs to fill in the gaps in the HMM's predictions. The HMM is written in Python, and the rest of the architecture is in Torch.
We also build a joint hybrid model, where the LSTM and HMM are simultaneously trained in Torch. We implemented an HMM Torch module, optimized using stochastic gradient descent rather than FFBS. Similarly to the sequential hybrid model, we concatenate the LSTM outputs with the HMM state probabilities.
## Experiments
We test the models on several text data sets on the character level: the Penn Tree Bank (5M characters), and two data sets used by BIBREF4 , Tiny Shakespeare (1M characters) and Linux Kernel (5M characters). We chose $k=20$ for the continuous HMM based on a PCA analysis of the LSTM states, as the first 20 components captured almost all the variance.
Table 1 shows the predictive log likelihood of the next text character for each method. On all text data sets, the hybrid algorithm performs a bit better than the standalone LSTM with the same LSTM state dimension. This effect gets smaller as we increase the LSTM size and the HMM makes less difference to the prediction (though it can still make a difference in terms of interpretability). The hybrid algorithm with 20 HMM states does better than the one with 10 HMM states. The joint hybrid algorithm outperforms the sequential hybrid on Shakespeare data, but does worse on PTB and Linux data, which suggests that the joint hybrid is more helpful for smaller data sets. The joint hybrid is an order of magnitude slower than the sequential hybrid, as the SGD-based HMM is slower to train than the FFBS-based HMM.
We interpret the HMM and LSTM states in the hybrid algorithm with 10 LSTM state dimensions and 10 HMM states in Figures 3 and 3 , showing which features are identified by the HMM and LSTM components. In Figures 3 and 3 , we color-code the training data with the 10 HMM states. In Figures 3 and 3 , we apply k-means clustering to the LSTM state vectors, and color-code the training data with the clusters. The HMM and LSTM states pick up on spaces, indentation, and special characters in the data (such as comment symbols in Linux data). We see some examples where the HMM and LSTM complement each other, such as learning different things about spaces and comments on Linux data, or punctuation on the Shakespeare data. In Figure 2 , we see that some individual LSTM hidden state dimensions identify similar features, such as comment symbols in the Linux data.
## Conclusion and future work
Hybrid HMM-RNN approaches combine the interpretability of HMMs with the predictive power of RNNs. Sometimes, a small hybrid model can perform better than a standalone LSTM of the same size. We use visualizations to show how the LSTM and HMM components of the hybrid algorithm complement each other in terms of features learned in the data.
| 7
|
1608.04917
|
Cohesion and Coalition Formation in the European Parliament: Roll-Call Votes and Twitter Activities
|
# Cohesion and Coalition Formation in the European Parliament: Roll-Call Votes and Twitter Activities
## Abstract
We study the cohesion within and the coalitions between political groups in the Eighth European Parliament (2014--2019) by analyzing two entirely different aspects of the behavior of the Members of the European Parliament (MEPs) in the policy-making processes. On one hand, we analyze their co-voting patterns and, on the other, their retweeting behavior. We make use of two diverse datasets in the analysis. The first one is the roll-call vote dataset, where cohesion is regarded as the tendency to co-vote within a group, and a coalition is formed when the members of several groups exhibit a high degree of co-voting agreement on a subject. The second dataset comes from Twitter; it captures the retweeting (i.e., endorsing) behavior of the MEPs and implies cohesion (retweets within the same group) and coalitions (retweets between groups) from a completely different perspective. We employ two different methodologies to analyze the cohesion and coalitions. The first one is based on Krippendorff's Alpha reliability, used to measure the agreement between raters in data-analysis scenarios, and the second one is based on Exponential Random Graph Models, often used in social-network analysis. We give general insights into the cohesion of political groups in the European Parliament, explore whether coalitions are formed in the same way for different policy areas, and examine to what degree the retweeting behavior of MEPs corresponds to their co-voting patterns. A novel and interesting aspect of our work is the relationship between the co-voting and retweeting patterns.
## Abstract
We study the cohesion within and the coalitions between political groups in the Eighth European Parliament (2014–2019) by analyzing two entirely different aspects of the behavior of the Members of the European Parliament (MEPs) in the policy-making processes. On one hand, we analyze their co-voting patterns and, on the other, their retweeting behavior. We make use of two diverse datasets in the analysis. The first one is the roll-call vote dataset, where cohesion is regarded as the tendency to co-vote within a group, and a coalition is formed when the members of several groups exhibit a high degree of co-voting agreement on a subject. The second dataset comes from Twitter; it captures the retweeting (i.e., endorsing) behavior of the MEPs and implies cohesion (retweets within the same group) and coalitions (retweets between groups) from a completely different perspective.
We employ two different methodologies to analyze the cohesion and coalitions. The first one is based on Krippendorff's Alpha reliability, used to measure the agreement between raters in data-analysis scenarios, and the second one is based on Exponential Random Graph Models, often used in social-network analysis. We give general insights into the cohesion of political groups in the European Parliament, explore whether coalitions are formed in the same way for different policy areas, and examine to what degree the retweeting behavior of MEPs corresponds to their co-voting patterns. A novel and interesting aspect of our work is the relationship between the co-voting and retweeting patterns.
## Introduction
Social-media activities often reflect phenomena that occur in other complex systems. By observing social networks and the content propagated through these networks, we can describe or even predict the interplay between the observed social-media activities and another complex system that is more difficult, if not impossible, to monitor. There are numerous studies reported in the literature that successfully correlate social-media activities to phenomena like election outcomes BIBREF0 , BIBREF1 or stock-price movements BIBREF2 , BIBREF3 .
In this paper we study the cohesion and coalitions exhibited by political groups in the Eighth European Parliament (2014–2019). We analyze two entirely different aspects of how the Members of the European Parliament (MEPs) behave in policy-making processes. On one hand, we analyze their co-voting patterns and, on the other, their retweeting (i.e., endorsing) behavior.
We use two diverse datasets in the analysis: the roll-call votes and the Twitter data. A roll-call vote (RCV) is a vote in the parliament in which the names of the MEPs are recorded along with their votes. The RCV data is available as part of the minutes of the parliament's plenary sessions. From this perspective, cohesion is seen as the tendency to co-vote (i.e., cast the same vote) within a group, and a coalition is formed when members of two or more groups exhibit a high degree of co-voting on a subject. The second dataset comes from Twitter. It captures the retweeting behavior of MEPs and implies cohesion (retweets within the same group) and coalitions (retweets between the groups) from a completely different perspective.
With over 300 million monthly active users and 500 million tweets posted daily, Twitter is one of the most popular social networks. Twitter allows its users to post short messages (tweets) and to follow other users. A user who follows another user is able to read his/her public tweets. Twitter also supports other types of interaction, such as user mentions, replies, and retweets. Of these, retweeting is the most important activity as it is used to share and endorse content created by other users. When a user retweets a tweet, the information about the original author as well as the tweet's content are preserved, and the tweet is shared with the user's followers. Typically, users retweet content that they agree with and thus endorse the views expressed by the original tweeter.
We apply two different methodologies to analyze the cohesion and coalitions. The first one is based on Krippendorff's INLINEFORM0 BIBREF4 which measures the agreement among observers, or voters in our case. The second one is based on Exponential Random Graph Models (ERGM) BIBREF5 . In contrast to the former, ERGM is a network-based approach and is often used in social-network analyses. Even though these two methodologies come with two different sets of techniques and are based on different assumptions, they provide consistent results.
The main contributions of this paper are as follows:
(i) We give general insights into the cohesion of political groups in the Eighth European Parliament, both overall and across different policy areas.
(ii) We explore whether coalitions are formed in the same way for different policy areas.
(iii) We explore to what degree the retweeting behavior of MEPs corresponds to their co-voting patterns.
(iv) We employ two statistically sound methodologies and examine the extent to which the results are sensitive to the choice of methodology. While the results are mostly consistent, we show that the difference are due to the different treatment of non-attending and abstaining MEPs by INLINEFORM0 and ERGM.
The most novel and interesting aspect of our work is the relationship between the co-voting and the retweeting patterns. The increased use of Twitter by MEPs on days with a roll-call vote session (see Fig FIGREF1 ) is an indicator that these two processes are related. In addition, the force-based layouts of the co-voting network and the retweet network reveal a very similar structure on the left-to-center side of the political spectrum (see Fig FIGREF2 ). They also show a discrepancy on the far-right side of the spectrum, which calls for a more detailed analysis.
## Related work
In this paper we study and relate two very different aspects of how MEPs behave in policy-making processes. First, we look at their co-voting behavior, and second, we examine their retweeting patterns. Thus, we draw related work from two different fields of science. On one hand, we look at how co-voting behavior is analyzed in the political-science literature and, on the other, we explore how Twitter is used to better understand political and policy-making processes. The latter has been more thoroughly explored in the field of data mining (specifically, text mining and network analysis).
To the best of our knowledge, this is the first paper that studies legislative behavior in the Eighth European Parliament. The legislative behavior of the previous parliaments was thoroughly studied by Hix, Attina, and others BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 . These studies found that voting behavior is determined to a large extent—and when viewed over time, increasingly so—by affiliation to a political group, as an organizational reflection of the ideological position. The authors found that the cohesion of political groups in the parliament has increased, while nationality has been less and less of a decisive factor BIBREF12 . The literature also reports that a split into political camps on the left and right of the political spectrum has recently replaced the `grand coalition' between the two big blocks of Christian Conservatives (EPP) and Social Democrats (S&D) as the dominant form of finding majorities in the parliament. The authors conclude that coalitions are to a large extent formed along the left-to-right axis BIBREF12 .
In this paper we analyze the roll-call vote data published in the minutes of the parliament's plenary sessions. For a given subject, the data contains the vote of each MEP present at the respective sitting. Roll-call vote data from the European Parliament has already been extensively studied by other authors, most notably by Hix et al. BIBREF10 , BIBREF13 , BIBREF11 . To be able to study the cohesion and coalitions, authors like Hix, Attina, and Rice BIBREF6 , BIBREF13 , BIBREF14 defined and employed a variety of agreement measures. The most prominent measure is the Agreement Index proposed by Hix et al. BIBREF13 . This measure computes the agreement score from the size of the majority class for a particular vote. The Agreement Index, however, exhibits two drawbacks: (i) it does not account for co-voting by chance, and (ii) without a proper adaptation, it does not accommodate the scenario in which the agreement is to be measured between two different political groups.
We employ two statistically sound methodologies developed in two different fields of science. The first one is based on Krippendorff's INLINEFORM0 BIBREF4 . INLINEFORM1 is a measure of the agreement among observers, coders, or measuring instruments that assign values to items or phenomena. It compares the observed agreement to the agreement expected by chance. INLINEFORM2 is used to measure the inter- and self-annotator agreement of human experts when labeling data, and the performance of classification models in machine learning scenarios BIBREF15 . In addition to INLINEFORM3 , we employ Exponential Random Graph Models (ERGM) BIBREF5 . In contrast to the former, ERGM is a network-based approach, often used in social-network analyses. ERGM can be employed to investigate how different network statistics (e.g., number of edges and triangles) or external factors (e.g., political group membership) govern the network-formation process.
The second important aspect of our study is related to analyzing the behavior of participants in social networks, specifically Twitter. Twitter is studied by researchers to better understand different political processes, and in some cases to predict their outcomes. Eom et al. BIBREF1 consider the number of tweets by a party as a proxy for the collective attention to the party, explore the dynamics of the volume, and show that this quantity contains information about an election's outcome. Other studies BIBREF16 reach similar conclusions. Conover et al. BIBREF17 predicted the political alignment of Twitter users in the run-up to the 2010 US elections based on content and network structure. They analyzed the polarization of the retweet and mention networks for the same elections BIBREF18 . Borondo et al. BIBREF19 analyzed user activity during the Spanish presidential elections. They additionally analyzed the 2012 Catalan elections, focusing on the interplay between the language and the community structure of the network BIBREF20 . Most existing research, as Larsson points out BIBREF21 , focuses on the online behavior of leading political figures during election campaigns.
This paper continues our research on communities that MEPs (and their followers) form on Twitter BIBREF22 . The goal of our research was to evaluate the role of Twitter in identifying communities of influence when the actual communities are known. We represent the influence on Twitter by the number of retweets that MEPs “receive”. We construct two networks of influence: (i) core, which consists only of MEPs, and (ii) extended, which also involves their followers. We compare the detected communities in both networks to the groups formed by the political, country, and language membership of MEPs. The results show that the detected communities in the core network closely match the political groups, while the communities in the extended network correspond to the countries of residence. This provides empirical evidence that analyzing retweet networks can reveal real-world relationships and can be used to uncover hidden properties of the networks.
Lazer BIBREF23 highlights the importance of network-based approaches in political science in general by arguing that politics is a relational phenomenon at its core. Some researchers have adopted the network-based approach to investigate the structure of legislative work in the US Congress, including committee and sub-committee membership BIBREF24 , bill co-sponsoring BIBREF25 , and roll-call votes BIBREF26 . More recently, Dal Maso et al. BIBREF27 examined the community structure with respect to political coalitions and government structure in the Italian Parliament. Scherpereel et al. BIBREF28 examined the constituency, personal, and strategic characteristics of MEPs that influence their tweeting behavior. They suggested that Twitter's characteristics, like immediacy, interactivity, spontaneity, personality, and informality, are likely to resonate with political parties across Europe. By fitting regression models, the authors find that MEPs from incohesive groups have a greater tendency to retweet.
In contrast to most of these studies, we focus on the Eighth European Parliament, and more importantly, we study and relate two entirely different behavioral aspects, co-voting and retweeting. The goal of this research is to better understand the cohesion and coalition formation processes in the European Parliament by quantifying and comparing the co-voting patterns and social behavior.
## Methods
In this section we present the methods to quantify cohesion and coalitions from the roll-call votes and Twitter activities.
## Co-voting measured by agreement
We first show how the co-voting behaviour of MEPs can be quantified by a measure of the agreement between them. We treat individual RCVs as observations, and MEPs as independent observers or raters. When they cast the same vote, there is a high level of agreement, and when they vote differently, there is a high level of disagreement. We define cohesion as the level of agreement within a political group, a coalition as a voting agreement between political groups, and opposition as a disagreement between different groups.
There are many well-known measures of agreement in the literature. We selected Krippendorff's Alpha-reliability ( INLINEFORM0 ) BIBREF4 , which is a generalization of several specialized measures. It works for any number of observers, and is applicable to different variable types and metrics (e.g., nominal, ordered, interval, etc.). In general, INLINEFORM1 is defined as follows: INLINEFORM2
where INLINEFORM0 is the actual disagreement between observers (MEPs), and INLINEFORM1 is disagreement expected by chance. When observers agree perfectly, INLINEFORM2 INLINEFORM3 , when the agreement equals the agreement by chance, INLINEFORM4 INLINEFORM5 , and when the observers disagree systematically, INLINEFORM6 INLINEFORM7 .
The two disagreement measures are defined as follows: INLINEFORM0
INLINEFORM0
The arguments INLINEFORM0 , and INLINEFORM1 are defined below and refer to the values in the coincidence matrix that is constructed from the RCVs data. In roll-call votes, INLINEFORM2 (and INLINEFORM3 ) is a nominal variable with two possible values: yes and no. INLINEFORM4 is a difference function between the values of INLINEFORM5 and INLINEFORM6 , defined as: INLINEFORM7
The RCVs data has the form of a reliability data matrix: INLINEFORM0
where INLINEFORM0 is the number of RCVs, INLINEFORM1 is the number of MEPs, INLINEFORM2 is the number of votes cast in the voting INLINEFORM3 , and INLINEFORM4 is the actual vote of an MEP INLINEFORM5 in voting INLINEFORM6 (yes or no).
A coincidence matrix is constructed from the reliability data matrix, and is in general a INLINEFORM0 -by- INLINEFORM1 square matrix, where INLINEFORM2 is the number of possible values of INLINEFORM3 . In our case, where only yes/no votes are relevant, the coincidence matrix is a 2-by-2 matrix of the following form: INLINEFORM4
A cell INLINEFORM0 accounts for all coincidences from all pairs of MEPs in all RCVs where one MEP has voted INLINEFORM1 and the other INLINEFORM2 . INLINEFORM3 and INLINEFORM4 are the totals for each vote outcome, and INLINEFORM5 is the grand total. The coincidences INLINEFORM6 are computed as: INLINEFORM7
where INLINEFORM0 is the number of INLINEFORM1 pairs in vote INLINEFORM2 , and INLINEFORM3 is the number of MEPs that voted in INLINEFORM4 . When computing INLINEFORM5 , each pair of votes is considered twice, once as a INLINEFORM6 pair, and once as a INLINEFORM7 pair. The coincidence matrix is therefore symmetrical around the diagonal, and the diagonal contains all the equal votes.
The INLINEFORM0 agreement is used to measure the agreement between two MEPs or within a group of MEPs. When applied to a political group, INLINEFORM1 corresponds to the cohesion of the group. The closer INLINEFORM2 is to 1, the higher the agreement of the MEPs in the group, and hence the higher the cohesion of the group.
We propose a modified version of INLINEFORM0 to measure the agreement between two different groups, INLINEFORM1 and INLINEFORM2 . In the case of a voting agreement between political groups, high INLINEFORM3 is interpreted as a coalition between the groups, whereas negative INLINEFORM4 indicates political opposition.
Suppose INLINEFORM0 and INLINEFORM1 are disjoint subsets of all the MEPs, INLINEFORM2 , INLINEFORM3 . The respective number of votes cast by both group members in vote INLINEFORM4 is INLINEFORM5 and INLINEFORM6 . The coincidences are then computed as: INLINEFORM7
where the INLINEFORM0 pairs come from different groups, INLINEFORM1 and INLINEFORM2 . The total number of such pairs in vote INLINEFORM3 is INLINEFORM4 . The actual number INLINEFORM5 of the pairs is multiplied by INLINEFORM6 so that the total contribution of vote INLINEFORM7 to the coincidence matrix is INLINEFORM8 .
## A network-based measure of co-voting
In this section we describe a network-based approach to analyzing the co-voting behavior of MEPs. For each roll-call vote we form a network, where the nodes in the network are MEPs, and an undirected edge between two MEPs is formed when they cast the same vote.
We are interested in the factors that determine the cohesion within political groups and coalition formation between political groups. Furthermore, we investigate to what extent communication in a different social context, i.e., the retweeting behavior of MEPs, can explain the co-voting of MEPs. For this purpose we apply an Exponential Random Graph Model BIBREF5 to individual roll-call vote networks, and aggregate the results by means of the meta-analysis.
ERGMs allow us to investigate the factors relevant for the network-formation process. Network metrics, as described in the abundant literature, serve to gain information about the structural properties of the observed network. A model investigating the processes driving the network formation, however, has to take into account that there can be a multitude of alternative networks. If we are interested in the parameters influencing the network formation we have to consider all possible networks and measure their similarity to the originally observed network. The family of ERGMs builds upon this idea.
Assume a random graph INLINEFORM0 , in the form of a binary adjacency matrix, made up of a set of INLINEFORM1 nodes and INLINEFORM2 edges INLINEFORM3 where, similar to a binary choice model, INLINEFORM4 if the nodes INLINEFORM5 are connected and INLINEFORM6 if not. Since network data is by definition relational and thus violates assumptions of independence, classical binary choice models, like logistic regression, cannot be applied in this context. Within an ERGM, the probability for a given network is modelled by DISPLAYFORM0
where INLINEFORM0 is the vector of parameters and INLINEFORM1 is the vector of network statistics (counts of network substructures), which are a function of the adjacency matrix INLINEFORM2 . INLINEFORM3 is a normalization constant corresponding to the sample of all possible networks, which ensures a proper probability distribution. Evaluating the above expression allows us to make assertions if and how specific nodal attributes influence the network formation process. These nodal attributes can be endogenous (dyad-dependent parameters) to the network, like the in- and out-degrees of a node, or exogenous (dyad-independent parameters), as the party affiliation, or the country of origin in our case.
An alternative formulation of the ERGM provides the interpretation of the coefficients. We introduce the change statistic, which is defined as the change in the network statistics when an edge between nodes INLINEFORM0 and INLINEFORM1 is added or not. If INLINEFORM2 and INLINEFORM3 denote the vectors of counts of network substructures when the edge is added or not, the change statistics is defined as follows: INLINEFORM4
With this at hand it can be shown that the distribution of the variable INLINEFORM0 , conditional on the rest of the graph INLINEFORM1 , corresponds to: INLINEFORM2
This implies on the one hand that the probability depends on INLINEFORM0 via the change statistic INLINEFORM1 , and on the other hand, that each coefficient within the vector INLINEFORM2 represents an increase in the conditional log-odds ( INLINEFORM3 ) of the graph when the corresponding element in the vector INLINEFORM4 increases by one. The need to condition the probability on the rest of the network can be illustrated by a simple example. The addition (removal) of a single edge alters the network statistics. If a network has only edges INLINEFORM5 and INLINEFORM6 , the creation of an edge INLINEFORM7 would not only add an additional edge but would also alter the count for other network substructures included in the model. In this example, the creation of the edge INLINEFORM8 also increases the number of triangles by one. The coefficients are transformed into probabilities with the logistic function: INLINEFORM9
For example, in the context of roll-call votes, the probability that an additional co-voting edge is formed between two nodes (MEPs) of the same political group is computed with that equation. In this context, the nodematch (nodemix) coefficients of the ERGM (described in detail bellow) therefore refer to the degree of homophilous (heterophilous) matching of MEPs with regard to their political affiliation, or, expressed differently, the propensity of MEPs to co-vote with other MEPs of their respective political group or another group.
A positive coefficient reflects an increased chance that an edge between two nodes with respective properties, like group affiliation, given all other parameters unchanged, is formed. Or, put differently, a positive coefficient implies that the probability of observing a network with a higher number of corresponding pairs relative to the hypothetical baseline network, is higher than to observe the baseline network itself BIBREF31 . For an intuitive interpretation, log-odds value of 0 corresponds to the even chance probability of INLINEFORM0 . Log-odds of INLINEFORM1 correspond to an increase of probability by INLINEFORM2 , whereas log-odds of INLINEFORM3 correspond to a decrease of probability by INLINEFORM4 .
The computational challenges of estimating ERGMs is to a large degree due to the estimation of the normalizing constant. The number of possible networks is already extremely large for very small networks and the computation is simply not feasible. Therefore, an appropriate sample has to be found, ideally covering the most probable areas of the probability distribution. For this we make use of a method from the Markov Chain Monte Carlo (MCMC) family, namely the Metropolis-Hastings algorithm.
The idea behind this algorithm is to generate and sample highly weighted random networks departing from the observed network. The Metropolis-Hastings algorithm is an iterative algorithm which samples from the space of possible networks by randomly adding or removing edges from the starting network conditional on its density. If the likelihood, in the ERGM context also denoted as weights, of the newly generated network is higher than that of the departure network it is retained, otherwise it is discarded. In the former case, the algorithm starts anew from the newly generated network. Otherwise departure network is used again. Repeating this procedure sufficiently often and summing the weights associated to the stored (sampled) networks allows to compute an approximation of the denominator in equation EQREF18 (normalizing constant).
The algorithm starts sampling from the originally observed network INLINEFORM0 . The optimization of the coefficients is done simultaneously, equivalently with the Metropolis-Hastings algorithm. At the beginning starting values have to be supplied. For the study at hand we used the “ergm” library from the statistical R software package BIBREF5 implementing the Gibbs-Sampling algorithm BIBREF32 which is a special case of the Metropolis-Hastings algorithm outlined.
In order to answer our question of the importance of the factors which drive the network formation process in the roll-call co-voting network, the ERGM is specified with the following parameters:
nodematch country: This parameter adds one network statistic to the model, i.e., the number of edges INLINEFORM0 where INLINEFORM1 . The coefficient indicates the homophilious mixing behavior of MEPs with respect to their country of origin. In other words, this coefficient indicates how relevant nationality is in the formation of edges in the co-voting network.
nodematch national party: This parameter adds one network statistic to the model: the number of edges INLINEFORM0 with INLINEFORM1 . The coefficient indicates the homophilious mixing behavior of the MEPs with regard to their party affiliation at the national level. In the context of this study, this coefficient can be interpreted as an indicator for within-party cohesion at the national level.
nodemix EP group: This parameter adds one network statistic for each pair of European political groups. These coefficients shed light on the degree of coalitions between different groups as well as the within group cohesion . Given that there are nine groups in the European Parliament, this coefficient adds in total 81 statistics to the model.
edge covariate Twitter: This parameter corresponds to a square matrix with the dimension of the adjacency matrix of the network, which corresponds to the number of mutual retweets between the MEPs. It provides an insight about the extent to which communication in one social context (Twitter), can explain cooperation in another social context (co-voting in RCVs).
An ERGM as specified above is estimated for each of the 2535 roll-call votes. Each roll-call vote is thereby interpreted as a binary network and as an independent study. It is assumed that a priori each MEP could possibly form an edge with each other MEP in the context of a roll-call vote. Assumptions over the presence or absence of individual MEPs in a voting session are not made. In other words the dimensions of the adjacency matrix (the node set), and therefore the distribution from which new networks are drawn, is kept constant over all RCVs and therefore for every ERGM. The ERGM results therefore implicitly generalize to the case where potentially all MEPs are present and could be voting. Not voting is incorporated implicitly by the disconnectedness of a node.
The coefficients of the 2535 roll-call vote studies are aggregated by means of a meta-analysis approach proposed by Lubbers BIBREF33 and Snijders et al. BIBREF34 . We are interested in average effect sizes of different matching patterns over different topics and overall. Considering the number of RCVs, it seems straightforward to interpret the different RCV networks as multiplex networks and collapse them into one weighted network, which could then be analysed by means of a valued ERGM BIBREF35 . There are, however, two reasons why we chose the meta-analysis approach instead. First, aggregating the RCV data results into an extremely dense network, leading to severe convergence (degeneracy) problems for the ERGM. Second, the RCV data contains information about the different policy areas the individual votes were about. Since we are interested in how the coalition formation in the European Parliament differs over different areas, a method is needed that allows for an ex-post analysis of the corresponding results. We therefore opted for the meta-analysis approach by Lubbers and Snijders et al. This approach allows us to summarize the results by decomposing the coefficients into average effects and (class) subject-specific deviations. The different ERGM runs for each RCV are thereby regarded as different studies with identical samples that are combined to obtain a general overview of effect sizes. The meta-regression model is defined as: INLINEFORM0
Here INLINEFORM0 is a parameter estimate for class INLINEFORM1 , and INLINEFORM2 is the average coefficient. INLINEFORM3 denotes the normally distributed deviation of the class INLINEFORM4 with a mean of 0 and a variance of INLINEFORM5 . INLINEFORM6 is the estimation error of the parameter value INLINEFORM7 from the ERGM. The meta-analysis model is fitted by an iterated, weighted, least-squares model in which the observations are weighted by the inverse of their variances. For the overall nodematch between political groups, we weighted the coefficients by group sizes. The results from the meta analysis can be interpreted as if they stemmed from an individual ERGM run.
In our study, the meta-analysis was performed using the RSiena library BIBREF36 , which implements the method proposed by Lubbers and Snijders et al. BIBREF33 , BIBREF34 .
## Measuring cohesion and coalitions on Twitter
The retweeting behavior of MEPs is captured by their retweet network. Each MEP active on Twitter is a node in this network. An edge in the network between two MEPs exists when one MEP retweeted the other. The weight of the edge is the number of retweets between the two MEPs. The resulting retweet network is an undirected, weighted network.
We measure the cohesion of a political group INLINEFORM0 as the average retweets, i.e., the ratio of the number of retweets between the MEPs in the group INLINEFORM1 to the number of MEPs in the group INLINEFORM2 . The higher the ratio, the more each MEP (on average) retweets the MEPs from the same political group, hence, the higher the cohesion of the political group. The definition of the average retweeets ( INLINEFORM3 ) of a group INLINEFORM4 is: INLINEFORM5
This measure of cohesion captures the aggregate retweeting behavior of the group. If we consider retweets as endorsements, a larger number of retweets within the group is an indicator of agreement between the MEPs in the group. It does not take into account the patterns of retweeting within the group, thus ignoring the social sub-structure of the group. This is a potentially interesting direction and we leave it for future work.
We employ an analogous measure for the strength of coalitions in the retweet network. The coalition strength between two groups INLINEFORM0 and INLINEFORM1 is the ratio of the number of retweets from one group to the other (but not within groups) INLINEFORM2 to the total number of MEPs in both groups, INLINEFORM3 . The definition of the average retweeets ( INLINEFORM4 ) between groups INLINEFORM5 and INLINEFORM6 is: INLINEFORM7
## Cohesion of political groups
In this section we first report on the level of cohesion of the European Parliament's groups by analyzing the co-voting through the agreement and ERGM measures. Next, we explore two important policy areas, namely Economic and monetary system and State and evolution of the Union. Finally, we analyze the cohesion of the European Parliament's groups on Twitter.
Existing research by Hix et al. BIBREF10 , BIBREF13 , BIBREF11 shows that the cohesion of the European political groups has been rising since the 1990s, and the level of cohesion remained high even after the EU's enlargement in 2004, when the number of MEPs increased from 626 to 732.
We measure the co-voting cohesion of the political groups in the Eighth European Parliament using Krippendorff's Alpha—the results are shown in Fig FIGREF30 (panel Overall). The Greens-EFA have the highest cohesion of all the groups. This finding is in line with an analysis of previous compositions of the Fifth and Sixth European Parliaments by Hix and Noury BIBREF11 , and the Seventh by VoteWatch BIBREF37 . They are closely followed by the S&D and EPP. Hix and Noury reported on the high cohesion of S&D in the Fifth and Sixth European Parliaments, and we also observe this in the current composition. They also reported a slightly less cohesive EPP-ED. This group split in 2009 into EPP and ECR. VoteWatch reports EPP to have cohesion on a par with Greens-EFA and S&D in the Seventh European Parliament. The cohesion level we observe in the current European Parliament is also similar to the level of Greens-EFA and S&D.
The catch-all group of the non-aligned (NI) comes out as the group with the lowest cohesion. In addition, among the least cohesive groups in the European Parliament are the Eurosceptics EFDD, which include the British UKIP led by Nigel Farage, and the ENL whose largest party are the French National Front, led by Marine Le Pen. Similarly, Hix and Noury found that the least cohesive groups in the Seventh European Parliament are the nationalists and Eurosceptics. The Eurosceptic IND/DEM, which participated in the Sixth European Parliament, transformed into the current EFDD, while the nationalistic UEN was dissolved in 2009.
We also measure the voting cohesion of the European Parliament groups using an ERGM, a network-based method—the results are shown in Fig FIGREF31 (panel Overall). The cohesion results obtained with ERGM are comparable to the results based on agreement. In this context, the parameters estimated by the ERGM refer to the matching of MEPs who belong to the same political group (one parameter per group). The parameters measure the homophilous matching between MEPs who have the same political affiliation. A positive value for the estimated parameter indicates that the co-voting of MEPs from that group is greater than what is expected by chance, where the expected number of co-voting links by chance in a group is taken to be uniformly random. A negative value indicates that there are fewer co-voting links within a group than expected by chance.
Even though INLINEFORM0 and ERGM compute scores relative to what is expected by chance, they refer to different interpretations of chance. INLINEFORM1 's concept of chance is based on the number of expected pair-wise co-votes between MEPs belonging to a group, knowing the votes of these MEPs on all RCVs. ERGM's concept of chance is based on the number of expected pair-wise co-votes between MEPs belonging to a group on a given RCV, knowing the network-related properties of the co-voting network on that particular RCV. The main difference between INLINEFORM2 and ERGM, though, is the treatment of non-voting and abstained MEPs. INLINEFORM3 considers only the yes/no votes, and consequently, agreements by the voting MEPs of the same groups are considerably higher than co-voting by chance. ERGM, on the other hand, always considers all MEPs, and non-voting and abstained MEPs are treated as disconnected nodes. The level of co-voting by chance is therefore considerably lower, since there is often a large fraction of MEPs that do not attand or abstain.
As with INLINEFORM0 , Greens-EFA, S&D, and EPP exhibit the highest cohesion, even though their ranking is permuted when compared to the ranking obtained with INLINEFORM1 . At the other end of the scale, we observe the same situation as with INLINEFORM2 . The non-aligned members NI have the lowest cohesion, followed by EFDD and ENL.
The only place where the two methods disagree is the level of cohesion of GUE-NGL. The Alpha attributes GUE-NGL a rather high level of cohesion, on a par with ALDE, whereas the ERGM attributes them a much lower cohesion. The reason for this difference is the relatively high abstention rate of GUE-NGL. Whereas the overall fraction of non-attending and abstaining MEPs across all RCVs and all political groups is 25%, the GUE-NGL abstention rate is 34%. This is reflected in an above average cohesion by INLINEFORM0 where only yes/no votes are considered, and in a relatively lower, below average cohesion by ERGM. In the later case, the non-attendance is interpreted as a non-cohesive voting of a political groups as a whole.
In addition to the overall cohesion, we also focus on two selected policy areas. The cohesion of the political groups related to these two policy areas is shown in the first two panels in Fig FIGREF30 ( INLINEFORM0 ) and Fig FIGREF31 (ERGM).
The most important observation is that the level of cohesion of the political groups is very stable across different policy areas. These results are corroborated by both methodologies. Similar to the overall cohesion, the most cohesive political groups are the S&D, Greens-EFA, and EPP. The least cohesive group is the NI, followed by the ENL and EFDD. The two methodologies agree on the level of cohesion for all the political groups, except for GUE-NGL, due to a lower attendance rate.
We determine the cohesion of political groups on Twitter by using the average number of retweets between MEPs within the same group. The results are shown in Fig FIGREF33 . The right-wing ENL and EFDD come out as the most cohesive groups, while all the other groups have a far lower average number of retweets. MEPs from ENL and EFDD post by far the largest number of retweets (over 240), and at the same time over 94% of their retweets are directed to MEPs from the same group. Moreover, these two groups stand out in the way the retweets are distributed within the group. A large portion of the retweets of EFDD (1755) go to Nigel Farage, the leader of the group. Likewise, a very large portion of retweets of ENL (2324) go to Marine Le Pen, the leader of the group. Farage and Le Pen are by far the two most retweeted MEPs, with the third one having only 666 retweets.
## Coalitions in the European Parliament
Coalition formation in the European Parliament is largely determined by ideological positions, reflected in the degree of cooperation of parties at the national and European levels. The observation of ideological inclinations in the coalition formation within the European Parliament was already made by other authors BIBREF11 and is confirmed in this study. The basic patterns of coalition formation in the European Parliament can already be seen in the co-voting network in Fig FIGREF2 A. It is remarkable that the degree of attachment between the political groups, which indicates the degree of cooperation in the European Parliament, nearly exactly corresponds to the left-to-right seating order.
The liberal ALDE seems to have an intermediator role between the left and right parts of the spectrum in the parliament. Between the extreme (GUE-NGL) and center left (S&D) groups, this function seems to be occupied by Greens-EFA. The non-aligned members NI, as well as the Eurosceptic EFFD and ENL, seem to alternately tip the balance on both poles of the political spectrum. Being ideologically more inclined to vote with other conservative and right-wing groups (EPP, ECR), they sometimes also cooperate with the extreme left-wing group (GUE-NGL) with which they share their Euroscepticism as a common denominator.
Figs FIGREF36 and FIGREF37 give a more detailed understanding of the coalition formation in the European Parliament. Fig FIGREF36 displays the degree of agreement or cooperation between political groups measured by Krippendorff's INLINEFORM0 , whereas Fig FIGREF37 is based on the result from the ERGM. We first focus on the overall results displayed in the right-hand plots of Figs FIGREF36 and FIGREF37 .
The strongest degrees of cooperation are observed, with both methods, between the two major parties (EPP and S&D) on the one hand, and the liberal ALDE on the other. Furthermore, we see a strong propensity for Greens-EFA to vote with the Social Democrats (5th strongest coalition by INLINEFORM0 , and 3rd by ERGM) and the GUE-NGL (3rd strongest coalition by INLINEFORM1 , and 5th by ERGM). These results underline the role of ALDE and Greens-EFA as intermediaries for the larger groups to achieve a majority. Although the two largest groups together have 405 seats and thus significantly more than the 376 votes needed for a simple majority, the degree of cooperation between the two major groups is ranked only as the fourth strongest by both methods. This suggests that these two political groups find it easier to negotiate deals with smaller counterparts than with the other large group. This observation was also made by Hix et al. BIBREF12 , who noted that alignments on the left and right of the political spectrum have in recent years replaced the “Grand Coalition” between the two large blocks of Christian Conservatives (EPP) and Social Democrats (S&D) as the dominant form of finding majorities in the parliament.
Next, we focus on the coalition formation within the two selected policy areas. The area State and Evolution of the Union is dominated by cooperation between the two major groups, S&D and EPP, as well as ALDE. We also observe a high degree of cooperation between groups that are generally regarded as integration friendly, like Greens-EFA and GUE-NGL. We see, particularly in Fig FIGREF36 , a relatively high degree of cooperation between groups considered as Eurosceptic, like ECR, EFFD, ENL, and the group of non-aligned members.
The dichotomy between supporters and opponents of European integration is even more pronounced within the policy area Economic and Monetary System. In fact, we are taking a closer look specifically at these two areas as they are, at the same time, both contentious and important. Both methods rank the cooperation between S&D and EPP on the one hand, and ALDE on the other, as the strongest.
We also observe a certain degree of unanimity among the Eurosceptic and right-wing groups (EFDD, ENL, and NI) in this policy area. This seems plausible, as these groups were (especially in the aftermath of the global financial crisis and the subsequent European debt crisis) in fierce opposition to further payments to financially troubled member states. However, we also observe a number of strong coalitions that might, at first glance, seem unusual, specifically involving the left-wing group GUE-NGL on the one hand, and the right-wing EFDD, ENL, and NI on the other. These links also show up in the network plot in Fig FIGREF2 A. This might be attributable to a certain degree of Euroscepticism on both sides: rooted in criticism of capitalism on the left, and at least partly a raison d'être on the right. Hix et al. BIBREF11 discovered this pattern as well, and proposed an additional explanation—these coalitions also relate to a form of government-opposition dynamic that is rooted at the national level, but is reflected in voting patterns at the European level.
In general, we observe two main differences between the INLINEFORM0 and ERGM results: the baseline cooperation as estimated by INLINEFORM1 is higher, and the ordering of coalitions from the strongest to the weakest is not exactly the same. The reason is the same as for the cohesion, namely different treatment of non-voting and abstaining MEPs. When they are ignored, as by INLINEFORM2 , the baseline level of inter-group co-voting is higher. When non-attending and abstaining is treated as voting differently, as by ERGM, it is considerably more difficult to achieve co-voting coalitions, specially when there are on average 25% MEPs that do not attend or abstain. Groups with higher non-attendance rates, such as GUE-NGL (34%) and NI (40%) are less likely to form coalitions, and therefore have relatively lower ERGM coefficients (Fig FIGREF37 ) than INLINEFORM3 scores (Fig FIGREF36 ).
The first insight into coalition formation on Twitter can be observed in the retweet network in Fig FIGREF2 B. The ideological left to right alignment of the political groups is reflected in the retweet network. Fig FIGREF40 shows the strength of the coalitions on Twitter, as estimated by the number of retweets between MEPs from different groups. The strongest coalitions are formed between the right-wing groups EFDD and ECR, as well as ENL and NI. At first, this might come as a surprise, since these groups do not form strong coalitions in the European Parliament, as can be seen in Figs FIGREF36 and FIGREF37 . On the other hand, the MEPs from these groups are very active Twitter users. As previously stated, MEPs from ENL and EFDD post the largest number of retweets. Moreover, 63% of the retweets outside of ENL are retweets of NI. This effect is even more pronounced with MEPs from EFDD, whose retweets of ECR account for 74% of their retweets from other groups.
In addition to these strong coalitions on the right wing, we find coalition patterns to be very similar to the voting coalitions observed in the European Parliament, seen in Figs FIGREF36 and FIGREF37 . The strongest coalitions, which come immediately after the right-wing coalitions, are between Greens-EFA on the one hand, and GUE-NGL and S&D on the other, as well as ALDE on the one hand, and EPP and S&D on the other. These results corroborate the role of ALDE and Greens-EFA as intermediaries in the European Parliament, not only in the legislative process, but also in the debate on social media.
To better understand the formation of coalitions in the European Parliament and on Twitter, we examine the strongest cooperation between political groups at three different thresholds. For co-voting coalitions in the European Parliament we choose a high threshold of INLINEFORM0 , a medium threshold of INLINEFORM1 , and a negative threshold of INLINEFORM2 (which corresponds to strong oppositions). In this way we observe the overall patterns of coalition and opposition formation in the European Parliament and in the two specific policy areas. For cooperation on Twitter, we choose a high threshold of INLINEFORM3 , a medium threshold of INLINEFORM4 , and a very low threshold of INLINEFORM5 .
The strongest cooperations in the European Parliament over all policy areas are shown in Fig FIGREF42 G. It comes as no surprise that the strongest cooperations are within the groups (in the diagonal). Moreover, we again observe GUE-NGL, S&D, Greens-EFA, ALDE, and EPP as the most cohesive groups. In Fig FIGREF42 H, we observe coalitions forming along the diagonal, which represents the seating order in the European Parliament. Within this pattern, we observe four blocks of coalitions: on the left, between GUE-NGL, S&D, and Greens-EFA; in the center, between S&D, Greens-EFA, ALDE, and EPP; on the right-center between ALDE, EPP, and ECR; and finally, on the far-right between ECR, EFDD, ENL, and NI. Fig FIGREF42 I shows the strongest opposition between groups that systematically disagree in voting. The strongest disagreements are between left- and right-aligned groups, but not between the left-most and right-most groups, in particular, between GUE-NGL and ECR, but also between S&D and Greens-EFA on one side, and ENL and NI on the other.
In the area of Economic and monetary system we see a strong cooperation between EPP and S&D (Fig FIGREF42 A), which is on a par with the cohesion of the most cohesive groups (GUE-NGL, S&D, Greens-EFA, ALDE, and EPP), and is above the cohesion of the other groups. As pointed out in the section “sec:coalitionpolicy”, there is a strong separation in two blocks between supporters and opponents of European integration, which is even more clearly observed in Fig FIGREF42 B. On one hand, we observe cooperation between S&D, ALDE, EPP, and ECR, and on the other, cooperation between GUE-NGL, Greens-EFA, EFDD, ENL, and NI. This division in blocks is seen again in Fig FIGREF42 C, which shows the strongest disagreements. Here, we observe two blocks composed of S&D, EPP, and ALDE on one hand, and GUE-NGL, EFDD, ENL, and NI on the other, which are in strong opposition to each other.
In the area of State and Evolution of the Union we again observe a strong division in two blocks (see Fig FIGREF42 E). This is different to the Economic and monetary system, however, where we observe a far-left and far-right cooperation, where the division is along the traditional left-right axis.
The patterns of coalitions forming on Twitter closely resemble those in the European Parliament. In Fig FIGREF42 J we see that the strongest degrees of cooperation on Twitter are within the groups. The only group with low cohesion is the NI, whose members have only seven retweets between them. The coalitions on Twitter follow the seating order in the European Parliament remarkably well (see Fig FIGREF42 K). What is striking is that the same blocks form on the left, center, and on the center-right, both in the European Parliament and on Twitter. The largest difference between the coalitions in the European Parliament and on Twitter is on the far-right, where we observe ENL and NI as isolated blocks.
The results shown in Fig FIGREF44 quantify the extent to which communication in one social context (Twitter) can explain cooperation in another social context (co-voting in the European Parliament). A positive value indicates that the matching behavior in the retweet network is similar to the one in the co-voting network, specific for an individual policy area. On the other hand, a negative value implies a negative “correlation” between the retweeting and co-voting of MEPs in the two different contexts.
The bars in Fig FIGREF44 correspond to the coefficients from the edge covariate terms of the ERGM, describing the relationship between the retweeting and co-voting behavior of MEPs. The coefficients are aggregated for individual policy areas by means of a meta-analysis.
Overall, we observe a positive correlation between retweeting and co-voting, which is significantly different from zero. The strongest positive correlations are in the areas Area of freedom, security and justice, External relations of the Union, and Internal markets. Weaker, but still positive, correlations are observed in the areas Economic, social and territorial cohesion, European citizenship, and State and evolution of the Union. The only exception, with a significantly negative coefficient, is the area Economic and monetary system. This implies that in the area Economic and monetary system we observe a significant deviation from the usual co-voting patterns. Results from section “sec:coalitionpolicy”, confirm that this is indeed the case. Especially noteworthy are the coalitions between GUE-NGL and Greens-EFA on the left wing, and EFDD and ENL on the right wing. In the section “sec:coalitionpolicy” we interpret these results as a combination of Euroscepticism on both sides, motivated on the left by a skeptical attitude towards the market orientation of the EU, and on the right by a reluctance to give up national sovereignty.
## Discussion
We study cohesion and coalitions in the Eighth European Parliament by analyzing, on one hand, MEPs' co-voting tendencies and, on the other, their retweeting behavior.
We reveal that the most cohesive political group in the European Parliament, when it comes to co-voting, is Greens-EFA, closely followed by S&D and EPP. This is consistent with what VoteWatch BIBREF37 reported for the Seventh European Parliament. The non-aligned (NI) come out as the least cohesive group, followed by the Eurosceptic EFDD. Hix and Noury BIBREF11 also report that nationalists and Eurosceptics form the least cohesive groups in the Sixth European Parliament. We reaffirm most of these results with both of the two employed methodologies. The only point where the two methodologies disagree is in the level of cohesion for the left-wing GUE-NGL, which is portrayed by ERGM as a much less cohesive group, due to their relatively lower attendance rate. The level of cohesion of the political groups is quite stable across different policy areas and similar conclusions apply.
On Twitter we can see results that are consistent with the RCV results for the left-to-center political spectrum. The exception, which clearly stands out, is the right-wing groups ENL and EFDD that seem to be the most cohesive ones. This is the direct opposite of what was observed in the RCV data. We speculate that this phenomenon can be attributed to the fact that European right-wing groups, on a European but also on a national level, rely to a large degree on social media to spread their narratives critical of European integration. We observed the same phenomenon recently during the Brexit campaign BIBREF38 . Along our interpretation the Brexit was “won” to some extent due to these social media activities, which are practically non-existent among the pro-EU political groups. The fact that ENL and EFDD are the least cohesive groups in the European Parliament can be attributed to their political focus. It seems more important for the group to agree on its anti-EU stance and to call for independence and sovereignty, and much less important to agree on other issues put forward in the parliament.
The basic pattern of coalition formation, with respect to co-voting, can already be seen in Fig FIGREF2 A: the force-based layout almost completely corresponds to the seating order in the European Parliament (from the left- to the right-wing groups). A more thorough examination shows that the strongest cooperation can be observed, for both methodologies, between EPP, S&D, and ALDE, where EPP and S&D are the two largest groups, while the liberal ALDE plays the role of an intermediary in this context. On the other hand, the role of an intermediary between the far-left GUE-NGL and its center-left neighbor, S&D, is played by the Greens-EFA. These three parties also form a strong coalition in the European Parliament. On the far right of the spectrum, the non-aligned, EFDD, and ENL form another coalition. This behavior was also observed by Hix et al. BIBREF12 , stating that alignments on the left and right have in recent years replaced the “Grand Coalition” between the two large blocks of Christian Conservatives (EPP) and Social Democrats (S&D) as the dominant form of finding majorities in the European Parliament. When looking at the policy area Economic and monetary system, we see the same coalitions. However, interestingly, EFDD, ENL, and NI often co-vote with the far-left GUE-NGL. This can be attributed to a certain degree of Euroscepticism on both sides: as a criticism of capitalism, on one hand, and as the main political agenda, on the other. This pattern was also discovered by Hix et al. BIBREF12 , who argued that these coalitions emerge from a form of government-opposition dynamics, rooted at the national level, but also reflected at the European level.
When studying coalitions on Twitter, the strongest coalitions can be observed on the right of the spectrum (between EFDD, ECR, ENL, and NI). This is, yet again, in contrast to what was observed in the RCV data. The reason lies in the anti-EU messages they tend to collectively spread (retweet) across the network. This behavior forms strong retweet ties, not only within, but also between, these groups. For example, MEPs of EFDD mainly retweet MEPs from ECR (with the exception of MEPs from their own group). In contrast to these right-wing coalitions, we find the other coalitions to be consistent with what is observed in the RCV data. The strongest coalitions on the left-to-center part of the axis are those between GUE-NGL, Greens-EFA, and S&D, and between S&D, ALDE, and EPP. These results reaffirm the role of Greens-EFA and ALDE as intermediaries, not only in the European Parliament but also in the debates on social media.
Last, but not least, with the ERGM methodology we measure the extent to which the retweet network can explain the co-voting activities in the European Parliament. We compute this for each policy area separately and also over all RCVs. We conclude that the retweet network indeed matches the co-voting behavior, with the exception of one specific policy area. In the area Economic and monetary system, the links in the (overall) retweet network do not match the links in the co-voting network. Moreover, the negative coefficients imply a radically different formation of coalitions in the European Parliament. This is consistent with the results in Figs FIGREF36 and FIGREF37 (the left-hand panels), and is also observed in Fig FIGREF42 (the top charts). From these figures we see that in this particular case, the coalitions are also formed between the right-wing groups and the far-left GUE-NGL. As already explained, we attribute this to the degree of Euroscepticism that these groups share on this particular policy issue.
## Conclusions
In this paper we analyze (co-)voting patterns and social behavior of members of the European Parliament, as well as the interaction between these two systems. More precisely, we analyze a set of 2535 roll-call votes as well as the tweets and retweets of members of the MEPs in the period from October 2014 to February 2016. The results indicate a considerable level of correlation between these two complex systems. This is consistent with previous findings of Cherepnalkoski et al. BIBREF22 , who reconstructed the adherence of MEPs to their respective political or national group solely from their retweeting behavior.
We employ two different methodologies to quantify the co-voting patterns: Krippendorff's INLINEFORM0 and ERGM. They were developed in different fields of research, use different techniques, and are based on different assumptions, but in general they yield consistent results. However, there are some differences which have consequences for the interpretation of the results.
INLINEFORM0 is a measure of agreement, designed as a generalization of several specialized measures, that can compare different numbers of observations, in our case roll-call votes. It only considers yes/no votes. Absence and abstention by MEPs is ignored. Its baseline ( INLINEFORM1 ), i.e., co-voting by chance, is computed from the yes/no votes of all MEPs on all RCVs.
ERGMs are used in social-network analyses to determine factors influencing the edge formation process. In our case an edge between two MEPs is formed when they cast the same yes/no vote within a RCV. It is assumed that a priori each MEP can form a link with any other MEP. No assumptions about the presence or absence of individual MEPs in a voting session are made. Each RCV is analyzed as a separate binary network. The node set is thereby kept constant for each RCV network. While the ERGM departs from the originally observed network, where MEPs who didn't vote or abstained appear as isolated nodes, links between these nodes are possible within the network sampling process which is part of the ERGM optimization process. The results of several RCVs are aggregated by means of the meta-analysis approach. The baseline (ERGM coefficients INLINEFORM0 ), i.e., co-voting by chance, is computed from a large sample of randomly generated networks.
These two different baselines have to be taken into account when interpreting the results of INLINEFORM0 and ERGM. In a typical voting session, 25% of the MEPs are missing or abstaining. When assessing cohesion of political groups, all INLINEFORM1 values are well above the baseline, and the average INLINEFORM2 . The average ERGM cohesion coefficients, on the other hand, are around the baseline. The difference is even more pronounced for groups with higher non-attendance/abstention rates like GUE-NGL (34%) and NI (40%). When assessing strength of coalitions between pairs of groups, INLINEFORM3 values are balanced around the baseline, while the ERGM coefficients are mostly negative. The ordering of coalitions from the strongest to the weakest is therefor different when groups with high non-attendance/abstention rates are involved.
The choice of the methodology to asses cohesion and coalitions is not obvious. Roll-call voting is used for decisions which demand a simple majority only. One might however argue that non-attendance/abstention corresponds to a no vote, or that absence is used strategically. Also, the importance of individual votes, i.e., how high on the agenda of a political group is the subject, affects their attendance, and consequently the perception of their cohesion and the potential to act as a reliable coalition partner.
## Acknowledgments
This work was supported in part by the EC projects SIMPOL (no. 610704) and DOLFINS (no. 640772), and by the Slovenian ARRS programme Knowledge Technologies (no. P2-103).
| 12
|
1610.00879
|
A Computational Approach to Automatic Prediction of Drunk Texting
|
# A Computational Approach to Automatic Prediction of Drunk Texting
## Abstract
Alcohol abuse may lead to unsociable behavior such as crime, drunk driving, or privacy leaks. We introduce automatic drunk-texting prediction as the task of identifying whether a text was written when under the influence of alcohol. We experiment with tweets labeled using hashtags as distant supervision. Our classifiers use a set of N-gram and stylistic features to detect drunk tweets. Our observations present the first quantitative evidence that text contains signals that can be exploited to detect drunk-texting.
## Introduction
The ubiquity of communication devices has made social media highly accessible. The content on these media reflects a user's day-to-day activities. This includes content created under the influence of alcohol. In popular culture, this has been referred to as `drunk-texting'. In this paper, we introduce automatic `drunk-texting prediction' as a computational task. Given a tweet, the goal is to automatically identify if it was written by a drunk user. We refer to tweets written under the influence of alcohol as `drunk tweets', and the opposite as `sober tweets'.
A key challenge is to obtain an annotated dataset. We use hashtag-based supervision so that the authors of the tweets mention if they were drunk at the time of posting a tweet. We create three datasets by using different strategies that are related to the use of hashtags. We then present SVM-based classifiers that use N-gram and stylistic features such as capitalisation, spelling errors, etc. Through our experiments, we make subtle points related to: (a) the performance of our features, (b) how our approach compares against human ability to detect drunk-texting, (c) most discriminative stylistic features, and (d) an error analysis that points to future work. To the best of our knowledge, this is a first study that shows the feasibility of text-based analysis for drunk-texting prediction.
## Motivation
Past studies show the relation between alcohol abuse and unsociable behaviour such as aggression BIBREF0 , crime BIBREF1 , suicide attempts BIBREF2 , drunk driving BIBREF3 , and risky sexual behaviour BIBREF4 . suicide state that “those responsible for assessing cases of attempted suicide should be adept at detecting alcohol misuse”. Thus, a drunk-texting prediction system can be used to identify individuals susceptible to these behaviours, or for investigative purposes after an incident.
Drunk-texting may also cause regret. Mail Goggles prompts a user to solve math questions before sending an email on weekend evenings. Some Android applications avoid drunk-texting by blocking outgoing texts at the click of a button. However, to the best of our knowledge, these tools require a user command to begin blocking. An ongoing text-based analysis will be more helpful, especially since it offers a more natural setting by monitoring stream of social media text and not explicitly seeking user input. Thus, automatic drunk-texting prediction will improve systems aimed to avoid regrettable drunk-texting. To the best of our knowledge, ours is the first study that does a quantitative analysis, in terms of prediction of the drunk state by using textual clues.
Several studies have studied linguistic traits associated with emotion expression and mental health issues, suicidal nature, criminal status, etc. BIBREF5 , BIBREF6 . NLP techniques have been used in the past to address social safety and mental health issues BIBREF7 .
## Definition and Challenges
Drunk-texting prediction is the task of classifying a text as drunk or sober. For example, a tweet `Feeling buzzed. Can't remember how the evening went' must be predicted as `drunk', whereas, `Returned from work late today, the traffic was bad' must be predicted as `sober'. The challenges are:
## Dataset Creation
We use hashtag-based supervision to create our datasets, similar to tasks like emotion classification BIBREF8 . The tweets are downloaded using Twitter API (https://dev.twitter.com/). We remove non-Unicode characters, and eliminate tweets that contain hyperlinks and also tweets that are shorter than 6 words in length. Finally, hashtags used to indicate drunk or sober tweets are removed so that they provide labels, but do not act as features. The dataset is available on request. As a result, we create three datasets, each using a different strategy for sober tweets, as follows:
The drunk tweets for Datasets 1 and 2 are the same. Figure FIGREF9 shows a word-cloud for these drunk tweets (with stop words and forms of the word `drunk' removed), created using WordItOut. The size of a word indicates its frequency. In addition to topical words such as `bar', `bottle' and `wine', the word-cloud shows sentiment words such as `love' or `damn', along with profane words.
Heuristics other than these hashtags could have been used for dataset creation. For example, timestamps were a good option to account for time at which a tweet was posted. However, this could not be used because user's local times was not available, since very few users had geolocation enabled.
## Feature Design
The complete set of features is shown in Table TABREF7 . There are two sets of features: (a) N-gram features, and (b) Stylistic features. We use unigrams and bigrams as N-gram features- considering both presence and count.
Table TABREF7 shows the complete set of stylistic features of our prediction system. POS ratios are a set of features that record the proportion of each POS tag in the dataset (for example, the proportion of nouns/adjectives, etc.). The POS tags and named entity mentions are obtained from NLTK BIBREF9 . Discourse connectors are identified based on a manually created list. Spelling errors are identified using a spell checker by enchant. The repeated characters feature captures a situation in which a word contains a letter that is repeated three or more times, as in the case of happpy. Since drunk-texting is often associated with emotional expression, we also incorporate a set of sentiment-based features. These features include: count/presence of emoticons and sentiment ratio. Sentiment ratio is the proportion of positive and negative words in the tweet. To determine positive and negative words, we use the sentiment lexicon in mpqa. To identify a more refined set of words that correspond to the two classes, we also estimated 20 topics for the dataset by estimating an LDA model BIBREF10 . We then consider top 10 words per topic, for both classes. This results in 400 LDA-specific unigrams that are then used as features.
## Evaluation
Using the two sets of features, we train SVM classifiers BIBREF11 . We show the five-fold cross-validation performance of our features on Datasets 1 and 2, in Section SECREF17 , and on Dataset H in Section SECREF21 . Section SECREF22 presents an error analysis. Accuracy, positive/negative precision and positive/negative recall are shown as A, PP/NP and PR/NR respectively. `Drunk' forms the positive class, while `Sober' forms the negative class.
## Performance for Datasets 1 and 2
Table TABREF14 shows the performance for five-fold cross-validation for Datasets 1 and 2. In case of Dataset 1, we observe that N-gram features achieve an accuracy of 85.5%. We see that our stylistic features alone exhibit degraded performance, with an accuracy of 75.6%, in the case of Dataset 1. Table TABREF16 shows top stylistic features, when trained on the two datasets. Spelling errors, POS ratios for nouns (POS_NOUN), length and sentiment ratios appear in both lists, in addition to LDA-based unigrams. However, negative recall reduces to a mere 3.2%. This degradation implies that our features capture a subset of drunk tweets and that there are properties of drunk tweets that may be more subtle. When both N-gram and stylistic features are used, there is negligible improvement. The accuracy for Dataset 2 increases from 77.9% to 78.1%. Precision/Recall metrics do not change significantly either. The best accuracy of our classifier is 78.1% for all features, and 75.6% for stylistic features. This shows that text-based clues can indeed be used for drunk-texting prediction.
## Performance for Held-out Dataset H
Using held-out dataset H, we evaluate how our system performs in comparison to humans. Three annotators, A1-A3, mark each tweet in the Dataset H as drunk or sober. Table TABREF19 shows a moderate agreement between our annotators (for example, it is 0.42 for A1 and A2). Table TABREF20 compares our classifier with humans. Our human annotators perform the task with an average accuracy of 68.8%, while our classifier (with all features) trained on Dataset 2 reaches 64%. The classifier trained on Dataset 2 is better than which is trained on Dataset 1.
## Error Analysis
Some categories of errors that occur are:
Incorrect hashtag supervision: The tweet `Can't believe I lost my bag last night, literally had everything in! Thanks god the bar man found it' was marked with`#Drunk'. However, this tweet is not likely to be a drunk tweet, but describes a drunk episode in retrospective. Our classifier predicts it as sober.
Seemingly sober tweets: Human annotators as well as our classifier could not identify whether `Will you take her on a date? But really she does like you' was drunk, although the author of the tweet had marked it so. This example also highlights the difficulty of drunk-texting prediction.
Pragmatic difficulty: The tweet `National dress of Ireland is one's one vomit.. my family is lovely' was correctly identified by our human annotators as a drunk tweet. This tweet contains an element of humour and topic change, but our classifier could not capture it.
## Conclusion & Future Work
In this paper, we introduce automatic drunk-texting prediction as the task of predicting a tweet as drunk or sober. First, we justify the need for drunk-texting prediction as means of identifying risky social behavior arising out of alcohol abuse, and the need to build tools that avoid privacy leaks due to drunk-texting. We then highlight the challenges of drunk-texting prediction: one of the challenges is selection of negative examples (sober tweets). Using hashtag-based supervision, we create three datasets annotated with drunk or sober labels. We then present SVM-based classifiers which use two sets of features: N-gram and stylistic features. Our drunk prediction system obtains a best accuracy of 78.1%. We observe that our stylistic features add negligible value to N-gram features. We use our heldout dataset to compare how our system performs against human annotators. While human annotators achieve an accuracy of 68.8%, our system reaches reasonably close and performs with a best accuracy of 64%.
Our analysis of the task and experimental findings make a case for drunk-texting prediction as a useful and feasible NLP application.
| 10
|
1610.05243
|
Pre-Translation for Neural Machine Translation
|
# Pre-Translation for Neural Machine Translation
## Abstract
Recently, the development of neural machine translation (NMT) has significantly improved the translation quality of automatic machine translation. While most sentences are more accurate and fluent than translations by statistical machine translation (SMT)-based systems, in some cases, the NMT system produces translations that have a completely different meaning. This is especially the case when rare words occur. When using statistical machine translation, it has already been shown that significant gains can be achieved by simplifying the input in a preprocessing step. A commonly used example is the pre-reordering approach. In this work, we used phrase-based machine translation to pre-translate the input into the target language. Then a neural machine translation system generates the final hypothesis using the pre-translation. Thereby, we use either only the output of the phrase-based machine translation (PBMT) system or a combination of the PBMT output and the source sentence. We evaluate the technique on the English to German translation task. Using this approach we are able to outperform the PBMT system as well as the baseline neural MT system by up to 2 BLEU points. We analyzed the influence of the quality of the initial system on the final result.
## Introduction
In the last years, statistical machine translation (SMT) system generated state-of-the-art performance for most language pairs. Recently, systems using neural machine translation (NMT) were able to outperform SMT systems in several evaluations. These models are able to generate more fluent and accurate translation for most of sentences.
Neural machine translation systems provide the output with high fluency. A weakness of NMT systems, however, is that they sometimes lose the original meaning of the source words during translation. One example from the first conference on machine translation (WMT16) test set is the segment in Table TABREF1 .
The English word goalie is not translated to the correct German word Torwart, but to the German word Gott, which means god. One problem could be that we need to limit the vocabulary size in order to train the model efficiently. We used Byte Pair Encoding (BPE) BIBREF0 to represent the text using a fixed size vocabulary. In our case the word goali is splitted into three parts go, al and ie. Then it is more difficult to transport the meaning to the translation.
In contrast to this, in phrase-based machine translation (PBMT), we do not need to limit the vocabulary and are often able to translate words even if we have seen them only very rarely in the training. In the example mentioned before, for instance, the PBMT system had no problems translating the expression correctly.
On the other hand, official evaluation campaigns BIBREF1 have shown that NMT system often create grammatically correct sentence and are able to model the morphologically agreement much better in German.
The goal of this work is to combine the advantages of neural and phrase-based machine translation systems. Handling of rare words is an essential aspect to consider when it comes to real-world applications. The pre-translation framework provides a straightforward way to support such applications. In our approach, we will first translate the input using a PBMT system, which can handle the rare words well. In a second step, we will generate the final translation using an NMT system. This NMT system is able to generate a more fluent and grammatically correct translation. Since the rare words are already handled by the PBMT system, there should be less problems to generate the translation of these words. Using this approach naturally introduces a necessity to handle the potential errors by the PBMT systems.
The remaining of the paper is structured as follows: In the next section we will review the related work. In Section SECREF3 , we will briefly review the phrase-based and neural approach to machine translation. Section SECREF4 will introduce the approach presented in this paper to pre-translate the input using a PBMT system. In the following section, we will evaluate the approach and analyze the errors. Finally, we will finish with a conclusion.
## Related Work
The idea of linear combining of machine translation systems using different paradigms has already been used successfully for SMT and rule-based machine translation (RBMT) BIBREF2 , BIBREF3 . They build an SMT system that is post-editing the output of an RBMT system. Using the combination of SMT and RBMT, they could outperform both single systems.
Those experiments promote the area of automatic post-editing BIBREF4 . Recently, it was shown that models based on neural MT are very successful in this task BIBREF5 .
For PBMT, there has been several attempts to apply preprocessing in order to improve the performance of the translation system. A commonly used preprocessing step is morphological splitting, like compound splitting in German BIBREF6 . Another example would be to use pre-reordering in order to achieve more monotone translation BIBREF7 .
In addition, the usefulness of using the translations of the training data of a PBMT system has been shown. The translations have been used to re-train the translation model BIBREF8 or to train additional discriminative translation models BIBREF9 .
In order to improve the translation of rare words in NMT, authors try to translate words that are not in the vocabulary in a post-processing step BIBREF10 . In BIBREF0 , a method to split words into sub-word units was presented to limit the vocabulary size. Also the integration of lexical probabilities into NMT was successfully investigated BIBREF11 .
## Phrase-based and Neural Machine Translation
Starting with the initial work on word-based translation system BIBREF12 , phrase-based machine translation BIBREF13 , BIBREF14 segments the sentence into continuous phrases that are used as basic translation units. This allows for many-to-many alignments.
Based on this segmentation, the probability of the translation is calculated using a log-linear combination of different features: DISPLAYFORM0
In the initial model, the features are based on language and translation model probabilities as well as a few count based features. In advanced PBMT systems, several additional features to better model the translation process have been developed. Especially models using neural networks were able to increase the translation performance.
Recently, state-of-the art performance in machine translation was significantly improved by using neural machine translation. In this approach to machine translation, a recurrent neural network (RNN)-based encoder-decoder architecture is used to transform the source sentence into the target sentence.
In the encoder, an RNN is used to encode the source sentence into a fixed size continuous space representation by inserting the source sentence word-by-word into the network. In a second step, the decoder is initialized by the representation of the source sentence and is then generating the target sequence one word after the other using the last generated word as input for the RNN BIBREF15 .
One main drawback of this approach is that the whole source sentence has to be stored in a fixed-size context vector. To overcome this problem, BIBREF16 introduced the soft attention mechanism. Instead of only considering the last state of the encoder RNN, they use a weighted sum of all hidden states. Using these weights, the model is able to put attention on different parts of the source sentence depending on the current status of the decoder RNN. In addition, they extended the encoder RNN to a bi-directional one to be able to get information from the whole sentence at every position of the encoder RNN. A detailed description of the NMT framework can be found in BIBREF16 .
## PBMT Pre-translation for NMT (PreMT)
In this work, we want to combine the advantages of PBMT and NMT. Using the combined system we should be able to generate a translation for all words that occur at least once in the training data, while maintaining high quality translations for most sentences from NMT. Motivated by several approaches to simplify the translation process for PBMT using preprocessing, we will translate the source as a preprocessing step using the phrase-base machine translation system.
The main translation task is done by the neural machine translation model, which can choose between using the output of the PBMT system or the original input when generate the translation.
## Pipeline
In our first attempt, we combined the phrase-based MT and the neural MT in one pipeline as shown in Figure FIGREF3 . The input is first processed by the phrase-based machine translation system from the input language INLINEFORM0 to the target language INLINEFORM1 . Since the machine translation system is not perfect, the output of the system may not be correct translation containing errors possibly. Therefore, we will call the output language of the PBMT system INLINEFORM2 .
In a second step, we will train a neural monolingual translation system, that translates from the output of the PBMT system INLINEFORM0 to a better target sentence INLINEFORM1 .
## Mixed Input
One drawback of the pipelined approach is that the PBMT system might introduce some errors in the translation that the NMT can not recover from. For example, it is possible that some information from the source sentence gets lost, since the word is entirely deleted during the translation of the PBMT system.
We try to overcome this problem by building an NMT system that does not only take the output of the PBMT system, but also the original source sentence. One advantage of NMT system is that we can easily encode different input information. The architecture of our system is shown in Figure FIGREF3 .
The implementation of the mixed input for the NMT system is straight forward. Given the source input INLINEFORM0 and the output of the PBMT system INLINEFORM1 , we generated the input for the NMT system. First, we ensured a non-overlapping vocabulary of INLINEFORM2 and INLINEFORM3 by marking each token in INLINEFORM4 by a character and INLINEFORM5 by different ones. Then both input sequences are concatenated to the input INLINEFORM6 of the NMT system.
Using this representation, the NMT can learn to focus on source word INLINEFORM0 and words INLINEFORM1 when generating a word INLINEFORM2 .
## Training
In both cases, we can no longer train the NMT system on the source language and target language data, but on the output of the PBMT system and the target language data. Therefore, we need to generate translations of the whole parallel training data using the PBMT system.
Due to its ability to use very long phrases, a PBMT system normally performs significantly better on the training data than on unseen test data. This of course will harm the performance of our approach, because the NMT system will underestimate the number of improvements it has to perform on the test data.
In order to limit this effect, we did not use the whole phrase tables when translating the training data. If a phrase pair only occurs once, we cannot learn it from a different sentence pair. Following BIBREF9 , we removed all phrase pairs that occur only once for the translation of the corpus.
## Experiments
We analyze the approach on the English to German news translation task of the Conference on Statistical Machine Translation (WMT). First, we will describe the system and analyze the translation quality measured in BLEU. Afterwards, we will analyze the performance depending on the frequency of the words and finally show some example translations.
## System description
For the pre-translation, we used a PBMT system. In order to analyze the influence of the quality of the PBMT system, we use two different systems, a baseline system and a system with advanced models. The systems were trained on all parallel data available for the WMT 2016. The news commentary corpus, the European parliament proceedings and the common crawl corpus sum up to 3.7M sentences and around 90M words.
In the baseline system, we use three language models, a word-based, a bilingual BIBREF17 and a cluster based language model, using 100 automatically generated clusters using MKCLS BIBREF18 .
The advanced system use pre-reodering BIBREF19 and lexicalized reordering. In addition, it uses a discriminative word lexicon BIBREF9 and a language model trained on the large monolingual data.
Both systems were optimized on the tst2014 using Minimum error rate training BIBREF20 . A detailed description of the systems can be found in BIBREF21 .
The neural machine translation was trained using Nematus. For the NMT system as well as for the PreMT system, we used the default configuration. In order to limit the vocabulary size, we use BPE as described in BIBREF0 with 40K operations. We run the NMT system for 420K iterations and stored a model every 30K iterations. We selected the model that performed best on the development data. For the ensemble system we took the last four models. We did not perform an additional fine-tuning.
The PreMT system was trained on translations of the PBMT system of the corpus and the target side of the corpus. For this translation, we only used the baseline PBMT system.
## English - German Machine Translation
The results of all systems are summarized in Table TABREF13 . It has to be noted, that the first set, tst2014, has been used as development data for the PBMT system and as validation set for the NMT-based systems.
Using the neural MT system, we reach a BLEU score of 23.34 and 27.65 on tst2015 and tst2016. Using an ensemble system, we can improve the performance to 24.03 and 28.89 respectively. The baseline PBMT system performs 1.5 to 1.2 BLEU points worse than the single NMT system. Using the PBMT system with advanced models, we get the same performance on the tst2015 and 0.5 BLEU points better on tst2016 compared to the NMT system.
First, we build a PreMT system using the pipeline method as described in Section SECREF6 . The system reaches a BLEU score of 22.04 and 26.75 on both test sets. While the PreMT can improve of the baseline PBMT system, the performance is worse than the pure NMT system. So the first approach to combine neural and statistical machine translation is not able the combine the strength of both system. In contrast, the NMT system seems to be not able to recover from the errors done by the SMT-based system.
In a second experiment, we use the advanced PBMT system to generate the translation of the test data. We did not use it to generate a new training corpus, since the translation is computationally very expensive. So the PreMT system stays the same, being trained on the translation of the baseline PBMT. However, it is getting better quality translation in testing. This also leads to an improvement of 0.9 BLEU points on both test sets. Although it is smaller then the difference between the two initial phrase-based translation systems of around 1.5 BLUE points, we are able to improve the translation quality by using a better pre-translation system. It is interesting to see that we can improve the quality of the PreMT system, but improving one component (SMT Pre-Translation), even if we do it only in evaluation and not in training. But the system does not improve over the pure NMT system and even the post editing of the NMT system lowers the performance compared to the initial PBMT system used for pre-translation.
After evaluating the pipelined system, we performed experiments using the mixed input system. This leads to an improvement in translation quality. Using the baseline PBMT system for per-translation, we perform 0.8 BLEU points better than the purely NMT system on tst2015 and 0.4 BLEU point better on tst2016. It also showed better performance than both PBMT systems on tst2015 and comparable performance with the advanced PBMT on tst2016. So by looking at the original input and the pre-translation, the NMT system is able to recover some of the errors done by the PBMT system and also to prevent errors the NMT does if it is directly translating the source sentence.
Using the advanced PBMT system for input, we can get additional gains of 0.3 and 1.6 BLEU points The system even outperforms the ensemble system on tst2016. The experiments showed that deploying a pre-translation PBMT system with a better quality improves the NMT quality in the mixed input scheme, even when it is used only in testing, not in training.
By using an ensemble of four model, we improve the model by one BLEU point on both test sets, leading to the best results of 25.35 and 30.67 BLEU points. This is 1.3 and 1.8 BLEU points better than the pure NMT ensemble system.
## System Comparison
After evaluating the approach, we further analyze the different techniques for machine translation. For this, we compared the single NMT system, the advanced PBMT system and the mixed system using the advanced PBMT system as input.
Out initial idea was that PBMT systems are better for translating rare words, while the NMT is generating more fluent translation. To confirm this assumption, we edited the output of all system. For all analyzed systems, we replaced all target words, which occur in the training data less than INLINEFORM0 times, by the UNK token. For large INLINEFORM1 , we have therefore only the most frequent words in the reference, while for lower INLINEFORM2 more and more words are used.
The results for INLINEFORM0 are shown in Figure FIGREF15 . Of course, with lower INLINEFORM1 we will have fewer UNK tokens in the output. Therefore, we normalized the BLEU scores by the performance of the PreMT system.
We can see in the figure, that when INLINEFORM0 , where only the common words are used, we perform best using the NMT system. The PreMT system performs similar and the PBMT system performs clearly worse. If we now decrease INLINEFORM1 , more and more less frequent words will be considered in the evaluation of the translation quality. Although the absolute BLEU scores raise for all systems, on these less frequent words the PBMT performs better than the NMT system and therefore, finally it even achieves a better performance.
In contrast to this, the PreMT is able to benefit from the pre-translation of the PBMT system and therefore stays better than the PBMT system.
## Examples
In Table TABREF17 we show the output of the PBMT, NMT and PreMT system. First, for the PBMT system, we see a typical error when translating from and to German. The verb of the subclause parried is located at the second position in English, but in the German sentence it has to be located at the end of the sentence. The PBMT system is often not able to perform this long-range reordering.
For the NMT system, we see two other errors. Both, the words goalie and parried are quite rarely in the training data and therefore, they are splitted into several parts by the BPE algorithm. In this case, the NMT makes more errors. For the first word, the NMT system generates a complete wrong translation Gott (engl. god) instead of Torwart. The second word is just dropped and does not appear in the translation.
The example shows that the pre-translation system prevents both errors. It is generating the correct words Torwart and pariert and putting them at the correct position in the German sentence.
To better understand how the pre-translation system is able to generate this translation, we also generated the alignment matrix of the attention model as shown in Figure FIGREF18 . The x-axis shows the input, where the words from the pre-translation are marked by D_ and the words from the original source by E_. The y-axis carries the translation. The symbol @@ marks subword units generated by the BPE algorithm. First, as indicated by the two diagonal lines the model is considering as both inputs, the original source and the pre-translation by the two diagonal lines.
Secondly, we see that the attention model is mainly focusing on the pre-translation for words that are not common and therefore got splitted into several parts by the BPE, such as shoot, goalie and parried.
A second example, which shows what happens with rare words occur in the source sentence, is shown in Table TABREF19 . In this case, the word riot is not translated but just passed to the target language. This behaviour is helpful for rare words like named entities, but the NMT system is using it also for many words that are not named entities. Other examples for words that were just passed through and not translated are crossbar or vigil.
## Conclusion
In this paper, we presented a technique to combine phrase-based and neural machine translation. Motivated by success in statistical machine translation, we used phrase-based machine translation to pre-translate the input and then we generate the final translation using neural machine translation.
While a simple serial combination of both models could not generate better translation than the neural machine translation system, we are able to improve over neural machine translation using a mixed input. By simple concatenation of the phrase-based translation and the original source as input for the neural machine translation, we can increase the machine translation quality measured in BLEU. The single pre-translated system could even outperform the ensemble NMT system. For the ensemble system, the PreMT system could outperform the NMT system by up to 1.8 BLEU points.
Using the combined approach, we can generate more fluent translation typical for the NMT system, but also translate rare words. These are often more easily translated by PBMT. Furthermore, we are able to improve the overall system performance by improving the individual components.
## Acknowledgments
The project leading to this application has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement n INLINEFORM0 645452. This work was supported by the Carl-Zeiss-Stiftung.
| 14
|
1610.08597
|
Word Embeddings to Enhance Twitter Gang Member Profile Identification
|
# Word Embeddings to Enhance Twitter Gang Member Profile Identification
## Abstract
Gang affiliates have joined the masses who use social media to share thoughts and actions publicly. Interestingly, they use this public medium to express recent illegal actions, to intimidate others, and to share outrageous images and statements. Agencies able to unearth these profiles may thus be able to anticipate, stop, or hasten the investigation of gang-related crimes. This paper investigates the use of word embeddings to help identify gang members on Twitter. Building on our previous work, we generate word embeddings that translate what Twitter users post in their profile descriptions, tweets, profile images, and linked YouTube content to a real vector format amenable for machine learning classification. Our experimental results show that pre-trained word embeddings can boost the accuracy of supervised learning algorithms trained over gang members social media posts.
## Introduction
Street gangs are defined as “a coalition of peers, united by mutual interests, with identifiable leadership and internal organization, who act collectively to conduct illegal activity and to control a territory, facility, or enterprise” BIBREF0 . They promote criminal activities such as drug trafficking, assault, robbery, and threatening or intimidating a neighborhood BIBREF1 . Today, over 1.4 million people, belonging to more than 33,000 gangs, are active in the United States BIBREF2 , of which 88% identify themselves as being members of a street gang. They are also active users of social media BIBREF2 ; according to 2007 National Assessment Center's survey of gang members, 25% of individuals in gangs use the Internet for at least 4 hours a week BIBREF3 . More recent studies report approximately 45% of gang members participate in online offending activities such as threatening, harassing individuals, posting violent videos or attacking someone on the street for something they said online BIBREF4 , BIBREF5 . They confirm that gang members use social media to express themselves in ways similar to their offline behavior on the streets BIBREF6 .
Despite its public nature, gang members post on social media without fear of consequences because there are only few tools law enforcement can presently use to surveil social media BIBREF7 . For example, the New York City police department employs over 300 detectives to combat teen violence triggered by insults, dares, and threats exchanged on social media, and the Toronto police department teaches officers about the use of social media in investigations BIBREF8 . From offline clues, the officers monitor just a selected set of social media accounts which are manually discovered and related to a specific investigation. Thus, developing tools to identify gang member profiles on social media is an important step in the direction of using machine intelligence to fight crime.
To help agencies monitor gang activity on social media, our past work investigated how features from Twitter profiles, including profile text, profile images, tweet text, emjoi use, and their links to YouTube, may be used to reliably find gang member profiles BIBREF9 . The diverse set of features, chosen to combat the fact that gang members often use local terms and hashtags in their posts, offered encouraging results. In this paper, we report our experience in integrating deep learning into our gang member profile classifier. Specifically, we investigate the effect of translating the features into a vector space using word embeddings BIBREF10 . This idea is motivated by the recent success of word embeddings-based methods to learn syntactic and semantic structures automatically when provided with large datasets. A dataset of over 3,000 gang and non-gang member profiles that we previously curated is used to train the word embeddings. We show that pre-trained word embeddings improve the machine learning models and help us obtain an INLINEFORM0 -score of INLINEFORM1 on gang member profiles (a 6.39% improvement in INLINEFORM2 -score compared to the baseline models which were not trained using word embeddings).
This paper is organized as follows. Section SECREF2 discusses the related literature and frames how this work differs from other related works. Section SECREF3 discusses our approach based on word embeddings to identify gang member profiles. Section SECREF4 reports on the evaluation of the proposed approach and the evaluation results in detail. Section SECREF5 concludes the work reported while discussing the future work planned.
## Related Work
Researchers have begun investigating the gang members' use of social media and have noticed the importance of identifying gang members' Twitter profiles a priori BIBREF6 , BIBREF7 . Before analyzing any textual context retrieved from their social media posts, knowing that a post has originated from a gang member could help systems to better understand the message conveyed by that post. Wijeratne et al. developed a framework to analyze what gang members post on social media BIBREF7 . Their framework could only extract social media posts from self identified gang members by searching for pre-identified gang names in a user's Twitter profile description. Patton et al. developed a method to collect tweets from a group of gang members operating in Detroit, MI BIBREF11 . However, their approach required the gang members' Twitter profile names to be known beforehand, and data collection was localized to a single city in the country. These studies investigated a small set of manually curated gang member profiles, often from a small geographic area that may bias their findings.
In our previous work BIBREF9 , we curated what may be the largest set of gang member profiles to study how gang member Twitter profiles can be automatically identified based on the content they share online. A data collection process involving location neutral keywords used by gang members, with an expanded search of their retweet, friends and follower networks, led to identifying 400 authentic gang member profiles on Twitter. Our study discovered that the text in their tweets and profile descriptions, their emoji use, their profile images, and music interests embodied by links to YouTube music videos, can help a classifier distinguish between gang and non-gang member profiles. While a very promising INLINEFORM0 measure with low false positive rate was achieved, we hypothesize that the diverse kinds and the multitude of features employed (e.g. unigrams of tweet text) could be amenable to an improved representation for classification. We thus explore the possibility of mapping these features into a considerably smaller feature space through the use of word embeddings.
Previous research has shown word embeddings-based methods can significantly improve short text classification BIBREF12 , BIBREF13 . For example, Lilleberget et al. showed that word embeddings weighted by INLINEFORM0 - INLINEFORM1 outperforms other variants of word embedding models discussed in BIBREF13 , after training word embedding models on over 18,000 newsgroup posts. Wang et al. showed that short text categorization can be improved by word embeddings with the help of a neural network model that feeds semantic cliques learned over word embeddings in to a convolutions neural network BIBREF12 . We believe our corpus of gang and non-gang member tweets, with nearly 64.6 million word tokens, could act as a rich resource to train word embeddings for distinguishing gang and non-gang member Twitter users. Our investigation differs from other word embeddings-based text classification systems such as BIBREF12 , BIBREF13 due to the fact that we use multiple feature types including emojis in tweets and image tags extracted from Twitter profile and cover images in our classification task.
## Word Embeddings
A word embedding model is a neural network that learns rich representations of words in a text corpus. It takes data from a large, INLINEFORM0 -dimensional `word space' (where INLINEFORM1 is the number of unique words in a corpus) and learns a transformation of the data into a lower INLINEFORM2 -dimensional space of real numbers. This transformation is developed in a way that similarities between the INLINEFORM3 -dimensional vector representation of two words reflects semantic relationships among the words themselves. These semantics are not captured by typical bag-of-words or INLINEFORM4 -gram models for classification tasks on text data BIBREF14 , BIBREF10 .
Word embeddings have led to the state-of-the-art results in many sequential learning tasks BIBREF15 . In fact, word embedding learning is an important step for many statistical language modeling tasks in text processing systems. Bengio et al. were the first ones to introduce the idea of learning a distributed representation for words over a text corpus BIBREF16 . They learned representations for each word in the text corpus using a neural network model that modeled the joint probability function of word sequences in terms of the feature vectors of the words in the sequence. Mikolov et al. showed that simple algebraic operations can be performed on word embeddings learned over a text corpus, which leads to findings such as the word embedding vector of the word “King” INLINEFORM0 the word embedding vectors of “Man” INLINEFORM1 “Woman” would results in a word embedding vector that is closest to the word embedding vector of the word “Queen” BIBREF14 . Recent successes in using word embeddings to improve text classification for short text BIBREF12 , BIBREF13 , encouraged us to explore how they can be used to improve gang and non-gang member Twitter profile classification.
Word embeddings can be performed under different neural network architectures; two popular ones are the Continuous Bag-of-Words (CBOW) and Continuous Skip-gram (Skip-gram) models BIBREF17 . The CBOW model learns a neural network such that given a set of context words surrounding a target word, it predict a target word. The Skip-gram model differs by predicting context words given a target word and by capturing the ordering of word occurrences. Recent improvements to Skip-gram model make it better able to handle less frequent words, especially when negative sampling is used BIBREF10 .
## Features considered
Gang member tweets and profile descriptions tend to have few textual indicators that demonstrate their gang affiliations or their tweets/profile text may carry acronyms which can only be deciphered by others involved in gang culture BIBREF9 . These gang-related terms are often local to gangs operating in neighborhoods and change rapidly when they form new gangs. Consequently, building a database of keywords, phrases, and other identifiers to find gang members nationally is not feasible. Instead, we use heterogeneous sets of features derived not only from profile and tweet text but also from the emoji usage, profile images, and links to YouTube videos reflecting their music preferences and affinity. In this section, we briefly discuss the feature types and their broad differences in gang and non-gang member profiles. An in-depth explanation of these feature selection can be found in BIBREF9 .
Tweet text: In our previous work, we observed that gang members use curse words nearly five times more than the average curse words use on Twitter BIBREF9 . Further, we noticed that gang members mainly use Twitter to discuss drugs and money using terms such as smoke, high, hit, money, got, and need while non-gang members mainly discuss their feelings using terms such as new, like, love, know, want, and look.
Twitter profile description: We found gang member profile descriptions to be rife with curse words (nigga, fuck, and shit) while non-gang members use words related to their feelings or interests (love, life, music, and book). We noticed that gang members use their profile descriptions as a space to grieve for their fallen or incarcerated gang members as about INLINEFORM0 of gang member Twitter profiles used terms such as rip and free.
Emoji features: We found that the fuel pump emoji was the most frequently used emoji by gang members, which is often used in the context of selling or consuming marijuana. The pistol emoji was the second most frequently used emoji, which is often used with the police cop emoji in an `emoji chain' to express their hatred towards law enforcement officers. The money bag emoji, money with wings emoji, unlock emoji, and a variety of the angry face emoji such as the devil face emoji and imp emoji were also common in gang members' but not in non-gang members' tweets.
Twitter profile and cover images: We noticed that gang members often pose holding or pointing weapons, seen in a group fashion which displays a gangster culture, show off graffiti, hand signs, tattoos, and bulk cash in their profile and cover images. We used Clarifai web service to tag the profile and cover images of the Twitter users in our dataset and used the image tags returned by Clarifai API to train word embeddings. Tags such as trigger, bullet, and worship were unique for gang member profiles while non-gang member images had unique tags such as beach, seashore, dawn, wildlife, sand, and pet.
YouTube videos: We found that 51.25% of the gang members in our dataset have a tweet that links to a YouTube video. Further, we found that 76.58% of the shared links are related to hip-hop music, gangster rap, and the culture that surrounds this music genre BIBREF9 . Moreover, we found that eight YouTube links are shared on average by a gang member. The top 5 terms used in YouTube videos shared by gang members were shit, like, nigga, fuck, and lil while like, love, peopl, song, and get were the top 5 terms in non-gang members' video data.
## Classification approach
Figure FIGREF11 gives an overview of the steps to learn word embeddings and to integrate them into a classifier. We first convert any non-textual features such as emoji and profile images into textual features. We use Emoji for Python and Clarifai services, respectively, to convert emoji and images into text. Prior to training the word embeddings, we remove all the seed words used to find gang member profiles and stopwords, and perform stemming across all tweets and profile descriptions. We then feed all the training data (word INLINEFORM0 in Figure FIGREF11 ) we collected from our Twitter dataset to Word2Vec tool and train it using a Skip-gram model with negative sampling. When training the Skip-gram model, we set the negative sampling rate to 10 sample words, which seems to work well with medium-sized datasets BIBREF10 . We set the context word window to be 5, so that it will consider 5 words to left and right of the target word (words INLINEFORM1 to INLINEFORM2 in Figure FIGREF11 ). This setting is suitable for sentences where average sentence length is less than 11 words, as is the case in tweets BIBREF18 . We ignore the words that occur less than 5 times in our training corpus.
We investigated how well the local language has been captured by the word embedding models we trained. We used the `most similar' functionality offered by Word2Vec tool to understand what the model has learned about few gang-related slang terms which are specific to Chicago area. For example, we analyzed the ten most similar words learned by the word embedding model for the term BDK (Black Desciples Killers). We noticed that out of the 10 most similar words, five were names of local Chicago gangs, which are rivals of the Black Disciples Gang, two were different syntactic variations of BDK (bdkk, bdkkk) and the other three were different syntactic variations of GDK (gdk, gdkk, gdkkk). GDK is a local gang slang for `Gangster Disciples Killer' which is used by rivals of Gangster Disciples gang to show their hatred towards them. We found similar results for the term GDK. Out of the ten most similar words, six were showing hatred towards six different Gangster Disciples gangs that operate in Chicago area. We believe that those who used the term GDK to show their hatred towards Gangster Disciples gangs might be also having rivalry with the six gangs we found.
We obtain word vectors of size 300 from the learned word embeddings. To represent a Twitter profile, we retrieve word vectors for all the words that appear in a particular profile including the words appear in tweets, profile description, words extracted from emoji, cover and profile images converted to textual formats, and words extracted from YouTube video comments and descriptions for all YouTube videos shared in the user's timeline. Those word vectors are combined to compute the final feature vector for the Twitter profile. To combine the word vectors, we consider five different methods. Letting the size of a word vector be INLINEFORM0 , for a Twitter profile INLINEFORM1 with INLINEFORM2 unique words and the vector of the INLINEFORM3 word in INLINEFORM4 denoted by INLINEFORM5 , we compute the feature vector for the Twitter profile INLINEFORM6 by:
Sum of word embeddings INLINEFORM0 . This is the sum the word embedding vectors obtained for all words in a Twitter profile: INLINEFORM1
Mean of word embeddings INLINEFORM0 . This is the mean of the word embedding vectors of all words found in a Twitter profile: INLINEFORM1
Sum of word embeddings weighted by term frequency INLINEFORM0 . This is each word embedding vector multiplied by the word's frequency for the Twitter profile: INLINEFORM1
where INLINEFORM0 is the term frequency for the INLINEFORM1 word in profile INLINEFORM2 .
Sum of word embeddings weighted by INLINEFORM0 - INLINEFORM1 INLINEFORM2 . This is each word vector multiplied by the word's INLINEFORM3 - INLINEFORM4 for the Twitter profile: INLINEFORM5
where INLINEFORM0 is the INLINEFORM1 - INLINEFORM2 value for the INLINEFORM3 word in profile INLINEFORM4 .
Mean of word embeddings weighted by term frequency INLINEFORM0 . This is the mean of the word embedding vectors weighted by term frequency: INLINEFORM1
## Evaluation
We evaluate the performance of using word embeddings to discover gang member profiles on Twitter. We first discuss the dataset, learning algorithms and baseline comparison models used in the experiments. Then we discuss the 10-fold cross validation experiments and the evaluation matrices used. Finally we present the results of the experiments.
## Evaluation setup
We consider a dataset of curated gang and non-gang members' Twitter profiles collected from our previous work BIBREF9 . It was developed by querying the Followerwonk Web service API with location-neutral seed words known to be used by gang members across the U.S. in their Twitter profiles. The dataset was further expanded by examining the friends, follower, and retweet networks of the gang member profiles found by searching for seed words. Specific details about our data curation procedure are discussed in BIBREF9 . Ultimately, this dataset consists of 400 gang member profiles and 2,865 non-gang member profiles. For each user profile, we collected up to most recent 3,200 tweets from their Twitter timelines, profile description text, profile and cover images, and the comments and video descriptions for every YouTube video shared by them. Table 1 provides statistics about the number of words found in each type of feature in the dataset. It includes a total of 821,412 tweets from gang members and 7,238,758 tweets from non-gang members.
To build the classifiers we used three different learning algorithms, namely Logistic Regression (LR), Random Forest (RF), and Support Vector Machines (SVM). We used version 0.17.1 of scikit-learn machine learning library for Python to implement the classifiers. An open source tool of Python, Gensim BIBREF19 was used to generate the word embeddings. We compare our results with the two best performing systems reported in BIBREF9 which are the two state-of-the-art models for identifying gang members in Twitter. Both baseline models are built from a random forest classifier trained over term frequencies for unigrams in tweet text, emoji, profile data, YouTube video data and image tags. Baseline Model(1) considers all 3,285 gang and non-gang member profiles in our dataset. Baseline Model(2) considers all Twitter profiles that contain every feature type discussed in Section SECREF2 . Because a Twitter profile may not have every feature type, baseline Model(1) represents a practical scenario where not every Twitter profile contains every type of feature. However, we compare our results to both baseline models and report the improvements.
## 10-fold cross validation
We conducted 10-fold cross validation experiments to evaluate the performance of our models. We used all Twitter profiles in the dataset to conduct experiments on the five methods we used to combine word embedding vectors. For each of the five vector combination methods (as mentioned in Section SECREF9 ), we trained classifiers using each learning algorithm we considered. In each fold, the training set was used to generate the word vectors, which were then used to compute features for both the training set and test set. For each 10-fold cross validation experiment, we report three evaluation metrics for the `gang' (positive) and `non-gang' (negative) classes, namely, the Precision = INLINEFORM0 , Recall = INLINEFORM1 , and INLINEFORM2 -score = INLINEFORM3 , where INLINEFORM4 is the number of true positives, INLINEFORM5 is the number of false positives, INLINEFORM6 is the number of true negatives, and INLINEFORM7 is the number of false negatives. We report these metrics for the `gang' and `non-gang' classes separately because of the class imbalance in the dataset.
## Experimental results
Table TABREF22 presents 10-fold cross validation results for the baseline models (first and second rows) and our word embeddings-based models (from third row to seventh row). As mentioned earlier both baseline models use a random forest classifier trained on term frequencies of unigram features extracted from all feature types. The two baseline models only differs on the training data filtering method used, which is based on the availability of features in the training dataset as described in BIBREF9 . The baseline Model(1) uses all profiles in the dataset and has a INLINEFORM0 -score of 0.7364 for `gang' class and 0.9690 for `non-gang' class. The baseline Model(2) which only uses profiles that contain each and every feature type has a INLINEFORM1 -score of 0.7755 for `gang' class and INLINEFORM2 -score of 0.9720 for `non-gang' class.
Vector sum is one of the basic operations we can perform on word embedding vectors. The random forest classifier performs the best among vector sum-based classifiers where logistic regression and SVM classifiers also perform comparatively well ( INLINEFORM0 ). Using vector mean ( INLINEFORM1 ) improves all classifier results and SVM classifier trained on the mean of word embeddings achieves very close results to the baseline Model(2). Multiplying vector sum with corresponding word counts for each word in word embeddings degrades the classifier accuracy for correctly identifying the positive class ( INLINEFORM2 ). When we multiply words by their corresponding INLINEFORM3 - INLINEFORM4 values before taking the vector sum, we again observe an increase in the classifiers' accuracy ( INLINEFORM5 ). We achieve the best performance by averaging the vector sum weighted by term frequency ( INLINEFORM6 ). Here we multiply the mean of the word embeddings by count of each word, which beats all other word embeddings-based models and the two baselines. In this setting, logistic regression classifier trained on word embeddings performs the best with a INLINEFORM7 -score of 0.7835. This is a 6.39% improvement in performance when compared to the baseline Model(1) and a 1.03% improvement in performance when compared to baseline Model(2). Overall, out of the five vector operations that we used to train machine learning classifiers, four gave us classifier models that beat baseline Model(1) and two vector based operations gave us classifier models that either achieved very similar results to baseline Model(2) or beat it. This evaluation demonstrates the promise of using pre-trained word embeddings to boost the accuracy of supervised learning algorithms for Twitter gang member profile classification.
## Conclusion and Future Work
This paper presented a word embeddings-based approach to address the problem of automatically identifying gang member profiles on Twitter. Using a Twitter user dataset that consist of 400 gang member and 2,865 non gang member profiles, we trained word embedding models based on users' tweets, profile descriptions, emoji, images, and videos shared on Twitter (textual features extracted from images, and videos). We then use the pre-trained word embedding models to train supervised machine learning classifiers, which showed superior performance when compared to the state-of-the-art baseline models reported in the literature. We plan to further extend our work by building our own image classification system specifically designed to identify images commonly shared by gang members such as guns, gang hand signs, stacks of cash and drugs. We would also like to experiment with automatically building dictionaries that contain gang names and gang-related slang using crowd-sourced gang-related knowledge-bases such as HipWiki. We also want to experiment with using such knowledge-bases to train word embeddings to understand whether having access to gang-related knowledge could boost the performance of our models. Finally, we would like to study how we can further use social networks of known gang members to identify new gang member profiles on Twitter.
| 10
|
1610.09516
|
Finding Street Gang Members on Twitter
|
# Finding Street Gang Members on Twitter
## Abstract
Most street gang members use Twitter to intimidate others, to present outrageous images and statements to the world, and to share recent illegal activities. Their tweets may thus be useful to law enforcement agencies to discover clues about recent crimes or to anticipate ones that may occur. Finding these posts, however, requires a method to discover gang member Twitter profiles. This is a challenging task since gang members represent a very small population of the 320 million Twitter users. This paper studies the problem of automatically finding gang members on Twitter. It outlines a process to curate one of the largest sets of verifiable gang member profiles that have ever been studied. A review of these profiles establishes differences in the language, images, YouTube links, and emojis gang members use compared to the rest of the Twitter population. Features from this review are used to train a series of supervised classifiers. Our classifier achieves a promising F1 score with a low false positive rate.
## Introduction and Motivation
The crime and violence street gangs introduce into neighborhoods is a growing epidemic in cities around the world. Today, over 1.23 million people in the United States are members of a street gang BIBREF0 , BIBREF1 , which is a coalition of peers, united by mutual interests, with identifiable leadership and internal organization, who act collectively to conduct illegal activity and to control a territory, facility, or enterprise BIBREF2 . They promote criminal activities such as drug trafficking, assault, robbery, and threatening or intimidating a neighborhood BIBREF1 . Moreover, data from the Centers for Disease Control in the United States suggests that the victims of at least 1.3% of all gang-related homicides are merely innocent bystanders who live in gang occupied neighborhoods BIBREF3 .
Street gang members have established online presences coinciding with their physical occupation of neighborhoods. The National Gang Threat Assessment Report confirms that at least tens of thousands of gang members are using social networking websites such as Twitter and video sharing websites such as YouTube in their daily life BIBREF0 . They are very active online; the 2007 National Assessment Center's survey of gang members found that 25% of individuals in gangs use the Internet for at least 4 hours a week BIBREF4 . Gang members typically use social networking sites and social media to develop online respect for their street gang BIBREF5 and to post intimidating, threatening images or videos BIBREF6 . This “Cyber-” or “Internet banging” BIBREF7 behavior is precipitated by the fact that an increasing number of young members of the society are joining gangs BIBREF8 , and these young members have become enamored with technology and with the notion of sharing information quickly and publicly through social media. Stronger police surveillance in the physical spaces where gangs congregate further encourages gang members to seek out virtual spaces such as social media to express their affiliation, to sell drugs, and to celebrate their illegal activities BIBREF9 .
Gang members are able to post publicly on Twitter without fear of consequences because there are few tools law enforcement can use to surveil this medium BIBREF10 . Police departments across the United States instead rely on manual processes to search social media for gang member profiles and to study their posts. For example, the New York City police department employs over 300 detectives to combat teen violence triggered by insults, dares, and threats exchanged on social media, and the Toronto police department teaches officers about the use of social media in investigations BIBREF11 . Officer training is broadly limited to understanding policies on using Twitter in investigations and best practices for data storage BIBREF12 . The safety and security of city neighborhoods can thus be improved if law enforcement were equipped with intelligent tools to study social media for gang activity.
The need for better tools for law enforcement cannot be underscored enough. Recent news reports have shown that many incidents involving gangs start on Twitter, escalate over time, and lead to an offline event that could have been prevented by an early warning. For example, the media reported on a possible connection between the death of a teenage rapper from Illinois and the final set of tweets he posted. One of his last tweets linked to a video of him shouting vulgar words at a rival gang member who, in return, replied “I'ma kill you” on social media. In a following tweet, the teenage rapper posted “im on 069”, revealing his location, and was shot dead soon after that post. Subsequent investigation revealed that the rivalry leading to his death began and was carried out entirely on social media. Other reporting has revealed how innocent bystanders have also become targets in online fights, leaving everyone in a neighborhood at risk.
This paper investigates whether gang member profiles can be identified automatically on Twitter, which can enable better surveillance of gang members on social media. Classifying Twitter profiles into particular types of users has been done in other contexts BIBREF13 , BIBREF14 , BIBREF15 , but gang member profiles pose unique challenges. For example, many Twitter profile classifiers search for contextual clues in tweets and profile descriptions BIBREF16 , but gang member profiles use a rapidly changing lexicon of keywords and phrases that often have only a local, geographic context. This is illustrated in Figure FIGREF6 , which shows the Twitter profile descriptions of two verified deceased gang members. The profile of @OsoArrogantJoJo provides evidence that he belongs to a rival gang of the Black Disciples by #BDK, a hashtag that is only known to those involved with gang culture in Chicago. @PappyNotPapi's profile mentions #PBG and our investigations revealed that this hashtag is newly founded and stands for the Pooh Bear Gang, a gang that was formerly known as the Insane Cutthroat Gangsters. Given the very local, rapidly changing lexicon of gang members on social media, building a database of keywords, phrases, and other identifiers to find gang members nationally is not feasible. Instead, this study proposes heterogeneous sets of features derived not only from profile and tweet text but also from the emoji usage, profile images, and links to YouTube videos reflecting their music culture. A large set of gang member profiles, obtained through a careful data collection process, is compared against non-gang member profiles to find contrasting features. Experimental results show that using these sets of features, we can build a classifier that has a low false positive rate and a promising INLINEFORM0 -score of 0.7755.
This paper is organized as follows. Section SECREF2 discusses the related literature and positions how this work differs from other related works. Section SECREF3 discusses the data collection, manual feature selection and our approach to identify gang member profiles. Section SECREF4 gives a detailed explanation for evaluation of the proposed method and the results in detail. Section SECREF5 concludes the work reported while discussing the future work planned.
## Related Work
Gang violence is a well studied social science topic dating back to 1927 BIBREF17 . However, the notions of “Cyber-” or “Internet banging”, which is defined as “the phenomenon of gang affiliates using social media sites to trade insults or make violent threats that lead to homicide or victimization” BIBREF7 , was only recently introduced BIBREF18 , BIBREF10 . Patton et al. introduced the concept of “Internet banging” and studied how social media is now being used as a tool for gang self-promotion and as a way for gang members to gain and maintain street credibility BIBREF7 . They also discussed the relationship between gang-related crime and hip-hop culture, giving examples on how hip-hop music shared on social media websites targeted at harassing rival gang members often ended up in real-world collisions among those gangs. Decker et al. and Patton et al. have also reported that street gangs perform Internet banging with social media posts of videos depicting their illegal behaviors, threats to rival gangs, and firearms BIBREF19 , BIBREF20 .
The ability to take action on these discoveries is limited by the tools available to discover gang members on social media and to analyze the content they post BIBREF18 . Recent attempts to improve our abilities include a proposed architecture for a surveillance system that can learn the structure, function, and operation of gangs through what they post on social media BIBREF10 . However, the architecture requires a set of gang member profiles for input, thus assuming that they have already been discovered. Patton et al. BIBREF20 devised a method to automatically collect tweets from a group of gang members operating in Detroit, MI. However, their approach required the profile names of the gang members to be known beforehand, and data collection was localized to a single city in the country.
This work builds upon existing methods to automatically discover gang member profiles on Twitter. This type of user profile classification problem has been explored in a diverse set of applications such as political affiliation BIBREF13 , ethnicity BIBREF13 , gender BIBREF15 , predicting brand loyalty BIBREF13 , and user occupations BIBREF16 . However, these approaches may utilize an abundance of positive examples in their training data, and only rely on a single feature type (typically, tweet text). Whereas most profile classifiers focus on a single type of feature (e.g. profile text), we consider the use of a variety of feature types, including emoji, YouTube links, and photo features.
## Discovering Gang Member Profiles
This section discusses the methodology we followed to study and classify the Twitter profiles of gang members automatically. It includes a semi-automatic data collection process to discover a large set of verifiable gang member profiles, an evaluation of the tweets of gang and non-gang member posts to identify promising features, and the deployment of multiple supervised learning algorithms to perform the classification.
## Data collection
Discovering gang member profiles on Twitter to build training and testing datasets is a challenging task. Past strategies to find these profiles were to search for keywords, phrases, and events that are known to be related to gang activity in a particular city a priori BIBREF10 , BIBREF20 . However, such approaches are unlikely to yield adequate data to train an automatic classifier since gang members from different geographic locations and cultures use local languages, location-specific hashtags, and share information related to activities in a local region BIBREF10 . Such region-specific tweets and profiles may be used to train a classifier to find gang members within a small region but not across the Twitterverse. To overcome these limitations, we adopted a semi-automatic workflow, illustrated in Figure FIGREF7 , to build a dataset of gang member profiles suitable for training a classifier. The steps of the workflow are:
1. Seed Term Discovery: Following the success of identifying gang member profiles from Chicago BIBREF10 , we began our data collection with discovering universal terms used by gang members. We first searched for profiles with hashtags for Chicago gangs noted in BIBREF10 , namely #BDK (Black Disciple Killers) and #GDK (Gangster Disciples Killers). Those profiles were analyzed and manually verified as explained in Step 3. Analysis of these profiles identified a small set of hashtags they all use in their profile descriptions. Searching Twitter profiles using those hashtags, we observed that gang members across the U.S. use them, thus we consider those terms to be location neutral. For example, gang members post #FreeDaGuys in their profile to support their fellow members who are in jail, #RIPDaGuys to convey the grieving for fallen gang members, and #FuckDaOpps to show their hatred towards police officers. We used these terms as keywords to discover Twitter profiles irrespective of geographical location. We used the Followerwonk Web service API and Twitter REST API to search Twitter profile descriptions by keywords #FreeDaGuys, #FreeMyNigga, #RIPDaGuys, and #FuckDaOpps. Since there are different informal ways people spell a word in social media, we also considered variations on the spelling of each keyword; for example, for #FreeDaGuys, we searched both #FreeDaGuys, and #FreeTheGuys.
2. Gang Affiliated Rappers' Twitter Profile Discovery: Finding profiles by a small set of keywords is unlikely to yield sufficient data. Thus, we sought additional gang member profiles with an observation from Patton et al. BIBREF7 that the influence of hip-hop music and culture on offline gang member activities can also be seen in their social media posts. We thus also consider the influence of hip-hop culture on Twitter by exploring the Twitter network of known gangster rappers who were murdered in 2015 due to gang-related incidents. We searched for these rapper profiles on Twitter and manually checked that the rapper was affiliated to a gang.
3. Manual verification of Twitter profiles: We verified each profile discovered manually by examining the profile picture, profile background image, recent tweets, and recent pictures posted by a user. During these checks, we searched for terms, activities, and symbols that we believed could be associated with a gang. For example, profiles whose image or background included guns in a threatening way, stacks of money, showing gang hand signs and gestures, and humans holding or posing with a gun, appeared likely to be from a gang member. Such images were often identified in profiles of users who submitted tweets that contain messages of support or sadness for prisoners or recently fallen gang members, or used a high volume of threatening and intimidating slang language. Only profiles where the images, words, and tweets all suggested gang affiliation were labeled as gang affiliates and added to our dataset. Although this manual verification does have a degree of subjectivity, in practice, the images and words used by gang members on social media are so pronounced that we believe any reasonable analyst would agree that they are gang members. We found that not all the profiles collected belonged to gang members; we observed relatives and followers of gang members posting the same hashtags as in Step 1 to convey similar feelings in their profile descriptions.
4. Using Retweets to discover more profiles: From the set of verified profiles, we explored their retweet and follower networks as a way to expand the dataset. We first considered authors of tweets which were retweeted by a gang member in our seed set. In Twitter, “retweeting” is a mechanism by which a user can share someone else's tweet to their follower audience. Assuming that a user only retweets things that they believe or their audience would be interested in, it may be reasonable to assume that gang members would only be interested in sharing what other gang members have to say, and hence, the authors of gang members' retweets could also be gang members.
5. Using Followers and Followees to discover more profiles: We analyzed followers and followees of our seed gang member profiles to find more gang member profiles. A Twitter user can follow other Twitter users so that the individual will be subscribed to their tweets as a follower and they will be able to start a private conversation by sending direct messages to the individual. Motivated by the sociological concept of homophily, which claims that individuals have a tendency to associate and bond with similar others, we hypothesized that the followers and followees of Twitter profiles from the seed set may also be gang members. Manual verification of Twitter profiles collected from retweets, followers, and followees of gang members showed that a majority of those profiles are non-gang members who are either family members, hip-hop artists, women or profiles with pornographic content. To ensure that our dataset is not biased towards a specific gang or geographic location, only a limited number of profiles were collected via retweets, followers and followees.
Table TABREF8 summarizes the number of profiles manually verified as gang members from Twitter profiles collected in step 1, 2, 4 and 5. Altogether we collected 400 gang member's Twitter profiles. This is a large number compared to previous studies of gang member activities on social media that curated a maximum of 91 profiles BIBREF10 . Moreover, we believe the profiles collected represent a diverse set of gang members that are not biased toward a particular geographic area or lingo as our data collection process used location-independent terms proven to be used by gang members when they express themselves.
## Data analysis
We next explore differences between gang and non-gang member Twitter usage to find promising features for classifying profiles. For this purpose, profiles of non-gang members were collected from the Twitter Streaming API. We collected a random sample of tweets and the profiles of the users who authored the tweets in the random sample. We manually verified that all Twitter profiles collected in this approach belong to non-gang members. The profiles selected were then filtered by location to remove non-U.S. profiles by reverse geo-coding the location stated in their profile description by the Google Maps API. Profiles with location descriptions that were unspecified or did not relate to a location in the U.S. were discarded. We collected 2,000 non-gang member profiles in this manner. In addition, we added 865 manually verified non-gang member profiles collected using the location neutral keywords discussed in Section SECREF3 . Introducing these profiles, which have some characteristics of gang members (such as cursing frequently or cursing at law enforcement) but are not, captures local languages used by family/friends of gang members and ordinary people in a neighborhood where gangs operate.
With the Twitter REST API, we collected the maximum number of most recent tweets that can be retrieved (3,200) along with profile descriptions and images (profile and cover photos) of every gang and non-gang member profile. The resulting dataset consists of 400 gang member Twitter profiles and 2,865 non-gang member Twitter profiles. The dataset has a total of 821,412 tweets from gang member profiles and 7,238,758 tweets from non-gang member profiles. Prior to analyzing any text content, we removed all of the seed words used to find gang member profiles, all stop words, and performed stemming across all tweets and profile descriptions.
Figure FIGREF14 summarizes the words seen most often in the gang and non-gang members' tweets as clouds. They show a clear difference in language. For example, we note that gang members more frequently use curse words in comparison to ordinary users. Although cursing is frequent in tweets, they represent just 1.15% of all words used BIBREF21 . In contrast, we found 5.72% of all words posted by gang member accounts to be classified as a curse word, which is nearly five times more than the average curse word usage on Twitter. The clouds also reflect the fact that gang members often talk about drugs and money with terms such as smoke, high, hit, and money, while ordinary users hardly speak about finances and drugs. We also noticed that gang members talk about material things with terms such as got, money, make, real, need whereas ordinary users tend to vocalize their feelings with terms such as new, like, love, know, want, look, make, us. These differences make it clear that the individual words used by gang and non-gang members will be relevant features for gang profile classification.
On Twitter, a user can give a self-description as a part of the user's profile. A comparison of the top 10 words in gang members' and non-gang members' Twitter profile descriptions is shown in Figure FIGREF21 . The first 10 words are the most frequently used words in non-gang members' profiles and the latter 10 words are the most frequently used words in gang members' profiles. Word comparison shows that gang members prefer to use curse words (nigga, fuck, shit) in their profile descriptions while non-gang members use words related to their feelings or interests (love, life, live, music, book). The terms rip and free which appear in approximately INLINEFORM0 of all gang member Twitter profiles, suggest that gang members use their profile descriptions as a space to grieve for their fallen or incarcerated gang members. The term gang in gang members' profile descriptions suggest that gang members like to self-identify themselves on Twitter. Such lexical features may therefore be of great importance for automatically identifying gang member profiles. We take counts of unigrams from gang and non-gang members' Twitter profile descriptions as classification features.
It has been recognized that music is a key cultural component in an urban lifestyle and that gang members often want to emulate the scenarios and activities the music conveys BIBREF7 . Our analysis confirms that the influence of gangster rap is expressed in gang members' Twitter posts. We found that 51.25% of the gang members collected have a tweet that links to a YouTube video. Following these links, a simple keyword search for the terms gangsta and hip-hop in the YouTube video description found that 76.58% of the shared links are related to hip-hop music, gangster rap, and the culture that surrounds this music genre. Moreover, this high proportion is not driven by a small number of profiles that prolifically share YouTube links; eight YouTube links are shared on average by a gang member.
Recognizing the frequency with which gang members post YouTube links on gangster rap and hip-hop, we consider the YouTube videos posted in a user's tweets as features for the classifier. In particular, for each YouTube video tweeted, we used the YouTube API to retrieve the video's description and its comments. Further analysis of YouTube data showed a difference between terms in gang members' YouTube data and non-gang members' YouTube data. For example, the top 5 terms (after stemming and stop word removal) used in YouTube videos shared by gang members are shit, like, nigga, fuck, lil while like, love, peopl, song, get are the top 5 terms in non-gang member video data. To represent a user profile based on their music interests, we generated a bag of words from the video descriptions and comments from all shared videos.
Motivated by recent work involving the use of emojis by gang members BIBREF22 , we also studied if and how gang and non-gang members use emoji symbols in their tweets. Our analysis found that gang members have a penchant for using just a small set of emoji symbols that convey their anger and violent behavior through their tweets. Figure FIGREF24 illustrates the emoji distribution for the top 20 most frequent emojis used by gang member profiles in our dataset. The fuel pump emoji was the most frequently used emoji by the gang members, which is often used in the context of selling or consuming marijuana. The pistol emoji is the second most frequent in our dataset, which is often used with the guardsman emoji or the police cop emoji in an `emoji chain'. Figure FIGREF28 presents some prototypical `chaining' of emojis used by gang members. The chains may reflect their anger at law enforcement officers, as a cop emoji is often followed up with the emoji of a weapon, bomb, or explosion. We found that 32.25% of gang members in our dataset have chained together the police and the pistol emoji, compared to just 1.14% of non-gang members. Moreover, only 1.71% of non-gang members have used the hundred points emoji and pistol emoji together in tweets while 53% of gang members have used them. A variety of the angry face emoji such as devil face emoji and imp emoji were also common in gang member tweets. The frequency of each emoji symbol used across the set of user's tweets are thus considered as features for our classifier.
In our profile verification process, we observed that most gang member profiles portray a context representative of gang culture. Some examples of these profile pictures are shown in Figure FIGREF32 , where the user holds or points weapons, is seen in a group fashion which displays a gangster culture, or is showing off graffiti, hand signs, tattoos and bulk cash. Descriptions of these images may thus empower our classifier. Thus, we translated profile images into features with the Clarifai web service. Clarifai offers a free API to query a deep learning system that tags images with a set of scored keywords that reflect what is seen in the image. We tagged the profile image and cover image for each profile using 20 tags identified by Clarifai. Figure FIGREF36 offers the 20 most often used tags applied to gang and non-gang member profiles. Since we take all the tags returned for an image, we see common words such as people and adult coming up in the top 20 tag set. However, gang member profile images were assigned unique tags such as trigger, bullet, worship while non-gang images were uniquely tagged with beach, seashore, dawn, wildlife, sand, pet. The set of tags returned by Clarifai were thus considered as features for the classifier.
## Learning algorithms
The unigrams of tweets, profile text, and linked YouTube video descriptions and comments, along with the distribution of emoji symbols and the profile image tags were used to train four different classification models: a Naive Bayes net, a Logistic Regression, a Random Forest, and a Support Vector Machine (SVM). These four models were chosen because they are known to perform well over text features, which is the dominant type of feature considered. The performance of the models are empirically compared to determine the most suitable classification technique for this problem. Data for the models are represented as a vector of term frequencies where the terms were collected from one or more feature sets described above.
## Evaluation
We next evaluate the performance of classifiers that use the above features to discover gang member profiles on Twitter. For this purpose, we use the training set discussed in Section SECREF3 with 400 gang member profiles (the `positive'/`gang' class) and 2,865 non-gang member profiles (the `negative'/`non-gang' class). We trained and evaluated the performance of the classifiers mentioned in Section SECREF31 under a 10-fold cross validation scheme. For each of the four learning algorithms, we consider variations involving only tweet text, emoji, profile, image, or music interest (YouTube comments and video description) features, and a final variant that considers all types of features together. The classifiers that use a single feature type were intended to help us study the quality of their predictive power by itself. When building these single-feature classifiers, we filtered the training dataset based on the availability of the single feature type in the training data. For example, we only used the twitter profiles that had at least a single emoji in their tweets to train classifiers that consider emoji features. We found 3,085 such profiles out of the 3,265 profiles in the training set. When all feature types were considered, we developed two different models:
Because a Twitter profile may not have every feature type, Model(1) represents a practical scenario where not every Twitter profile contains every type of feature. In this model, the non-occurrence of a feature is represented by `zeroing out' the feature value during model training. Model(2) represents the ideal scenario where all profiles contain every feature type. For this model, we used 1,358 training instances (42% of all training instances), out of which 172 were gang members (43% of all gang members) and 1,186 were non-gang members (41% of all non-gang members). We used version 0.17.1 of scikit-learn machine learning library to implement the classifiers.
For each 10-fold cross validation experiment, we report three evaluation metrics for the `gang' and `non-gang' classes, namely, the Precision = INLINEFORM0 , Recall = INLINEFORM1 , and INLINEFORM2 -score = INLINEFORM3 , where INLINEFORM4 is the number of true positives, INLINEFORM5 is the number of false positives, INLINEFORM6 is the number of true negatives, and INLINEFORM7 is the number of false negatives. We report these metrics for the positive `gang' and negative `non-gang' classes separately because of class imbalance in our dataset.
## Experimental results
Table TABREF37 presents the average precision, recall, and INLINEFORM0 -score over the 10 folds for the single-feature and combined feature classifiers. The table includes, in braces (`{ }'), the number of gang and non-gang profiles that contain a particular feature type, and hence the number of profiles used for the 10-fold cross validation. It is reasonable to expect that any Twitter profile is not that of a gang member, predicting a Twitter user as a non-gang member is much easier than predicting a Twitter user as a gang member. Moreover false positive classifications of the `gang' class may be detrimental to law enforcement investigations, which may go awry as they surveil an innocent person based on the classifier's suggestion. We thus believe that a small false positive rate of the `gang' class to be an especially important evaluation metric. We say that a classifier is `ideal' if it demonstrates high precision, recall, and INLINEFORM1 -score for the `gang' class while performing well on the `non-gang' class as well.
The best performing classifier that considers single features is a Random Forest model over tweet features (T), with a reasonable INLINEFORM0 -score of 0.7229 for the `gang' class. It also features the highest INLINEFORM1 -score for the `non-gang' class (0.9671). Its strong performance is intuitive given the striking differences in language as shown in Figure FIGREF14 and discussed in Section UID22 . We also noted that music features offer promising results, with an INLINEFORM2 -score of 0.6505 with a Naive Bayes classifier, as well as emoji features with an INLINEFORM3 -score of 0.6067 also achieved by a Naive Bayes classifier. However, the use of profile data and image tags by themselves yield relatively poor INLINEFORM4 -scores no matter which classifier considered. There may be two reasons for this despite the differences we observed in Section SECREF17 . First, these two feature types did not generate a large number of specific features for learning. For example, descriptions are limited to just 160 characters per profile, leading to a limited number of unigrams (in our dataset, 10 on average) that can be used to train the classifiers. Second, the profile images were tagged by a third party Web service which is not specifically designed to identify gang hand signs, drugs and guns, which are often shared by gang members. This led to a small set of image tags in their profiles that were fairly generic, i.e., the image tags in Figure FIGREF36 such as `people', `man', and `adult'.
Combining these diverse sets of features into a single classifier yields even better results. Our results for Model(1) show that the Random Forest achieves the highest INLINEFORM0 -scores for both `gang' (0.7364) and `non-gang' (0.9690) classes and yields the best precision of 0.8792, which corresponds to a low false positive rate when labeling a profile as a gang member. Despite the fact that it has lower positive recall compared to the second best performing classifier (a Random Forest trained over only tweet text features (T)), for this problem setting, we should be willing to increase the chance that a gang member will go unclassified if it means reducing the chance of applying a `gang' label to a non-gang member. When we tested Model(2), a Random Forrest classifier achieved an INLINEFORM1 -score of 0.7755 (improvement of 7.28% with respect to the best performing single feature type classifier (T)) for `gang' class with a precision of 0.8961 (improvement of 6.26% with respect to (T)) and a recall of 0.6994 (improvement of 9.26% with respect to (T)). Model(2) thus outperforms Model(1), and we expect its performance to improve with the availability of more training data with all feature types.
px
## Evaluation Over Unseen Profiles
We also tested the trained classifiers using a set of Twitter profiles from a separate data collection process that may emulate the classifier's operation in a real-time setting. For this experiment, we captured real-time tweets from Los Angeles, CA and from ten South Side, Chicago neighborhoods that are known for gang-related activities BIBREF10 using the Twitter streaming API. We consider these areas with known gang presence on social media to ensure that some positive profiles would appear in our test set. We ultimately collected 24,162 Twitter profiles: 15,662 from Los Angeles, and 8,500 from Chicago. We populated data for each profile by using the 3,200 most recent tweets (the maximum that can be collected from Twitter's API) for each profile. Since the 24,162 profiles are far too many to label manually, we qualitatively study those profiles the classifier placed into the `gang' class.
We used the training dataset to train our best performing random forest classifier (which use all feature types) and tested it on the test dataset. We then analyzed the Twitter profiles that our classifier labeled as belonging to the `gang' class. Each of those profiles had several features which overlap with gang members such as displaying hand signs and weapons in their profile images or in videos posted by them, gang names or gang-related hashtags in their profile descriptions, frequent use of curse words, and the use of terms such as “my homie" to refer to self-identified gang members. Representative tweets extracted from those profiles are depicted in Figure FIGREF41 . The most frequent words found in tweets from those profiles were shit, nigga, got, bitch, go, fuck etc. and their user profiles had terms such as free, artist, shit, fuck, freedagang, and ripthefallen. They had frequently used emojis such as face with tears of joy, hundred points symbol, fire, skull, money bag, and pistol. For some profiles, it was less obvious that the classifier correctly identified a gang member. Such profiles used the same emojis and curse words commonly found in gang members profiles, but their profile picture and tweet content was not indicative of a gang affiliation. In conclusion, we find that in a real-time-like setting, the classifier to be able to extract profiles with features that strongly suggest gang affiliation. Of course, these profiles demand further investigation and extensive evidence from other sources in order to draw a concrete conclusion, especially in the context of a law enforcement investigation. We refrain from reporting any profile names or specific details about the profiles labeled as a `gang' member to comply with the applicable IRB governing this human subject research.
px
## Conclusion and Future Work
This paper presented an approach to address the problem of automatically identifying gang member profiles on Twitter. Despite the challenges in developing such automated systems, mainly due to difficulties in finding online gang member profiles for developing training datasets, we proposed an approach that uses features extracted from textual descriptions, emojis, images and videos shared on Twitter (textual features extracted from images, and videos). Exploratory analysis of these types of features revealed interesting, and sometimes striking differences in the ways gang and non-gang members use Twitter. Classifiers trained over features that highlight these differences, were evaluated under 10-fold cross validation. Our best classifier achieved a promising INLINEFORM0 -score of 0.7755 over the `gang' profiles when all types of features were considered.
Future work will strengthen our training dataset by including more gang member Twitter profiles by searching for more location-independent keywords. We also plan to develop our own image classification system specifically designed to classify images found on gang member profiles. We would also like to experiment with building dictionaries that contain gang names to understand whether “having a gang name in the profile description” as a feature can improve our results. Finally, we would also like to study how can we further improve our classifier models using word embeddings BIBREF23 and social networks of known gang members.
px
## Acknowledgement
We are thankful to Uday Kiran Yeda for helping us with data collection. We acknowledge partial support from the National Science Foundation (NSF) award: CNS-1513721: “Context-Aware Harassment Detection on Social Media”, National Institutes of Health (NIH) award: MH105384-01A1: “Modeling Social Behavior for Healthcare Utilization in Depression” and Grant No. 2014-PS-PSN-00006 awarded by the Bureau of Justice Assistance. The Bureau of Justice Assistance is a component of the U.S. Department of Justice's Office of Justice Programs, which also includes the Bureau of Justice Statistics, the National Institute of Justice, the Office of Juvenile Justice and Delinquency Prevention, the Office for Victims of Crime, and the SMART Office. Points of view or opinions in this document are those of the authors and do not necessarily represent the official position or policies of the U.S. Department of Justice, NSF or NIH.
px
| 11
|
1611.01400
|
Learning to Rank Scientific Documents from the Crowd
|
# Learning to Rank Scientific Documents from the Crowd
## Abstract
Finding related published articles is an important task in any science, but with the explosion of new work in the biomedical domain it has become especially challenging. Most existing methodologies use text similarity metrics to identify whether two articles are related or not. However biomedical knowledge discovery is hypothesis-driven. The most related articles may not be ones with the highest text similarities. In this study, we first develop an innovative crowd-sourcing approach to build an expert-annotated document-ranking corpus. Using this corpus as the gold standard, we then evaluate the approaches of using text similarity to rank the relatedness of articles. Finally, we develop and evaluate a new supervised model to automatically rank related scientific articles. Our results show that authors' ranking differ significantly from rankings by text-similarity-based models. By training a learning-to-rank model on a subset of the annotated corpus, we found the best supervised learning-to-rank model (SVM-Rank) significantly surpassed state-of-the-art baseline systems.
[block]I.1em
[block]i.1em
Learning to Rank Scientific Documents from the CrowdLearning to Rank Scientific Documents from the Crowd
-4
[1]1
## Introduction
The number of biomedical research papers published has increased dramatically in recent years. As of October, 2016, PubMed houses over 26 million citations, with almost 1 million from the first 3 quarters of 2016 alone . It has become impossible for any one person to actually read all of the work being published. We require tools to help us determine which research articles would be most informative and related to a particular question or document. For example, a common task when reading articles is to find articles that are most related to another. Major research search engines offer such a “related articles” feature. However, we propose that instead of measuring relatedness by text-similarity measures, we build a model that is able to infer relatedness from the authors' judgments.
BIBREF0 consider two kinds of queries important to bibliographic information retrieval: the first is a search query written by the user and the second is a request for documents most similar to a document already judged relevant by the user. Such a query-by-document (or query-by-example) system has been implemented in the de facto scientific search engine PubMed—called Related Citation Search. BIBREF1 show that 19% of all PubMed searches performed by users have at least one click on a related article. Google Scholar provides a similar Related Articles system. Outside of bibliographic retrieval, query-by-document systems are commonly used for patent retrieval, Internet search, and plagiarism detection, amongst others. Most work in the area of query-by-document uses text-based similarity measures ( BIBREF2 , BIBREF3 , BIBREF4 ). However, scientific research is hypothesis driven and therefore we question whether text-based similarity alone is the best model for bibliographic retrieval. In this study we asked authors to rank documents by “closeness” to their work. The definition of “closeness” was left for the authors to interpret, as the goal is to model which documents the authors subjectively feel are closest to their own. Throughout the paper we will use “closeness” and “relatedness” interchangeably.
We found that researchers' ranking by closeness differs significantly from the ranking provided by a traditional IR system. Our contributions are three fold:
The principal ranking algorithms of query-by-document in bibliographic information retrieval rely mainly on text similarity measures ( BIBREF1 , BIBREF0 ). For example, the foundational work of BIBREF0 introduced the concept of a “document neighborhood” in which they pre-compute a text-similarity based distance between each pair of documents. When a user issues a query, first an initial set of related documents is retrieved. Then, the neighbors of each of those documents is retrieved, i.e., documents with the highest text similarity to those in the initial set. In a later work, BIBREF1 develop the PMRA algorithm for PubMed related article search. PMRA is an unsupervised probabilistic topic model that is trained to model “relatedness” between documents. BIBREF5 introduce the competing algorithm Find-Similar for this task, treating the full text of documents as a query and selecting related documents from the results.
Outside bibliographic IR, prior work in query-by-document includes patent retrieval ( BIBREF6 , BIBREF3 ), finding related documents given a manuscript ( BIBREF1 , BIBREF7 ), and web page search ( BIBREF8 , BIBREF9 ). Much of the work focuses on generating shorter queries from the lengthy document. For example, noun-phrase extraction has been used for extracting short, descriptive phrases from the original lengthy text ( BIBREF10 ). Topic models have been used to distill a document into a set of topics used to form query ( BIBREF11 ). BIBREF6 generated queries using the top TF*IDF weighted terms in each document. BIBREF4 suggested extracting phrasal concepts from a document, which are then used to generate queries. BIBREF2 combined query extraction and pseudo-relevance feedback for patent retrieval. BIBREF9 employ supervised machine learning model (i.e., Conditional Random Fields) ( BIBREF12 ) for query generation. BIBREF13 explored ontology to identify chemical concepts for queries.
There are also many biomedical-document specific search engines available. Many information retrieval systems focus on question answering systems such as those developed for the TREC Genomics Track ( BIBREF14 ) or BioASQ Question-Answer ( BIBREF15 ) competitions. Systems designed for question-answering use a combination of natural language processing techniques to identify biomedical entities, and then information retrieval systems to extract relevant answers to questions. Systems like those detailed in BIBREF16 can provide answers to yes/no biomedical questions with high precision. However what we propose differs from these systems in a fundamental way: given a specific document, suggest the most important documents that are related to it.
The body of work most related to ours is that of citation recommendation. The goal of citation recommendation is to suggest a small number of publications that can be used as high quality references for a particular article ( BIBREF17 , BIBREF1 ). Topic models have been used to rank articles based on the similarity of latent topic distribution ( BIBREF11 , BIBREF18 , BIBREF1 ). These models attempt to decompose a document into a few important keywords. Specifically, these models attempt to find a latent vector representation of a document that has a much smaller dimensionality than the document itself and compare the reduced dimension vectors.
Citation networks have also been explored for ranking articles by importance, i.e., authority ( BIBREF19 , BIBREF20 ). BIBREF17 introduced heterogeneous network models, called meta-path based models, to incorporate venues (the conference where a paper is published) and content (the term which links two articles, for citation recommendation). Another highly relevant work is BIBREF8 who decomposed a document to represent it with a compact vector, which is then used to measure the similarity with other documents. Note that we exclude the work of context-aware recommendation, which analyze each citation's local context, which is typically short and does not represent a full document.
One of the key contributions of our study is an innovative approach for automatically generating a query-by-document gold standard. Crowd-sourcing has generated large databases, including Wikipedia and Freebase. Recently, BIBREF21 concluded that unpaid participants performed better than paid participants for question answering. They attribute this to unpaid participants being more intrinsically motivated than the paid test takers: they performed the task for fun and already had knowledge about the subject being tested. In contrast, another study, BIBREF22 , compared unpaid workers found through Google Adwords (GA) to paid workers found through Amazon Mechanical Turk (AMT). They found that the paid participants from AMT outperform the unpaid ones. This is attributed to the paid workers being more willing to look up information they didn't know. In the bibliographic domain, authors of scientific publications have contributed annotations ( BIBREF23 ). They found that authors are more willing to annotate their own publications ( BIBREF23 ) than to annotate other publications ( BIBREF24 ) even though they are paid. In this work, our annotated dataset was created by the unpaid authors of the articles.
## Benchmark Datasets
In order to develop and evaluate ranking algorithms we need a benchmark dataset. However, to the best of our knowledge, we know of no openly available benchmark dataset for bibliographic query-by-document systems. We therefore created such a benchmark dataset.
The creation of any benchmark dataset is a daunting labor-intensive task, and in particular, challenging in the scientific domain because one must master the technical jargon of a scientific article, and such experts are not easy to find when using traditional crowd-sourcing technologies (e.g., AMT). For our task, the ideal annotator for each of our articles are the authors themselves. The authors of a publication typically have a clear knowledge of the references they cite and their scientific importance to their publication, and therefore may be excellent judges for ranking the reference articles.
Given the full text of a scientific publication, we want to rank its citations according to the author's judgments. We collected recent publications from the open-access PLoS journals and asked the authors to rank by closeness five citations we selected from their paper. PLoS articles were selected because its journals cover a wide array of topics and the full text articles are available in XML format. We selected the most recent publications as previous work in crowd-sourcing annotation shows that authors' willingness to participate in an unpaid annotation task declines with the age of publication ( BIBREF23 ). We then extracted the abstract, citations, full text, authors, and corresponding author email address from each document. The titles and abstracts of the citations were retrieved from PubMed, and the cosine similarity between the PLoS abstract and the citation's abstract was calculated. We selected the top five most similar abstracts using TF*IDF weighted cosine similarity, shuffled their order, and emailed them to the corresponding author for annotation. We believe that ranking five articles (rather than the entire collection of the references) is a more manageable task for an author compared to asking them to rank all references. Because the documents to be annotated were selected based on text similarity, they also represent a challenging baseline for models based on text-similarity features. In total 416 authors were contacted, and 92 responded (22% response rate). Two responses were removed from the dataset for incomplete annotation.
We asked authors to rank documents by how “close to your work” they were. The definition of closeness was left to the discretion of the author. The dataset is composed of 90 annotated documents with 5 citations each ranked 1 to 5, where 1 is least relevant and 5 is most relevant for a total of 450 annotated citations.
## Learning to Rank
Learning-to-rank is a technique for reordering the results returned from a search engine query. Generally, the initial query to a search engine is concerned more with recall than precision: the goal is to obtain a subset of potentially related documents from the corpus. Then, given this set of potentially related documents, learning-to-rank algorithms reorder the documents such that the most relevant documents appear at the top of the list. This process is illustrated in Figure FIGREF6 .
There are three basic types of learning-to-rank algorithms: point-wise, pair-wise, and list-wise. Point-wise algorithms assign a score to each retrieved document and rank them by their scores. Pair-wise algorithms turn learning-to-rank into a binary classification problem, obtaining a ranking by comparing each individual pair of documents. List-wise algorithms try to optimize an evaluation parameter over all queries in the dataset.
Support Vector Machine (SVM) ( BIBREF25 ) is a commonly used supervised classification algorithm that has shown good performance over a range of tasks. SVM can be thought of as a binary linear classifier where the goal is to maximize the size of the gap between the class-separating line and the points on either side of the line. This helps avoid over-fitting on the training data. SVMRank is a modification to SVM that assigns scores to each data point and allows the results to be ranked ( BIBREF26 ). We use SVMRank in the experiments below. SVMRank has previously been used in the task of document retrieval in ( BIBREF27 ) for a more traditional short query task and has been shown to be a top-performing system for ranking.
SVMRank is a point-wise learning-to-rank algorithm that returns scores for each document. We rank the documents by these scores. It is possible that sometimes two documents will have the same score, resulting in a tie. In this case, we give both documents the same rank, and then leave a gap in the ranking. For example, if documents 2 and 3 are tied, their ranked list will be [5, 3, 3, 2, 1].
Models are trained by randomly splitting the dataset into 70% training data and 30% test data. We apply a random sub-sampling approach where the dataset is randomly split, trained, and tested 100 times due to the relatively small size of the data. A model is learned for each split and a ranking is produced for each annotated document.
We test three different supervised models. The first supervised model uses only text similarity features, the second model uses all of the features, and the third model runs forward feature selection to select the best performing combination of features. We also test using two different models trained on two different datasets: one trained using the gold standard annotations, and another trained using the judgments based on text similarity that were used to select the citations to give to the authors.
We tested several different learning to rank algorithms for this work. We found in preliminary testing that SVMRank had the best performance, so it will be used in the following experiments.
## Features
Each citation is turned into a feature vector representing the relationship between the published article and the citation. Four types of features are used: text similarity, citation count and location, age of the citation, and the number of times the citation has appeared in the literature (citation impact). Text similarity features measure the similarity of the words used in different parts of the document. In this work, we calculate the similarity between a document INLINEFORM0 and a document it cites INLINEFORM1 by transforming the their text into term vectors. For example, to calculate the similarity of the abstracts between INLINEFORM2 and INLINEFORM3 we transform the abstracts into two term vectors, INLINEFORM4 and INLINEFORM5 . The length of each of the term vectors is INLINEFORM6 . We then weight each word by its Term-frequency * Inverse-document frequency (TF*IDF) weight. TF*IDF is a technique to give higher weight to words that appear frequently in a document but infrequently in the corpus. Term frequency is simply the number of times that a word INLINEFORM7 appears in a document. Inverse-document frequency is the logarithmically-scaled fraction of documents in the corpus in which the word INLINEFORM8 appears. Or, more specifically: INLINEFORM9
where INLINEFORM0 is the total number of documents in the corpus, and the denominator is the number of documents in which a term INLINEFORM1 appears in the corpus INLINEFORM2 . Then, TF*IDF is defined as: INLINEFORM3
where INLINEFORM0 is a term, INLINEFORM1 is the document, and INLINEFORM2 is the corpus. For example, the word “the” may appear often in a document, but because it also appears in almost every document in the corpus it is not useful for calculating similarity, thus it receives a very low weight. However, a word such as “neurogenesis” may appear often in a document, but does not appear frequently in the corpus, and so it receives a high weight. The similarity between term vectors is then calculated using cosine similarity: INLINEFORM3
where INLINEFORM0 and INLINEFORM1 are two term vectors. The cosine similarity is a measure of the angle between the two vectors. The smaller the angle between the two vectors, i.e., the more similar they are, then the closer the value is to 1. Conversely, the more dissimilar the vectors, the closer the cosine similarity is to 0.
We calculate the text similarity between several different sections of the document INLINEFORM0 and the document it cites INLINEFORM1 . From the citing article INLINEFORM2 , we use the title, full text, abstract, the combined discussion/conclusion sections, and the 10 words on either side of the place in the document where the actual citation occurs. From the document it cites INLINEFORM3 we only use the title and the abstract due to limited availability of the full text. In this work we combine the discussion and conclusion sections of each document because some documents have only a conclusion section, others have only a discussion, and some have both. The similarity between each of these sections from the two documents is calculated and used as features in the model.
The age of the citation may be relevant to its importance. As a citation ages, we hypothesize that it is more likely to become a “foundational” citation rather than one that directly influenced the development of the article. Therefore more recent citations may be more likely relevant to the article. Similarly, “citation impact”, that is, the number of times a citation has appeared in the literature (as measured by Google Scholar) may be an indicator of whether or not an article is foundational rather than directly related. We hypothesize that the fewer times an article is cited in the literature, the more impact it had on the article at hand.
We also keep track of the number of times a citation is mentioned in both the full text and discussion/conclusion sections. We hypothesize that if a citation is mentioned multiple times, it is more important than citations that are mentioned only once. Further, citations that appear in the discussion/conclusion sections are more likely to be crucial to understanding the results. We normalize the counts of the citations by the total number of citations in that section. In total we select 15 features, shown in Table TABREF15 . The features are normalized within each document so that each of citation features is on a scale from 0 to 1, and are evenly distributed within that range. This is done because some of the features (such as years since citation) are unbounded.
## Baseline Systems
We compare our system to a variety of baselines. (1) Rank by the number of times a citation is mentioned in the document. (2) Rank by the number of times the citation is cited in the literature (citation impact). (3) Rank using Google Scholar Related Articles. (4) Rank by the TF*IDF weighted cosine similarity. (5) Rank using a learning-to-rank model trained on text similarity rankings. The first two baseline systems are models where the values are ordered from highest to lowest to generate the ranking. The idea behind them is that the number of times a citation is mentioned in an article, or the citation impact may already be good indicators of their closeness. The text similarity model is trained using the same features and methods used by the annotation model, but trained using text similarity rankings instead of the author's judgments.
We also compare our rankings to those found on the popular scientific article search engine Google Scholar. Google Scholar is a “black box” IR system: they do not release details about which features they are using and how they judge relevance of documents. Google Scholar provides a “Related Articles” feature for each document in its index that shows the top 100 related documents for each article. To compare our rankings, we search through these related documents and record the ranking at which each of the citations we selected appeared. We scale these rankings such that the lowest ranked article from Google Scholar has the highest relevance ranking in our set. If the cited document does not appear in the set, we set its relevance-ranking equal to one below the lowest relevance ranking found.
Four comparisons are performed with the Google Scholar data. (1) We first train a model using our gold standard and see if we can predict Google Scholar's ranking. (2) We compare to a baseline of using Google Scholar's rankings to train and compare with their own rankings using our feature set. (3) Then we train a model using Google Scholar's rankings and try to predict our gold standard. (4) We compare it to the model trained on our gold standard to predict our gold standard.
## Evaluation Measures
Normalized Discounted Cumulative Gain (NDCG) is a common measure for comparing a list of estimated document relevance judgments with a list of known judgments ( BIBREF28 ). To calculate NDCG we first calculate a ranking's Discounted Cumulative Gain (DCG) as: DISPLAYFORM0
where rel INLINEFORM0 is the relevance judgment at position INLINEFORM1 . Intuitively, DCG penalizes retrieval of documents that are not relevant (rel INLINEFORM2 ). However, DCG is an unbounded value. In order to compare the DCG between two models, we must normalize it. To do this, we use the ideal DCG (IDCG), i.e., the maximum possible DCG given the relevance judgments. The maximum possible DCG occurs when the relevance judgments are in the correct order. DISPLAYFORM0
The NDCG value is in the range of 0 to 1, where 0 means that no relevant documents were retrieved, and 1 means that the relevant documents were retrieved and in the correct order of their relevance judgments.
Kendall's INLINEFORM0 is a measure of the correlation between two ranked lists. It compares the number of concordant pairs with the number of discordant pairs between each list. A concordant pair is defined over two observations INLINEFORM1 and INLINEFORM2 . If INLINEFORM3 and INLINEFORM4 , then the pair at indices INLINEFORM5 is concordant, that is, the ranking at INLINEFORM6 in both ranking sets INLINEFORM7 and INLINEFORM8 agree with each other. Similarly, a pair INLINEFORM9 is discordant if INLINEFORM10 and INLINEFORM11 or INLINEFORM12 and INLINEFORM13 . Kendall's INLINEFORM14 is then defined as: DISPLAYFORM0
where C is the number of concordant pairs, D is the number of discordant pairs, and the denominator represents the total number of possible pairs. Thus, Kendall's INLINEFORM0 falls in the range of INLINEFORM1 , where -1 means that the ranked lists are perfectly negatively correlated, 0 means that they are not significantly correlated, and 1 means that the ranked lists are perfectly correlated. One downside of this measure is that it does not take into account where in the ranked list an error occurs. Information retrieval, in general, cares more about errors near the top of the list rather than errors near the bottom of the list.
Average-Precision INLINEFORM0 ( BIBREF29 ) (or INLINEFORM1 ) extends on Kendall's INLINEFORM2 by incorporating the position of errors. If an error occurs near the top of the list, then that is penalized heavier than an error occurring at the bottom of the list. To achieve this, INLINEFORM3 incorporates ideas from the popular Average Precision measure, were we calculate the precision at each index of the list and then average them together. INLINEFORM4 is defined as: DISPLAYFORM0
Intuitively, if an error occurs at the top of the list, then that error is propagated into each iteration of the summation, meaning that it's penalty is added multiple times. INLINEFORM0 's range is between -1 and 1, where -1 means the lists are perfectly negatively correlated, 0 means that they are not significantly correlated, and 1 means that they are perfectly correlated.
## Forward Feature Selection
Forward feature selection was performed by iteratively testing each feature one at a time. The highest performing feature is kept in the model, and another sweep is done over the remaining features. This continues until all features have been selected. This approach allows us to explore the effect of combinations of features and the effect of having too many or too few features. It also allows us to evaluate which features and combinations of features are the most powerful.
## Results
We first compare our gold standard to the baselines. A random baseline is provided for reference. Because all of the documents that we rank are relevant, NDCG will be fairly high simply by chance. We find that the number of times a document is mentioned in the annotated document is significantly better than the random baseline or the citation impact. The more times a document is mentioned in a paper, the more likely the author was to annotate it as important. Interestingly, we see a negative correlation with the citation impact. The more times a document is mentioned in the literature, the less likely it is to be important. These results are shown in Table TABREF14 .
Next we rank the raw values of the features and compare them to our gold standard to obtain a baseline (Table TABREF15 ). The best performing text similarity feature is the similarity between the abstract of the annotated document and the abstract of the cited document. However, the number of times that a cited document is mentioned in the text of the annotated document are also high-scoring features, especially in the INLINEFORM0 correlation coefficient. These results indicate that text similarity alone may not be a good measure for judging the rank of a document.
Next we test three different feature sets for our supervised learning-to-rank models. The model using only the text similarity features performs poorly: NDCG stays at baseline and the correlation measures are low. Models that incorporate information about the age, number of times a cited document was referenced, and the citation impact of that document in addition to the text similarity features significantly outperformed models that used only text similarity features INLINEFORM0 . Because INLINEFORM1 takes into account the position in the ranking of the errors, this indicates that the All Features model was able to better correctly place highly ranked documents above lower ranked ones. Similarly, because Kendall's INLINEFORM2 is an overall measure of correlation that does not take into account the position of errors, the higher value here means that more rankings were correctly placed. Interestingly, feature selection (which is optimized for NDCG) does not outperform the model using all of the features in terms of our correlation measures. The features chosen during forward feature selection are (1) the citation impact, (2) number of mentions in the full text, (3) text similarity between the annotated document's title and the referenced document's abstract, (4) the text similarity between the annotated document's discussion/conclusion section and the referenced document's title. These results are shown in Table TABREF16 . The models trained on the text similarity judgments perform worse than the models trained on the annotated data. However, in terms of both NDCG and the correlation measures, they perform significantly better than the random baseline.
Next we compare our model to Google Scholar's rankings. Using the ranking collected from Google Scholar, we build a training set to try to predict our authors' rankings. We find that Google Scholar performs similarly to the text-only features model. This indicates that the rankings we obtained from the authors are substantially different than the rankings that Google Scholar provides. Results appear in Table TABREF17 .
## Discussion
We found that authors rank the references they cite substantially differently from rankings based on text-similarity. Our results show that decomposing a document into a set of features that is able to capture that difference is key. While text similarity is indeed important (as evidenced by the Similarity(a,a) feature in Table TABREF15 ), we also found that the number of times a document is referenced in the text and the number of times a document is referenced in the literature are also both important features (via feature selection). The more often a citation is mentioned in the text, the more likely it is to be important. This feature is often overlooked in article citation recommendation. We also found that recency is important: the age of the citation is negatively correlated with the rank. Newer citations are more likely to be directly important than older, more foundational citations. Additionally, the number of times a document is cited in the literature is negatively correlated with rank. This is likely due to highly cited documents being more foundational works; they may be older papers that are important to the field but not directly influential to the new work.
The model trained using the author's judgments does significantly better than the model trained using the text-similarity-based judgments. An error analysis was performed to find out why some of the rankings disagreed with the author's annotations. We found that in some cases our features were unable to capture the relationship: for example a biomedical document applying a model developed in another field to the dataset may use very different language to describe the model than the citation. Previous work adopting topic models to query document search may prove useful for such cases.
A small subset of features ended up performing as well as the full list of features. The number of times a citation was mentioned and the citation impact score in the literature ended up being two of the most important features. Indeed, without the citation-based features, the model performs as though it were trained with the text-similarity rankings. Feature engineering is a part of any learning-to-rank system, especially in domain-specific contexts. Citations are an integral feature of our dataset. For learning-to-rank to be applied to other datasets feature engineering must also occur to exploit the unique properties of those datasets. However, we show that combining the domain-specific features with more traditional text-based features does improve the model's scores over simply using the domain-specific features themselves.
Interestingly, citation impact and age of the citation are both negatively correlated with rank. We hypothesize that this is because both measures can be indicators of recency: a new publication is more likely to be directly influenced by more recent work. Many other related search tools, however, treat the citation impact as a positive feature of relatedness: documents with a higher citation impact appear higher on the list of related articles than those with lower citation impacts. This may be the opposite of what the user actually desires.
We also found that rankings from our text-similarity based IR system or Google Scholar's IR system were unable to rank documents by the authors' annotations as well as our system. In one sense, this is reasonable: the rankings coming from these systems were from a different system than the author annotations. However, in domain-specific IR, domain experts are the best judges. We built a system that exploits these expert judgments. The text similarity and Google Scholar models were able to do this to some extent, performing above the random baseline, but not on the level of our model.
Additionally, we observe that NDCG may not be the most appropriate measure for comparing short ranked lists where all of the documents are relevant to some degree. NDCG gives a lot of credit to relevant documents that occur in the highest ranks. However, all of the documents here are relevant, just to varying degrees. Thus, NDCG does not seem to be the most appropriate measure, as is evident in our scores. The correlation coefficients from Kendall's INLINEFORM0 and INLINEFORM1 seem to be far more appropriate for this case, as they are not concerned with relevance, only ranking.
One limitation of our work is that we selected a small set of references based on their similarities to the article that cites them. Ideally, we would have had authors rank all of their citations for us, but this would have been a daunting task for authors to perform. We chose to use the Google Scholar dataset in order to attempt to mitigate this: we obtain a ranking for the set of references from a system that is also ranking many other documents. The five citations selected by TF*IDF weighted cosine similarity represent a “hard” gold standard: we are attempting to rank documents that are known to all be relevant by their nature, and have high similarity with the text. Additionally, there are plethora of other, more expensive features we could explore to improve the model. Citation network features, phrasal concepts, and topic models could all be used to help improve our results, at the cost of computational complexity.
We have developed a model for fast related-document ranking based on crowd-sourced data. The model, data, and data collection software are all publicly available and can easily be used in future applications as an automatic search to help users find the most important citations given a particular document. The experimental setup is portable to other datasets with some feature engineering. We were able to identify that several domain-specific features were crucial to our model, and that we were able to improve on the results of simply using those features alone by adding more traditional features.
Query-by-document is a complicated and challenging task. We provide an approach with an easily obtained dataset and a computationally inexpensive model. By working with biomedical researchers we were able to build a system that ranks documents in a quantitatively different way than previous systems, and to provide a tool that helps researchers find related documents.
## Acknowledgments
We would like to thank all of the authors who took the time to answer our citation ranking survey. This work is supported by National Institutes of Health with the grant number 1R01GM095476. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
| 11
|
1611.02550
|
Discriminative Acoustic Word Embeddings: Recurrent Neural Network-Based Approaches
|
# Discriminative Acoustic Word Embeddings: Recurrent Neural Network-Based Approaches
## Abstract
Acoustic word embeddings --- fixed-dimensional vector representations of variable-length spoken word segments --- have begun to be considered for tasks such as speech recognition and query-by-example search. Such embeddings can be learned discriminatively so that they are similar for speech segments corresponding to the same word, while being dissimilar for segments corresponding to different words. Recent work has found that acoustic word embeddings can outperform dynamic time warping on query-by-example search and related word discrimination tasks. However, the space of embedding models and training approaches is still relatively unexplored. In this paper we present new discriminative embedding models based on recurrent neural networks (RNNs). We consider training losses that have been successful in prior work, in particular a cross entropy loss for word classification and a contrastive loss that explicitly aims to separate same-word and different-word pairs in a"Siamese network"training setting. We find that both classifier-based and Siamese RNN embeddings improve over previously reported results on a word discrimination task, with Siamese RNNs outperforming classification models. In addition, we present analyses of the learned embeddings and the effects of variables such as dimensionality and network structure.
## Introduction
Many speech processing tasks – such as automatic speech recognition or spoken term detection – hinge on associating segments of speech signals with word labels. In most systems developed for such tasks, words are broken down into sub-word units such as phones, and models are built for the individual units. An alternative, which has been considered by some researchers, is to consider each entire word segment as a single unit, without assigning parts of it to sub-word units. One motivation for the use of whole-word approaches is that they avoid the need for sub-word models. This is helpful since, despite decades of work on sub-word modeling BIBREF0 , BIBREF1 , it still poses significant challenges. For example, speech processing systems are still hampered by differences in conversational pronunciations BIBREF2 . A second motivation is that considering whole words at once allows us to consider a more flexible set of features and reason over longer time spans.
Whole-word approaches typically involve, at some level, template matching. For example, in template-based speech recognition BIBREF3 , BIBREF4 , word scores are computed from dynamic time warping (DTW) distances between an observed segment and training segments of the hypothesized word. In query-by-example search, putative matches are typically found by measuring the DTW distance between the query and segments of the search database BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 . In other words, whole-word approaches often boil down to making decisions about whether two segments are examples of the same word or not.
An alternative to DTW that has begun to be explored is the use of acoustic word embeddings (AWEs), or vector representations of spoken word segments. AWEs are representations that can be learned from data, ideally such that the embeddings of two segments corresponding to the same word are close, while embeddings of segments corresponding to different words are far apart. Once word segments are represented via fixed-dimensional embeddings, computing distances is as simple as measuring a cosine or Euclidean distance between two vectors.
There has been some, thus far limited, work on acoustic word embeddings, focused on a number of embedding models, training approaches, and tasks BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 . In this paper we explore new embedding models based on recurrent neural networks (RNNs), applied to a word discrimination task related to query-by-example search. RNNs are a natural model class for acoustic word embeddings, since they can handle arbitrary-length sequences. We compare several types of RNN-based embeddings and analyze their properties. Compared to prior embeddings tested on the same task, our best models achieve sizable improvements in average precision.
## Related work
We next briefly describe the most closely related prior work.
Maas et al. BIBREF9 and Bengio and Heigold BIBREF10 used acoustic word embeddings, based on convolutional neural networks (CNNs), to generate scores for word segments in automatic speech recognition. Maas et al. trained CNNs to predict (continuous-valued) embeddings of the word labels, and used the resulting embeddings to define feature functions in a segmental conditional random field BIBREF17 rescoring system. Bengio and Heigold also developed CNN-based embeddings for lattice rescoring, but with a contrastive loss to separate embeddings of a given word from embeddings of other words.
Levin et al. BIBREF11 developed unsupervised embeddings based on representing each word as a vector of DTW distances to a collection of reference word segments. This representation was subsequently used in several applications: a segmental approach for query-by-example search BIBREF12 , lexical clustering BIBREF18 , and unsupervised speech recognition BIBREF19 . Voinea et al. BIBREF15 developed a representation also based on templates, in their case phone templates, designed to be invariant to specific transformations, and showed their robustness on digit classification.
Kamper et al. BIBREF13 compared several types of acoustic word embeddings for a word discrimination task related to query-by-example search, finding that embeddings based on convolutional neural networks (CNNs) trained with a contrastive loss outperformed the reference vector approach of Levin et al. BIBREF11 as well as several other CNN and DNN embeddings and DTW using several feature types. There have now been a number of approaches compared on this same task and data BIBREF11 , BIBREF20 , BIBREF21 , BIBREF22 . For a direct comparison with this prior work, in this paper we use the same task and some of the same training losses as Kamper et al., but develop new embedding models based on RNNs.
The only prior work of which we are aware using RNNs for acoustic word embeddings is that of Chen et al. BIBREF16 and Chung et al. BIBREF14 . Chen et al. learned a long short-term memory (LSTM) RNN for word classification and used the resulting hidden state vectors as a word embedding in a query-by-example task. The setting was quite specific, however, with a small number of queries and speaker-dependent training. Chung et al. BIBREF14 worked in an unsupervised setting and trained single-layer RNN autoencoders to produce embeddings for a word discrimination task. In this paper we focus on the supervised setting, and compare a variety of RNN-based structures trained with different losses.
## Approach
An acoustic word embedding is a function that takes as input a speech segment corresponding to a word, INLINEFORM0 , where each INLINEFORM1 is a vector of frame-level acoustic features, and outputs a fixed-dimensional vector representing the segment, INLINEFORM2 . The basic embedding model structure we use is shown in Fig. FIGREF1 . The model consists of a deep RNN with some number INLINEFORM3 of stacked layers, whose final hidden state vector is passed as input to a set of INLINEFORM4 of fully connected layers; the output of the final fully connected layer is the embedding INLINEFORM5 .
The RNN hidden state at each time frame can be viewed as a representation of the input seen thus far, and its value in the last time frame INLINEFORM0 could itself serve as the final word embedding. The fully connected layers are added to account for the fact that some additional transformation may improve the representation. For example, the hidden state may need to be larger than the desired word embedding dimension, in order to be able to "remember" all of the needed intermediate information. Some of that information may not be needed in the final embedding. In addition, the information maintained in the hidden state may not necessarily be discriminative; some additional linear or non-linear transformation may help to learn a discriminative embedding.
Within this class of embedding models, we focus on Long Short-Term Memory (LSTM) networks BIBREF23 and Gated Recurrent Unit (GRU) networks BIBREF24 . These are both types of RNNs that include a mechanism for selectively retaining or discarding information at each time frame when updating the hidden state, in order to better utilize long-term context. Both of these RNN variants have been used successfully in speech recognition BIBREF25 , BIBREF26 , BIBREF27 , BIBREF28 .
In an LSTM RNN, at each time frame both the hidden state INLINEFORM0 and an associated “cell memory" vector INLINEFORM1 , are updated and passed on to the next time frame. In other words, each forward edge in Figure FIGREF1 can be viewed as carrying both the cell memory and hidden state vectors. The updates are modulated by the values of several gating vectors, which control the degree to which the cell memory and hidden state are updated in light of new information in the current frame. For a single-layer LSTM network, the updates are as follows:
INLINEFORM0
where INLINEFORM0 , and INLINEFORM1 are all vectors of the same dimensionality, INLINEFORM2 , and INLINEFORM3 are learned weight matrices of the appropriate sizes, INLINEFORM4 and INLINEFORM5 are learned bias vectors, INLINEFORM6 is a componentwise logistic activation, and INLINEFORM7 refers to the Hadamard (componentwise) product.
Similarly, in a GRU network, at each time step a GRU cell determines what components of old information are retained, overwritten, or modified in light of the next step in the input sequence. The output from a GRU cell is only the hidden state vector. A GRU cell uses a reset gate INLINEFORM0 and an update gate INLINEFORM1 as described below for a single-layer network: INLINEFORM2
where INLINEFORM0 , and INLINEFORM1 are all the same dimensionality, INLINEFORM2 , and INLINEFORM3 are learned weight matrices of the appropriate size, and INLINEFORM4 , INLINEFORM5 and INLINEFORM6 are learned bias vectors.
All of the above equations refer to single-layer networks. In a deep network, with multiple stacked layers, the same update equations are used in each layer, with the state, cell, and gate vectors replaced by layer-specific vectors INLINEFORM0 and so on for layer INLINEFORM1 . For all but the first layer, the input INLINEFORM2 is replaced by the hidden state vector from the previous layer INLINEFORM3 .
For the fully connected layers, we use rectified linear unit (ReLU) BIBREF29 activation, except for the final layer which depends on the form of supervision and loss used in training.
## Training
We train the RNN-based embedding models using a set of pre-segmented spoken words. We use two main training approaches, inspired by prior work but with some differences in the details. As in BIBREF13 , BIBREF10 , our first approach is to use the word labels of the training segments and train the networks to classify the word. In this case, the final layer of INLINEFORM0 is a log-softmax layer. Here we are limited to the subset of the training set that has a sufficient number of segments per word to train a good classifier, and the output dimensionality is equal to the number of words (but see BIBREF13 for a study of varying the dimensionality in such a classifier-based embedding model by introducing a bottleneck layer). This model is trained end-to-end and is optimized with a cross entropy loss. Although labeled data is necessarily limited, the hope is that the learned models will be useful even when applied to spoken examples of words not previously seen in the training data. For words not seen in training, the embeddings should correspond to some measure of similarity of the word to the training words, measured via the posterior probabilities of the previously seen words. In the experiments below, we examine this assumption by analyzing performance on words that appear in the training data compared to those that do not.
The second training approach, based on earlier work of Kamper et al. BIBREF13 , is to train "Siamese" networks BIBREF30 . In this approach, full supervision is not needed; rather, we use weak supervision in the form of pairs of segments labeled as same or different. The base model remains the same as before—an RNN followed by a set of fully connected layers—but the final layer is no longer a softmax but rather a linear activation layer of arbitrary size. In order to learn the parameters, we simultaneously feed three word segments through three copies of our model (i.e. three networks with shared weights). One input segment is an “anchor", INLINEFORM0 , the second is another segment with the same word label, INLINEFORM1 , and the third is a segment corresponding to a different word label, INLINEFORM2 . Then, the network is trained using a “cos-hinge" loss:
DISPLAYFORM0
where INLINEFORM0 is the cosine distance between INLINEFORM1 . Unlike cross entropy training, here we directly aim to optimize relative (cosine) distance between same and different word pairs. For tasks such as query-by-example search, this training loss better respects our end objective, and can use more data since neither fully labeled data nor any minimum number of examples of each word should be needed.
## EXPERIMENTS
Our end goal is to improve performance on downstream tasks requiring accurate word discrimination. In this paper we use an intermediate task that more directly tests whether same- and different-word pairs have the expected relationship. and that allows us to compare to a variety of prior work. Specifically, we use the word discrimination task of Carlin et al. BIBREF20 , which is similar to a query-by-example task where the word segmentations are known. The evaluation consists of determining, for each pair of evaluation segments, whether they are examples of the same or different words, and measuring performance via the average precision (AP). We do this by measuring the cosine similarity between their acoustic word embeddings and declaring them to be the same if the distance is below a threshold. By sweeping the threshold, we obtain a precision-recall curve from which we compute the AP.
The data used for this task is drawn from the Switchboard conversational English corpus BIBREF31 . The word segments range from 50 to 200 frames in length. The acoustic features in each frame (the input to the word embedding models INLINEFORM0 ) are 39-dimensional MFCCs+ INLINEFORM1 + INLINEFORM2 . We use the same train, development, and test partitions as in prior work BIBREF13 , BIBREF11 , and the same acoustic features as in BIBREF13 , for as direct a comparison as possible. The train set contains approximately 10k example segments, while dev and test each contain approximately 11k segments (corresponding to about 60M pairs for computing the dev/test AP). As in BIBREF13 , when training the classification-based embeddings, we use a subset of the training set containing all word types with a minimum of 3 occurrences, reducing the training set size to approximately 9k segments.
When training the Siamese networks, the training data consists of all of the same-word pairs in the full training set (approximately 100k pairs). For each such training pair, we randomly sample a third example belonging to a different word type, as required for the INLINEFORM0 loss.
## Classification network details
Our classifier-based embeddings use LSTM or GRU networks with 2–4 stacked layers and 1–3 fully connected layers. The final embedding dimensionality is equal to the number of unique word labels in the training set, which is 1061. The recurrent hidden state dimensionality is fixed at 512 and dropout BIBREF32 between stacked recurrent layers is used with probability INLINEFORM0 . The fully connected hidden layer dimensionality is fixed at 1024. Rectified linear unit (ReLU) non-linearities and dropout with INLINEFORM1 are used between fully-connected layers. However, between the final recurrent hidden state output and the first fully-connected layer no non-linearity or dropout is applied. These settings were determined through experiments on the development set.
The classifier network is trained with a cross entropy loss and optimized using stochastic gradient descent (SGD) with Nesterov momentum BIBREF33 . The learning rate is initialized at 0.1 and is reduced by a factor of 10 according to the following heuristic: If 99% of the current epoch's average batch loss is greater than the running average of batch losses over the last 3 epochs, this is considered a plateau; if there are 3 consecutive plateau epochs, then the learning rate is reduced. Training stops when reducing the learning rate no longer improves dev set AP. Then, the model from the epoch corresponding to the the best dev set AP is chosen. Several other optimizers—Adagrad BIBREF34 , Adadelta BIBREF35 , and Adam BIBREF36 —were explored in initial experiments on the dev set, but all reported results were obtained using SGD with Nesterov momentum.
## Siamese network details
For experiments with Siamese networks, we initialize (warm-start) the networks with the tuned classification network, removing the final log-softmax layer and replacing it with a linear layer of size equal to the desired embedding dimensionality. We explored embeddings with dimensionalities between 8 and 2048. We use a margin of 0.4 in the cos-hinge loss.
In training the Siamese networks, each training mini-batch consists of INLINEFORM0 triplets. INLINEFORM1 triplets are of the form INLINEFORM2 where INLINEFORM3 and INLINEFORM4 are examples of the same class (a pair from the 100k same-word pair set) and INLINEFORM5 is a randomly sampled example from a different class. Then, for each of these INLINEFORM6 triplets INLINEFORM7 , an additional triplet INLINEFORM8 is added to the mini-batch to allow all segments to serve as anchors. This is a slight departure from earlier work BIBREF13 , which we found to improve stability in training and performance on the development set.
In preliminary experiments, we compared two methods for choosing the negative examples INLINEFORM0 during training, a uniform sampling approach and a non-uniform one. In the case of uniform sampling, we sample INLINEFORM1 uniformly at random from the full set of training examples with labels different from INLINEFORM2 . This sampling method requires only word-pair supervision. In the case of non-uniform sampling, INLINEFORM3 is sampled in two steps. First, we construct a distribution INLINEFORM4 over word labels INLINEFORM5 and sample a different label from it. Second, we sample an example uniformly from within the subset with the chosen label. The goal of this method is to speed up training by targeting pairs that violate the margin constraint. To construct the multinomial PMF INLINEFORM6 , we maintain an INLINEFORM7 matrix INLINEFORM8 , where INLINEFORM9 is the number of unique word labels in training. Each word label corresponds to an integer INLINEFORM10 INLINEFORM11 [1, INLINEFORM12 ] and therefore a row in INLINEFORM13 . The values in a row of INLINEFORM14 are considered similarity scores, and we can retrieve the desired PMF for each row by normalizing by its sum.
At the start of each epoch, we initialize INLINEFORM0 with 0's along the diagonal and 1's elsewhere (which reduces to uniform sampling). For each training pair INLINEFORM1 , we update INLINEFORM2 for both INLINEFORM3 and INLINEFORM4 :
INLINEFORM0
The PMFs INLINEFORM0 are updated after the forward pass of an entire mini-batch. The constant INLINEFORM1 enforces a potentially stronger constraint than is used in the INLINEFORM2 loss, in order to promote diverse sampling. In all experiments, we set INLINEFORM3 . This is a heuristic approach, and it would be interesting to consider various alternatives. Preliminary experiments showed that the non-uniform sampling method outperformed uniform sampling, and in the following we report results with non-uniform sampling.
We optimize the Siamese network model using SGD with Nesterov momentum for 15 epochs. The learning rate is initialized to 0.001 and dropped every 3 epochs until no improvement is seen on the dev set. The final model is taken from the epoch with the highest dev set AP. All models were implemented in Torch BIBREF37 and used the rnn library of BIBREF38 .
## Results
Based on development set results, our final embedding models are LSTM networks with 3 stacked layers and 3 fully connected layers, with output dimensionality of 1024 in the case of Siamese networks. Final test set results are given in Table TABREF7 . We include a comparison with the best prior results on this task from BIBREF13 , as well as the result of using standard DTW on the input MFCCs (reproduced from BIBREF13 ) and the best prior result using DTW, obtained with frame features learned with correlated autoencoders BIBREF21 . Both classifier and Siamese LSTM embedding models outperform all prior results on this task of which we are aware.
We next analyze the effects of model design choices, as well as the learned embeddings themselves.
## Effect of model structure
Table TABREF10 shows the effect on development set performance of the number of stacked layers INLINEFORM0 , the number of fully connected layers INLINEFORM1 , and LSTM vs. GRU cells, for classifier-based embeddings. The best performance in this experiment is achieved by the LSTM network with INLINEFORM2 . However, performance still seems to be improving with additional layers, suggesting that we may be able to further improve performance by adding even more layers of either type. However, we fixed the model to INLINEFORM3 in order to allow for more experimentation and analysis within a reasonable time.
Table TABREF10 reveals an interesting trend. When only one fully connected layer is used, the GRU networks outperform the LSTMs given a sufficient number of stacked layers. On the other hand, once we add more fully connected layers, the LSTMs outperform the GRUs. In the first few lines of Table TABREF10 , we use 2, 3, and 4 layer stacks of LSTMs and GRUs while holding fixed the number of fully-connected layers at INLINEFORM0 . There is clear utility in stacking additional layers; however, even with 4 stacked layers the RNNs still underperform the CNN-based embeddings of BIBREF13 until we begin adding fully connected layers.
After exploring a variety of stacked RNNs, we fixed the stack to 3 layers and varied the number of fully connected layers. The value of each additional fully connected layer is clearly greater than that of adding stacked layers. All networks trained with 2 or 3 fully connected layers obtain more than 0.4 AP on the development set, while stacked RNNs with 1 fully connected layer are at around 0.3 AP or less. This may raise the question of whether some simple fully connected model may be all that is needed; however, previous work has shown that this approach is not competitive BIBREF13 , and convolutional or recurrent layers are needed to summarize arbitrary-length segments into a fixed-dimensional representation.
## Effect of embedding dimensionality
For the Siamese networks, we varied the output embedding dimensionality, as shown in Fig. FIGREF11 . This analysis shows that the embeddings learned by the Siamese RNN network are quite robust to reduced dimensionality, outperforming the classifier model for all dimensionalities 32 or higher and outperforming previously reported dev set performance with CNN-based embeddings BIBREF13 for all dimensionalities INLINEFORM0 .
## Effect of training vocabulary
We might expect the learned embeddings to be more accurate for words that are seen in training than for ones that are not. Fig. FIGREF11 measures this effect by showing performance as a function of the number of occurrences of the dev words in the training set. Indeed, both model types are much more successful for in-vocabulary words, and their performance improves the higher the training frequency of the words. However, performance increases more quickly for the Siamese network than for the classifier as training frequency increases. This may be due to the fact that, if a word type occurs at least INLINEFORM0 times in the classifier training set, then it occurs at least INLINEFORM1 times in the Siamese paired training data.
## Visualization of embeddings
In order to gain a better qualitative understanding of the differences between clasiffier and Siamese-based embeddings, and of the learned embedding space more generally, we plot a two-dimensional visualization of some of our learned embeddings via t-SNE BIBREF40 in Fig. FIGREF12 . For both classifier and Siamese embeddings, there is a marked difference in the quality of clusters formed by embeddings of words that were previously seen vs. previously unseen in training. However, the Siamese network embeddings appear to have better relative distances between word clusters with similar and dissimilar pronunciations. For example, the word programs appears equidistant from problems and problem in the classifier-based embedding space, but in the Siamese embedding space problems falls between problem and programs. Similarly, the cluster for democracy shifts with respect to actually and especially to better respect differences in pronunciation. More study of learned embeddings, using more data and word types, is needed to confirm such patterns in general. Improvements in unseen word embeddings from the classifier embedding space to the Siamese embedding space (such as for democracy, morning, and basketball) are a likely result of optimizing the model for relative distances between words.
## Conclusion
Our main finding is that RNN-based acoustic word embeddings outperform prior approaches, as measured via a word discrimination task related to query-by-example search. Our best results are obtained with deep LSTM RNNs with a combination of several stacked layers and several fully connected layers, optimized with a contrastive Siamese loss. Siamese networks have the benefit that, for any given training data set, they are effectively trained on a much larger set, in the sense that they measure a loss and gradient for every possible pair of data points. Our experiments suggest that the models could still be improved with additional layers. In addition, we have found that, for the purposes of acoustic word embeddings, fully connected layers are very important and have a more significant effect per layer than stacked layers, particularly when trained with the cross entropy loss function.
These experiments represent an initial exploration of sequential neural models for acoustic word embeddings. There are a number of directions for further work. For example, while our analyses suggest that Siamese networks are better than classifier-based models at embedding previously unseen words, our best embeddings are still much poorer for unseen words. Improvements in this direction may come from larger training sets, or may require new models that better model the shared structure between words. Other directions for future work include additional forms of supervision and training, as well as application to downstream tasks.
| 13
|
1611.03599
|
UTCNN: a Deep Learning Model of Stance Classificationon on Social Media Text
|
# UTCNN: a Deep Learning Model of Stance Classificationon on Social Media Text
## Abstract
Most neural network models for document classification on social media focus on text infor-mation to the neglect of other information on these platforms. In this paper, we classify post stance on social media channels and develop UTCNN, a neural network model that incorporates user tastes, topic tastes, and user comments on posts. UTCNN not only works on social media texts, but also analyzes texts in forums and message boards. Experiments performed on Chinese Facebook data and English online debate forum data show that UTCNN achieves a 0.755 macro-average f-score for supportive, neutral, and unsupportive stance classes on Facebook data, which is significantly better than models in which either user, topic, or comment information is withheld. This model design greatly mitigates the lack of data for the minor class without the use of oversampling. In addition, UTCNN yields a 0.842 accuracy on English online debate forum data, which also significantly outperforms results from previous work as well as other deep learning models, showing that UTCNN performs well regardless of language or platform.
## Introduction
This work is licenced under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/ Deep neural networks have been widely used in text classification and have achieved promising results BIBREF0 , BIBREF1 , BIBREF2 . Most focus on content information and use models such as convolutional neural networks (CNN) BIBREF3 or recursive neural networks BIBREF4 . However, for user-generated posts on social media like Facebook or Twitter, there is more information that should not be ignored. On social media platforms, a user can act either as the author of a post or as a reader who expresses his or her comments about the post.
In this paper, we classify posts taking into account post authorship, likes, topics, and comments. In particular, users and their “likes” hold strong potential for text mining. For example, given a set of posts that are related to a specific topic, a user's likes and dislikes provide clues for stance labeling. From a user point of view, users with positive attitudes toward the issue leave positive comments on the posts with praise or even just the post's content; from a post point of view, positive posts attract users who hold positive stances. We also investigate the influence of topics: different topics are associated with different stance labeling tendencies and word usage. For example we discuss women's rights and unwanted babies on the topic of abortion, but we criticize medicine usage or crime when on the topic of marijuana BIBREF5 . Even for posts on a specific topic like nuclear power, a variety of arguments are raised: green energy, radiation, air pollution, and so on. As for comments, we treat them as additional text information. The arguments in the comments and the commenters (the users who leave the comments) provide hints on the post's content and further facilitate stance classification.
In this paper, we propose the user-topic-comment neural network (UTCNN), a deep learning model that utilizes user, topic, and comment information. We attempt to learn user and topic representations which encode user interactions and topic influences to further enhance text classification, and we also incorporate comment information. We evaluate this model on a post stance classification task on forum-style social media platforms. The contributions of this paper are as follows: 1. We propose UTCNN, a neural network for text in modern social media channels as well as legacy social media, forums, and message boards — anywhere that reveals users, their tastes, as well as their replies to posts. 2. When classifying social media post stances, we leverage users, including authors and likers. User embeddings can be generated even for users who have never posted anything. 3. We incorporate a topic model to automatically assign topics to each post in a single topic dataset. 4. We show that overall, the proposed method achieves the highest performance in all instances, and that all of the information extracted, whether users, topics, or comments, still has its contributions.
## Extra-Linguistic Features for Stance Classification
In this paper we aim to use text as well as other features to see how they complement each other in a deep learning model. In the stance classification domain, previous work has showed that text features are limited, suggesting that adding extra-linguistic constraints could improve performance BIBREF6 , BIBREF7 , BIBREF8 . For example, Hasan and Ng as well as Thomas et al. require that posts written by the same author have the same stance BIBREF9 , BIBREF10 . The addition of this constraint yields accuracy improvements of 1–7% for some models and datasets. Hasan and Ng later added user-interaction constraints and ideology constraints BIBREF7 : the former models the relationship among posts in a sequence of replies and the latter models inter-topic relationships, e.g., users who oppose abortion could be conservative and thus are likely to oppose gay rights.
For work focusing on online forum text, since posts are linked through user replies, sequential labeling methods have been used to model relationships between posts. For example, Hasan and Ng use hidden Markov models (HMMs) to model dependent relationships to the preceding post BIBREF9 ; Burfoot et al. use iterative classification to repeatedly generate new estimates based on the current state of knowledge BIBREF11 ; Sridhar et al. use probabilistic soft logic (PSL) to model reply links via collaborative filtering BIBREF12 . In the Facebook dataset we study, we use comments instead of reply links. However, as the ultimate goal in this paper is predicting not comment stance but post stance, we treat comments as extra information for use in predicting post stance.
## Deep Learning on Extra-Linguistic Features
In recent years neural network models have been applied to document sentiment classification BIBREF13 , BIBREF4 , BIBREF14 , BIBREF15 , BIBREF2 . Text features can be used in deep networks to capture text semantics or sentiment. For example, Dong et al. use an adaptive layer in a recursive neural network for target-dependent Twitter sentiment analysis, where targets are topics such as windows 7 or taylor swift BIBREF16 , BIBREF17 ; recursive neural tensor networks (RNTNs) utilize sentence parse trees to capture sentence-level sentiment for movie reviews BIBREF4 ; Le and Mikolov predict sentiment by using paragraph vectors to model each paragraph as a continuous representation BIBREF18 . They show that performance can thus be improved by more delicate text models.
Others have suggested using extra-linguistic features to improve the deep learning model. The user-word composition vector model (UWCVM) BIBREF19 is inspired by the possibility that the strength of sentiment words is user-specific; to capture this they add user embeddings in their model. In UPNN, a later extension, they further add a product-word composition as product embeddings, arguing that products can also show different tendencies of being rated or reviewed BIBREF20 . Their addition of user information yielded 2–10% improvements in accuracy as compared to the above-mentioned RNTN and paragraph vector methods. We also seek to inject user information into the neural network model. In comparison to the research of Tang et al. on sentiment classification for product reviews, the difference is two-fold. First, we take into account multiple users (one author and potentially many likers) for one post, whereas only one user (the reviewer) is involved in a review. Second, we add comment information to provide more features for post stance classification. None of these two factors have been considered previously in a deep learning model for text stance classification. Therefore, we propose UTCNN, which generates and utilizes user embeddings for all users — even for those who have not authored any posts — and incorporates comments to further improve performance.
## Method
In this section, we first describe CNN-based document composition, which captures user- and topic-dependent document-level semantic representation from word representations. Then we show how to add comment information to construct the user-topic-comment neural network (UTCNN).
## User- and Topic-dependent Document Composition
As shown in Figure FIGREF4 , we use a general CNN BIBREF3 and two semantic transformations for document composition . We are given a document with an engaged user INLINEFORM0 , a topic INLINEFORM1 , and its composite INLINEFORM2 words, each word INLINEFORM3 of which is associated with a word embedding INLINEFORM4 where INLINEFORM5 is the vector dimension. For each word embedding INLINEFORM6 , we apply two dot operations as shown in Equation EQREF6 : DISPLAYFORM0
where INLINEFORM0 models the user reading preference for certain semantics, and INLINEFORM1 models the topic semantics; INLINEFORM2 and INLINEFORM3 are the dimensions of transformed user and topic embeddings respectively. We use INLINEFORM4 to model semantically what each user prefers to read and/or write, and use INLINEFORM5 to model the semantics of each topic. The dot operation of INLINEFORM6 and INLINEFORM7 transforms the global representation INLINEFORM8 to a user-dependent representation. Likewise, the dot operation of INLINEFORM9 and INLINEFORM10 transforms INLINEFORM11 to a topic-dependent representation.
After the two dot operations on INLINEFORM0 , we have user-dependent and topic-dependent word vectors INLINEFORM1 and INLINEFORM2 , which are concatenated to form a user- and topic-dependent word vector INLINEFORM3 . Then the transformed word embeddings INLINEFORM4 are used as the CNN input. Here we apply three convolutional layers on the concatenated transformed word embeddings INLINEFORM5 : DISPLAYFORM0
where INLINEFORM0 is the index of words; INLINEFORM1 is a non-linear activation function (we use INLINEFORM2 ); INLINEFORM5 is the convolutional filter with input length INLINEFORM6 and output length INLINEFORM7 , where INLINEFORM8 is the window size of the convolutional operation; and INLINEFORM9 and INLINEFORM10 are the output and bias of the convolution layer INLINEFORM11 , respectively. In our experiments, the three window sizes INLINEFORM12 in the three convolution layers are one, two, and three, encoding unigram, bigram, and trigram semantics accordingly.
After the convolutional layer, we add a maximum pooling layer among convolutional outputs to obtain the unigram, bigram, and trigram n-gram representations. This is succeeded by an average pooling layer for an element-wise average of the three maximized convolution outputs.
## UTCNN Model Description
Figure FIGREF10 illustrates the UTCNN model. As more than one user may interact with a given post, we first add a maximum pooling layer after the user matrix embedding layer and user vector embedding layer to form a moderator matrix embedding INLINEFORM0 and a moderator vector embedding INLINEFORM1 for moderator INLINEFORM2 respectively, where INLINEFORM3 is used for the semantic transformation in the document composition process, as mentioned in the previous section. The term moderator here is to denote the pseudo user who provides the overall semantic/sentiment of all the engaged users for one document. The embedding INLINEFORM4 models the moderator stance preference, that is, the pattern of the revealed user stance: whether a user is willing to show his preference, whether a user likes to show impartiality with neutral statements and reasonable arguments, or just wants to show strong support for one stance. Ideally, the latent user stance is modeled by INLINEFORM5 for each user. Likewise, for topic information, a maximum pooling layer is added after the topic matrix embedding layer and topic vector embedding layer to form a joint topic matrix embedding INLINEFORM6 and a joint topic vector embedding INLINEFORM7 for topic INLINEFORM8 respectively, where INLINEFORM9 models the semantic transformation of topic INLINEFORM10 as in users and INLINEFORM11 models the topic stance tendency. The latent topic stance is also modeled by INLINEFORM12 for each topic.
As for comments, we view them as short documents with authors only but without likers nor their own comments. Therefore we apply document composition on comments although here users are commenters (users who comment). It is noticed that the word embeddings INLINEFORM0 for the same word in the posts and comments are the same, but after being transformed to INLINEFORM1 in the document composition process shown in Figure FIGREF4 , they might become different because of their different engaged users. The output comment representation together with the commenter vector embedding INLINEFORM2 and topic vector embedding INLINEFORM3 are concatenated and a maximum pooling layer is added to select the most important feature for comments. Instead of requiring that the comment stance agree with the post, UTCNN simply extracts the most important features of the comment contents; they could be helpful, whether they show obvious agreement or disagreement. Therefore when combining comment information here, the maximum pooling layer is more appropriate than other pooling or merging layers. Indeed, we believe this is one reason for UTCNN's performance gains.
Finally, the pooled comment representation, together with user vector embedding INLINEFORM0 , topic vector embedding INLINEFORM1 , and document representation are fed to a fully connected network, and softmax is applied to yield the final stance label prediction for the post.
## Experiment
We start with the experimental dataset and then describe the training process as well as the implementation of the baselines. We also implement several variations to reveal the effects of features: authors, likers, comment, and commenters. In the results section we compare our model with related work.
## Dataset
We tested the proposed UTCNN on two different datasets: FBFans and CreateDebate. FBFans is a privately-owned, single-topic, Chinese, unbalanced, social media dataset, and CreateDebate is a public, multiple-topic, English, balanced, forum dataset. Results using these two datasets show the applicability and superiority for different topics, languages, data distributions, and platforms.
The FBFans dataset contains data from anti-nuclear-power Chinese Facebook fan groups from September 2013 to August 2014, including posts and their author and liker IDs. There are a total of 2,496 authors, 505,137 likers, 33,686 commenters, and 505,412 unique users. Two annotators were asked to take into account only the post content to label the stance of the posts in the whole dataset as supportive, neutral, or unsupportive (hereafter denoted as Sup, Neu, and Uns). Sup/Uns posts were those in support of or against anti-reconstruction; Neu posts were those evincing a neutral standpoint on the topic, or were irrelevant. Raw agreement between annotators is 0.91, indicating high agreement. Specifically, Cohen’s Kappa for Neu and not Neu labeling is 0.58 (moderate), and for Sup or Uns labeling is 0.84 (almost perfect). Posts with inconsistent labels were filtered out, and the development and testing sets were randomly selected from what was left. Posts in the development and testing sets involved at least one user who appeared in the training set. The number of posts for each stance is shown on the left-hand side of Table TABREF12 . About twenty percent of the posts were labeled with a stance, and the number of supportive (Sup) posts was much larger than that of the unsupportive (Uns) ones: this is thus highly skewed data, which complicates stance classification. On average, 161.1 users were involved in one post. The maximum was 23,297 and the minimum was one (the author). For comments, on average there were 3 comments per post. The maximum was 1,092 and the minimum was zero.
To test whether the assumption of this paper – posts attract users who hold the same stance to like them – is reliable, we examine the likes from authors of different stances. Posts in FBFans dataset are used for this analysis. We calculate the like statistics of each distinct author from these 32,595 posts. As the numbers of authors in the Sup, Neu and Uns stances are largely imbalanced, these numbers are normalized by the number of users of each stance. Table TABREF13 shows the results. Posts with stances (i.e., not neutral) attract users of the same stance. Neutral posts also attract both supportive and neutral users, like what we observe in supportive posts, but just the neutral posts can attract even more neutral likers. These results do suggest that users prefer posts of the same stance, or at least posts of no obvious stance which might cause annoyance when reading, and hence support the user modeling in our approach.
The CreateDebate dataset was collected from an English online debate forum discussing four topics: abortion (ABO), gay rights (GAY), Obama (OBA), and marijuana (MAR). The posts are annotated as for (F) and against (A). Replies to posts in this dataset are also labeled with stance and hence use the same data format as posts. The labeling results are shown in the right-hand side of Table TABREF12 . We observe that the dataset is more balanced than the FBFans dataset. In addition, there are 977 unique users in the dataset. To compare with Hasan and Ng's work, we conducted five-fold cross-validation and present the annotation results as the average number of all folds BIBREF9 , BIBREF5 .
The FBFans dataset has more integrated functions than the CreateDebate dataset; thus our model can utilize all linguistic and extra-linguistic features. For the CreateDebate dataset, on the other hand, the like and comment features are not available (as there is a stance label for each reply, replies are evaluated as posts as other previous work) but we still implemented our model using the content, author, and topic information.
## Settings
In the UTCNN training process, cross-entropy was used as the loss function and AdaGrad as the optimizer. For FBFans dataset, we learned the 50-dimensional word embeddings on the whole dataset using GloVe BIBREF21 to capture the word semantics; for CreateDebate dataset we used the publicly available English 50-dimensional word embeddings, pre-trained also using GloVe. These word embeddings were fixed in the training process. The learning rate was set to 0.03. All user and topic embeddings were randomly initialized in the range of [-0.1 0.1]. Matrix embeddings for users and topics were sized at 250 ( INLINEFORM0 ); vector embeddings for users and topics were set to length 10.
We applied the LDA topic model BIBREF22 on the FBFans dataset to determine the latent topics with which to build topic embeddings, as there is only one general known topic: nuclear power plants. We learned 100 latent topics and assigned the top three topics for each post. For the CreateDebate dataset, which itself constitutes four topics, the topic labels for posts were used directly without additionally applying LDA.
For the FBFans data we report class-based f-scores as well as the macro-average f-score ( INLINEFORM0 ) shown in equation EQREF19 . DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are the average precision and recall of the three class. We adopted the macro-average f-score as the evaluation metric for the overall performance because (1) the experimental dataset is severely imbalanced, which is common for contentious issues; and (2) for stance classification, content in minor-class posts is usually more important for further applications. For the CreateDebate dataset, accuracy was adopted as the evaluation metric to compare the results with related work BIBREF7 , BIBREF9 , BIBREF12 .
## Baselines
We pit our model against the following baselines: 1) SVM with unigram, bigram, and trigram features, which is a standard yet rather strong classifier for text features; 2) SVM with average word embedding, where a document is represented as a continuous representation by averaging the embeddings of the composite words; 3) SVM with average transformed word embeddings (the INLINEFORM0 in equation EQREF6 ), where a document is represented as a continuous representation by averaging the transformed embeddings of the composite words; 4) two mature deep learning models on text classification, CNN BIBREF3 and Recurrent Convolutional Neural Networks (RCNN) BIBREF0 , where the hyperparameters are based on their work; 5) the above SVM and deep learning models with comment information; 6) UTCNN without user information, representing a pure-text CNN model where we use the same user matrix and user embeddings INLINEFORM1 and INLINEFORM2 for each user; 7) UTCNN without the LDA model, representing how UTCNN works with a single-topic dataset; 8) UTCNN without comments, in which the model predicts the stance label given only user and topic information. All these models were trained on the training set, and parameters as well as the SVM kernel selections (linear or RBF) were fine-tuned on the development set. Also, we adopt oversampling on SVMs, CNN and RCNN because the FBFans dataset is highly imbalanced.
## Results on FBFans Dataset
In Table TABREF22 we show the results of UTCNN and the baselines on the FBFans dataset. Here Majority yields good performance on Neu since FBFans is highly biased to the neutral class. The SVM models perform well on Sup and Neu but perform poorly for Uns, showing that content information in itself is insufficient to predict stance labels, especially for the minor class. With the transformed word embedding feature, SVM can achieve comparable performance as SVM with n-gram feature. However, the much fewer feature dimension of the transformed word embedding makes SVM with word embeddings a more efficient choice for modeling the large scale social media dataset. For the CNN and RCNN models, they perform slightly better than most of the SVM models but still, the content information is insufficient to achieve a good performance on the Uns posts. As to adding comment information to these models, since the commenters do not always hold the same stance as the author, simply adding comments and post contents together merely adds noise to the model.
Among all UTCNN variations, we find that user information is most important, followed by topic and comment information. UTCNN without user information shows results similar to SVMs — it does well for Sup and Neu but detects no Uns. Its best f-scores on both Sup and Neu among all methods show that with enough training data, content-based models can perform well; at the same time, the lack of user information results in too few clues for minor-class posts to either predict their stance directly or link them to other users and posts for improved performance. The 17.5% improvement when adding user information suggests that user information is especially useful when the dataset is highly imbalanced. All models that consider user information predict the minority class successfully. UCTNN without topic information works well but achieves lower performance than the full UTCNN model. The 4.9% performance gain brought by LDA shows that although it is satisfactory for single topic datasets, adding that latent topics still benefits performance: even when we are discussing the same topic, we use different arguments and supporting evidence. Lastly, we get 4.8% improvement when adding comment information and it achieves comparable performance to UTCNN without topic information, which shows that comments also benefit performance. For platforms where user IDs are pixelated or otherwise hidden, adding comments to a text model still improves performance. In its integration of user, content, and comment information, the full UTCNN produces the highest f-scores on all Sup, Neu, and Uns stances among models that predict the Uns class, and the highest macro-average f-score overall. This shows its ability to balance a biased dataset and supports our claim that UTCNN successfully bridges content and user, topic, and comment information for stance classification on social media text. Another merit of UTCNN is that it does not require a balanced training data. This is supported by its outperforming other models though no oversampling technique is applied to the UTCNN related experiments as shown in this paper. Thus we can conclude that the user information provides strong clues and it is still rich even in the minority class.
We also investigate the semantic difference when a user acts as an author/liker or a commenter. We evaluated a variation in which all embeddings from the same user were forced to be identical (this is the UTCNN shared user embedding setting in Table TABREF22 ). This setting yielded only a 2.5% improvement over the model without comments, which is not statistically significant. However, when separating authors/likers and commenters embeddings (i.e., the UTCNN full model), we achieved much greater improvements (4.8%). We attribute this result to the tendency of users to use different wording for different roles (for instance author vs commenter). This is observed when the user, acting as an author, attempts to support her argument against nuclear power by using improvements in solar power; when acting as a commenter, though, she interacts with post contents by criticizing past politicians who supported nuclear power or by arguing that the proposed evacuation plan in case of a nuclear accident is ridiculous. Based on this finding, in the final UTCNN setting we train two user matrix embeddings for one user: one for the author/liker role and the other for the commenter role.
## Results on CreateDebate Dataset
Table TABREF24 shows the results of UTCNN, baselines as we implemented on the FBFans datset and related work on the CreateDebate dataset. We do not adopt oversampling on these models because the CreateDebate dataset is almost balanced. In previous work, integer linear programming (ILP) or linear-chain conditional random fields (CRFs) were proposed to integrate text features, author, ideology, and user-interaction constraints, where text features are unigram, bigram, and POS-dependencies; the author constraint tends to require that posts from the same author for the same topic hold the same stance; the ideology constraint aims to capture inferences between topics for the same author; the user-interaction constraint models relationships among posts via user interactions such as replies BIBREF7 , BIBREF9 .
The SVM with n-gram or average word embedding feature performs just similar to the majority. However, with the transformed word embedding, it achieves superior results. It shows that the learned user and topic embeddings really capture the user and topic semantics. This finding is not so obvious in the FBFans dataset and it might be due to the unfavorable data skewness for SVM. As for CNN and RCNN, they perform slightly better than most SVMs as we found in Table TABREF22 for FBFans.
Compared to the ILP BIBREF7 and CRF BIBREF9 methods, the UTCNN user embeddings encode author and user-interaction constraints, where the ideology constraint is modeled by the topic embeddings and text features are modeled by the CNN. The significant improvement achieved by UTCNN suggests the latent representations are more effective than overt model constraints.
The PSL model BIBREF12 jointly labels both author and post stance using probabilistic soft logic (PSL) BIBREF23 by considering text features and reply links between authors and posts as in Hasan and Ng's work. Table TABREF24 reports the result of their best AD setting, which represents the full joint stance/disagreement collective model on posts and is hence more relevant to UTCNN. In contrast to their model, the UTCNN user embeddings represent relationships between authors, but UTCNN models do not utilize link information between posts. Though the PSL model has the advantage of being able to jointly label the stances of authors and posts, its performance on posts is lower than the that for the ILP or CRF models. UTCNN significantly outperforms these models on posts and has the potential to predict user stances through the generated user embeddings.
For the CreateDebate dataset, we also evaluated performance when not using topic embeddings or user embeddings; as replies in this dataset are viewed as posts, the setting without comment embeddings is not available. Table TABREF24 shows the same findings as Table TABREF22 : the 21% improvement in accuracy demonstrates that user information is the most vital. This finding also supports the results in the related work: user constraints are useful and can yield 11.2% improvement in accuracy BIBREF7 . Further considering topic information yields 3.4% improvement, suggesting that knowing the subject of debates provides useful information. In sum, Table TABREF22 together with Table TABREF24 show that UTCNN achieves promising performance regardless of topic, language, data distribution, and platform.
## Conclusion
We have proposed UTCNN, a neural network model that incorporates user, topic, content and comment information for stance classification on social media texts. UTCNN learns user embeddings for all users with minimum active degree, i.e., one post or one like. Topic information obtained from the topic model or the pre-defined labels further improves the UTCNN model. In addition, comment information provides additional clues for stance classification. We have shown that UTCNN achieves promising and balanced results. In the future we plan to explore the effectiveness of the UTCNN user embeddings for author stance classification.
## Acknowledgements
Research of this paper was partially supported by Ministry of Science and Technology, Taiwan, under the contract MOST 104-2221-E-001-024-MY2.
| 14
|
1611.09441
|
Sentiment Analysis for Twitter : Going Beyond Tweet Text
|
# Sentiment Analysis for Twitter : Going Beyond Tweet Text
## Abstract
Analysing sentiment of tweets is important as it helps to determine the users' opinion. Knowing people's opinion is crucial for several purposes starting from gathering knowledge about customer base, e-governance, campaigning and many more. In this report, we aim to develop a system to detect the sentiment from tweets. We employ several linguistic features along with some other external sources of information to detect the sentiment of a tweet. We show that augmenting the 140 character-long tweet with information harvested from external urls shared in the tweet as well as Social Media features enhances the sentiment prediction accuracy significantly.
## Introduction
Analysing sentiment from text is a well-known NLP problem. Several state-of-the-art tools exist that can achieve this with reasonable accuracy. However most of the existing tools perform well on well-formatted text. In case of tweets, the user generated content is short, noisy, and in many cases ( INLINEFORM0 ) doesn't follow proper grammatical structure. Additionally, numerous internet slangs, abbreviations, urls, emoticons, and unconventional style of capitalization are found in the tweets. As a result, the accuracy of the state-of-the art NLP tools decreases sharply. In this project, we develop new features to incorporate the styles salient in short, informal user generated contents like tweets. We achieve an F1-accuracy of INLINEFORM1 for predicting the sentiment of tweets in our data-set. We also propose a method to discover new sentiment terms from the tweets.
In section SECREF2 we present analysis of the data-set. We describe the data-preprocessing that we have done in section SECREF3 . In section SECREF4 we describe how the feature-set was extracted, the classification framework, and also the tuning of the parameters for reasonable accuracy. In section SECREF5 we report the performance of our system. We also report how the different features affect the accuracy of the system. We describe how we harvest new sentiment terms using our framework in section SECREF6 . In this section we also present how we predict strength of sentiment from the tweets. We finally conclude with some possible future directions of work in section SECREF7 .
## Data-analysis
Tweets are short messages, restricted to 140 characters in length. Due to the nature of this microblogging service (quick and short messages), people use acronyms, make spelling mistakes, use emoticons and other characters that express special meanings. Following is a brief terminology associated with tweets:
Our dataset contains tweets about `ObamaCare' in USA collected during march 2010. It is divided into three subsets (train, dev, and test). Some tweets are manually annotated with one of the following classes.
positive, negative, neutral, unsure, and irrelevant
We ignore the tweets which are annotated unsure, or irrelevant. We present some preliminary statistics about the training data and test data in Table TABREF5 . We observe that there is an imbalance in the dataset. In training dataset, the ratio of positive tweets to negative ones is almost 1:2. In test set, it is heavily skewed with the ratio being less than 1:3. We handle this data imbalance problem using class prior parameters of the learning algorithm. We discuss this is detail in section SECREF38 .
## Data pre-processing
Since tweets are informal in nature, some pre-processing is required. Consider the following tweet.
“#Healthcare #Ins. Cigna denies #MD prescribed #tx 2 customers 20% of the time. - http://bit.ly/5PoQfo #HCR #Passit #ILDems #p2 PLS RT"
It is difficult to understand what is the content of the tweet unless it is normalized. We process all the tweets through the following stages.
## Normalization
Normalization is done as follows:
Removing patterns like 'RT', '@user_name', url.
Tokenizing tweet text using NLTK BIBREF0 word tokenizer.
Making use of the stopwords list by NLTK to remove them from the tweet text.
Rectifying informal/misspelled words using normalization dictionary BIBREF1 . For example, “foundation" for “foudation", “forgot" for “forgt".
Expanding abbreviations using slang dictionary. For example, “btw" is expanded to “by the way".
Removing emoticons. However we keep the number of positive and negative emoticons in each tweet as feature. We make use of the emoticon dictionary(Table TABREF14 ) presented in BIBREF2 .
## Hashtag Segmentation
We segment a hashtag into meaningful English phrases. The `#' character is removed from the tweet text. As for example, #killthebill is transformed into kill the bill.
In order to achieve this, we use a dictionary of English words. We recursively break the hashtagged phrase into segments and match the segments in the dictionary until we get a complete set of meaningful words. This is important since many users tend to post tweets where the actual message of the tweet is expressed in form of terse hashtagged phrases.
## Processing URLs
The urls embedded in the tweet are a good source of additional context to the actual short tweet content. Sometimes tweets are too terse to comprehend just from the text content of it alone. However if there is a url embedded in the tweet, that can help us understand the context of it – perhaps the sentiment expressed as well.
In order to leverage this additional source of information, we identify all the urls present in the tweets and crawl the web-pages using AlchemyAPI. The API retrieves only the textual body of the article in a web-page. We analyze the article texts later on to get more context for the tweet.
## Algorithmic Framework
We employ a supervised learning model using the manually labeled data as training set and a collection of handcrafted features. In this section we describe the features and the classification model used in this task.
## Feature Extraction
Table TABREF19 presents the set of features we use in our experiment. We have used some basic features (that are commonly used for text classification task) as well as some advanced ones suitable for this particular domain.
We use two basic features:
Parts of Speech (POS) tags: We use the POS tagger of NLTK to tag the tweet texts BIBREF0 . We use counts of noun, adjective, adverb, verb words in a tweet as POS features.
Prior polarity of the words: We use a polarity dictionary BIBREF3 to get the prior polarity of words. The dictionary contains positive, negative and neutral words along with their polarity strength (weak or strong). The polarity of a word is dependent on its POS tag. For example, the word `excuse' is negative when used as `noun' or `adjective', but it carries a positive sense when used as a `verb'. We use the tags produced by NLTK postagger while selecting the prior polarity of a word from the dictionary. We also employ stemming (Porter Stemmer implementation from NLTK) while performing the dictionary lookup to increase number of matches. We use the counts of weak positive words, weak negative words, strong positive words and strong negative words in a tweet as features.
We have also explored some advanced features that helps improve detecting sentiment of tweets.
Emoticons: We use the emoticon dictionary from BIBREF2 , and count the positive and negtive emocicons for each tweet.
The sentiment of url: Since almost all the articles are written in well-formatted english, we analyze the sentiment of the first paragraph of the article using Standford Sentiment Analysis tool BIBREF4 . It predicts sentiment for each sentence within the article. We calculate the fraction of sentences that are negative, positive, and neutral and use these three values as features.
Hashtag: We count the number of hashtags in each tweet.
Capitalization: We assume that capitalization in the tweets has some relationship with the degree of sentiment. We count the number of words with capitalization in the tweets.
Retweet: This is a boolean feature indicating whether the tweet is a retweet or not.
User Mention: A boolean feature indicating whether the tweet contains a user mention.
Negation: Words like `no', `not', `won't' are called negation words since they negate the meaning of the word that is following it. As for example `good' becomes `not good'. We detect all the negation words in the tweets. If a negation word is followed by a polarity word, then we negate the polarity of that word. For example, if `good' is preceeded by a `not', we change the polarity from `weak positive' to `weak negative'.
Text Feature: We use tf-idf based text features to predict the sentiment of a tweet. We perform tf-idf based scoring of words in a tweet and the hashtags present in the tweets. We use the tf-idf vectors to train a classifier and predict the sentiment. This is then used as a stacked prediction feature in the final classifier.
Target: We use the target of the tweet as categorical feature for our classifier.
User: On a particular topic one particular user will generally have a single viewpoint (either positive or negative or neutral). If there are multiple posts within a short period of time from a user, then possibly the posts will contain the same sentiment. We use the user id as a categorical feature. On an average there are INLINEFORM0 tweets per user in the dataset, and over INLINEFORM1 users in the train set have expressed a single viewpoint for all their tweets (either positive or negative). Hence we believe this feature should be able to capture a user's viewpoint on the topic.
.
## Classifier
We experiment with the following set of machine learning classifiers. We train the model with manually labeled data and used the above described features to predict the sentiment. We consider only positive, negative and neutral classes.
Multinomial Naive Bayes : Naive Bayes have been one of the most commonly used classifiers for text classification problems over the years. Naive Bayes classifier makes the assumption that the value of a particular feature is independent of the value of any other feature, given the class variable. This independence assumption makes the classifier both simple and scalable. Bayes classifier assigns a class label INLINEFORM0 for some k according to the following equation: DISPLAYFORM0
The assumptions on distributions of features define the event model of the Naive Bayes classifier. We use multinomial Naive Bayes classifer, which is suitable for discrete features (like counts and frequencies).
Linear SVM : Support Vector Machines are linear non-probabilistic learning algorithms that given training examples, depending on features, build a model to classify new data points to one of the probable classes. We have used support vector machine with stochastic gradient descent learning where gradient of loss is estimated and model is updated at each sample with decreasing strength.
. For this task we found Multinomial Naive Bayes performs slightly better than Linear SVM, hence in the evaluation we report accuracy with this classifier.
## Parameter Tuning
Parameter tuning or hyperparameter optimization is an important step in model selection since it prevents the model from overfitting and optimize the performance of a model on an independent dataset. We perform hyperparameter optimization by using grid search, i.e. an exhaustive searching through a manually specified subset of the hyperparameter space for a learning algorithm. We do grid search and set the `best parameters' by doing cross validation on training set and verified the improvement of accuracy on the validation set. Finally we use the model with best hyperparameters to make predictions on the test set.
## Evaluation and Analysis
Table TABREF39 shows the test results when features are added incrementally. We start with our basic model (with only POS tag features and word polarity features) and subsequently add various sets of features. First we add emoticon features, it has not much effect. This is reasonable since only 8 positive emoticons and 3 negative emoticons are detected(Table TABREF5 ) out of 40049 tokens. So the significance of emoticon can be neglected in this dataset. Then we add hashtag and capitalization features, and obtain an overall gain of 2% over the basic model. By adding the sentiment features from URL articles, we get overall 6% improvement over baseline. Further twitter specific features and user features improve the f1 by 12%. Last, we add TF-IDF feature, and the result improves a lot, and our sentiment classifier reaches the best classification results with an F1-accuracy of INLINEFORM0 as shown in the table.
Analyzing the results for different classes, we observe that the classifier works best for negative tweets. This can be explained by the number of training tweets for each classes, since proportion of negative tweets were considerably higher in both train and test sets as mentioned in Section SECREF2 .
## Comparison with Stanford Sentiment Analysis Tool
In this section we compare the performance of our framework with an openly available state-of-the-art sentiment analysis tool. We choose Stanford coreNLP package as the baseline. It uses recursive deep models to do sentiment analysis and achieves good accuracy ( INLINEFORM0 ) for formal corpora BIBREF4 . However for noisy and informal texts like tweets, their performance decreases sharply. We present the performance of Stanford coreNLP tool over the test dataset.
Comparing table TABREF41 with table TABREF39 we observe that our framework outperforms stanford coreNLP by a significant margin ( INLINEFORM0 ). This owes to the fact that stanford coreNLP is not able to handle text with lot of noise, lack of formality, and slangs/abbreviations. This proves the effectiveness of our framework.
## Enhancements
Apart from sentiment prediction, we also present some extensions to our system.
## Harvest New Sentiment Terms
We have used a static dictionary to get prior polarity of a word, which helps detect the overall sentiment of a sentence. However the usage of words varies depending on conversation medium (e.g. : informal social media, blogs, news media), context and topic. For instance, the word `simple' is generally used in positive sense, but consider its use while describing the storyline of a movie. In this context, a `simple storyline' will probably hint at a negative sentiment. For a dynamic media like Twitter, where the topic mix and word mix change often, having a static dictionary of words with fixed polarity will not suffice. To get temporal and topic-specific sentiment terms, we make use of the tweets classified by our classifier.
We consider the words that appear in the positive, neutral and negative tweets. A word that very frequently occurs in tweets with positive (negative) sentiment and hardly occurs with tweets with a negative (positive) sentiment, will probably have a positive (negative) orientation for that particular topic. To implement this hypothesis, we first count the word frequency in each tweet collection. Then for each collection, we select top INLINEFORM0 most frequent words and deduct from top INLINEFORM1 words from other two collections. For example, in Algorithm SECREF42 , if we want to get new negative words, we find the words in top INLINEFORM2 from negative collection. And we compare the words that appear in top INLINEFORM3 of the other two, remove words that co-appear. Part of the new negative terms we find are shown in Table TABREF43 . We use same procedure to find new positive and neutral words.
Harvest New Negative Words Algorithm negativeCol, positiveCol, neutralCol new negative words from data collection
INLINEFORM0 INLINEFORM1 INLINEFORM2 INLINEFORM3 INLINEFORM4 INLINEFORM5 drop word
## Predicting Strength of Sentiment
Apart from predicting the sentiment class of tweets we are also interested in predicting the strength or intensity of the sentiment associated. Consider the following tweets.
t1: `GO TO YOUR US REPS OFFICE ON SATURDAY AND SAY VOTE NO! ON #HCR #Obama #cnn #killthebill #p2 #msnbc #foxnews #congress #tcot'
t2: `Thankfully the Democrat Party isn't too big to fail. #tcot #hcr'
Although both the tweets have negative sentiment towards `ObamaCare', the intensity in both are not the same. The first tweet (t1) is quite aggressive whereas the other one (t2) is not that much. Here we propose a technique to predict the strength of sentiment.
We consider few features from the tweet in order to do this. If our classifier predicts the sentiment to be neutral we say that the strength of sentiment is 0. However if it is not i.e., if it is either positive or negative, we increase strength of sentiment for each of the following features of the tweet.
Number of capitalized words.
Number of strong positive words.
Number of strong negative words.
Number of weak positive words.
Number of weak negative words.
Each of these features contributes to the strength score of a tweet. Once calculated, we normalize the score within [0-5]. Finally we assign sentiment polarity depending on the overall sentiment of the tweet. As for example, if a tweet has score of 3 and the overall predicted sentiment is negative then we give it a score of `-3'. It denotes that the tweet is moderately negative. Having said that, strength of sentiment is highly subjective. A tweet can appear to be very much aggressive to some person whereas the same may appear to not to be that aggressive to some other person.
## Conclusion
In this report we have presented a sentiment analysis tool for Twitter posts. We have discussed the characteristics of Twitter that make existing sentiment analyzers perform poorly. The model proposed in this report has addressed the challenges by using normalization methods and features specific to this media. We show that using external knowledge outside the tweet text (from landing pages of URLs) and user features can significantly improve performance. We have presented experimental results and comparison with state-of-the-art tools.
We have presented two enhanced functionalities, i.e. discovering new sentiment terms and predicting strength of the sentiment. Due to the absence of labelled data we couldn't discuss the accuracies of these two enhancements. In the future, we plan to use these as feedback mechanism to classify new tweets.
| 16
|
1612.05270
|
A Simple Approach to Multilingual Polarity Classification in Twitter
|
# A Simple Approach to Multilingual Polarity Classification in Twitter
## Abstract
Recently, sentiment analysis has received a lot of attention due to the interest in mining opinions of social media users. Sentiment analysis consists in determining the polarity of a given text, i.e., its degree of positiveness or negativeness. Traditionally, Sentiment Analysis algorithms have been tailored to a specific language given the complexity of having a number of lexical variations and errors introduced by the people generating content. In this contribution, our aim is to provide a simple to implement and easy to use multilingual framework, that can serve as a baseline for sentiment analysis contests, and as starting point to build new sentiment analysis systems. We compare our approach in eight different languages, three of them have important international contests, namely, SemEval (English), TASS (Spanish), and SENTIPOLC (Italian). Within the competitions our approach reaches from medium to high positions in the rankings; whereas in the remaining languages our approach outperforms the reported results.
## Introduction
Sentiment analysis is a crucial task in opinion mining field where the goal is to extract opinions, emotions, or attitudes to different entities (person, objects, news, among others). Clearly, this task is of interest for all languages; however, there exists a significant gap between English state-of-the-art methods and other languages. It is expected that some researchers decided to test the straightforward approach which consists in, first, translating the messages to English, and, then, use a high performing English sentiment classifier (for instance, see BIBREF0 and BIBREF1 ) instead of creating a sentiment classifier optimized for a given language. However, the advantages of a properly tuned sentiment classifier have been studied for different languages (for instance, see BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 ).
This manuscript focuses on the particular case of multilingual sentiment analysis of short informal texts such as Twitter messages. Our aim is to provide an easy-to-use tool to create sentiment classifiers based on supervised learning (i.e., labeled dataset) where the classifier should be competitive to those sentiment classifiers carefully tuned by some given languages. Furthermore, our second contribution is to create a well-performing baseline to compare new sentiment classifiers in a broad range of languages or to bootstrap new sentiment analysis systems. Our approach is based on selecting the best text-transforming techniques that optimize some performance measures where the chosen techniques are robust to typical writing errors.
In this context, we propose a robust multilingual sentiment analysis method, tested in eight different languages: Spanish, English, Italian, Arabic, German, Portuguese, Russian and Swedish. We compare our approach ranking in three international contests: TASS'15, SemEval'15-16 and SENTIPOLC'14, for Spanish, English and Italian respectively; the remaining languages are compared directly with the results reported in the literature. The experimental results locate our approach in good positions for all considered competitions; and excellent results in the other five languages tested. Finally, even when our method is almost cross-language, it can be extended to take advantage of language dependencies; we also provide experimental evidence of the advantages of using these language-dependent techniques.
The rest of the manuscript is organized as follows. Section SECREF2 describes our proposed Sentiment Analysis method. Section SECREF3 describes the datasets and contests used to test our approach; whereas, the experimental results, and, the discussion are presented on Section SECREF4 . Finally, Section SECREF5 concludes.
## Our Approach: Multilingual Polarity Classification
We propose a method for multilingual polarity classification that can serve as a baseline as well as a framework to build more complex sentiment analysis systems due to its simplicity and availability as an open source software. As we mentioned, this baseline algorithm for multilingual Sentiment Analysis (B4MSA) was designed with the purpose of being multilingual and easy to implement. B4MSA is not a naïve baseline which is experimentally proved by evaluating it on several international competitions.
In a nutshell, B4MSA starts by applying text-transformations to the messages, then transformed text is represented in a vector space model (see Subsection SECREF13 ), and finally, a Support Vector Machine (with linear kernel) is used as the classifier. B4MSA uses a number of text transformations that are categorized in cross-language features (see Subsection SECREF3 ) and language dependent features (see Subsection SECREF9 ). It is important to note that, all the text-transformations considered are either simple to implement or there is a well-known library (e.g. BIBREF6 , BIBREF7 ) to use them. It is important to note that to maintain the cross-language property, we limit ourselves to not use additional knowledge, this include knowledge from affective lexicons or models based on distributional semantics.
To obtain the best performance, one needs to select those text-transformations that work best for a particular dataset, therefore, B4MSA uses a simple random search and hill-climbing (see Subsection SECREF14 ) in space of text-transformations to free the user from this delicate and time-consuming task. Before going into the details of each text-transformation, Table TABREF2 gives a summary of the text-transformations used as well as their parameters associated.
## Cross-language Features
We defined cross-language features as a set of features that could be applied in most similar languages, not only related language families such as Germanic languages (English, German, etc.), Romance languages (Spanish, Italian, etc.), among others; but also similar surface features such as punctuation, diacritics, symbol duplication, case sensitivity, etc. Later, the combination of these features will be explored to find the best configuration for a given classifier.
Generally, Twitter messages are full of slang, misspelling, typographical and grammatical errors among others; in order to tackle these aspects we consider different parameters to study this effect. The following points are the parameters to be considered as spelling features. Punctuation (del-punc) considers the use of symbols such as question mark, period, exclamation point, commas, among other spelling marks. Diacritic symbols (del-diac) are commonly used in languages such as Spanish, Italian, Russian, etc., and its wrong usage is one of the main sources of orthographic errors in informal texts; this parameter considers the use or absence of diacritical marks. Symbol reduction (del-d1), usually, twitter messages use repeated characters to emphasize parts of the word to attract user's attention. This aspect makes the vocabulary explodes. We applied the strategy of replacing the repeated symbols by one occurrence of the symbol. Case sensitivity (lc) considers letters to be normalized in lowercase or to keep the original source; the aim is to cut the words that are the same in uppercase and lowercase.
We classified around 500 most popular emoticons, included text emoticons, and the whole set of unicode emoticons (around INLINEFORM0 ) defined by BIBREF8 into three classes: positive, negative and neutral, which are grouped under its corresponding polarity word defined by the class name.
Table TABREF6 shows an excerpt of the dictionary that maps emoticons to their corresponding polarity class.
N-words (word sequences) are widely used in many NLP tasks, and they have also been used in Sentiment Analysis BIBREF9 and BIBREF10 . To compute the N-words, the text is tokenized and N-words are calculated from tokens. For example, let INLINEFORM0 be the text, so its 1-words (unigrams) are each word alone, and its 2-words (bigrams) set are the sequences of two words, the set ( INLINEFORM1 ), and so on. INLINEFORM2 = {the lights, lights and, and shadows, shadows of, of your, your future}, so, given text of size INLINEFORM3 words, we obtain a set containing at most INLINEFORM4 elements. Generally, N-words are used up to 2 or 3-words because it is uncommon to find, between texts, good matches of word sequences greater than three or four words BIBREF11 .
In addition to the traditional N-words representation, we represent the resulting text as q-grams. A q-grams is an agnostic language transformation that consists in representing a document by all its substring of length INLINEFORM0 . For example, let INLINEFORM1 be the text, its 3-grams set are INLINEFORM2
so, given text of size INLINEFORM0 characters, we obtain a set with at most INLINEFORM1 elements. Notice that this transformation handles white-spaces as part of the text. Since there will be q-grams connecting words, in some sense, applying q-grams to the entire text can capture part of the syntactic and contextual information in the sentence. The rationale of q-grams is also to tackle misspelled sentences from the approximate pattern matching perspective BIBREF12 .
## Language Dependent Features
The following features are language dependent because they use specific information from the language concerned. Usually, the use of stopwords, stemming and negations are traditionally used in Sentiment Analysis. The users of this approach could add other features such as part of speech, affective lexicons, etc. to improve the performance BIBREF13 .
In many languages, there is a set of extremely common words such as determiners or conjunctions ( INLINEFORM0 or INLINEFORM1 ) which help to build sentences but do not carry any meaning for themselves. These words are known as Stopwords, and they are removed from text before any attempt to classify them. Generally, a stopword list is built using the most frequent terms from a huge document collection. We used the Spanish, English and Italian stopword lists included in the NLTK Python package BIBREF6 in order to identify them.
Stemming is a well-known heuristic process in Information Retrieval field that chops off the end of words and often includes the removal of derivational affixes. This technique uses the morphology of the language coded in a set of rules that are applied to find out word stems and reduce the vocabulary collapsing derivationally related words. In our study, we use the Snowball Stemmer for Spanish and Italian, and the Porter Stemmer for English that are implemented in NLTK package BIBREF6 .
Negation markers might change the polarity of the message. Thus, we attached the negation clue to the nearest word, similar to the approaches used in BIBREF9 . A set of rules was designed for common negation structures that involve negation markers for Spanish, English and Italian. For instance, negation markers used for Spanish are no (not), nunca, jamás (never), and sin (without). The rules (regular expressions) are processed in order, and their purpose is to negate the nearest word to the negation marker using only the information on the text, e.g., avoiding mainly pronouns and articles. For example, in the sentence El coche no es bonito (The car is not nice), the negation marker no and not (for English) is attached to its adjective no_bonito (not_nice).
## Text Representation
After text-transformations, it is needed to represent the text in suitable form in order to use a traditional classifier such as SVM. It was decided to select the well known vector representation of a text given its simplicity and powerful representation. Particularly, it is used the Term Frequency-Inverse Document Frequency which is a well-known weighting scheme in NLP. TF-IDF computes a weight that represents the importance of words or terms inside a document in a collection of documents, i.e., how frequently they appear across multiple documents. Therefore, common words such as the and in, which appear in many documents, will have a low score, and words that appear frequently in a single document will have high score. This weighting scheme selects the terms that represent a document.
## Parameter Optimization
The model selection, sometimes called hyper-parameter optimization, is essential to ensure the performance of a sentiment classifier. In particular, our approach is highly parametric; in fact, we use such property to adapt to several languages. Table TABREF2 summarizes the parameters and their valid values. The search space contains more than 331 thousand configurations when limited to multilingual and language independent parameters; while the search space reaches close to 4 million configurations when we add our three language-dependent parameters. Depending on the size of the training set, each configuration needs several minutes on a commodity server to be evaluated; thus, an exhaustive exploration of the parameter space can be quite expensive making the approach useless in practice. To tackle the efficiency problems, we perform the model selection using two hyper-parameter optimization algorithms.
The first corresponds to Random Search, described in depth in BIBREF14 . Random search consists on randomly sampling the parameter space and select the best configuration among the sample. The second algorithm consists on a Hill Climbing BIBREF15 , BIBREF16 implemented with a memory to avoid testing a configuration twice. The main idea behind hill climbing H+M is to take a pivoting configuration, explore the configuration's neighborhood, and greedily moves to the best neighbor. The process is repeated until no improvement is possible. The configuration neighborhood is defined as the set of configurations such that these differ in just one parameter's value. This rule is strengthened for tokenizer (see Table TABREF2 ) to differ in a single internal value not in the whole parameter value. More precisely, let INLINEFORM0 be a valid value for tokenizer and INLINEFORM1 the set of valid values for neighborhoods of INLINEFORM2 , then INLINEFORM3 and INLINEFORM4 for any INLINEFORM5 .
To guarantee a better or equal performance than random search, the H+M process starts with the best configuration found in the random search. By using H+M, sample size can be set to 32 or 64, as rule of thumb, and even reach improvements in most cases (see § SECREF4 ). Nonetheless, this simplification and performance boosting comes along with possible higher optimization times. Finally, the performance of each configuration is obtained using a cross-validation technique on the training data, and the metrics are usually used in classification such as: accuracy, score INLINEFORM0 , and recall, among others.
## Datasets and contests
Nowadays, there are several international competitions related to text mining, which include diverse tasks such as: polarity classification (at different levels), subjectivity classification, entity detection, and iron detection, among others. These competitions are relevant to measure the potential of different proposed techniques. In this case, we focused on polarity classification task, hence, we developed a baseline method with an acceptable performance achieved in three different contests, namely, TASS'15 (Spanish) BIBREF17 , SemEval'15-16 (English) BIBREF18 , BIBREF19 , and SENTIPOLC'14 (Italian) BIBREF20 . In addition, our approach was tested with other languages (Arabic, German, Portuguese, Russian, and Swedish) to show that is feasible to use our framework as basis for building more complex sentiment analysis systems. From these languages, datasets and results can be seen in BIBREF21 , BIBREF3 and BIBREF2 .
Table TABREF15 presents the details of each of the competitions considered as well as the other languages tested. It can be observed, from the table, the number of examples as well as the number of instances for each polarity level, namely, positive, neutral, negative and none. The training and development (only in SemEval) sets are used to train the sentiment classifier, and the gold set is used to test the classifier. In the case there dataset was not split in training and gold (Arabic, German, Portuguese, Russian, and Swedish) then a cross-validation (10 folds) technique is used to test the classifier. The performance of the classifier is presented using different metrics depending the competition. SemEval uses the average of score INLINEFORM0 of positive and negative labels, TASS uses the accuracy and SENTIPOLC uses a custom metric (see BIBREF17 , BIBREF18 , BIBREF19 , BIBREF20 ).
## Experimental Results
We tested our framework on two kinds of datasets. On one hand, we compare our performance on three languages having well known sentiment analysis contests; here, we compare our work against competitors of those challenges. On the other hand, we selected five languages without popular opinion mining contests; for these languages, we compare our approach against research works reporting the used corpus.
## Performance on sentiment analysis contests
Figure FIGREF17 shows the performance on four contests, corresponding to three different languages. The performance corresponds to the multilingual set of features, i.e., we do not used language-dependent techniques.
Figures UID18 - UID21 illustrates the results on each challenge, all competitors are ordered in score's descending order (higher is better). The achieved performance of our approach is marked with a horizontal line on each figure. Figure UID22 briefly describes each challenge and summarizes our performance on each contest; also, we added three standard measures to simplify the insight's creation of the reader.
The winner method in SENTIPOLC'14 (Italian) is reported in BIBREF22 . This method uses three groups of features: keyword and micro-blogging characteristics, Sentiment Lexicons, SentiWordNet and MultiWordNet, and Distributional Semantic Model (DSM) with a SVM classifier. In contrast with our method, in BIBREF22 three external sentiment lexicons dictionaries were employed; that is, external information.
In TASS'15 (Spanish) competition, the winner reported method was BIBREF23 , which proposed an adaptation based on a tokenizer of tweets Tweetmotif BIBREF24 , Freeling BIBREF25 as lemmatizer, entity detector, morphosyntactic labeler and a translation of the Afinn dictionary. In contrast with our method, BIBREF23 employs several complex and expensive tools. In this task we reached the fourteenth position with an accuracy of INLINEFORM0 . Figure UID19 shows the B4MSA performance to be over two thirds of the competitors.
The remaining two contests correspond to the SemEval'15-16. The B4MSA performance in SemEval is depicted in Figures UID20 and UID21 ; here, B4MSA does not perform as well as in other challenges, mainly because, contrary to other challenges, SemEval rules promotes the enrichment of the official training set. To be consistent with the rest of the experiments, B4MSA uses only the official training set. The results can be significantly improved using larger training datasets; for example, joining SemEval'13 and SemEval'16 training sets, we can reach INLINEFORM0 for SemEval'16, which improves the B4MSA's performance (see Table FIGREF17 ).
In SemEval'15, the winner method is BIBREF26 , which combines three approaches among the participants of SemEval'13, teams: NRC-Canada, GU-MLT-LT and KLUE, and from SemEval'14 the participant TeamX all of them employing external information. In SemEval'16, the winner method was BIBREF27 is composed with an ensemble of two subsystems based on convolutional neural networks, the first subsystem is created using 290 million tweets, and the second one is feeded with 150 million tweets. All these tweets were selected from a very large unlabeled dataset through distant supervision techniques.
Table TABREF23 shows the multilingual set of techniques and the set with language-dependent techniques; for each, we optimized the set of parameters through Random Search and INLINEFORM0 (see Subsection SECREF14 ). The reached performance is reported using both cross-validation and the official gold-standard. Please notice how INLINEFORM1 consistently reaches better performances, even on small sampling sizes. The sampling size is indicated with subscripts in Table TABREF23 . Note that, in SemEval challenges, the cross-validation performances are higher than those reached by evaluating the gold-standard, mainly because the gold-standard does not follow the distribution of training set. This can be understood because the rules of SemEval promote the use of external knowledge.
Table TABREF24 compares our performance on five different languages; we do not apply language-dependent techniques. For each comparison, we took a labeled corpus from BIBREF3 (Arabic) and BIBREF21 (the remaining languages). According to author's reports, all tweets were manually labeled by native speakers as pos, neg, or neu. The Arabic dataset contains INLINEFORM0 items; the other datasets contain from 58 thousand tweets to more than 157 thousand tweets. We were able to fetch a fraction of the original datasets; so, we drop the necessary items to hold the original class-population ratio. The ratio of tweets in our training dataset, respect to the original dataset, is indicated beside the name. As before, we evaluate our algorithms through a 10-fold cross validation.
In BIBREF3 , BIBREF2 , the authors study the effect of translation in sentiment classifiers; they found better to use native Arabic speakers as annotators than fine-tuned translators plus fine-tuned English sentiment classifiers. In BIBREF21 , the idea is to measure the effect of the agreement among annotators on the production of a sentiment-analysis corpus. Both, on the technical side, both papers use fine tuned classifiers plus a variety of pre-processing techniques to prove their claims. Table TABREF24 supports the idea of choosing B4MSA as a bootstrapping sentiment classifier because, in the overall, B4MSA reaches superior performances regardless of the language. Our approach achieves those performance's levels since it optimizes a set of parameters carefully selected to work on a variety of languages and being robust to informal writing. The latter problem is not properly tackled in many cases.
## Conclusions
We presented a simple to implement multilingual framework for polarity classification whose main contributions are in two aspects. On one hand, our approach can serve as a baseline to compare other classification systems. It considers techniques for text representation such as spelling features, emoticons, word-based n-grams, character-based q-grams and language dependent features. On the other hand, our approach is a framework for practitioners or researchers looking for a bootstrapping sentiment classifier method in order to build more elaborated systems.
Besides the text-transformations, the proposed framework uses a SVM classifier (with linear kernel), and, hyper-parameter optimization using random search and H+M over the space of text-transformations. The experimental results show good overall performance in all international contests considered, and the best results in the other five languages tested.
It is important to note that all the methods that outperformed B4MSA in the sentiment analysis contests use extra knowledge (lexicons included) meanwhile B4MSA uses only the information provided by each contests. In future work, we will extend our methodology to include extra-knowledge in order to improve the performance.
## Acknowledgements
We would like to thank Valerio Basile, Julio Villena-Roman, and Preslav Nakov for kindly give us access to the gold-standards of SENTIPOLC'14, TASS'15 and SemEval 2015 & 2016, respectively. The authors also thank Elio Villaseñor for the helpful discussions in early stages of this research.
| 11
|
1612.05310
|
Modeling Trolling in Social Media Conversations
|
# Modeling Trolling in Social Media Conversations
## Abstract
Social media websites, electronic newspapers and Internet forums allow visitors to leave comments for others to read and interact. This exchange is not free from participants with malicious intentions, who troll others by positing messages that are intended to be provocative, offensive, or menacing. With the goal of facilitating the computational modeling of trolling, we propose a trolling categorization that is novel in the sense that it allows comment-based analysis from both the trolls' and the responders' perspectives, characterizing these two perspectives using four aspects, namely, the troll's intention and his intention disclosure, as well as the responder's interpretation of the troll's intention and her response strategy. Using this categorization, we annotate and release a dataset containing excerpts of Reddit conversations involving suspected trolls and their interactions with other users. Finally, we identify the difficult-to-classify cases in our corpus and suggest potential solutions for them.
## Introduction
In contrast to traditional content distribution channels like television, radio and newspapers, Internet opened the door for direct interaction between the content creator and its audience. Young people are now gaining more frequent access to online, networked media. Although most of the time, their Internet use is harmless, there are some risks associated with these online activities, such as the use of social networking sites (e.g., Twitter, Facebook, Reddit). The anonymity and freedom provided by social networks makes them vulnerable to threatening situations on the Web, such as trolling.
Trolling is “the activity of posting messages via a communications network that are intended to be provocative, offensive or menacing” BIBREF0 . People who post such comments are known as trolls. According to hardaker2010trolling, a troll's “real intention(s) is/are to cause disruption and/or trigger or exacerbate conflict for the purpose of their own amusement”. Worse still, the troll's comments may have a negative psychological impact on his target/victim and possibly others who participated in the same conversation. It is therefore imperative to identify such comments and perhaps even terminate the conversation before it evolves into something psychological disruptive for the participants. Monitoring conversations is a labor-intensive task: it can potentially place a severe burden on the moderators, and it may not be an effective solution when traffic is heavy. This calls for the need to develop automatic methods for identifying malicious comments, which we will refer to as trolling attempts in this paper.
In fact, there have recently been some attempts to automatically identify comments containing cyberbullying (e.g., van2015detection), which corresponds to the most severe cases of trolling BIBREF0 . However, we believe that it is important not only to identify trolling attempts, but also comments that could have a negative psychological impact on their recipients. As an example, consider the situation where a commenter posts a comment with the goal of amusing others. However, it is conceivable that not everybody would be aware of these playful intentions, and these people may disagree or dislike the mocking comments and take them as inappropriate, prompting a negative reaction or psychological impact on themselves.
In light of this discussion, we believe that there is a need to identify not only the trolling attempts, but also comments that could have a negative psychological impact on its receipts. To this end, we seek to achieve the following goals in this paper. First, we propose a comprehensive categorization of trolling that allows us to model not only the troll's intention given his trolling attempt, but also the recipients' perception of the troll's intention and subsequently their reaction to the trolling attempt. This categorization gives rise to very interesting problems in pragmatics that involve the computational modeling of intentions, perceived intentions, and reactions to perceived intentions. Second, we create a new annotated resource for computational modeling of trolling. Each instance in this resource corresponds to a suspected trolling attempt taken from a Reddit conversation, it's surrounding context, and its immediate responses and will be manually coded with information such as the troll's intention and the recipients' reactions using our proposed categorization of trolling. Finally, we identify the instances that are difficult to classify with the help of a classifier trained with features taken from the state of the art, and subsequently present an analysis of these instances.
To our knowledge, our annotated resource is the first one of its sort that allows computational modeling on both the troll's side and the recipients' side. By making it publicly available, we hope to stimulate further research on this task. We believe that it will be valuable to any NLP researcher who is interested in the computational modeling of trolling.
## Related Work
In this section, we discuss related work in the areas of trolling, bullying, abusive language detection and politeness, as they intersect in their scope and at least partially address the problem presented in this work.
In the realm of psychology, bishop2013effect and bishop2014representations elaborate a deep description of a troll's personality, motivations, effects on the community that trolls interfere in and the criminal and psychological aspects of trolls. Their main focus are flaming (trolls), and hostile and aggressive interactions between users BIBREF1 .
On the computational side, mihaylov2015finding address the problem of identifying manipulation trolls in news community forums. Not only do they focus solely on troll identification, but the major difference with this work is that all their predictions are based on non-linguistic information such as number of votes, dates, number of comments and so on. In a networks related framework, kumar2014accurately and guha2004propagation present a methodology to identify malicious individuals in a network based solely on the network's properties rather than on the textual content of comments. cambria2010not propose a method that involves NLP components, but fail to provide an evaluation of their system.
There is extensive work on detecting offensive and abusive language in social media BIBREF2 and BIBREF3 . There are two clear differences between their work and ours. One is that trolling is concerned about not only abusive language but also a much larger range of language styles and addresses the intentions and interpretations of the commenters, which goes beyond the linguistic dimension. The other is that we are additionally interested in the reactions to trolling attempts, real or perceived, because we argued that this is a phenomenon that occurs in pairs through the interaction of at least two individuals, which is different from abusive language detection. Also, xu2012learning, xu2012fast and xu2013examination address bullying traces. Bullying traces are self-reported events of individuals describing being part of bullying events, but we believe that the real impact of computational trolling research is not on analyzing retrospective incidents, but on analyzing real-time conversations. chen2012detecting use lexical and semantic features to determine sentence offensiveness levels to identify cyberbullying, offensive or abusive comments on Youtube. On Youtube as well, dinakar2012common identified sensitive topics for cyberbullying. dadvar2014experts used expert systems to classify between bullying and no bullying in posts. van2015detection predict fine-grained categories for cyberbullying, distinguishing between insults and threats and identified user roles in the exchanges. Finally, hardaker2010trolling argues that trolling cannot be studied using established politeness research categories.
## Trolling Categorization
In this section, we describe our proposal of a comprehensive trolling categorization. While there have been attempts in the realm of psychology to provide a working definition of trolling (e.g., hardaker2010trolling, bishop2014representations), their focus is mostly on modeling the troll's behavior. For instance, bishop2014representations constructed a “trolling magnitude” scale focused on the severity of abuse and misuse of internet mediated communications. bishop2013effect also categorized trolls based on psychological characteristics focused on pathologies and possible criminal behaviors. In contrast, our trolling categorization seeks to model not only the troll's behavior but also the impact on the recipients, as described below.
Since one of our goals is to identify trolling events, our datasets will be composed of suspected trolling attempts (i.e., comments that are suspected to be trolling attempts). In other words, some of these suspected trolling attempts will be real trolling attempts, and some of them won't. So, if a suspected trolling attempt is in fact not a trolling attempt, then its author will not be a troll.
To cover both the troll and the recipients, we define a (suspected trolling attempt, responses) pair as the basic unit that we consider for the study of trolling, where “responses” are all the direct responses to the suspected trolling attempt. We characterize a (suspected trolling attempt, responses) pair using four aspects. Two aspects describe the trolling attempt: (1) Intention (I) (what is its author's purpose?), and (2) Intention Disclosure (D) (is its author trying to deceive its readers by hiding his real (i.e., malicious) intentions?). The remaining two aspects are defined on each of the (direct) responses to the trolling attempt: (1) Intention Interpretation (R) (what is the responder's perception of the troll's intention?), and (2) the Response strategy (B) (what is the responder's reaction?). Two points deserve mention. First, R can be different from I due to misunderstanding and the fact that the troll may be trying to hide his intention. Second, B is influenced by R, and the responder's comment can itself be a trolling attempt. We believe that these four aspects constitute interesting, under-studied pragmatics tasks for NLP researchers.
The possible values of each aspect are described in Table TABREF1 . As noted before, since these are suspected trolling attempts, if an attempt turns out not to be a trolling attempt, its author will not be a troll.
For a given (suspected trolling attempt, responses) pair, not all of the 189 (= INLINEFORM0 ) combinations of values of the four aspects are possible. There are logical constraints that limit plausible combinations: a) Trolling or Playing Intentions (I) must have Hidden or Exposed Intention Disclosure (D), b) Normal intentions (I) can only have None Intention disclosure (D) and c) Trolling or Playing interpretation (R) cannot have Normal response strategy (B).
## Conversation Excerpts
To enable the reader to better understand this categorization, we present two example excerpts taken from the original (Reddit) conversations. The first comment on each excerpt, generated by author C0, is given as a minimal piece of context. The second comment, written by the author C1 in italics, is the suspected trolling attempt. The rest of the comments comprise all direct responses to the suspected trolling comment.
Example 1.
[noitemsep,nolistsep]
Yeah, cause that's what usually happens. Also, quit following me around, I don't want a boyfriend.
[noitemsep,nolistsep]
I wasn't aware you were the same person.... I've replied to a number of stupid people recently, my bad
[noitemsep,nolistsep]
Trollname trollpost brotroll
In this example, C1 is teasing of C0, expecting to provoke or irritate irritate, and he is clearly disclosing her trolling intentions. In C0's response, we see that he clearly believe that C1 is trolling, since is directly calling him a “brotroll” and his response strategy is frustrate the trolling attempt by denouncing C1 troll's intentions “trollpost” and true identity “brotroll”.
Example 2.
[noitemsep,nolistsep]
Please post a video of your dog doing this. The way I'm imagining this is adorable.
[noitemsep,nolistsep]
I hope the dog gets run over by a truck on the way out of the childrens playground.
[noitemsep,nolistsep]
If you're going to troll, can you at least try to be a bit more
Haha I hope the cancer kills you. convincing?
In this example, we observe that C0's first comment is making a polite request (Please). In return, C1 answer is a mean spirited comment whose intention is to disrupt and possible hurtful C0. Also, C1's comment is not subtle at all, so his intention is clearly disclosed. As for C2, she is clearly acknowledging C1's trolling intention and her response strategy is a criticism which we categorize as frustrate. Now, in C0's second comment, we observe that his interpretation is clear, he believes that C1 is trolling and the negative effect is so tangible, that his response strategy is to troll back or counter-troll by replying with a comparable mean comment.
## Corpus and Annotation
Reddit is popular website that allows registered users (without identity verification) to participate in fora grouped by topic or interest. Participation consists of posting stories that can be seen by other users, voting stories and comments, and comments in the story's comment section, in the form of a forum. The forums are arranged in the form of a tree, allowing nested conversations, where the replies to a comment are its direct responses. We collected all comments in the stories' conversation in Reddit that were posted in August 2015. Since it is infeasible to manually annotate all of the comments, we process this dataset with the goal of extracting threads that involve suspected trolling attempts and the direct responses to them. To do so, we used Lucene to create an inverted index from the comments and queried it for comments containing the word “troll” with an edit distance of 1 in order to include close variations of this word, hypothesizing that such comments would be reasonable candidates of real trolling attempts. We did observe, however, that sometimes people use the word troll to point out that another user is trolling. Other times, people use the term to express their frustration about a particular user, but there is no trolling attempt. Yet other times people simply discuss trolling and trolls without actually observing one. Nonetheless, we found that this search produced a dataset in which 44.3% of the comments are real trolling attempts. Moreover, it is possible for commenters to believe that they are witnessing a trolling attempt and respond accordingly even where there is none due to misunderstanding. Therefore, the inclusion of comments that do not involve trolling would allow us to learn what triggers a user's interpretation of trolling when it is not present and what kind of response strategies are used.
For each retrieved comment, we reconstructed the original conversation tree it appears in, from the original post (i.e., the root) to the leaves, so that its parent and children can be recovered. We consider a comment in our dataset a suspected trolling attempt if at least one of its immediate children contains the word troll. For annotation purposes, we created snippets of conversations exactly like the ones shown in Example 1 and Example 2, each of which consists of the parent of the suspected trolling attempt, the suspected trolling attempt, and all of the direct responses to the suspected trolling attempt.
We had two human annotators who were trained on snippets (i.e., (suspected trolling attempt, responses) pairs) taken from 200 conversations and were allowed to discuss their findings. After this training stage, we asked them to independently label the four aspects for each snippet. We recognize that this limited amount of information is not always sufficient to recover the four aspects we are interested in, so we give the annotators the option to discard instances for which they couldn't determine the labels confidently. The final annotated dataset consists of 1000 conversations composed of 6833 sentences and 88047 tokens. The distribution over the classes per trolling aspect is shown in the table TABREF19 in the column “Size”.
Due to the subjective nature of the task we did not expect perfect agreement. However, on the 100 doubly-annotated snippets, we obtained substantial inter-annotator agreement according to Cohen's kappa statistic BIBREF4 for each of the four aspects: Intention: 0.788, Intention Disclosure: 0.780, Interpretation: 0.797 and Response 0.776. In the end, the annotators discussed their discrepancies and managed to resolve all of them.
## Trolling Attempt Prediction
In this section, we make predictions on the four aspects of our task, with the primary goal of identifying the errors our classifier makes (i.e., the hard-to-classify instances) and hence the directions for future work, and the secondary goal of estimating the state of the art on this new task using only shallow (i.e., lexical and wordlist-based) features.
## Feature Sets
For prediction we define two sets of features: (1) a basic feature set taken from Van Hee's van2015detection paper on cyberbullying prediction, and (2) an extended feature set that we designed using primarily information extracted from wordlists and dictionaries.
N-gram features. We encode each lemmatized and unlemmatized unigram and bigram collected from the training comments as a binary feature. In a similar manner, we include the unigram and bigram along with their POS tag as in BIBREF5 . To extract these features we used Stanford CoreNLP BIBREF6 .
Sentiment Polarity. The overall comment's emotion could be useful to identify the response and intention in a trolling attempt. So, we apply the Vader Sentiment Polarity Analyzer BIBREF7 and include four features, one per each measurement given by the analyzer: positive, neutral, negative and a composite metric, each as a real number value.
Emoticons. Reddit's comments make extensive use of emoticons. We argue that some emoticons are specifically used in trolling attempts to express a variety of emotions, which we hypothesize would be useful to identify a comment's intention, interpretation and response. For that reason, we use the emoticon dictionary developed hogenboom2015exploiting. We create a binary feature whose value is one if at least one of these emoticons is found in the comment.
Harmful Vocabulary. In their research on bullying, nitta2013detecting identified a small set of words that are highly offensive. We create a binary feature whose value is one if the comment contains at least one of these words.
Emotions Synsets. As in xu2012fast, we extracted all lemmas associated with each WordNet BIBREF8 synset involving seven emotions (anger, embarrassment, empathy, fear, pride, relief and sadness) as well as the synonyms of these emotion words extracted from the English merriam2004merriam dictionary. We create a binary feature whose value is one if any of these synsets or synonyms appears in the comment.
Swearing Vocabulary. We manually collected 1061 swear words and short phrases from the internet, blogs, forums and smaller repositories . The informal nature of this dictionary resembles the type of language used by flaming trolls and agitated responses, so we encode a binary feature whose value is one when at least one such swear word is found in the comment.
Swearing Vocabulary in Username. An interesting feature that is suggestive of the intention of a comment is the author's username. We found that abusive and annoying commenters contained cursing words in their usernames. So, we create a binary feature whose value is one if a swear word from the swearing vocabulary is found in their usernames.
Framenet. We apply the SEMAFOR parser BIBREF9 to each sentence in every comment, and construct three different types of binary features: every frame name that is present in the sentence, the frame name and the target word associated with it, and the argument name along with the token or lexical unit in the sentence associated with it. We believe that some frames are especially interesting from the trolling perspective. We hypothesize that these features are useful for identifying trolling attempts in which semantic and not just syntactic information is required.
Politeness cues. danescu2013computational identified cues that signal polite and impolite interactions among groups of people collaborating online. Based on our observations of trolling examples, it is clear that flaming, hostile and aggressive interactions between users BIBREF1 and engaged or emotional responses would use impolite cues. In contrast, neutralizing and frustrating responses to the troll avoid falling in confrontation and their vocabulary tends to be more polite. So we create a binary feature whose value is one if at least one cue appears in the comment.
GloVe Embeddings. All the aforementioned features constitute a high dimensional bag of words (BOW). Word embeddings were created to overcome certain problems with the BOW representation, like sparsity, and weight in correlations of semantically similar words. For this reason, and following nobata2016abusive, we create a distributed representation of the comments by averaging the word vector of each lowercase token in the comment found in the Twitter corpus pre-trained GloVe vectors BIBREF10 . The resulting comment vector representation is a 200 dimensional array that is concatenated with the existing BOW.
## Results
Using the features described in the previous subsection, we train four independent classifiers using logistic regression, one per each of the four prediction tasks. All the results are obtained using 5-fold cross-validation experiments. In each fold experiment, we use three folds for training, one fold for development, and one fold for testing. All learning parameters are set to their default values except for the regularization parameter, which we tuned on the development set. In Table TABREF19 the leftmost results column reports F1 score based on majority class prediction. The next section (Single Feature Group) reports F1 scores obtained by using one feature group at a time. The goal of the later set of experiments is to gain insights about feature predictive effectiveness. The right side section (All features) shows the system performance measured using recall, precision, and F-1 as shown when all features described in section SECREF13 are used.
The majority class prediction experiment is simplest baseline to which we can can compare the rest of the experiments. In order to illustrate the prediction power of each feature group independent from all others, we perform the “Single Feature Group”, experiments. As we can observe in Table TABREF19 there are groups of features that independently are not better than the majority baseline, for example, the emoticons, politeness cues and polarity are not better disclosure predictors than the majority base. Also, we observe that only n-grams and GloVe features are the only group of features that contribute to more than a class type for the different tasks. Now, the “All Features” experiment shows how the interaction between feature sets perform than any of the other features groups in isolation. The accuracy metric for each trolling task is meant to provide an overall performance for all the classes within a particular task, and allow comparison between different experiments. In particular, we observe that GloVe vectors are the most powerful feature set, accuracy-wise, even better than the experiments with all features for all tasks except interpretation.
The overall Total Accuracy score reported in table TABREF19 using the entire feature set is 549. This result is what makes this dataset interesting: there is still lots of room for research on this task. Again, the primary goal of this experiment is to help identify the difficult-to-classify instances for analysis in the next section.
## Error Analysis
In order to provide directions for future work, we analyze the errors made by the classifier trained on the extended features on the four prediction tasks.
Errors on Intention (I) prediction: The lack of background is a major problem when identifying trolling comments. For example, “your comments fit well in Stormfront” seems inoffensive on the surface. However, people who know that Stormfront is a white supremacist website will realize that the author of this comment had an annoying or malicious intention. But our system had no knowledge about it and simply predicted it as non-trolling. These kind of errors reduces recall on the prediction of trolling comments. A solution would be to include additional knowledge from anthologies along with a sentiment or polarity. One could modify NELL BIBREF12 to broaden the understanding of entities in the comments.
Non-cursing aggressions and insults This is a challenging problem, since the majority of abusive and insulting comments rely on profanity and swearing. The problem arises with subtler aggressions and insults that are equally or even more annoying, such as “Troll? How cute.” and “settle down drama queen”. The classifier has a more difficult task of determining that these are indeed aggressions or insults. This error also decreases the recall of trolling intention. A solution would be to exploit all the comments made by the suspected troll in the entire conversation in order to increase the chances of finding curse words or other cues that lead the classifier to correctly classify the comment as trolling.
Another source of error is the presence of controversial topic words such as “black”,“feminism”, “killing”, “racism”, “brown”, etc. that are commonly used by trolls. The classifier seems too confident to classify a comment as trolling in the presence of these words, but in many cases they do not. In order to ameliorate this problem, one could create ad-hoc word embeddings by training glove or other type of distributed representation on a large corpus for the specific social media platform in consideration. From these vectors one could expect a better representation of controversial topics and their interactions with other words so they might help to reduce these errors.
Errors on Disclosure (D) prediction: A major source of error that affects disclosure is the shallow meaning representation obtained from the BOW model even when augmented with the distributional features given by the glove vectors. For example, the suspected troll's comment “how to deal with refugees? How about a bullet to the head” is clearly mean-spirited and is an example of disclosed trolling. However, to reach that conclusion the reader need to infer the meaning of “bullet to the head” and that this action is desirable for a vulnerable group like migrants or refugees. This problem produces low recall for the disclosed prediction task. A solution for this problem may be the use of deeper semantics, where we represent the comments and sentences in their logical form and infer from them the intended meaning.
Errors on Interpretation (R) prediction: it is a common practice from many users to directly ask the suspected troll if he/she is trolling or not. There are several variations of this question, such as “Are you a troll?” and “not sure if trolling or not”. While the presence of a question like these seems to give us a hint of the responder's interpretation, we cannot be sure of his interpretation without also considering the context. One way to improve interpretation is to exploit the response strategy, but the response strategy in our model is predicted independently of interpretation. So one solution could be similar to the one proposed above for the disclosure task problem: jointly learning classifiers that predict both variables simultaneously. Another possibility is to use the temporal sequence of response comments and make use of older response interpretation as input features for later comments. This could be useful since commenters seem to influence each other as they read through the conversation.
Errors on Response Strategy (B) prediction: In some cases there is a blurry line between “Frustrate” and “Neutralize”. The key distinction between them is that there exists some criticism in the Frustrate responses towards the suspected troll's comment, while “Neutralizing” comments acknowledge that the suspected troll has trolling intentions, but gives no importance to them. For example, response comments such as “oh, you are a troll” and “you are just a lame troll” are examples of this subtle difference. The first is a case of “neutralize” while the second is indeed criticizing the suspected troll's comment and therefore a “frustrate” response strategy. This kind of error affects both precision and recall for these two classes. A possible solution could be to train a specialized classifier to disambiguate between “frustrate” and “neutralize” only.
Another challenging problem is the distinction between the classes “Troll” and “Engage”. This is true when the direct responder is intensely flared up with the suspected comment to the point that his own comment becomes a trolling attempt. A useful indicator for distinguishing these cases are the presence of insults, and to detect them we look for swear words, but as we noted before, there is no guarantee that swear words are used for insulting. This kind of error affects the precision and recall for the “troll” and “engage” classes. A solution to this problem may be the inclusion of longer parts of the conversation. It is typical in a troll-engaged comment scheme to observe longer than usual exchanges between two users, and the comments evolve in very agitated remarks. One may then use this information to disambiguate between the two classes.
## Conclusion and Future Work
We presented a new view on the computational modeling of trolling in Internet fora where we proposed a comprehensive categorization of trolling attempts that for the first time considers trolling from not only the troll's perspective but also the responders' perspectives. This categorization gives rise to four interesting pragmatics tasks that involve modeling intensions, perceived intensions, and reactions. Perhaps most importantly, we create an annotated dataset that we believe is the first of its sort. We intend to make publicly available with the hope of stimulating research on trolling.
| 10
|
1612.08205
|
Predicting the Industry of Users on Social Media
|
# Predicting the Industry of Users on Social Media
## Abstract
Automatic profiling of social media users is an important task for supporting a multitude of downstream applications. While a number of studies have used social media content to extract and study collective social attributes, there is a lack of substantial research that addresses the detection of a user's industry. We frame this task as classification using both feature engineering and ensemble learning. Our industry-detection system uses both posted content and profile information to detect a user's industry with 64.3% accuracy, significantly outperforming the majority baseline in a taxonomy of fourteen industry classes. Our qualitative analysis suggests that a person's industry not only affects the words used and their perceived meanings, but also the number and type of emotions being expressed.
## Introduction
Over the past two decades, the emergence of social media has enabled the proliferation of traceable human behavior. The content posted by users can reflect who their friends are, what topics they are interested in, or which company they are working for. At the same time, users are listing a number of profile fields to define themselves to others. The utilization of such metadata has proven important in facilitating further developments of applications in advertising BIBREF0 , personalization BIBREF1 , and recommender systems BIBREF2 . However, profile information can be limited, depending on the platform, or it is often deliberately omitted BIBREF3 . To uncloak this information, a number of studies have utilized social media users' footprints to approximate their profiles.
This paper explores the potential of predicting a user's industry –the aggregate of enterprises in a particular field– by identifying industry indicative text in social media. The accurate prediction of users' industry can have a big impact on targeted advertising by minimizing wasted advertising BIBREF4 and improved personalized user experience. A number of studies in the social sciences have associated language use with social factors such as occupation, social class, education, and income BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 . An additional goal of this paper is to examine such findings, and in particular the link between language and occupational class, through a data-driven approach.
In addition, we explore how meaning changes depending on the occupational context. By leveraging word embeddings, we seek to quantify how, for example, cloud might mean a separate concept (e.g., condensed water vapor) in the text written by users that work in environmental jobs while it might be used differently by users in technology occupations (e.g., Internet-based computing).
Specifically, this paper makes four main contributions. First, we build a large, industry-annotated dataset that contains over 20,000 blog users. In addition to their posted text, we also link a number of user metadata including their gender, location, occupation, introduction and interests.
Second, we build content-based classifiers for the industry prediction task and study the effect of incorporating textual features from the users' profile metadata using various meta-classification techniques, significantly improving both the overall accuracy and the average per industry accuracy.
Next, after examining which words are indicative for each industry, we build vector-space representations of word meanings and calculate one deviation for each industry, illustrating how meaning is differentiated based on the users' industries. We qualitatively examine the resulting industry-informed semantic representations of words by listing the words per industry that are most similar to job related and general interest terms.
Finally, we rank the different industries based on the normalized relative frequencies of emotionally charged words (positive and negative) and, in addition, discover that, for both genders, these frequencies do not statistically significantly correlate with an industry's gender dominance ratio.
After discussing related work in Section SECREF2 , we present the dataset used in this study in Section SECREF3 . In Section SECREF4 we evaluate two feature selection methods and examine the industry inference problem using the text of the users' postings. We then augment our content-based classifier by building an ensemble that incorporates several metadata classifiers. We list the most industry indicative words and expose how each industrial semantic field varies with respect to a variety of terms in Section SECREF5 . We explore how the frequencies of emotionally charged words in each gender correlate with the industries and their respective gender dominance ratio and, finally, conclude in Section SECREF6 .
## Related Work
Alongside the wide adoption of social media by the public, researchers have been leveraging the newly available data to create and refine models of users' behavior and profiling. There exists a myriad research that analyzes language in order to profile social media users. Some studies sought to characterize users' personality BIBREF9 , BIBREF10 , while others sequenced the expressed emotions BIBREF11 , studied mental disorders BIBREF12 , and the progression of health conditions BIBREF13 . At the same time, a number of researchers sought to predict the social media users' age and/or gender BIBREF14 , BIBREF15 , BIBREF16 , while others targeted and analyzed the ethnicity, nationality, and race of the users BIBREF17 , BIBREF18 , BIBREF19 . One of the profile fields that has drawn a great deal of attention is the location of a user. Among others, Hecht et al. Hecht11 predicted Twitter users' locations using machine learning on nationwide and state levels. Later, Han et al. Han14 identified location indicative words to predict the location of Twitter users down to the city level.
As a separate line of research, a number of studies have focused on discovering the political orientation of users BIBREF15 , BIBREF20 , BIBREF21 . Finally, Li et al. Li14a proposed a way to model major life events such as getting married, moving to a new place, or graduating. In a subsequent study, BIBREF22 described a weakly supervised information extraction method that was used in conjunction with social network information to identify the name of a user's spouse, the college they attended, and the company where they are employed.
The line of work that is most closely related to our research is the one concerned with understanding the relation between people's language and their industry. Previous research from the fields of psychology and economics have explored the potential for predicting one's occupation from their ability to use math and verbal symbols BIBREF23 and the relationship between job-types and demographics BIBREF24 . More recently, Huang et al. Huang15 used machine learning to classify Sina Weibo users to twelve different platform-defined occupational classes highlighting the effect of homophily in user interactions. This work examined only users that have been verified by the Sina Weibo platform, introducing a potential bias in the resulting dataset. Finally, Preotiuc-Pietro et al. Preoctiuc15 predicted the occupational class of Twitter users using the Standard Occupational Classification (SOC) system, which groups the different jobs based on skill requirements. In that work, the data collection process was limited to only users that specifically mentioned their occupation in their self-description in a way that could be directly mapped to a SOC occupational class. The mapping between a substring of their self-description and a SOC occupational class was done manually. Because of the manual annotation step, their method was not scalable; moreover, because they identified the occupation class inside a user self-description, only a very small fraction of the Twitter users could be included (in their case, 5,191 users).
Both of these recent studies are based on micro-blogging platforms, which inherently restrict the number of characters that a post can have, and consequently the way that users can express themselves.
Moreover, both studies used off-the-shelf occupational taxonomies (rather than self-declared occupation categories), resulting in classes that are either too generic (e.g., media, welfare and electronic are three of the twelve Sina Weibo categories), or too intermixed (e.g., an assistant accountant is in a different class from an accountant in SOC). To address these limitations, we investigate the industry prediction task in a large blog corpus consisting of over 20K American users, 40K web-blogs, and 560K blog posts.
## Dataset
We compile our industry-annotated dataset by identifying blogger profiles located in the U.S. on the profile finder on http://www.blogger.com, and scraping only those users that had the industry profile element completed.
For each of these bloggers, we retrieve all their blogs, and for each of these blogs we download the 21 most recent blog postings. We then clean these blog posts of HTML tags and tokenize them, and drop those bloggers whose cumulative textual content in their posts is less than 600 characters. Following these guidelines, we identified all the U.S. bloggers with completed industry information.
Traditionally, standardized industry taxonomies organize economic activities into groups based on similar production processes, products or services, delivery systems or behavior in financial markets. Following such assumptions and regardless of their many similarities, a tomato farmer would be categorized into a distinct industry from a tobacco farmer. As demonstrated in Preotiuc-Pietro et al. Preoctiuc15 such groupings can cause unwarranted misclassifications.
The Blogger platform provides a total of 39 different industry options. Even though a completed industry value is an implicit text annotation, we acknowledge the same problem noted in previous studies: some categories are too broad, while others are very similar. To remedy this and following Guibert et al. Guibert71, who argued that the denominations used in a classification must reflect the purpose of the study, we group the different Blogger industries based on similar educational background and similar technical terminology. To do that, we exclude very general categories and merge conceptually similar ones. Examples of broad categories are the Education and the Student options: a teacher could be teaching in any concentration, while a student could be enrolled in any discipline. Examples of conceptually similar categories are the Investment Banking and the Banking options.
The final set of categories is shown in Table TABREF1 , along with the number of users in each category. The resulting dataset consists of 22,880 users, 41,094 blogs, and 561,003 posts. Table TABREF2 presents additional statistics of our dataset.
## Text-based Industry Modeling
After collecting our dataset, we split it into three sets: a train set, a development set, and a test set. The sizes of these sets are 17,880, 2,500, and 2,500 users, respectively, with users randomly assigned to these sets. In all the experiments that follow, we evaluate our classifiers by training them on the train set, configure the parameters and measure performance on the development set, and finally report the prediction accuracy and results on the test set. Note that all the experiments are performed at user level, i.e., all the data for one user is compiled into one instance in our data sets.
To measure the performance of our classifiers, we use the prediction accuracy. However, as shown in Table TABREF1 , the available data is skewed across categories, which could lead to somewhat distorted accuracy numbers depending on how well a model learns to predict the most populous classes. Moreover, accuracy alone does not provide a great deal of insight into the individual performance per industry, which is one of the main objectives in this study. Therefore, in our results below, we report: (1) micro-accuracy ( INLINEFORM0 ), calculated as the percentage of correctly classified instances out of all the instances in the development (test) data; and (2) macro-accuracy ( INLINEFORM1 ), calculated as the average of the per-category accuracies, where the per-category accuracy is the percentage of correctly classified instances out of the instances belonging to one category in the development (test) data.
## Leveraging Blog Content
In this section, we seek the effectiveness of using solely textual features obtained from the users' postings to predict their industry.
The industry prediction baseline Majority is set by discovering the most frequently featured class in our training set and picking that class in all predictions in the respective development or testing set.
After excluding all the words that are not used by at least three separate users in our training set, we build our AllWords model by counting the frequencies of all the remaining words and training a multinomial Naive Bayes classifier. As seen in Figure FIGREF3 , we can far exceed the Majority baseline performance by incorporating basic language signals into machine learning algorithms (173% INLINEFORM0 improvement).
We additionally explore the potential of improving our text classification task by applying a number of feature ranking methods and selecting varying proportions of top ranked features in an attempt to exclude noisy features. We start by ranking the different features, w, according to their Information Gain Ratio score (IGR) with respect to every industry, i, and training our classifier using different proportions of the top features. INLINEFORM0 INLINEFORM1
Even though we find that using the top 95% of all the features already exceeds the performance of the All Words model on the development data, we further experiment with ranking our features with a more aggressive formula that heavily promotes the features that are tightly associated with any industry category. Therefore, for every word in our training set, we define our newly introduced ranking method, the Aggressive Feature Ranking (AFR), as: INLINEFORM0
In Figure FIGREF3 we illustrate the performance of all four methods in our industry prediction task on the development data. Note that for each method, we provide both the accuracy ( INLINEFORM0 ) and the average per-class accuracy ( INLINEFORM1 ). The Majority and All Words methods apply to all the features; therefore, they are represented as a straight line in the figure. The IGR and AFR methods are applied to varying subsets of the features using a 5% step.
Our experiments demonstrate that the word choice that the users make in their posts correlates with their industry. The first observation in Figure FIGREF3 is that the INLINEFORM0 is proportional to INLINEFORM1 ; as INLINEFORM2 increases, so does INLINEFORM3 . Secondly, the best result on the development set is achieved by using the top 90% of the features using the AFR method. Lastly, the improvements of the IGR and AFR feature selections are not substantially better in comparison to All Words (at most 5% improvement between All Words and AFR), which suggest that only a few noisy features exist and most of the words play some role in shaping the “language" of an industry.
As a final evaluation, we apply on the test data the classifier found to work best on the development data (AFR feature selection, top 90% features), for an INLINEFORM0 of 0.534 and INLINEFORM1 of 0.477.
## Leveraging User Metadata
Together with the industry information and the most recent postings of each blogger, we also download a number of accompanying profile elements. Using these additional elements, we explore the potential of incorporating users' metadata in our classifiers.
Table TABREF7 shows the different user metadata we consider together with their coverage percentage (not all users provide a value for all of the profile elements). With the exception of the gender field, the remaining metadata elements shown in Table TABREF7 are completed by the users as a freely editable text field. This introduces a considerable amount of noise in the set of possible metadata values. Examples of noise in the occupation field include values such as “Retired”, “I work.”, or “momma” which are not necessarily informative for our industry prediction task.
To examine whether the metadata fields can help in the prediction of a user's industry, we build classifiers using the different metadata elements. For each metadata element that has a textual value, we use all the words in the training set for that field as features. The only two exceptions are the state field, which is encoded as one feature that can take one out of 50 different values representing the 50 U.S. states; and the gender field, which is encoded as a feature with a distinct value for each user gender option: undefined, male, or female.
As shown in Table TABREF9 , we build four different classifiers using the multinomial NB algorithm: Occu (which uses the words found in the occupation profile element), Intro (introduction), Inter (interests), and Gloc (combined gender, city, state).
In general, all the metadata classifiers perform better than our majority baseline ( INLINEFORM0 of 18.88%). For the Gloc classifier, this result is in alignment with previous studies BIBREF24 . However, the only metadata classifier that outperforms the content classifier is the Occu classifier, which despite missing and noisy occupation values exceeds the content classifier's performance by an absolute 3.2%.
To investigate the promise of combining the five different classifiers we have built so far, we calculate their inter-prediction agreement using Fleiss's Kappa BIBREF25 , as well as the lower prediction bounds using the double fault measure BIBREF26 . The Kappa values, presented in the lower left side of Table TABREF10 , express the classification agreement for categorical items, in this case the users' industry. Lower values, especially values below 30%, mean smaller agreement. Since all five classifiers have better-than-baseline accuracy, this low agreement suggests that their predictions could potentially be combined to achieve a better accumulated result.
Moreover, the double fault measure values, which are presented in the top-right hand side of Table TABREF10 , express the proportion of test cases for which both of the two respective classifiers make false predictions, essentially providing the lowest error bound for the pairwise ensemble classifier performance. The lower those numbers are, the greater the accuracy potential of any meta-classification scheme that combines those classifiers. Once again, the low double fault measure values suggest potential gain from a combination of the base classifiers into an ensemble of models.
After establishing the promise of creating an ensemble of classifiers, we implement two meta-classification approaches. First, we combine our classifiers using features concatenation (or early fusion). Starting with our content-based classifier (Text), we successively add the features derived from each metadata element. The results, both micro- and macro-accuracy, are presented in Table TABREF12 . Even though all these four feature concatenation ensembles outperform the content-based classifier in the development set, they fail to outperform the Occu classifier.
Second, we explore the potential of using stacked generalization (or late fusion) BIBREF27 . The base classifiers, referred to as L0 classifiers, are trained on different folds of the training set and used to predict the class of the remaining instances. Those predictions are then used together with the true label of the training instances to train a second classifier, referred to as the L1 classifier: this L1 is used to produce the final prediction on both the development data and the test data. Traditionally, stacking uses different machine learning algorithms on the same training data. However in our case, we use the same algorithm (multinomial NB) on heterogeneous data (i.e., different types of data such as content, occupation, introduction, interests, gender, city and state) in order to exploit all available sources of information.
The ensemble learning results on the development set are shown in Table TABREF12 . We notice a constant improvement for both metrics when adding more classifiers to our ensemble except for the Gloc classifier, which slightly reduces the performance. The best result is achieved using an ensemble of the Text, Occu, Intro, and Inter L0 classifiers; the respective performance on the test set is an INLINEFORM0 of 0.643 and an INLINEFORM1 of 0.564. Finally, we present in Figure FIGREF11 the prediction accuracy for the final classifier for each of the different industries in our test dataset. Evidently, some industries are easier to predict than others. For example, while the Real Estate and Religion industries achieve accuracy figures above 80%, other industries, such as the Banking industry, are predicted correctly in less than 17% of the time. Anecdotal evidence drawn from the examination of the confusion matrix does not encourage any strong association of the Banking class with any other. The misclassifications are roughly uniform across all other classes, suggesting that the users in the Banking industry use language in a non-distinguishing way.
## Qualitative Analysis
In this section, we provide a qualitative analysis of the language of the different industries.
## Top-Ranked Words
To conduct a qualitative exploration of which words indicate the industry of a user, Table TABREF14 shows the three top-ranking content words for the different industries using the AFR method.
Not surprisingly, the top ranked words align well with what we would intuitively expect for each industry. Even though most of these words are potentially used by many users regardless of their industry in our dataset, they are still distinguished by the AFR method because of the different frequencies of these words in the text of each industry.
## Industry-specific Word Similarities
Next, we examine how the meaning of a word is shaped by the context in which it is uttered. In particular, we qualitatively investigate how the speakers' industry affects meaning by learning vector-space representations of words that take into account such contextual information. To achieve this, we apply the contextualized word embeddings proposed by Bamman et al. Bamman14, which are based on an extension of the “skip-gram" language model BIBREF28 .
In addition to learning a global representation for each word, these contextualized embeddings compute one deviation from the common word embedding representation for each contextual variable, in this case, an industry option. These deviations capture the terms' meaning variations (shifts in the INLINEFORM0 -dimensional space of the representations, where INLINEFORM1 in our experiments) in the text of the different industries, however all the embeddings are in the same vector space to allow for comparisons to one another.
Using the word representations learned for each industry, we present in Table TABREF16 the terms in the Technology and the Tourism industries that have the highest cosine similarity with a job-related word, customers. Similarly, Table TABREF17 shows the words in the Environment and the Tourism industries that are closest in meaning to a general interest word, food. More examples are given in the Appendix SECREF8 .
The terms that rank highest in each industry are noticeably different. For example, as seen in Table TABREF17 , while food in the Environment industry is similar to nutritionally and locally, in the Tourism industry the same word relates more to terms such as delicious and pastries. These results not only emphasize the existing differences in how people in different industries perceive certain terms, but they also demonstrate that those differences can effectively be captured in the resulting word embeddings.
## Emotional Orientation per Industry and Gender
As a final analysis, we explore how words that are emotionally charged relate to different industries. To quantify the emotional orientation of a text, we use the Positive Emotion and Negative Emotion categories in the Linguistic Inquiry and Word Count (LIWC) dictionary BIBREF29 . The LIWC dictionary contains lists of words that have been shown to correlate with the psychological states of people that use them; for example, the Positive Emotion category contains words such as “happy,” “pretty,” and “good.”
For the text of all the users in each industry we measure the frequencies of Positive Emotion and Negative Emotion words normalized by the text's length. Table TABREF20 presents the industries' ranking for both categories of words based on their relative frequencies in the text of each industry.
We further perform a breakdown per-gender, where we once again calculate the proportion of emotionally charged words in each industry, but separately for each gender. We find that the industry rankings of the relative frequencies INLINEFORM0 of emotionally charged words for the two genders are statistically significantly correlated, which suggests that regardless of their gender, users use positive (or negative) words with a relative frequency that correlates with their industry. (In other words, even if e.g., Fashion has a larger number of women users, both men and women working in Fashion will tend to use more positive words than the corresponding gender in another industry with a larger number of men users such as Automotive.)
Finally, motivated by previous findings of correlations between job satisfaction and gender dominance in the workplace BIBREF30 , we explore the relationship between the usage of Positive Emotion and Negative Emotion words and the gender dominance in an industry. Although we find that there are substantial gender imbalances in each industry (Appendix SECREF9 ), we did not find any statistically significant correlation between the gender dominance ratio in the different industries and the usage of positive (or negative) emotional words in either gender in our dataset.
## Conclusion
In this paper, we examined the task of predicting a social media user's industry. We introduced an annotated dataset of over 20,000 blog users and applied a content-based classifier in conjunction with two feature selection methods for an overall accuracy of up to 0.534, which represents a large improvement over the majority class baseline of 0.188.
We also demonstrated how the user metadata can be incorporated in our classifiers. Although concatenation of features drawn both from blog content and profile elements did not yield any clear improvements over the best individual classifiers, we found that stacking improves the prediction accuracy to an overall accuracy of 0.643, as measured on our test dataset. A more in-depth analysis showed that not all industries are equally easy to predict: while industries such as Real Estate and Religion are clearly distinguishable with accuracy figures over 0.80, others such as Banking are much harder to predict.
Finally, we presented a qualitative analysis to provide some insights into the language of different industries, which highlighted differences in the top-ranked words in each industry, word semantic similarities, and the relative frequency of emotionally charged words.
## Acknowledgments
This material is based in part upon work supported by the National Science Foundation (#1344257) and by the John Templeton Foundation (#48503). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or the John Templeton Foundation.
## Additional Examples of Word Similarities
| 13
|
1701.00185
|
Self-Taught Convolutional Neural Networks for Short Text Clustering
|
# Self-Taught Convolutional Neural Networks for Short Text Clustering
## Abstract
Short text clustering is a challenging problem due to its sparseness of text representation. Here we propose a flexible Self-Taught Convolutional neural network framework for Short Text Clustering (dubbed STC^2), which can flexibly and successfully incorporate more useful semantic features and learn non-biased deep text representation in an unsupervised manner. In our framework, the original raw text features are firstly embedded into compact binary codes by using one existing unsupervised dimensionality reduction methods. Then, word embeddings are explored and fed into convolutional neural networks to learn deep feature representations, meanwhile the output units are used to fit the pre-trained binary codes in the training process. Finally, we get the optimal clusters by employing K-means to cluster the learned representations. Extensive experimental results demonstrate that the proposed framework is effective, flexible and outperform several popular clustering methods when tested on three public short text datasets.
## Introduction
Short text clustering is of great importance due to its various applications, such as user profiling BIBREF0 and recommendation BIBREF1 , for nowaday's social media dataset emerged day by day. However, short text clustering has the data sparsity problem and most words only occur once in each short text BIBREF2 . As a result, the Term Frequency-Inverse Document Frequency (TF-IDF) measure cannot work well in short text setting. In order to address this problem, some researchers work on expanding and enriching the context of data from Wikipedia BIBREF3 or an ontology BIBREF4 . However, these methods involve solid Natural Language Processing (NLP) knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another way to overcome these issues is to explore some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Yet how to design an effective model is an open question, and most of these methods directly trained based on Bag-of-Words (BoW) are shallow structures which cannot preserve the accurate semantic similarities.
Recently, with the help of word embedding, neural networks demonstrate their great performance in terms of constructing text representation, such as Recursive Neural Network (RecNN) BIBREF6 , BIBREF7 and Recurrent Neural Network (RNN) BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the hidden layer computed at the last word to represent the text, is a biased model where later words are more dominant than earlier words BIBREF9 . Whereas for the non-biased models, the learned representation of one text can be extracted from all the words in the text with non-dominant learned weights. More recently, Convolution Neural Network (CNN), as the most popular non-biased model and applying convolutional filters to capture local features, has achieved a better performance in many NLP applications, such as sentence modeling BIBREF10 , relation classification BIBREF11 , and other traditional NLP tasks BIBREF12 . Most of the previous works focus CNN on solving supervised NLP tasks, while in this paper we aim to explore the power of CNN on one unsupervised NLP task, short text clustering.
We systematically introduce a simple yet surprisingly powerful Self-Taught Convolutional neural network framework for Short Text Clustering, called STC INLINEFORM0 . An overall architecture of our proposed approach is illustrated in Figure FIGREF5 . We, inspired by BIBREF13 , BIBREF14 , utilize a self-taught learning framework into our task. In particular, the original raw text features are first embedded into compact binary codes INLINEFORM1 with the help of one traditional unsupervised dimensionality reduction function. Then text matrix INLINEFORM2 projected from word embeddings are fed into CNN model to learn the deep feature representation INLINEFORM3 and the output units are used to fit the pre-trained binary codes INLINEFORM4 . After obtaining the learned features, K-means algorithm is employed on them to cluster texts into clusters INLINEFORM5 . Obviously, we call our approach “self-taught” because the CNN model is learnt from the pseudo labels generated from the previous stage, which is quite different from the term “self-taught” in BIBREF15 . Our main contributions can be summarized as follows:
This work is an extension of our conference paper BIBREF16 , and they differ in the following aspects. First, we put forward a general a self-taught CNN framework in this paper which can flexibly couple various semantic features, whereas the conference version can be seen as a specific example of this work. Second, in this paper we use a new short text dataset, Biomedical, in the experiment to verify the effectiveness of our approach. Third, we put much effort on studying the influence of various different semantic features integrated in our self-taught CNN framework, which is not involved in the conference paper.
For the purpose of reproducibility, we make the datasets and software used in our experiments publicly available at the website.
The remainder of this paper is organized as follows: In Section SECREF2 , we first briefly survey several related works. In Section SECREF3 , we describe the proposed approach STC INLINEFORM0 and implementation details. Experimental results and analyses are presented in Section SECREF4 . Finally, conclusions are given in the last Section.
## Related Work
In this section, we review the related work from the following two perspectives: short text clustering and deep neural networks.
## Short Text Clustering
There have been several studies that attempted to overcome the sparseness of short text representation. One way is to expand and enrich the context of data. For example, Banerjee et al. BIBREF3 proposed a method of improving the accuracy of short text clustering by enriching their representation with additional features from Wikipedia, and Fodeh et al. BIBREF4 incorporate semantic knowledge from an ontology into text clustering. However, these works need solid NLP knowledge and still use high-dimensional representation which may result in a waste of both memory and computation time. Another direction is to map the original features into reduced space, such as Latent Semantic Analysis (LSA) BIBREF17 , Laplacian Eigenmaps (LE) BIBREF18 , and Locality Preserving Indexing (LPI) BIBREF19 . Even some researchers explored some sophisticated models to cluster short texts. For example, Yin and Wang BIBREF5 proposed a Dirichlet multinomial mixture model-based approach for short text clustering. Moreover, some studies even focus the above both two streams. For example, Tang et al. BIBREF20 proposed a novel framework which enrich the text features by employing machine translation and reduce the original features simultaneously through matrix factorization techniques.
Despite the above clustering methods can alleviate sparseness of short text representation to some extent, most of them ignore word order in the text and belong to shallow structures which can not fully capture accurate semantic similarities.
## Deep Neural Networks
Recently, there is a revival of interest in DNN and many researchers have concentrated on using Deep Learning to learn features. Hinton and Salakhutdinov BIBREF21 use DAE to learn text representation. During the fine-tuning procedure, they use backpropagation to find codes that are good at reconstructing the word-count vector.
More recently, researchers propose to use external corpus to learn a distributed representation for each word, called word embedding BIBREF22 , to improve DNN performance on NLP tasks. The Skip-gram and continuous bag-of-words models of Word2vec BIBREF23 propose a simple single-layer architecture based on the inner product between two word vectors, and Pennington et al. BIBREF24 introduce a new model for word representation, called GloVe, which captures the global corpus statistics.
In order to learn the compact representation vectors of sentences, Le and Mikolov BIBREF25 directly extend the previous Word2vec BIBREF23 by predicting words in the sentence, which is named Paragraph Vector (Para2vec). Para2vec is still a shallow window-based method and need a larger corpus to yield better performance. More neural networks utilize word embedding to capture true meaningful syntactic and semantic regularities, such as RecNN BIBREF6 , BIBREF7 and RNN BIBREF8 . However, RecNN exhibits high time complexity to construct the textual tree, and RNN, using the layer computed at the last word to represent the text, is a biased model. Recently, Long Short-Term Memory (LSTM) BIBREF26 and Gated Recurrent Unit (GRU) BIBREF27 , as sophisticated recurrent hidden units of RNN, has presented its advantages in many sequence generation problem, such as machine translation BIBREF28 , speech recognition BIBREF29 , and text conversation BIBREF30 . While, CNN is better to learn non-biased implicit features which has been successfully exploited for many supervised NLP learning tasks as described in Section SECREF1 , and various CNN based variants are proposed in the recent works, such as Dynamic Convolutional Neural Network (DCNN) BIBREF10 , Gated Recursive Convolutional Neural Network (grConv) BIBREF31 and Self-Adaptive Hierarchical Sentence model (AdaSent) BIBREF32 .
In the past few days, Visin et al. BIBREF33 have attempted to replace convolutional layer in CNN to learn non-biased features for object recognition with four RNNs, called ReNet, that sweep over lower-layer features in different directions: (1) bottom to top, (2) top to bottom, (3) left to right and (4) right to left. However, ReNet does not outperform state-of-the-art convolutional neural networks on any of the three benchmark datasets, and it is also a supervised learning model for classification. Inspired by Skip-gram of word2vec BIBREF34 , BIBREF23 , Skip-thought model BIBREF35 describe an approach for unsupervised learning of a generic, distributed sentence encoder. Similar as Skip-gram model, Skip-thought model trains an encoder-decoder model that tries to reconstruct the surrounding sentences of an encoded sentence and released an off-the-shelf encoder to extract sentence representation. Even some researchers introduce continuous Skip-gram and negative sampling to CNN for learning visual representation in an unsupervised manner BIBREF36 . This paper, from a new perspective, puts forward a general self-taught CNN framework which can flexibly couple various semantic features and achieve a good performance on one unsupervised learning task, short text clustering.
## Methodology
Assume that we are given a dataset of INLINEFORM0 training texts denoted as: INLINEFORM1 , where INLINEFORM2 is the dimensionality of the original BoW representation. Denote its tag set as INLINEFORM3 and the pre-trained word embedding set as INLINEFORM4 , where INLINEFORM5 is the dimensionality of word vectors and INLINEFORM6 is the vocabulary size. In order to learn the INLINEFORM7 -dimensional deep feature representation INLINEFORM8 from CNN in an unsupervised manner, some unsupervised dimensionality reduction methods INLINEFORM9 are employed to guide the learning of CNN model. Our goal is to cluster these texts INLINEFORM10 into clusters INLINEFORM11 based on the learned deep feature representation while preserving the semantic consistency.
As depicted in Figure FIGREF5 , the proposed framework consist of three components, deep convolutional neural network (CNN), unsupervised dimensionality reduction function and K-means module. In the rest sections, we first present the first two components respectively, and then give the trainable parameters and the objective function to learn the deep feature representation. Finally, the last section describe how to perform clustering on the learned features.
## Deep Convolutional Neural Networks
In this section, we briefly review one popular deep convolutional neural network, Dynamic Convolutional Neural Network (DCNN) BIBREF10 as an instance of CNN in the following sections, which as the foundation of our proposed method has been successfully proposed for the completely supervised learning task, text classification.
Taking a neural network with two convolutional layers in Figure FIGREF9 as an example, the network transforms raw input text to a powerful representation. Particularly, each raw text vector INLINEFORM0 is projected into a matrix representation INLINEFORM1 by looking up a word embedding INLINEFORM2 , where INLINEFORM3 is the length of one text. We also let INLINEFORM4 and INLINEFORM5 denote the weights of the neural networks. The network defines a transformation INLINEFORM6 INLINEFORM7 which transforms an input raw text INLINEFORM8 to a INLINEFORM9 -dimensional deep representation INLINEFORM10 . There are three basic operations described as follows:
Wide one-dimensional convolution This operation INLINEFORM0 is applied to an individual row of the sentence matrix INLINEFORM1 , and yields a resulting matrix INLINEFORM2 , where INLINEFORM3 is the width of convolutional filter.
Folding In this operation, every two rows in a feature map are simply summed component-wisely. For a map of INLINEFORM0 rows, folding returns a map of INLINEFORM1 rows, thus halving the size of the representation and yielding a matrix feature INLINEFORM2 . Note that folding operation does not introduce any additional parameters.
Dynamic INLINEFORM0 -max pooling Assuming the pooling parameter as INLINEFORM1 , INLINEFORM2 -max pooling selects the sub-matrix INLINEFORM3 of the INLINEFORM4 highest values in each row of the matrix INLINEFORM5 . For dynamic INLINEFORM6 -max pooling, the pooling parameter INLINEFORM7 is dynamically selected in order to allow for a smooth extraction of higher-order and longer-range features BIBREF10 . Given a fixed pooling parameter INLINEFORM8 for the topmost convolutional layer, the parameter INLINEFORM9 of INLINEFORM10 -max pooling in the INLINEFORM11 -th convolutional layer can be computed as follows: DISPLAYFORM0
where INLINEFORM0 is the total number of convolutional layers in the network.
## Unsupervised Dimensionality Reduction
As described in Figure FIGREF5 , the dimensionality reduction function is defined as follows: DISPLAYFORM0
where, INLINEFORM0 are the INLINEFORM1 -dimensional reduced latent space representations. Here, we take four popular dimensionality reduction methods as examples in our framework.
Average Embedding (AE): This method directly averages the word embeddings which are respectively weighted with TF and TF-IDF. Huang et al. BIBREF37 used this strategy as the global context in their task, and Socher et al. BIBREF7 and Lai et al. BIBREF9 used this method for text classification. The weighted average of all word vectors in one text can be computed as follows: DISPLAYFORM0
where INLINEFORM0 can be any weighting function that captures the importance of word INLINEFORM1 in the text INLINEFORM2 .
Latent Semantic Analysis (LSA): LSA BIBREF17 is the most popular global matrix factorization method, which applies a dimension reducing linear projection, Singular Value Decomposition (SVD), of the corresponding term/document matrix. Suppose the rank of INLINEFORM0 is INLINEFORM1 , LSA decompose INLINEFORM2 into the product of three other matrices: DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are the singular values of INLINEFORM2 , INLINEFORM3 is a set of left singular vectors and INLINEFORM4 is a set of right singular vectors. LSA uses the top INLINEFORM5 vectors in INLINEFORM6 as the transformation matrix to embed the original text features into a INLINEFORM7 -dimensional subspace INLINEFORM8 BIBREF17 .
Laplacian Eigenmaps (LE): The top eigenvectors of graph Laplacian, defined on the similarity matrix of texts, are used in the method, which can discover the manifold structure of the text space BIBREF18 . In order to avoid storing the dense similarity matrix, many approximation techniques are proposed to reduce the memory usage and computational complexity for LE. There are two representative approximation methods, sparse similarity matrix and Nystr INLINEFORM0 m approximation. Following previous studies BIBREF38 , BIBREF13 , we select the former technique to construct the INLINEFORM1 local similarity matrix INLINEFORM2 by using heat kernel as follows: DISPLAYFORM0
where, INLINEFORM0 is a tuning parameter (default is 1) and INLINEFORM1 represents the set of INLINEFORM2 -nearest-neighbors of INLINEFORM3 . By introducing a diagonal INLINEFORM4 matrix INLINEFORM5 whose entries are given by INLINEFORM6 , the graph Laplacian INLINEFORM7 can be computed by ( INLINEFORM8 ). The optimal INLINEFORM9 real-valued matrix INLINEFORM10 can be obtained by solving the following objective function: DISPLAYFORM0
where INLINEFORM0 is the trace function, INLINEFORM1 requires the different dimensions to be uncorrelated, and INLINEFORM2 requires each dimension to achieve equal probability as positive or negative).
Locality Preserving Indexing (LPI): This method extends LE to deal with unseen texts by approximating the linear function INLINEFORM0 BIBREF13 , and the subspace vectors are obtained by finding the optimal linear approximations to the eigenfunctions of the Laplace Beltrami operator on the Riemannian manifold BIBREF19 . Similar as LE, we first construct the local similarity matrix INLINEFORM1 , then the graph Laplacian INLINEFORM2 can be computed by ( INLINEFORM3 ), where INLINEFORM4 measures the local density around INLINEFORM5 and is equal to INLINEFORM6 . Compute the eigenvectors INLINEFORM7 and eigenvalues INLINEFORM8 of the following generalized eigen-problem: DISPLAYFORM0
The mapping function INLINEFORM0 can be obtained and applied to the unseen data BIBREF38 .
All of the above methods claim a better performance in capturing semantic similarity between texts in the reduced latent space representation INLINEFORM0 than in the original representation INLINEFORM1 , while the performance of short text clustering can be further enhanced with the help of our framework, self-taught CNN.
## Learning
The last layer of CNN is an output layer as follows: DISPLAYFORM0
where, INLINEFORM0 is the deep feature representation, INLINEFORM1 is the output vector and INLINEFORM2 is weight matrix.
In order to incorporate the latent semantic features INLINEFORM0 , we first binary the real-valued vectors INLINEFORM1 to the binary codes INLINEFORM2 by setting the threshold to be the media vector INLINEFORM3 . Then, the output vector INLINEFORM4 is used to fit the binary codes INLINEFORM5 via INLINEFORM6 logistic operations as follows: DISPLAYFORM0
All parameters to be trained are defined as INLINEFORM0 . DISPLAYFORM0
Given the training text collection INLINEFORM0 , and the pre-trained binary codes INLINEFORM1 , the log likelihood of the parameters can be written down as follows: DISPLAYFORM0
Following the previous work BIBREF10 , we train the network with mini-batches by back-propagation and perform the gradient-based optimization using the Adagrad update rule BIBREF39 . For regularization, we employ dropout with 50% rate to the penultimate layer BIBREF10 , BIBREF40 .
## K-means for Clustering
With the given short texts, we first utilize the trained deep neural network to obtain the semantic representations INLINEFORM0 , and then employ traditional K-means algorithm to perform clustering.
## Datasets
We test our proposed approach on three public short text datasets. The summary statistics and semantic topics of these datasets are described in Table TABREF24 and Table TABREF25 .
SearchSnippets. This dataset was selected from the results of web search transaction using predefined phrases of 8 different domains by Phan et al. BIBREF41 .
StackOverflow. We use the challenge data published in Kaggle.com. The raw dataset consists 3,370,528 samples through July 31st, 2012 to August 14, 2012. In our experiments, we randomly select 20,000 question titles from 20 different tags as in Table TABREF25 .
Biomedical. We use the challenge data published in BioASQ's official website. In our experiments, we randomly select 20, 000 paper titles from 20 different MeSH major topics as in Table TABREF25 . As described in Table TABREF24 , the max length of selected paper titles is 53.
For these datasets, we randomly select 10% of data as the development set. Since SearchSnippets has been pre-processed by Phan et al. BIBREF41 , we do not further process this dataset. In StackOverflow, texts contain lots of computer terminology, and symbols and capital letters are meaningful, thus we do not do any pre-processed procedures. For Biomedical, we remove the symbols and convert letters into lower case.
## Pre-trained Word Vectors
We use the publicly available word2vec tool to train word embeddings, and the most parameters are set as same as Mikolov et al. BIBREF23 to train word vectors on Google News setting, except of vector dimensionality using 48 and minimize count using 5. For SearchSnippets, we train word vectors on Wikipedia dumps. For StackOverflow, we train word vectors on the whole corpus of the StackOverflow dataset described above which includes the question titles and post contents. For Biomedical, we train word vectors on all titles and abstracts of 2014 training articles. The coverage of these learned vectors on three datasets are listed in Table TABREF32 , and the words not present in the set of pre-trained words are initialized randomly.
## Comparisons
In our experiment, some widely used text clustering methods are compared with our approach. Besides K-means, Skip-thought Vectors, Recursive Neural Network and Paragraph Vector based clustering methods, four baseline clustering methods are directly based on the popular unsupervised dimensionality reduction methods as described in Section SECREF11 . We further compare our approach with some other non-biased neural networks, such as bidirectional RNN. More details are listed as follows:
K-means K-means BIBREF42 on original keyword features which are respectively weighted with term frequency (TF) and term frequency-inverse document frequency (TF-IDF).
Skip-thought Vectors (SkipVec) This baseline BIBREF35 gives an off-the-shelf encoder to produce highly generic sentence representations. The encoder is trained using a large collection of novels and provides three encoder modes, that are unidirectional encoder (SkipVec (Uni)) with 2,400 dimensions, bidirectional encoder (SkipVec (Bi)) with 2,400 dimensions and combined encoder (SkipVec (Combine)) with SkipVec (Uni) and SkipVec (Bi) of 2,400 dimensions each. K-means is employed on the these vector representations respectively.
Recursive Neural Network (RecNN) In BIBREF6 , the tree structure is firstly greedy approximated via unsupervised recursive autoencoder. Then, semi-supervised recursive autoencoders are used to capture the semantics of texts based on the predicted structure. In order to make this recursive-based method completely unsupervised, we remove the cross-entropy error in the second phrase to learn vector representation and subsequently employ K-means on the learned vectors of the top tree node and the average of all vectors in the tree.
Paragraph Vector (Para2vec) K-means on the fixed size feature vectors generated by Paragraph Vector (Para2vec) BIBREF25 which is an unsupervised method to learn distributed representation of words and paragraphs. In our experiments, we use the open source software released by Mesnil et al. BIBREF43 .
Average Embedding (AE) K-means on the weighted average vectors of the word embeddings which are respectively weighted with TF and TF-IDF. The dimension of average vectors is equal to and decided by the dimension of word vectors used in our experiments.
Latent Semantic Analysis (LSA) K-means on the reduced subspace vectors generated by Singular Value Decomposition (SVD) method. The dimension of subspace is default set to the number of clusters, we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 10 on SearchSnippets, 20 on StackOverflow and 20 on Biomedical in our experiments.
Laplacian Eigenmaps (LE) This baseline, using Laplacian Eigenmaps and subsequently employing K-means algorithm, is well known as spectral clustering BIBREF44 . The dimension of subspace is default set to the number of clusters BIBREF18 , BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 70 on StackOverflow and 30 on Biomedical in our experiments.
Locality Preserving Indexing (LPI) This baseline, projecting the texts into a lower dimensional semantic space, can discover both the geometric and discriminating structures of the original feature space BIBREF38 . The dimension of subspace is default set to the number of clusters BIBREF38 , we also iterate the dimensions ranging from 10:10:200 to get the best performance, that is 20 on SearchSnippets, 80 on StackOverflow and 30 on Biomedical in our experiments.
bidirectional RNN (bi-RNN) We replace the CNN model in our framework as in Figure FIGREF5 with some bi-RNN models. Particularly, LSTM and GRU units are used in the experiments. In order to generate the fixed-length document representation from the variable-length vector sequences, for both bi-LSTM and bi-GRU based clustering methods, we further utilize three pooling methods: last pooling (using the last hidden state), mean pooling and element-wise max pooling. These pooling methods are respectively used in the previous works BIBREF45 , BIBREF27 , BIBREF46 and BIBREF9 . For regularization, the training gradients of all parameters with an INLINEFORM0 2 norm larger than 40 are clipped to 40, as the previous work BIBREF47 .
## Evaluation Metrics
The clustering performance is evaluated by comparing the clustering results of texts with the tags/labels provided by the text corpus. Two metrics, the accuracy (ACC) and the normalized mutual information metric (NMI), are used to measure the clustering performance BIBREF38 , BIBREF48 . Given a text INLINEFORM0 , let INLINEFORM1 and INLINEFORM2 be the obtained cluster label and the label provided by the corpus, respectively. Accuracy is defined as: DISPLAYFORM0
where, INLINEFORM0 is the total number of texts, INLINEFORM1 is the indicator function that equals one if INLINEFORM2 and equals zero otherwise, and INLINEFORM3 is the permutation mapping function that maps each cluster label INLINEFORM4 to the equivalent label from the text data by Hungarian algorithm BIBREF49 .
Normalized mutual information BIBREF50 between tag/label set INLINEFORM0 and cluster set INLINEFORM1 is a popular metric used for evaluating clustering tasks. It is defined as follows: DISPLAYFORM0
where, INLINEFORM0 is the mutual information between INLINEFORM1 and INLINEFORM2 , INLINEFORM3 is entropy and the denominator INLINEFORM4 is used for normalizing the mutual information to be in the range of [0, 1].
## Hyperparameter Settings
The most of parameters are set uniformly for these datasets. Following previous study BIBREF38 , the number of nearest neighbors in Eqn. ( EQREF15 ) is fixed to 15 when constructing the graph structures for LE and LPI. For CNN model, the networks has two convolutional layers. The widths of the convolutional filters are both 3. The value of INLINEFORM0 for the top INLINEFORM1 -max pooling in Eqn. ( EQREF10 ) is 5. The number of feature maps at the first convolutional layer is 12, and 8 feature maps at the second convolutional layer. Both those two convolutional layers are followed by a folding layer. We further set the dimension of word embeddings INLINEFORM2 as 48. Finally, the dimension of the deep feature representation INLINEFORM3 is fixed to 480. Moreover, we set the learning rate INLINEFORM4 as 0.01 and the mini-batch training size as 200. The output size INLINEFORM5 in Eqn. ( EQREF19 ) is set same as the best dimensions of subspace in the baseline method, as described in Section SECREF37 .
For initial centroids have significant impact on clustering results when utilizing the K-means algorithms, we repeat K-means for multiple times with random initial centroids (specifically, 100 times for statistical significance) as Huang BIBREF48 . The all subspace vectors are normalized to 1 before applying K-means and the final results reported are the average of 5 trials with all clustering methods on three text datasets.
## Results and Analysis
In Table TABREF43 and Table TABREF44 , we report the ACC and NMI performance of our proposed approaches and four baseline methods, K-means, SkipVec, RecNN and Para2vec based clustering methods. Intuitively, we get a general observation that (1) BoW based approaches, including K-means (TF) and K-means (TF-IDF), and SkipVec based approaches perform not well; (2) RecNN based approaches, both RecNN (Ave.) and RecNN (Top+Ave.), do better; (3) Para2vec makes a comparable performance with the most baselines; and (4) the evaluation clearly demonstrate the superiority of our proposed methods STC INLINEFORM0 . It is an expected results. For SkipVec based approaches, the off-the-shelf encoders are trained on the BookCorpus datasets BIBREF51 , and then applied to our datasets to extract the sentence representations. The SkipVec encoders can produce generic sentence representations but may not perform well for specific datasets, in our experiments, StackOverflow and Biomedical datasets consist of many computer terms and medical terms, such as “ASP.NET”, “XML”, “C#”, “serum” and “glycolytic”. When we take a more careful look, we find that RecNN (Top) does poorly, even worse than K-means (TF-IDF). The reason maybe that although recursive neural models introduce tree structure to capture compositional semantics, the vector of the top node mainly captures a biased semantic while the average of all vectors in the tree nodes, such as RecNN (Ave.), can be better to represent sentence level semantic. And we also get another observation that, although our proposed STC INLINEFORM1 -LE and STC INLINEFORM2 -LPI outperform both BoW based and RecNN based approaches across all three datasets, STC INLINEFORM3 -AE and STC INLINEFORM4 -LSA do just exhibit some similar performances as RecNN (Ave.) and RecNN (Top+Ave.) do in the datasets of StackOverflow and Biomedical.
We further replace the CNN model in our framework as in Figure FIGREF5 with some other non-biased models, such as bi-LSTM and bi-GRU, and report the results in Table TABREF46 and Table TABREF47 . As an instance, the binary codes generated from LPI are used to guide the learning of bi-LSTM/bi-GRU models. From the results, we can see that bi-GRU and bi-LSTM based clustering methods do equally well, no clear winner, and both achieve great enhancements compared with LPI (best). Compared with these bi-LSTM/bi-GRU based models, the evaluation results still demonstrate the superiority of our approach methods, CNN based clustering model, in the most cases. As the results reported by Visin et al. BIBREF33 , despite bi-directional or multi-directional RNN models perform a good non-biased feature extraction, they yet do not outperform state-of-the-art CNN on some tasks.
In order to make clear what factors make our proposed method work, we report the bar chart results of ACC and MNI of our proposed methods and the corresponding baseline methods in Figure FIGREF49 and Figure FIGREF53 . It is clear that, although AE and LSA does well or even better than LE and LPI, especially in dataset of both StackOverflow and Biomedical, STC INLINEFORM0 -LE and STC INLINEFORM1 -LPI achieve a much larger performance enhancements than STC INLINEFORM2 -AE and STC INLINEFORM3 -LSA do. The possible reason is that the information the pseudo supervision used to guide the learning of CNN model that make difference. Especially, for AE case, the input features fed into CNN model and the pseudo supervision employed to guide the learning of CNN model are all come from word embeddings. There are no different semantic features to be used into our proposed method, thus the performance enhancements are limited in STC INLINEFORM4 -AE. For LSA case, as we known, LSA is to make matrix factorization to find the best subspace approximation of the original feature space to minimize the global reconstruction error. And as BIBREF24 , BIBREF52 recently point out that word embeddings trained with word2vec or some variances, is essentially to do an operation of matrix factorization. Therefore, the information between input and the pseudo supervision in CNN is not departed very largely from each other, and the performance enhancements of STC INLINEFORM5 -AE is also not quite satisfactory. For LE and LPI case, as we known that LE extracts the manifold structure of the original feature space, and LPI extracts both geometric and discriminating structure of the original feature space BIBREF38 . We guess that our approach STC INLINEFORM6 -LE and STC INLINEFORM7 -LPI achieve enhancements compared with both LE and LPI by a large margin, because both of LE and LPI get useful semantic features, and these features are also different from word embeddings used as input of CNN. From this view, we say that our proposed STC has potential to behave more effective when the pseudo supervision is able to get semantic meaningful features, which is different enough from the input of CNN.
Furthermore, from the results of K-means and AE in Table TABREF43 - TABREF44 and Figure FIGREF49 - FIGREF53 , we note that TF-IDF weighting gives a more remarkable improvement for K-means, while TF weighting works better than TF-IDF weighting for Average Embedding. Maybe the reason is that pre-trained word embeddings encode some useful information from external corpus and are able to get even better results without TF-IDF weighting. Meanwhile, we find that LE get quite unusual good performance than LPI, LSA and AE in SearchSnippets dataset, which is not found in the other two datasets. To get clear about this, and also to make a much better demonstration about our proposed approaches and other baselines, we further report 2-dimensional text embeddings on SearchSnippets in Figure FIGREF58 , using t-SNE BIBREF53 to get distributed stochastic neighbor embedding of the feature representations used in the clustering methods. We can see that the results of from AE and LSA seem to be fairly good or even better than the ones from LE and LPI, which is not the same as the results from ACC and NMI in Figure FIGREF49 - FIGREF53 . Meanwhile, RecNN (Ave.) performs better than BoW (both TF and TF-IDF) while RecNN (Top) does not, which is the same as the results from ACC and NMI in Table TABREF43 and Table TABREF44 . Then we guess that both ”the same as” and ”not the same as” above, is just a good example to illustrate that visualization tool, such as t-SNE, get some useful information for measuring results, which is different from the ones of ACC and NMI. Moreover, from this complementary view of t-SNE, we can see that our STC INLINEFORM0 -AE, STC INLINEFORM1 -LSA, STC INLINEFORM2 -LE, and STC INLINEFORM3 -LPI show more clear-cut margins among different semantic topics (that is, tags/labels), compared with AE, LSA, LE and LPI, respectively, as well as compared with both baselines, BoW and RecNN based ones.
From all these results, with three measures of ACC, NMI and t-SNE under three datasets, we can get a solid conclusion that our proposed approaches is an effective approaches to get useful semantic features for short text clustering.
## Conclusions
With the emergence of social media, short text clustering has become an increasing important task. This paper explores a new perspective to cluster short texts based on deep feature representation learned from the proposed self-taught convolutional neural networks. Our framework can be successfully accomplished without using any external tags/labels and complicated NLP pre-processing, and and our approach is a flexible framework, in which the traditional dimension reduction approaches could be used to get performance enhancement. Our extensive experimental study on three short text datasets shows that our approach can achieve a significantly better performance. In the future, how to select and incorporate more effective semantic features into the proposed framework would call for more research.
## Acknowledgments
We would like to thank reviewers for their comments, and acknowledge Kaggle and BioASQ for making the datasets available. This work is supported by the National Natural Science Foundation of China (No. 61602479, No. 61303172, No. 61403385) and the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB02070005).
| 17
|
1701.06538
|
Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
|
# Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
## Abstract
The capacity of a neural network to absorb information is limited by its number of parameters. Conditional computation, where parts of the network are active on a per-example basis, has been proposed in theory as a way of dramatically increasing model capacity without a proportional increase in computation. In practice, however, there are significant algorithmic and performance challenges. In this work, we address these challenges and finally realize the promise of conditional computation, achieving greater than 1000x improvements in model capacity with only minor losses in computational efficiency on modern GPU clusters. We introduce a Sparsely-Gated Mixture-of-Experts layer (MoE), consisting of up to thousands of feed-forward sub-networks. A trainable gating network determines a sparse combination of these experts to use for each example. We apply the MoE to the tasks of language modeling and machine translation, where model capacity is critical for absorbing the vast quantities of knowledge available in the training corpora. We present model architectures in which a MoE with up to 137 billion parameters is applied convolutionally between stacked LSTM layers. On large language modeling and machine translation benchmarks, these models achieve significantly better results than state-of-the-art at lower computational cost.
## Conditional Computation
Exploiting scale in both training data and model size has been central to the success of deep learning. When datasets are sufficiently large, increasing the capacity (number of parameters) of neural networks can give much better prediction accuracy. This has been shown in domains such as text BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , images BIBREF4 , BIBREF5 , and audio BIBREF6 , BIBREF7 . For typical deep learning models, where the entire model is activated for every example, this leads to a roughly quadratic blow-up in training costs, as both the model size and the number of training examples increase. Unfortunately, the advances in computing power and distributed computation fall short of meeting such demand.
Various forms of conditional computation have been proposed as a way to increase model capacity without a proportional increase in computational costs BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 . In these schemes, large parts of a network are active or inactive on a per-example basis. The gating decisions may be binary or sparse and continuous, stochastic or deterministic. Various forms of reinforcement learning and back-propagation are proposed for trarining the gating decisions.
While these ideas are promising in theory, no work to date has yet demonstrated massive improvements in model capacity, training time, or model quality. We blame this on a combination of the following challenges:
Modern computing devices, especially GPUs, are much faster at arithmetic than at branching. Most of the works above recognize this and propose turning on/off large chunks of the network with each gating decision.
Large batch sizes are critical for performance, as they amortize the costs of parameter transfers and updates. Conditional computation reduces the batch sizes for the conditionally active chunks of the network.
Network bandwidth can be a bottleneck. A cluster of GPUs may have computational power thousands of times greater than the aggregate inter-device network bandwidth. To be computationally efficient, the relative computational versus network demands of an algorithm must exceed this ratio. Embedding layers, which can be seen as a form of conditional computation, are handicapped by this very problem. Since the embeddings generally need to be sent across the network, the number of (example, parameter) interactions is limited by network bandwidth instead of computational capacity.
Depending on the scheme, loss terms may be necessary to achieve the desired level of sparsity per-chunk and/or per example. BIBREF13 use three such terms. These issues can affect both model quality and load-balancing.
Model capacity is most critical for very large data sets. The existing literature on conditional computation deals with relatively small image recognition data sets consisting of up to 600,000 images. It is hard to imagine that the labels of these images provide a sufficient signal to adequately train a model with millions, let alone billions of parameters.
In this work, we for the first time address all of the above challenges and finally realize the promise of conditional computation. We obtain greater than 1000x improvements in model capacity with only minor losses in computational efficiency and significantly advance the state-of-the-art results on public language modeling and translation data sets.
## Our Approach: The Sparsely-Gated Mixture-of-Experts Layer
Our approach to conditional computation is to introduce a new type of general purpose neural network component: a Sparsely-Gated Mixture-of-Experts Layer (MoE). The MoE consists of a number of experts, each a simple feed-forward neural network, and a trainable gating network which selects a sparse combination of the experts to process each input (see Figure FIGREF8 ). All parts of the network are trained jointly by back-propagation.
While the introduced technique is generic, in this paper we focus on language modeling and machine translation tasks, which are known to benefit from very large models. In particular, we apply a MoE convolutionally between stacked LSTM layers BIBREF15 , as in Figure FIGREF8 . The MoE is called once for each position in the text, selecting a potentially different combination of experts at each position. The different experts tend to become highly specialized based on syntax and semantics (see Appendix SECREF84 Table TABREF92 ). On both language modeling and machine translation benchmarks, we improve on best published results at a fraction of the computational cost.
## Related work on Mixtures of Experts
Since its introduction more than two decades ago BIBREF16 , BIBREF17 , the mixture-of-experts approach has been the subject of much research. Different types of expert architectures hae been proposed such as SVMs BIBREF18 , Gaussian Processes BIBREF19 , BIBREF20 , BIBREF21 , Dirichlet Processes BIBREF22 , and deep networks. Other work has focused on different expert configurations such as a hierarchical structure BIBREF23 , infinite numbers of experts BIBREF24 , and adding experts sequentially BIBREF25 . BIBREF26 suggest an ensemble model in the format of mixture of experts for machine translation. The gating network is trained on a pre-trained ensemble NMT model.
The works above concern top-level mixtures of experts. The mixture of experts is the whole model. BIBREF10 introduce the idea of using multiple MoEs with their own gating networks as parts of a deep model. It is intuitive that the latter approach is more powerful, since complex problems may contain many sub-problems each requiring different experts. They also allude in their conclusion to the potential to introduce sparsity, turning MoEs into a vehicle for computational computation.
Our work builds on this use of MoEs as a general purpose neural network component. While BIBREF10 uses two stacked MoEs allowing for two sets of gating decisions, our convolutional application of the MoE allows for different gating decisions at each position in the text. We also realize sparse gating and demonstrate its use as a practical way to massively increase model capacity.
## The Structure of the Mixture-of-Experts layer
The Mixture-of-Experts (MoE) layer consists of a set of INLINEFORM0 “expert networks" INLINEFORM1 , and a “gating network" INLINEFORM2 whose output is a sparse INLINEFORM3 -dimensional vector. Figure FIGREF8 shows an overview of the MoE module. The experts are themselves neural networks, each with their own parameters. Although in principle we only require that the experts accept the same sized inputs and produce the same-sized outputs, in our initial investigations in this paper, we restrict ourselves to the case where the models are feed-forward networks with identical architectures, but with separate parameters.
Let us denote by INLINEFORM0 and INLINEFORM1 the output of the gating network and the output of the INLINEFORM2 -th expert network for a given input INLINEFORM3 . The output INLINEFORM4 of the MoE module can be written as follows: DISPLAYFORM0
We save computation based on the sparsity of the output of INLINEFORM0 . Wherever INLINEFORM1 , we need not compute INLINEFORM2 . In our experiments, we have up to thousands of experts, but only need to evaluate a handful of them for every example. If the number of experts is very large, we can reduce the branching factor by using a two-level hierarchical MoE. In a hierarchical MoE, a primary gating network chooses a sparse weighted combination of “experts", each of which is itself a secondary mixture-of-experts with its own gating network. In the following we focus on ordinary MoEs. We provide more details on hierarchical MoEs in Appendix SECREF60 .
Our implementation is related to other models of conditional computation. A MoE whose experts are simple weight matrices is similar to the parameterized weight matrix proposed in BIBREF12 . A MoE whose experts have one hidden layer is similar to the block-wise dropout described in BIBREF13 , where the dropped-out layer is sandwiched between fully-activated layers.
## Gating Network
A simple choice of non-sparse gating function BIBREF17 is to multiply the input by a trainable weight matrix INLINEFORM0 and then apply the INLINEFORM1 function. DISPLAYFORM0
We add two components to the Softmax gating network: sparsity and noise. Before taking the softmax function, we add tunable Gaussian noise, then keep only the top k values, setting the rest to INLINEFORM0 (which causes the corresponding gate values to equal 0). The sparsity serves to save computation, as described above. While this form of sparsity creates some theoretically scary discontinuities in the output of gating function, we have not yet observed this to be a problem in practice. The noise term helps with load balancing, as will be discussed in Appendix SECREF51 . The amount of noise per component is controlled by a second trainable weight matrix INLINEFORM1 . DISPLAYFORM0 DISPLAYFORM1
We train the gating network by simple back-propagation, along with the rest of the model. If we choose INLINEFORM0 , the gate values for the top k experts have nonzero derivatives with respect to the weights of the gating network. This type of occasionally-sensitive behavior is described in BIBREF9 with respect to noisy rectifiers. Gradients also back-propagate through the gating network to its inputs. Our method differs here from BIBREF13 who use boolean gates and a REINFORCE-style approach to train the gating network.
## The Shrinking Batch Problem
On modern CPUs and GPUs, large batch sizes are necessary for computational efficiency, so as to amortize the overhead of parameter loads and updates. If the gating network chooses INLINEFORM0 out of INLINEFORM1 experts for each example, then for a batch of INLINEFORM2 examples, each expert receives a much smaller batch of approximately INLINEFORM3 examples. This causes a naive MoE implementation to become very inefficient as the number of experts increases. The solution to this shrinking batch problem is to make the original batch size as large as possible. However, batch size tends to be limited by the memory necessary to store activations between the forwards and backwards passes. We propose the following techniques for increasing the batch size:
In a conventional distributed training setting, multiple copies of the model on different devices asynchronously process distinct batches of data, and parameters are synchronized through a set of parameter servers. In our technique, these different batches run synchronously so that they can be combined for the MoE layer. We distribute the standard layers of the model and the gating network according to conventional data-parallel schemes, but keep only one shared copy of each expert. Each expert in the MoE layer receives a combined batch consisting of the relevant examples from all of the data-parallel input batches. The same set of devices function as data-parallel replicas (for the standard layers and the gating networks) and as model-parallel shards (each hosting a subset of the experts). If the model is distributed over INLINEFORM0 devices, and each device processes a batch of size INLINEFORM1 , each expert receives a batch of approximately INLINEFORM2 examples. Thus, we achieve a factor of INLINEFORM3 improvement in expert batch size.
In the case of a hierarchical MoE (Section SECREF60 ), the primary gating network employs data parallelism, and the secondary MoEs employ model parallelism. Each secondary MoE resides on one device.
This technique allows us to increase the number of experts (and hence the number of parameters) by proportionally increasing the number of devices in the training cluster. The total batch size increases, keeping the batch size per expert constant. The memory and bandwidth requirements per device also remain constant, as do the step times, as does the amount of time necessary to process a number of training examples equal to the number of parameters in the model. It is our goal to train a trillion-parameter model on a trillion-word corpus. We have not scaled our systems this far as of the writing of this paper, but it should be possible by adding more hardware.
In our language models, we apply the same MoE to each time step of the previous layer. If we wait for the previous layer to finish, we can apply the MoE to all the time steps together as one big batch. Doing so increases the size of the input batch to the MoE layer by a factor of the number of unrolled time steps.
We suspect that even more powerful models may involve applying a MoE recurrently. For example, the weight matrices of a LSTM or other RNN could be replaced by a MoE. Sadly, such models break the convolutional trick from the last paragraph, since the input to the MoE at one timestep depends on the output of the MoE at the previous timestep. BIBREF27 describe a technique for drastically reducing the number of stored activations in an unrolled RNN, at the cost of recomputing forward activations. This would allow for a large increase in batch size.
## Network Bandwidth
Another major performance concern in distributed computing is network bandwidth. Since the experts are stationary (see above) and the number of gating parameters is small, most of the communication involves sending the inputs and outputs of the experts across the network. To maintain computational efficiency, the ratio of an expert's computation to the size of its input and output must exceed the ratio of computational to network capacity of the computing device. For GPUs, this may be thousands to one. In our experiments, we use experts with one hidden layer containing thousands of RELU-activated units. Since the weight matrices in the expert have sizes INLINEFORM0 _ INLINEFORM1 _ INLINEFORM2 and INLINEFORM3 _ INLINEFORM4 _ INLINEFORM5 , the ratio of computation to input and output is equal to the size of the hidden layer. Conveniently, we can increase computational efficiency simply by using a larger hidden layer, or more hidden layers.
## Balancing Expert Utilization
We have observed that the gating network tends to converge to a state where it always produces large weights for the same few experts. This imbalance is self-reinforcing, as the favored experts are trained more rapidly and thus are selected even more by the gating network. BIBREF10 describe the same phenomenon, and use a hard constraint at the beginning of training to avoid this local minimum. BIBREF13 include a soft constraint on the batch-wise average of each gate.
We take a soft constraint approach. We define the importance of an expert relative to a batch of training examples to be the batchwise sum of the gate values for that expert. We define an additional loss INLINEFORM0 , which is added to the overall loss function for the model. This loss is equal to the square of the coefficient of variation of the set of importance values, multiplied by a hand-tuned scaling factor INLINEFORM1 . This additional loss encourages all experts to have equal importance. DISPLAYFORM0 DISPLAYFORM1
While this loss function can ensure equal importance, experts may still receive very different numbers of examples. For example, one expert may receive a few examples with large weights, and another may receive many examples with small weights. This can cause memory and performance problems on distributed hardware. To solve this problem, we introduce a second loss function, INLINEFORM0 , which ensures balanced loads. Appendix SECREF51 contains the definition of this function, along with experimental results.
## 1 Billion Word Language Modeling Benchmark
This dataset, introduced by BIBREF28 consists of shuffled unique sentences from news articles, totaling approximately 829 million words, with a vocabulary of 793,471 words.
The best previously published results BIBREF2 use models consisting of one or more stacked Long Short-Term Memory (LSTM) layers BIBREF15 , BIBREF29 . The number of parameters in the LSTM layers of these models vary from 2 million to 151 million. Quality increases greatly with parameter count, as do computational costs. Results for these models form the top line of Figure FIGREF32 -right.
Our models consist of two stacked LSTM layers with a MoE layer between them (see Figure FIGREF8 ). We vary the sizes of the layers and the number of experts. For full details on model architecture, training regimen, additional baselines and results, see Appendix SECREF65 .
To investigate the effects of adding capacity, we trained a series of MoE models all with roughly equal computational costs: about 8 million multiply-and-adds per training example per timestep in the forwards pass, excluding the softmax layer. We call this metric (ops/timestep). We trained models with flat MoEs containing 4, 32, and 256 experts, and models with hierarchical MoEs containing 256, 1024, and 4096 experts. Each expert had about 1 million parameters. For all the MoE layers, 4 experts were active per input.
The results of these models are shown in Figure FIGREF32 -left. The model with 4 always-active experts performed (unsurprisingly) similarly to the computationally-matched baseline models, while the largest of the models (4096 experts) achieved an impressive 24% lower perplexity on the test set.
In addition to the largest model from the previous section, we trained two more MoE models with similarly high capacity (4 billion parameters), but higher computation budgets. These models had larger LSTMs, and fewer but larger and experts. Details can be found in Appendix UID77 . Results of these three models form the bottom line of Figure FIGREF32 -right. Table TABREF33 compares the results of these models to the best previously-published result on this dataset . Even the fastest of these models beats the best published result (when controlling for the number of training epochs), despite requiring only 6% of the computation.
We trained our models using TensorFlow BIBREF30 on clusters containing 16-32 Tesla K40 GPUs. For each of our models, we determine computational efficiency in TFLOPS/GPU by dividing the number of floating point operations required to process one training batch by the observed step time and the number of GPUs in the cluster. The operation counts used here are higher than the ones we report in our ops/timestep numbers in that we include the backwards pass, we include the importance-sampling-based training of the softmax layer, and we count a multiply-and-add as two separate operations. For all of our MoE models, the floating point operations involved in the experts represent between 37% and 46% of the total.
For our baseline models wtih no MoE, observed computational efficiency ranged from 1.07-1.29 TFLOPS/GPU. For our low-computation MoE models, computation efficiency ranged from 0.74-0.90 TFLOPS/GPU, except for the 4-expert model which did not make full use of the available parallelism. Our highest-computation MoE model was more efficient at 1.56 TFLOPS/GPU, likely due to the larger matrices. These numbers represent a significant fraction of the theoretical maximum of 4.29 TFLOPS/GPU claimed by NVIDIA. Detailed results are in Appendix SECREF65 , Table TABREF76 .
## 100 Billion Word Google News Corpus
On the 1-billion-word corpus, adding additional capacity seems to produce diminishing returns as the number of parameters in the MoE layer exceeds 1 billion, as can be seen in Figure FIGREF32 -left. We hypothesized that for a larger training set, even higher capacities would produce significant quality improvements.
We constructed a similar training set consisting of shuffled unique sentences from Google's internal news corpus, totalling roughly 100 billion words. Similarly to the previous section, we tested a series of models with similar computational costs of about 8 million ops/timestep. In addition to a baseline LSTM model, we trained models augmented with MoE layers containing 32, 256, 1024, 4096, 16384, 65536, and 131072 experts. This corresponds to up to 137 billion parameters in the MoE layer. Details on architecture, training, and results are given in Appendix SECREF78 .
Figure FIGREF37 shows test perplexity as a function of capacity after training on 10 billion words (top line) and 100 billion words (bottom line). When training over the full 100 billion words, test perplexity improves significantly up to 65536 experts (68 billion parameters), dropping 39% lower than the computationally matched baseline, but degrades at 131072 experts, possibly a result of too much sparsity. The widening gap between the two lines demonstrates (unsurprisingly) that increased model capacity helps more on larger training sets.
Even at 65536 experts (99.994% layer sparsity), computational efficiency for the model stays at a respectable 0.72 TFLOPS/GPU.
## Machine Translation (Single Language Pair)
Our model was a modified version of the GNMT model described in BIBREF3 . To reduce computation, we decreased the number of LSTM layers in the encoder and decoder from 9 and 8 to 3 and 2 respectively. We inserted MoE layers in both the encoder (between layers 2 and 3) and the decoder (between layers 1 and 2). Each MoE layer contained up to 2048 experts each with about two million parameters, adding a total of about 8 billion parameters to the models. Further details on model architecture, testing procedure and results can be found in Appendix SECREF84 .
We benchmarked our method on the WMT'14 En INLINEFORM0 Fr and En INLINEFORM1 De corpora, whose training sets have 36M sentence pairs and 5M sentence pairs, respectively. The experimental protocols were also similar to those in BIBREF3 : newstest2014 was used as the test set to compare against previous work BIBREF31 , BIBREF32 , BIBREF3 , while the combination of newstest2012 and newstest2013 was used as the development set. We also tested the same model on a Google's Production English to French data.
Tables TABREF42 , TABREF43 , and TABREF44 show the results of our largest models, compared with published results. Our approach achieved BLEU scores of 40.56 and 26.03 on the WMT'14 En INLINEFORM0 Fr and En INLINEFORM1 De benchmarks. As our models did not use RL refinement, these results constitute significant gains of 1.34 and 1.12 BLEU score on top of the strong baselines in BIBREF3 . The perplexity scores are also better. On the Google Production dataset, our model achieved 1.01 higher test BLEU score even after training for only one sixth of the time.
## Multilingual Machine Translation
BIBREF35 train a single GNMT BIBREF3 model on a very large combined dataset of twelve language pairs. Results are somewhat worse than those for 12 separately trained single-pair GNMT models. This is not surprising, given that the twelve models have 12 times the capacity and twelve times the aggregate training of the one model. We repeat this experiment with a single MoE-augmented model. See Appendix SECREF84 for details on model architecture. We train our model on the same dataset as BIBREF35 and process the same number of training examples (about 3 billion sentence pairs). Our training time was shorter due to the lower computational budget of our model.
Results for the single-pair GNMT models, the multilingual GNMT model and the multilingual MoE model are given in Table TABREF50 . The MoE model achieves 19% lower perplexity on the dev set than the multilingual GNMT model. On BLEU score, the MoE model significantly beats the multilingual GNMT model on 11 of the 12 language pairs (by as much as 5.84 points), and even beats the monolingual GNMT models on 8 of 12 language pairs. The poor performance on English INLINEFORM0 Korean seems to be a result of severe overtraining, as for the rarer language pairs a small number of real examples were highly oversampled in the training corpus.
## Conclusion
This work is the first to demonstrate major wins from conditional computation in deep networks. We carefully identified the design considerations and challenges of conditional computing and addressed them with a combination of algorithmic and engineering solutions. While we focused on text, conditional computation may help in other domains as well, provided sufficiently large training sets. We look forward to seeing many novel implementations and applications of conditional computation in the years to come.
## Appendices
tocsectionAppendices
## Load-Balancing Loss
As discussed in section SECREF4 , for load-balancing purposes, we want to define an additional loss function to encourage experts to receive roughly equal numbers of training examples. Unfortunately, the number of examples received by an expert is a discrete quantity, so it can not be used in back-propagation. Instead, we define a smooth estimator INLINEFORM0 of the number of examples assigned to each expert for a batch INLINEFORM1 of inputs. The smoothness allows us to back-propagate gradients through the estimator. This is the purpose of the noise term in the gating function. We define INLINEFORM2 as the probability that INLINEFORM3 is nonzero, given a new random choice of noise on element INLINEFORM4 , but keeping the already-sampled choices of noise on the other elements. To compute INLINEFORM5 , we note that the INLINEFORM6 is nonzero if and only if INLINEFORM7 is greater than the INLINEFORM8 -greatest element of INLINEFORM9 excluding itself. The probability works out to be: DISPLAYFORM0
Where INLINEFORM0 means the kth highest component of INLINEFORM1 , excluding component INLINEFORM2 . Simplifying, we get: DISPLAYFORM0
Where INLINEFORM0 is the CDF of the standard normal distribution. DISPLAYFORM0
We can now define the load loss to be the square of the coefficient of variation of the load vector, multiplied by a hand-tuned scaling factor INLINEFORM0 . DISPLAYFORM0
To avoid out-of-memory errors, we need to initialize the network in a state of approximately equal expert load (since the soft constraints need some time to work). To accomplish this, we initialize the matrices INLINEFORM0 and INLINEFORM1 to all zeros, which yields no signal and some noise.
We trained a set of models with identical architecture (the MoE-256 model described in Appendix SECREF65 ), using different values of INLINEFORM0 and INLINEFORM1 . We trained each model for 10 epochs, then measured perplexity on the test set. We also measured the coefficients of variation in INLINEFORM2 and INLINEFORM3 , as well as ratio of the load on the most overloaded expert to the average load. This last value is significant for load balancing purposes on distributed hardware. All of these metrics were averaged over several training batches.
Results are reported in Table TABREF58 . All the combinations containing at least one the two losses led to very similar model quality, where having no loss was much worse. Models with higher values of INLINEFORM0 had lower loads on the most overloaded expert.
## Hierachical Mixture of Experts
If the number of experts is very large, we can reduce the branching factor by using a two-level hierarchical MoE. In a hierarchical MoE, a primary gating network chooses a sparse weighted combination of “experts", each of which is itself a secondary mixture-of-experts with its own gating network. If the hierarchical MoE consists of INLINEFORM0 groups of INLINEFORM1 experts each, we denote the primary gating network by INLINEFORM2 , the secondary gating networks by INLINEFORM3 , and the expert networks by INLINEFORM4 . The output of the MoE is given by: DISPLAYFORM0
Our metrics of expert utilization change to the following: DISPLAYFORM0 DISPLAYFORM1
INLINEFORM0 and INLINEFORM1 deonte the INLINEFORM2 functions for the primary gating network and INLINEFORM3 secondary gating network respectively. INLINEFORM4 denotes the subset of INLINEFORM5 for which INLINEFORM6 .
It would seem simpler to let INLINEFORM0 , but this would not have a gradient with respect to the primary gating network, so we use the formulation above.
## 1 Billion Word Language Modeling Benchmark - Experimental Details
Our model consists of five layers: a word embedding layer, a recurrent Long Short-Term Memory (LSTM) layer BIBREF15 , BIBREF29 , a MoE layer, a second LSTM layer, and a softmax layer. The dimensionality of the embedding layer, the number of units in each LSTM layer, and the input and output dimensionality of the MoE layer are all equal to 512. For every layer other than the softmax, we apply drouput BIBREF43 to the layer output, dropping each activation with probability INLINEFORM0 , otherwise dividing by INLINEFORM1 . After dropout, the output of the previous layer is added to the layer output. This residual connection encourages gradient flow BIBREF37 .
Each expert in the MoE layer is a feed forward network with one ReLU-activated hidden layer of size 1024 and an output layer of size 512. Thus, each expert contains INLINEFORM0 parameters. The output of the MoE layer is passed through a sigmoid function before dropout. We varied the number of experts between models, using ordinary MoE layers with 4, 32 and 256 experts and hierarchical MoE layers with 256, 1024 and 4096 experts. We call the resulting models MoE-4, MoE-32, MoE-256, MoE-256-h, MoE-1024-h and MoE-4096-h. For the hierarchical MoE layers, the first level branching factor was 16, corresponding to the number of GPUs in our cluster. We use Noisy-Top-K Gating (see Section UID14 ) with INLINEFORM1 for the ordinary MoE layers and INLINEFORM2 at each level of the hierarchical MoE layers. Thus, each example is processed by exactly 4 experts for a total of 4M ops/timestep. The two LSTM layers contribute 2M ops/timestep each for the desired total of 8M.
The MoE-4 model does not employ sparsity, since all 4 experts are always used. In addition, we trained four more computationally-matched baseline models with no sparsity:
MoE-1-Wide: The MoE layer consists of a single "expert" containing one ReLU-activated hidden layer of size 4096.
MoE-1-Deep: The MoE layer consists of a single "expert" containing four ReLU-activated hidden layers, each with size 1024.
4xLSTM-512: We replace the MoE layer with two additional 512-unit LSTM layers.
LSTM-2048-512: The model contains one 2048-unit LSTM layer (and no MoE). The output of the LSTM is projected down to 512 dimensions BIBREF41 . The next timestep of the LSTM receives the projected output. This is identical to one of the models published in BIBREF2 . We re-ran it to account for differences in training regimen, and obtained results very similar to the published ones.
The models were trained on a cluster of 16 K40 GPUs using the synchronous method described in Section SECREF3 . Each batch consisted of a set of sentences totaling roughly 300,000 words. In the interest of time, we limited training to 10 epochs, (27,000 steps). Training took 12-16 hours for all models, except for MoE-4, which took 18 hours (since all the expert computation was performed on only 4 of 16 GPUs). We used the Adam optimizer BIBREF39 . The base learning rate was increased linearly for the first 1000 training steps, and decreased after that so as to be proportional to the inverse square root of the step number. The Softmax output layer was trained efficiently using importance sampling similarly to the models in BIBREF2 . For each model, we performed a hyper-parmeter search to find the best dropout probability, in increments of 0.1.
To ensure balanced expert utilization we set INLINEFORM0 and INLINEFORM1 , as described in Section SECREF4 and Appendix SECREF51 .
We evaluate our model using perplexity on the holdout dataset, used by BIBREF28 , BIBREF2 . We follow the standard procedure and sum over all the words including the end of sentence symbol. Results are reported in Table TABREF76 . For each model, we report the test perplexity, the computational budget, the parameter counts, the value of INLINEFORM0 , and the computational efficiency.
We ran two additional models (MoE-34M and MoE-143M) to investigate the effects of adding more computation in the presence of a large MoE layer. These models have computation budgets of 34M and 143M ops/timestep. Similar to the models above, these models use a MoE layer between two LSTM layers. The dimensionality of the embedding layer, and the input and output dimensionality of the MoE layer are set to 1024 instead of 512. For MoE-34M, the LSTM layers have 1024 units. For MoE-143M, the LSTM layers have 4096 units and an output projection of size 1024 BIBREF41 . MoE-34M uses a hierarchical MoE layer with 1024 experts, each with a hidden layer of size 2048. MoE-143M uses a hierarchical MoE layer with 256 experts, each with a hidden layer of size 8192. Both models have 4B parameters in the MoE layers. We searched for the best INLINEFORM0 for each model, and trained each model for 10 epochs.
The two models achieved test perplexity of INLINEFORM0 and INLINEFORM1 respectively, showing that even in the presence of a large MoE, more computation is still useful. Results are reported at the bottom of Table TABREF76 . The larger of the two models has a similar computational budget to the best published model from the literature, and training times are similar. Comparing after 10 epochs, our model has a lower test perplexity by INLINEFORM2 .
## 100 Billion Word Google News Corpus - Experimental Details
The models are similar in structure to the 8-million-operations-per-timestep models described in the previous section. We vary the number of experts between models, using an ordinary MoE layer with 32 experts and hierarchical MoE layers with 256, 1024, 4096, 16384, 65536 and 131072 experts. For the hierarchical MoE layers, the first level branching factors are 32, 32, 64, 128, 256 and 256, respectively.
Models are trained on a cluster of 32 Tesla K40 GPUs, except for the last two models, which are trained on clusters of 64 and 128 GPUs so as to have enough memory for all the parameters. For all models, training batch sizes are approximately 2.5 million words. Models are trained once-through over about 100 billion words.
We implement several memory optimizations in order to fit up to 1 billion parameters per GPU. First, we do not store the activations of the hidden layers of the experts, but instead recompute them on the backwards pass. Secondly, we modify the optimizer on the expert parameters to require less auxiliary storage:
The Adam optimizer BIBREF39 keeps first and second moment estimates of the per-parameter gradients. This triples the required memory. To avoid keeping a first-moment estimator, we set INLINEFORM0 . To reduce the size of the second moment estimator, we replace it with a factored approximation. For a matrix of parameters, instead of maintaining a full matrix of second-moment estimators, we maintain vectors of row-wise and column-wise averages of that matrix. At each step, the matrix of estimators is taken to be the outer product of those two vectors divided by the mean of either one. This technique could similarly be applied to Adagrad BIBREF36 .
We evaluate our model using perplexity on a holdout dataset. Results are reported in Table TABREF81 . Perplexity after 100 billion training words is 39% lower for the 68-billion-parameter MoE model than for the baseline model. It is notable that the measured computational efficiency of the largest model (0.30 TFLOPS/GPU) is very low compared to the other models. This is likely a result of the fact that, for purposes of comparison to the other models, we did not increase the training batch size proportionally to the number of GPUs. For comparison, we include results for a computationally matched baseline model consisting of 4 LSTMs, and for an unpruned 5-gram model with Kneser-Ney smoothing BIBREF40 .
## Machine Translation - Experimental Details
Our model is a modified version of the GNMT model described in BIBREF3 . To reduce computation, we decrease the number of LSTM layers in the encoder and decoder from 9 and 8 to 3 and 2 respectively. We insert MoE layers in both the encoder (between layers 2 and 3) and the decoder (between layers 1 and 2). We use an attention mechanism between the encoder and decoder, with the first decoder LSTM receiving output from and providing input for the attention . All of the layers in our model have input and output dimensionality of 512. Our LSTM layers have 2048 hidden units, with a 512-dimensional output projection. We add residual connections around all LSTM and MoE layers to encourage gradient flow BIBREF37 . Similar to GNMT, to effectively deal with rare words, we used sub-word units (also known as “wordpieces") BIBREF42 for inputs and outputs in our system.
We use a shared source and target vocabulary of 32K wordpieces. We also used the same beam search technique as proposed in BIBREF3 .
We train models with different numbers of experts in the MoE layers. In addition to a baseline model with no MoE layers, we train models with flat MoE layers containing 32 experts, and models with hierarchical MoE layers containing 512 and 2048 experts. The flat MoE layers use INLINEFORM0 and the hierarchical MoE models use INLINEFORM1 at each level of the gating network. Thus, each input is processed by exactly 4 experts in each MoE layer. Each expert in the MoE layer is a feed forward network with one hidden layer of size 2048 and ReLU activation. Thus, each expert contains INLINEFORM2 parameters. The output of the MoE layer is passed through a sigmoid function. We use the strictly-balanced gating function described in Appendix SECREF93 .
We used the same model architecture as for the single-language-pair models, with the following exceptions: We used noisy-top-k gating as described in Section UID14 , not the scheme from Appendix SECREF93 . The MoE layers in the encoder and decoder are non-hierarchical MoEs with INLINEFORM0 experts, and INLINEFORM1 . Each expert has a larger hidden layer of size 8192. This doubles the amount of computation in the MoE layers, raising the computational budget of the entire model from 85M to 102M ops/timestep.
We trained our networks using the Adam optimizer BIBREF39 . The base learning rate was increased linearly for the first 2000 training steps, held constant for an additional 8000 steps, and decreased after that so as to be proportional to the inverse square root of the step number. For the single-language-pair models, similarly to BIBREF3 , we applied dropout BIBREF43 to the output of all embedding, LSTM and MoE layers, using INLINEFORM0 . Training was done synchronously on a cluster of up to 64 GPUs as described in section SECREF3 . Each training batch consisted of a set of sentence pairs containing roughly 16000 words per GPU.
To ensure balanced expert utilization we set INLINEFORM0 and INLINEFORM1 , as described in Section SECREF4 and Appendix SECREF51 .
We evaluated our models using the perplexity and the standard BLEU score metric. We reported tokenized BLEU score as computed by the multi-bleu.pl script, downloaded from the public implementation of Moses (on Github), which was also used in BIBREF31 .
Tables TABREF42 , TABREF43 and TABREF44 in Section SECREF39 show comparisons of our results to other published methods. Figure FIGREF91 shows test perplexity as a function of number of words in the (training data's) source sentences processed for models with different numbers of experts. As can be seen from the Figure, as we increased the number of experts to approach 2048, the test perplexity of our model continued to improve.
We found that the experts indeed become highly specialized by syntax and/or semantics, as can be seen in Table TABREF92 . For example, one expert is used when the indefinite article “a" introduces the direct object in a verb phrase indicating importance or leadership.
## Strictly Balanced Gating
Due to some peculiarities in our infrastructure which have since been fixed, at the time we ran some of the machine translation experiments, our models ran faster if every expert received exactly the same batch size. To accommodate this, we used a different gating function which we describe below.
Recall that we define the softmax gating function to be: DISPLAYFORM0
To obtain a sparse gating vector, we multiply INLINEFORM0 component-wise with a sparse mask INLINEFORM1 and normalize the output. The mask itself is a function of INLINEFORM2 and specifies which experts are assigned to each input example: DISPLAYFORM0
To implement top-k gating in this formulation, we would let INLINEFORM0 , where: DISPLAYFORM0
To force each expert to receive the exact same number of examples, we introduce an alternative mask function, INLINEFORM0 , which operates over batches of input vectors. Instead of keeping the top INLINEFORM1 values per example, we keep the top INLINEFORM2 values per expert across the training batch, where INLINEFORM3 , so that each example is sent to an average of INLINEFORM4 experts. DISPLAYFORM0
As our experiments suggest and also observed in BIBREF38 , using a batchwise function during training (such as INLINEFORM0 ) requires modifications to the inference when we may not have a large batch of examples. Our solution to this is to train a vector INLINEFORM1 of per-expert threshold values to approximate the effects of the batchwise mask. We use the following mask at inference time: DISPLAYFORM0
To learn the threshold values, we apply an additional loss at training time which is minimized when the batchwise mask and the threshold mask are identical. DISPLAYFORM0
## Attention Function
The attention mechanism described in GNMT BIBREF3 involves a learned “Attention Function" INLINEFORM0 which takes a “source vector" INLINEFORM1 and a “target vector" INLINEFORM2 , and must be computed for every source time step INLINEFORM3 and target time step INLINEFORM4 . In GNMT, the attention function is implemented as a feed forward neural network with a hidden layer of size INLINEFORM5 . It can be expressed as: DISPLAYFORM0
Where INLINEFORM0 and INLINEFORM1 are trainable weight matrices and INLINEFORM2 is a trainable weight vector.
For performance reasons, in our models, we used a slightly different attention function: DISPLAYFORM0
With our attention function, we can simultaneously compute the attention function on multiple source time steps and multiple target time steps using optimized matrix multiplications. We found little difference in quality between the two functions.
| 21
|
1703.04617
|
Exploring Question Understanding and Adaptation in Neural-Network-Based Question Answering
|
# Exploring Question Understanding and Adaptation in Neural-Network-Based Question Answering
## Abstract
The last several years have seen intensive interest in exploring neural-network-based models for machine comprehension (MC) and question answering (QA). In this paper, we approach the problems by closely modelling questions in a neural network framework. We first introduce syntactic information to help encode questions. We then view and model different types of questions and the information shared among them as an adaptation task and proposed adaptation models for them. On the Stanford Question Answering Dataset (SQuAD), we show that these approaches can help attain better results over a competitive baseline.
## Introduction
Enabling computers to understand given documents and answer questions about their content has recently attracted intensive interest, including but not limited to the efforts as in BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 . Many specific problems such as machine comprehension and question answering often involve modeling such question-document pairs.
The recent availability of relatively large training datasets (see Section "Related Work" for more details) has made it more feasible to train and estimate rather complex models in an end-to-end fashion for these problems, in which a whole model is fit directly with given question-answer tuples and the resulting model has shown to be rather effective.
In this paper, we take a closer look at modeling questions in such an end-to-end neural network framework, since we regard question understanding is of importance for such problems. We first introduced syntactic information to help encode questions. We then viewed and modelled different types of questions and the information shared among them as an adaptation problem and proposed adaptation models for them. On the Stanford Question Answering Dataset (SQuAD), we show that these approaches can help attain better results on our competitive baselines.
## Related Work
Recent advance on reading comprehension and question answering has been closely associated with the availability of various datasets. BIBREF0 released the MCTest data consisting of 500 short, fictional open-domain stories and 2000 questions. The CNN/Daily Mail dataset BIBREF1 contains news articles for close style machine comprehension, in which only entities are removed and tested for comprehension. Children's Book Test (CBT) BIBREF2 leverages named entities, common nouns, verbs, and prepositions to test reading comprehension. The Stanford Question Answering Dataset (SQuAD) BIBREF3 is more recently released dataset, which consists of more than 100,000 questions for documents taken from Wikipedia across a wide range of topics. The question-answer pairs are annotated through crowdsourcing. Answers are spans of text marked in the original documents. In this paper, we use SQuAD to evaluate our models.
Many neural network models have been studied on the SQuAD task. BIBREF6 proposed match LSTM to associate documents and questions and adapted the so-called pointer Network BIBREF7 to determine the positions of the answer text spans. BIBREF8 proposed a dynamic chunk reader to extract and rank a set of answer candidates. BIBREF9 focused on word representation and presented a fine-grained gating mechanism to dynamically combine word-level and character-level representations based on the properties of words. BIBREF10 proposed a multi-perspective context matching (MPCM) model, which matched an encoded document and question from multiple perspectives. BIBREF11 proposed a dynamic decoder and so-called highway maxout network to improve the effectiveness of the decoder. The bi-directional attention flow (BIDAF) BIBREF12 used the bi-directional attention to obtain a question-aware context representation.
In this paper, we introduce syntactic information to encode questions with a specific form of recursive neural networks BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 . More specifically, we explore a tree-structured LSTM BIBREF13 , BIBREF14 which extends the linear-chain long short-term memory (LSTM) BIBREF17 to a recursive structure, which has the potential to capture long-distance interactions over the structures.
Different types of questions are often used to seek for different types of information. For example, a "what" question could have very different property from that of a "why" question, while they may share information and need to be trained together instead of separately. We view this as a "adaptation" problem to let different types of questions share a basic model but still discriminate them when needed. Specifically, we are motivated by the ideas "i-vector" BIBREF18 in speech recognition, where neural network based adaptation is performed among different (groups) of speakers and we focused instead on different types of questions here.
## The Baseline Model
Our baseline model is composed of the following typical components: word embedding, input encoder, alignment, aggregation, and prediction. Below we discuss these components in more details.
We concatenate embedding at two levels to represent a word: the character composition and word-level embedding. The character composition feeds all characters of a word into a convolutional neural network (CNN) BIBREF19 to obtain a representation for the word. And we use the pre-trained 300-D GloVe vectors BIBREF20 (see the experiment section for details) to initialize our word-level embedding. Each word is therefore represented as the concatenation of the character-composition vector and word-level embedding. This is performed on both questions and documents, resulting in two matrices: the $\mathbf {Q}^e \in \mathbb {R} ^{N\times d_w}$ for a question and the $\mathbf {D}^e \in \mathbb {R} ^{M\times d_w}$ for a document, where $N$ is the question length (number of word tokens), $M$ is the document length, and $d_w$ is the embedding dimensionality.
The above word representation focuses on representing individual words, and an input encoder here employs recurrent neural networks to obtain the representation of a word under its context. We use bi-directional GRU (BiGRU) BIBREF21 for both documents and questions.
$${\mathbf {Q}^c_i}&=\text{BiGRU}(\mathbf {Q}^e_i,i),\forall i \in [1, \dots , N] \\
{\mathbf {D}^c_j}&=\text{BiGRU}(\mathbf {D}^e_j,j),\forall j \in [1, \dots , M]$$ (Eq. 5)
A BiGRU runs a forward and backward GRU on a sequence starting from the left and the right end, respectively. By concatenating the hidden states of these two GRUs for each word, we obtain the a representation for a question or document: $\mathbf {Q}^c \in \mathbb {R} ^{N\times d_c}$ for a question and $\mathbf {D}^c \in \mathbb {R} ^{M\times d_c}$ for a document.
Questions and documents interact closely. As in most previous work, our framework use both soft attention over questions and that over documents to capture the interaction between them. More specifically, in this soft-alignment layer, we first feed the contextual representation matrix $\mathbf {Q}^c$ and $\mathbf {D}^c$ to obtain alignment matrix $\mathbf {U} \in \mathbb {R} ^{N\times M}$ :
$$\mathbf {U}_{ij} =\mathbf {Q}_i^c \cdot \mathbf {D}_j^{c\mathrm {T}}, \forall i \in [1, \dots , N], \forall j \in [1, \dots , M]$$ (Eq. 7)
Each $\mathbf {U}_{ij}$ represents the similarity between a question word $\mathbf {Q}_i^c$ and a document word $\mathbf {D}_j^c$ .
Word-level Q-code Similar as in BIBREF12 , we obtain a word-level Q-code. Specifically, for each document word $w_j$ , we find which words in the question are relevant to it. To this end, $\mathbf {a}_j\in \mathbb {R} ^{N}$ is computed with the following equation and used as a soft attention weight:
$$\mathbf {a}_j = softmax(\mathbf {U}_{:j}), \forall j \in [1, \dots , M]$$ (Eq. 8)
With the attention weights computed, we obtain the encoding of the question for each document word $w_j$ as follows, which we call word-level Q-code in this paper:
$$\mathbf {Q}^w=\mathbf {a}^{\mathrm {T}} \cdot \mathbf {Q}^{c} \in \mathbb {R} ^{M\times d_c}$$ (Eq. 9)
Question-based filtering To better explore question understanding, we design this question-based filtering layer. As detailed later, different question representation can be easily incorporated to this layer in addition to being used as a filter to find key information in the document based on the question. This layer is expandable with more complicated question modeling.
In the basic form of question-based filtering, for each question word $w_i$ , we find which words in the document are associated. Similar to $\mathbf {a}_j$ discussed above, we can obtain the attention weights on document words for each question word $w_i$ :
$$\mathbf {b}_i=softmax(\mathbf {U}_{i:})\in \mathbb {R} ^{M}, \forall i \in [1, \dots , N]$$ (Eq. 10)
By pooling $\mathbf {b}\in \mathbb {R} ^{N\times M}$ , we can obtain a question-based filtering weight $\mathbf {b}^f$ :
$$\mathbf {b}^f=norm(pooling(\mathbf {b})) \in \mathbb {R} ^{M}$$ (Eq. 11)
$$norm(\mathbf {x})=\frac{\mathbf {x}}{\sum _i x_i}$$ (Eq. 12)
where the specific pooling function we used include max-pooling and mean-pooling. Then the document softly filtered based on the corresponding question $\mathbf {D}^f$ can be calculated by:
$$\mathbf {D}_j^{f_{max}}=b^{f_{max}}_j \mathbf {D}_j^{c}, \forall j \in [1, \dots , M]$$ (Eq. 13)
$$\mathbf {D}_j^{f_{mean}}=b^{f_{mean}}_j \mathbf {D}_j^{c}, \forall j \in [1, \dots , M]$$ (Eq. 14)
Through concatenating the document representation $\mathbf {D}^c$ , word-level Q-code $\mathbf {Q}^w$ and question-filtered document $\mathbf {D}^f$ , we can finally obtain the alignment layer representation:
$$\mathbf {I}=[\mathbf {D}^c, \mathbf {Q}^w,\mathbf {D}^c \circ \mathbf {Q}^w,\mathbf {D}^c - \mathbf {Q}^w, \mathbf {D}^f, \mathbf {b}^{f_{max}}, \mathbf {b}^{f_{mean}}] \in \mathbb {R} ^{M \times (6d_c+2)}$$ (Eq. 16)
where " $\circ $ " stands for element-wise multiplication and " $-$ " is simply the vector subtraction.
After acquiring the local alignment representation, key information in document and question has been collected, and the aggregation layer is then performed to find answers. We use three BiGRU layers to model the process that aggregates local information to make the global decision to find the answer spans. We found a residual architecture BIBREF22 as described in Figure 2 is very effective in this aggregation process:
$$\mathbf {I}^1_i=\text{BiGRU}(\mathbf {I}_i)$$ (Eq. 18)
$$\mathbf {I}^2_i=\mathbf {I}^1_i + \text{BiGRU}(\mathbf {I}^1_i)$$ (Eq. 19)
The SQuAD QA task requires a span of text to answer a question. We use a pointer network BIBREF7 to predict the starting and end position of answers as in BIBREF6 . Different from their methods, we use a two-directional prediction to obtain the positions. For one direction, we first predict the starting position of the answer span followed by predicting the end position, which is implemented with the following equations:
$$P(s+)=softmax(W_{s+}\cdot I^3)$$ (Eq. 23)
$$P(e+)=softmax(W_{e+} \cdot I^3 + W_{h+} \cdot h_{s+})$$ (Eq. 24)
where $\mathbf {I}^3$ is inference layer output, $\mathbf {h}_{s+}$ is the hidden state of the first step, and all $\mathbf {W}$ are trainable matrices. We also perform this by predicting the end position first and then the starting position:
$$P(e-)=softmax(W_{e-}\cdot I^3)$$ (Eq. 25)
$$P(s-)=softmax(W_{s-} \cdot I^3 + W_{h-} \cdot h_{e-})$$ (Eq. 26)
We finally identify the span of an answer with the following equation:
$$P(s)=pooling([P(s+), P(s-)])$$ (Eq. 27)
$$P(e)=pooling([P(e+), P(e-)])$$ (Eq. 28)
We use the mean-pooling here as it is more effective on the development set than the alternatives such as the max-pooling.
## Question Understanding and Adaptation
The interplay of syntax and semantics of natural language questions is of interest for question representation. We attempt to incorporate syntactic information in questions representation with TreeLSTM BIBREF13 , BIBREF14 . In general a TreeLSTM could perform semantic composition over given syntactic structures.
Unlike the chain-structured LSTM BIBREF17 , the TreeLSTM captures long-distance interaction on a tree. The update of a TreeLSTM node is described at a high level with Equation ( 31 ), and the detailed computation is described in (–). Specifically, the input of a TreeLSTM node is used to configure four gates: the input gate $\mathbf {i}_t$ , output gate $\mathbf {o}_t$ , and the two forget gates $\mathbf {f}_t^L$ for the left child input and $\mathbf {f}_t^R$ for the right. The memory cell $\mathbf {c}_t$ considers each child's cell vector, $\mathbf {c}_{t-1}^L$ and $\mathbf {c}_{t-1}^R$ , which are gated by the left forget gate $\mathbf {f}_t^L$ and right forget gate $\mathbf {f}_t^R$ , respectively.
$$\mathbf {h}_t &= \text{TreeLSTM}(\mathbf {x}_t, \mathbf {h}_{t-1}^L, \mathbf {h}_{t-1}^R), \\
\mathbf {h}_t &= \mathbf {o}_t \circ \tanh (\mathbf {c}_{t}),\\
\mathbf {o}_t &= \sigma (\mathbf {W}_o \mathbf {x}_t + \mathbf {U}_o^L \mathbf {h}_{t-1}^L + \mathbf {U}_o^R \mathbf {h}_{t-1}^R), \\\mathbf {c}_t &= \mathbf {f}_t^L \circ \mathbf {c}_{t-1}^L + \mathbf {f}_t^R \circ \mathbf {c}_{t-1}^R + \mathbf {i}_t \circ \mathbf {u}_t, \\\mathbf {f}_t^L &= \sigma (\mathbf {W}_f \mathbf {x}_t + \mathbf {U}_f^{LL} \mathbf {h}_{t-1}^L + \mathbf {U}_f^{LR} \mathbf {h}_{t-1}^R),\\
\mathbf {f}_t^R &= \sigma (\mathbf {W}_f \mathbf {x}_t + \mathbf {U}_f^{RL} \mathbf {h}_{t-1}^L + \mathbf {U}_f^{RR} \mathbf {h}_{t-1}^R), \\\mathbf {i}_t &= \sigma (\mathbf {W}_i \mathbf {x}_t + \mathbf {U}_i^L \mathbf {h}_{t-1}^L + \mathbf {U}_i^R \mathbf {h}_{t-1}^R), \\\mathbf {u}_t &= \tanh (\mathbf {W}_c \mathbf {x}_t + \mathbf {U}_c^L \mathbf {h}_{t-1}^L + \mathbf {U}_c^R \mathbf {h}_{t-1}^R),$$ (Eq. 31)
where $\sigma $ is the sigmoid function, $\circ $ is the element-wise multiplication of two vectors, and all $\mathbf {W}$ , $\mathbf {U}$ are trainable matrices.
To obtain the parse tree information, we use Stanford CoreNLP (PCFG Parser) BIBREF23 , BIBREF24 to produce a binarized constituency parse for each question and build the TreeLSTM based on the parse tree. The root node of TreeLSTM is used as the representation for the whole question. More specifically, we use it as TreeLSTM Q-code $\mathbf {Q}^{TL}\in \mathbb {R} ^{d_c}$ , by not only simply concatenating it to the alignment layer output but also using it as a question filter, just as we discussed in the question-based filtering section:
$$\mathbf {Q}^{TL}=\text{TreeLSTM}(\mathbf {Q}^e) \in \mathbb {R} ^{d_c}$$ (Eq. 32)
$$\mathbf {b}^{TL}=norm(\mathbf {Q}^{TL} \cdot \mathbf {D}^{c\mathrm {T}}) \in \mathbb {R} ^{M}$$ (Eq. 33)
where $\mathbf {I}_{new}$ is the new output of alignment layer, and function $repmat$ copies $\mathbf {Q}^{TL}$ for M times to fit with $\mathbf {I}$ .
Questions by nature are often composed to fulfill different types of information needs. For example, a "when" question seeks for different types of information (i.e., temporal information) than those for a "why" question. Different types of questions and the corresponding answers could potentially have different distributional regularity.
The previous models are often trained for all questions without explicitly discriminating different question types; however, for a target question, both the common features shared by all questions and the specific features for a specific type of question are further considered in this paper, as they could potentially obey different distributions. In this paper we further explicitly model different types of questions in the end-to-end training. We start from a simple way to first analyze the word frequency of all questions, and obtain top-10 most frequent question types: what, how, who, when, which, where, why, be, whose, and whom, in which be stands for the questions beginning with different forms of the word be such as is, am, and are. We explicitly encode question-type information to be an 11-dimensional one-hot vector (the top-10 question types and "other" question type). Each question type is with a trainable embedding vector. We call this explicit question type code, $\mathbf {ET}\in \mathbb {R} ^{d_{ET}}$ . Then the vector for each question type is tuned during training, and is added to the system with the following equation:
$$\mathbf {I}_{new}=[\mathbf {I}, repmat(\mathbf {ET})]$$ (Eq. 38)
As discussed, different types of questions and their answers may share common regularity and have separate property at the same time. We also view this as an adaptation problem in order to let different types of questions share a basic model but still discriminate them when needed. Specifically, we borrow ideas from speaker adaptation BIBREF18 in speech recognition, where neural-network-based adaptation is performed among different groups of speakers.
Conceptually we regard a type of questions as a group of acoustically similar speakers. Specifically we propose a question discriminative block or simply called a discriminative block (Figure 3 ) below to perform question adaptation. The main idea is described below:
$$\mathbf {x^\prime } = f([\mathbf {x}, \mathbf {\bar{x}}^c, \mathbf {\delta _x}])$$ (Eq. 40)
For each input question $\mathbf {x}$ , we can decompose it to two parts: the cluster it belong(i.e., question type) and the diverse in the cluster. The information of the cluster is encoded in a vector $\mathbf {\bar{x}}^c$ . In order to keep calculation differentiable, we compute the weight of all the clusters based on the distances of $\mathbf {x}$ and each cluster center vector, in stead of just choosing the closest cluster. Then the discriminative vector $\mathbf {\delta _x}$ with regard to these most relevant clusters are computed. All this information is combined to obtain the discriminative information. In order to keep the full information of input, we also copy the input question $\mathbf {x}$ , together with the acquired discriminative information, to a feed-forward layer to obtain a new representation $\mathbf {x^\prime }$ for the question.
More specifically, the adaptation algorithm contains two steps: adapting and updating, which is detailed as follows:
Adapting In the adapting step, we first compute the similarity score between an input question vector $\mathbf {x}\in \mathbb {R} ^{h}$ and each centroid vector of $K$ clusters $~\mathbf {\bar{x}}\in \mathbb {R} ^{K \times h}$ . Each cluster here models a question type. Unlike the explicit question type modeling discussed above, here we do not specify what question types we are modeling but let the system to learn. Specifically, we only need to pre-specific how many clusters, $K$ , we are modeling. The similarity between an input question and cluster centroid can be used to compute similarity weight $\mathbf {w}^a$ :
$$w_k^a = softmax(cos\_sim(\mathbf {x}, \mathbf {\bar{x}}_k), \alpha ), \forall k \in [1, \dots , K]$$ (Eq. 43)
$$cos\_sim(\mathbf {u}, \mathbf {v}) = \frac{<\mathbf {u},\mathbf {v}>}{||\mathbf {u}|| \cdot ||\mathbf {v}||}$$ (Eq. 44)
We set $\alpha $ equals 50 to make sure only closest class will have a high weight while maintain differentiable. Then we acquire a soft class-center vector $\mathbf {\bar{x}}^c$ :
$$\mathbf {\bar{x}}^c = \sum _k w^a_k \mathbf {\bar{x}}_k \in \mathbb {R} ^{h}$$ (Eq. 46)
We then compute a discriminative vector $\mathbf {\delta _x}$ between the input question with regard to the soft class-center vector:
$$\mathbf {\delta _x} = \mathbf {x} - \mathbf {\bar{x}}^c$$ (Eq. 47)
Note that $\bar{\mathbf {x}}^c$ here models the cluster information and $\mathbf {\delta _x}$ represents the discriminative information in the cluster. By feeding $\mathbf {x}$ , $\bar{\mathbf {x}}^c$ and $\mathbf {\delta _x}$ into feedforward layer with Relu, we obtain $\mathbf {x^{\prime }}\in \mathbb {R} ^{K}$ :
$$\mathbf {x^{\prime }} = Relu(\mathbf {W} \cdot [\mathbf {x},\bar{\mathbf {x}}^c,\mathbf {\delta _x}])$$ (Eq. 48)
With $\mathbf {x^{\prime }}$ ready, we can apply Discriminative Block to any question code and obtain its adaptation Q-code. In this paper, we use TreeLSTM Q-code as the input vector $\mathbf {x}$ , and obtain TreeLSTM adaptation Q-code $\mathbf {Q}^{TLa}\in \mathbb {R} ^{d_c}$ . Similar to TreeLSTM Q-code $\mathbf {Q}^{TL}$ , we concatenate $\mathbf {Q}^{TLa}$ to alignment output $\mathbf {I}$ and also use it as a question filter:
$$\mathbf {Q}^{TLa} = Relu(\mathbf {W} \cdot [\mathbf {Q}^{TL},\overline{\mathbf {Q}^{TL}}^c,\mathbf {\delta _{\mathbf {Q}^{TL}}}])$$ (Eq. 49)
$$\mathbf {b}^{TLa}=norm(\mathbf {Q}^{TLa} \cdot \mathbf {D}^{c\mathrm {T}}) \in \mathbb {R} ^{M}$$ (Eq. 50)
Updating The updating stage attempts to modify the center vectors of the $K$ clusters in order to fit each cluster to model different types of questions. The updating is performed according to the following formula:
$$\mathbf {\bar{x}^{\prime }}_k = (1-\beta \text{w}_k^a)\mathbf {\bar{x}}_k+\beta \text{w}_k^a\mathbf {x}, \forall k \in [1, \dots , K]$$ (Eq. 54)
In the equation, $\beta $ is an updating rate used to control the amount of each updating, and we set it to 0.01. When $\mathbf {x}$ is far away from $K$ -th cluster center $\mathbf {\bar{x}}_k$ , $\text{w}_k^a$ is close to be value 0 and the $k$ -th cluster center $\mathbf {\bar{x}}_k$ tends not to be updated. If $\mathbf {x}$ is instead close to the $j$ -th cluster center $\mathbf {\bar{x}}_j$ , $\mathbf {x}$0 is close to the value 1 and the centroid of the $\mathbf {x}$1 -th cluster $\mathbf {x}$2 will be updated more aggressively using $\mathbf {x}$3 .
## Set-Up
We test our models on Stanford Question Answering Dataset (SQuAD) BIBREF3 . The SQuAD dataset consists of more than 100,000 questions annotated by crowdsourcing workers on a selected set of Wikipedia articles, and the answer to each question is a span of text in the Wikipedia articles. Training data includes 87,599 instances and validation set has 10,570 instances. The test data is hidden and kept by the organizer. The evaluation of SQuAD is Exact Match (EM) and F1 score.
We use pre-trained 300-D Glove 840B vectors BIBREF20 to initialize our word embeddings. Out-of-vocabulary (OOV) words are initialized randomly with Gaussian samples. CharCNN filter length is 1,3,5, each is 50 dimensions. All vectors including word embedding are updated during training. The cluster number K in discriminative block is 100. The Adam method BIBREF25 is used for optimization. And the first momentum is set to be 0.9 and the second 0.999. The initial learning rate is 0.0004 and the batch size is 32. We will half learning rate when meet a bad iteration, and the patience is 7. Our early stop evaluation is the EM and F1 score of validation set. All hidden states of GRUs, and TreeLSTMs are 500 dimensions, while word-level embedding $d_w$ is 300 dimensions. We set max length of document to 500, and drop the question-document pairs beyond this on training set. Explicit question-type dimension $d_{ET}$ is 50. We apply dropout to the Encoder layer and aggregation layer with a dropout rate of 0.5.
## Results
Table 1 shows the official leaderboard on SQuAD test set when we submitted our system. Our model achieves a 68.73% EM score and 77.39% F1 score, which is ranked among the state of the art single models (without model ensembling).
Table 2 shows the ablation performances of various Q-code on the development set. Note that since the testset is hidden from us, we can only perform such an analysis on the development set. Our baseline model using no Q-code achieved a 68.00% and 77.36% EM and F1 scores, respectively. When we added the explicit question type T-code into the baseline model, the performance was improved slightly to 68.16%(EM) and 77.58%(F1). We then used TreeLSTM introduce syntactic parses for question representation and understanding (replacing simple question type as question understanding Q-code), which consistently shows further improvement. We further incorporated the soft adaptation. When letting the number of hidden question types ( $K$ ) to be 20, the performance improves to 68.73%/77.74% on EM and F1, respectively, which corresponds to the results of our model reported in Table 1 . Furthermore, after submitted our result, we have experimented with a large value of $K$ and found that when $K=100$ , we can achieve a better performance of 69.10%/78.38% on the development set.
Figure UID61 shows the EM/F1 scores of different question types while Figure UID62 is the question type amount distribution on the development set. In Figure UID61 we can see that the average EM/F1 of the "when" question is highest and those of the "why" question is the lowest. From Figure UID62 we can see the "what" question is the major class.
Figure 5 shows the composition of F1 score. Take our best model as an example, we observed a 78.38% F1 score on the whole development set, which can be separated into two parts: one is where F1 score equals to 100%, which means an exact match. This part accounts for 69.10% of the entire development set. And the other part accounts for 30.90%, of which the average F1 score is 30.03%. For the latter, we can further divide it into two sub-parts: one is where the F1 score equals to 0%, which means that predict answer is totally wrong. This part occupies 14.89% of the total development set. The other part accounts for 16.01% of the development set, of which average F1 score is 57.96%. From this analysis we can see that reducing the zero F1 score (14.89%) is potentially an important direction to further improve the system.
## Conclusions
Closely modelling questions could be of importance for question answering and machine reading. In this paper, we introduce syntactic information to help encode questions in neural networks. We view and model different types of questions and the information shared among them as an adaptation task and proposed adaptation models for them. On the Stanford Question Answering Dataset (SQuAD), we show that these approaches can help attain better results over a competitive baseline.
| 7
|
1703.10344
|
Automated News Suggestions for Populating Wikipedia Entity Pages
|
# Automated News Suggestions for Populating Wikipedia Entity Pages
## Abstract
Wikipedia entity pages are a valuable source of information for direct consumption and for knowledge-base construction, update and maintenance. Facts in these entity pages are typically supported by references. Recent studies show that as much as 20\% of the references are from online news sources. However, many entity pages are incomplete even if relevant information is already available in existing news articles. Even for the already present references, there is often a delay between the news article publication time and the reference time. In this work, we therefore look at Wikipedia through the lens of news and propose a novel news-article suggestion task to improve news coverage in Wikipedia, and reduce the lag of newsworthy references. Our work finds direct application, as a precursor, to Wikipedia page generation and knowledge-base acceleration tasks that rely on relevant and high quality input sources. We propose a two-stage supervised approach for suggesting news articles to entity pages for a given state of Wikipedia. First, we suggest news articles to Wikipedia entities (article-entity placement) relying on a rich set of features which take into account the \emph{salience} and \emph{relative authority} of entities, and the \emph{novelty} of news articles to entity pages. Second, we determine the exact section in the entity page for the input article (article-section placement) guided by class-based section templates. We perform an extensive evaluation of our approach based on ground-truth data that is extracted from external references in Wikipedia. We achieve a high precision value of up to 93\% in the \emph{article-entity} suggestion stage and upto 84\% for the \emph{article-section placement}. Finally, we compare our approach against competitive baselines and show significant improvements.
## Introduction
Wikipedia is the largest source of open and collaboratively curated knowledge in the world. Introduced in 2001, it has evolved into a reference work with around 5m pages for the English Wikipedia alone. In addition, entities and event pages are updated quickly via collaborative editing and all edits are encouraged to include source citations, creating a knowledge base which aims at being both timely as well as authoritative. As a result, it has become the preferred source of information consumption about entities and events. Moreso, this knowledge is harvested and utilized in building knowledge bases like YAGO BIBREF0 and DBpedia BIBREF1 , and used in applications like text categorization BIBREF2 , entity disambiguation BIBREF3 , entity ranking BIBREF4 and distant supervision BIBREF5 , BIBREF6 .
However, not all Wikipedia pages referring to entities (entity pages) are comprehensive: relevant information can either be missing or added with a delay. Consider the city of New Orleans and the state of Odisha which were severely affected by cyclones Hurricane Katrina and Odisha Cyclone, respectively. While Katrina finds extensive mention in the entity page for New Orleans, Odisha Cyclone which has 5 times more human casualties (cf. Figure FIGREF2 ) is not mentioned in the page for Odisha. Arguably Katrina and New Orleans are more popular entities, but Odisha Cyclone was also reported extensively in national and international news outlets. This highlights the lack of important facts in trunk and long-tail entity pages, even in the presence of relevant sources. In addition, previous studies have shown that there is an inherent delay or lag when facts are added to entity pages BIBREF7 .
To remedy these problems, it is important to identify information sources that contain novel and salient facts to a given entity page. However, not all information sources are equal. The online presence of major news outlets is an authoritative source due to active editorial control and their articles are also a timely container of facts. In addition, their use is in line with current Wikipedia editing practice, as is shown in BIBREF7 that almost 20% of current citations in all entity pages are news articles. We therefore propose news suggestion as a novel task that enhances entity pages and reduces delay while keeping its pages authoritative.
Existing efforts to populate Wikipedia BIBREF8 start from an entity page and then generate candidate documents about this entity using an external search engine (and then post-process them). However, such an approach lacks in (a) reproducibility since rankings vary with time with obvious bias to recent news (b) maintainability since document acquisition for each entity has to be periodically performed. To this effect, our news suggestion considers a news article as input, and determines if it is valuable for Wikipedia. Specifically, given an input news article INLINEFORM0 and a state of Wikipedia, the news suggestion problem identifies the entities mentioned in INLINEFORM1 whose entity pages can improve upon suggesting INLINEFORM2 . Most of the works on knowledge base acceleration BIBREF9 , BIBREF10 , BIBREF11 , or Wikipedia page generation BIBREF8 rely on high quality input sources which are then utilized to extract textual facts for Wikipedia page population. In this work, we do not suggest snippets or paraphrases but rather entire articles which have a high potential importance for entity pages. These suggested news articles could be consequently used for extraction, summarization or population either manually or automatically – all of which rely on high quality and relevant input sources.
We identify four properties of good news recommendations: salience, relative authority, novelty and placement. First, we need to identify the most salient entities in a news article. This is done to avoid pollution of entity pages with only marginally related news. Second, we need to determine whether the news is important to the entity as only the most relevant news should be added to a precise reference work. To do this, we compute the relative authority of all entities in the news article: we call an entity more authoritative than another if it is more popular or noteworthy in the real world. Entities with very high authority have many news items associated with them and only the most relevant of these should be included in Wikipedia whereas for entities of lower authority the threshold for inclusion of a news article will be lower. Third, a good recommendation should be able to identify novel news by minimizing redundancy coming from multiple news articles. Finally, addition of facts is facilitated if the recommendations are fine-grained, i.e., recommendations are made on the section level rather than the page level (placement).
Approach and Contributions. We propose a two-stage news suggestion approach to entity pages. In the first stage, we determine whether a news article should be suggested for an entity, based on the entity's salience in the news article, its relative authority and the novelty of the article to the entity page. The second stage takes into account the class of the entity for which the news is suggested and constructs section templates from entities of the same class. The generation of such templates has the advantage of suggesting and expanding entity pages that do not have a complete section structure in Wikipedia, explicitly addressing long-tail and trunk entities. Afterwards, based on the constructed template our method determines the best fit for the news article with one of the sections.
We evaluate the proposed approach on a news corpus consisting of 351,982 articles crawled from the news external references in Wikipedia from 73,734 entity pages. Given the Wikipedia snapshot at a given year (in our case [2009-2014]), we suggest news articles that might be cited in the coming years. The existing news references in the entity pages along with their reference date act as our ground-truth to evaluate our approach. In summary, we make the following contributions.
## Related Work
As we suggest a new problem there is no current work addressing exactly the same task. However, our task has similarities to Wikipedia page generation and knowledge base acceleration. In addition, we take inspiration from Natural Language Processing (NLP) methods for salience detection.
Wikipedia Page Generation is the problem of populating Wikipedia pages with content coming from external sources. Sauper and Barzilay BIBREF8 propose an approach for automatically generating whole entity pages for specific entity classes. The approach is trained on already-populated entity pages of a given class (e.g. `Diseases') by learning templates about the entity page structure (e.g. diseases have a treatment section). For a new entity page, first, they extract documents via Web search using the entity title and the section title as a query, for example `Lung Cancer'+`Treatment'. As already discussed in the introduction, this has problems with reproducibility and maintainability. However, their main focus is on identifying the best paragraphs extracted from the collected documents. They rank the paragraphs via an optimized supervised perceptron model for finding the most representative paragraph that is the least similar to paragraphs in other sections. This paragraph is then included in the newly generated entity page. Taneva and Weikum BIBREF12 propose an approach that constructs short summaries for the long tail. The summaries are called `gems' and the size of a `gem' can be user defined. They focus on generating summaries that are novel and diverse. However, they do not consider any structure of entities, which is present in Wikipedia.
In contrast to BIBREF8 and BIBREF12 , we actually focus on suggesting entire documents to Wikipedia entity pages. These are authoritative documents (news), which are highly relevant for the entity, novel for the entity and in which the entity is salient. Whereas relevance in Sauper and Barzilay is implicitly computed by web page ranking we solve that problem by looking at relative authority and salience of an entity, using the news article and entity page only. As Sauper and Barzilay concentrate on empty entity pages, the problem of novelty of their content is not an issue in their work whereas it is in our case which focuses more on updating entities. Updating entities will be more and more important the bigger an existing reference work is. Both the approaches in BIBREF8 and BIBREF12 (finding paragraphs and summarization) could then be used to process the documents we suggest further. Our concentration on news is also novel.
Knowledge Base Acceleration. In this task, given specific information extraction templates, a given corpus is analyzed in order to find worthwhile mentions of an entity or snippets that match the templates. Balog BIBREF9 , BIBREF10 recommend news citations for an entity. Prior to that, the news articles are classified for their appropriateness for an entity, where as features for the classification task they use entity, document, entity-document and temporal features. The best performing features are those that measure similarity between an entity and the news document. West et al. BIBREF13 consider the problem of knowledge base completion, through question answering and complete missing facts in Freebase based on templates, i.e. Frank_Zappa bornIn Baltymore, Maryland.
In contrast, we do not extract facts for pre-defined templates but rather suggest news articles based on their relevance to an entity. In cases of long-tail entities, we can suggest to add a novel section through our abstraction and generation of section templates at entity class level.
Entity Salience. Determining which entities are prominent or salient in a given text has a long history in NLP, sparked by the linguistic theory of Centering BIBREF14 . Salience has been used in pronoun and co-reference resolution BIBREF15 , or to predict which entities will be included in an abstract of an article BIBREF11 . Frequent features to measure salience include the frequency of an entity in a document, positioning of an entity, grammatical function or internal entity structure (POS tags, head nouns etc.). These approaches are not currently aimed at knowledge base generation or Wikipedia coverage extension but we postulate that an entity's salience in a news article is a prerequisite to the news article being relevant enough to be included in an entity page. We therefore use the salience features in BIBREF11 as part of our model. However, these features are document-internal — we will show that they are not sufficient to predict news inclusion into an entity page and add features of entity authority, news authority and novelty that measure the relations between several entities, between entity and news article as well as between several competing news articles.
## Terminology and Problem Definition
We are interested in named entities mentioned in documents. An entity INLINEFORM0 can be identified by a canonical name, and can be mentioned differently in text via different surface forms. We canonicalize these mentions to entity pages in Wikipedia, a method typically known as entity linking. We denote the set of canonicalized entities extracted and linked from a news article INLINEFORM1 as INLINEFORM2 . For example, in Figure FIGREF7 , entities are canonicalized into Wikipedia entity pages (e.g. Odisha is canonicalized to the corresponding article). For a collection of news articles INLINEFORM3 , we further denote the resulting set of entities by INLINEFORM4 .
Information in an entity page is organized into sections and evolves with time as more content is added. We refer to the state of Wikipedia at a time INLINEFORM0 as INLINEFORM1 and the set of sections for an entity page INLINEFORM2 as its entity profile INLINEFORM3 . Unlike news articles, text in Wikipedia could be explicitly linked to entity pages through anchors. The set of entities explicitly referred in text from section INLINEFORM4 is defined as INLINEFORM5 . Furthermore, Wikipedia induces a category structure over its entities, which is exploited by knowledge bases like YAGO (e.g. Barack_Obama isA Person). Consequently, each entity page belongs to one or more entity categories or classes INLINEFORM6 . Now we can define our news suggestion problem below:
Definition 1 (News Suggestion Problem) Given a set of news articles INLINEFORM0 and set of Wikipedia entity pages INLINEFORM1 (from INLINEFORM2 ) we intend to suggest a news article INLINEFORM3 published at time INLINEFORM4 to entity page INLINEFORM5 and additionally to the most relevant section for the entity page INLINEFORM6 .
## Approach Overview
We approach the news suggestion problem by decomposing it into two tasks:
AEP: Article–Entity placement
ASP: Article–Section placement
In this first step, for a given entity-news pair INLINEFORM0 , we determine whether the given news article INLINEFORM1 should be suggested (we will refer to this as `relevant') to entity INLINEFORM2 . To generate such INLINEFORM3 pairs, we perform the entity linking process, INLINEFORM4 , for INLINEFORM5 .
The article–entity placement task (described in detail in Section SECREF16 ) for a pair INLINEFORM0 outputs a binary label (either `non-relevant' or `relevant') and is formalized in Equation EQREF14 . DISPLAYFORM0
In the second step, we take into account all `relevant' pairs INLINEFORM0 and find the correct section for article INLINEFORM1 in entity INLINEFORM2 , respectively its profile INLINEFORM3 (see Section SECREF30 ). The article–section placement task, determines the correct section for the triple INLINEFORM4 , and is formalized in Equation EQREF15 . DISPLAYFORM0
In the subsequent sections we describe in details how we approach the two tasks for suggesting news articles to entity pages.
## News Article Suggestion
In this section, we provide an overview of the news suggestion approach to Wikipedia entity pages (see Figure FIGREF7 ). The approach is split into two tasks: (i) article-entity (AEP) and (ii) article-section (ASP) placement. For a Wikipedia snapshot INLINEFORM0 and a news corpus INLINEFORM1 , we first determine which news articles should be suggested to an entity INLINEFORM2 . We will denote our approach for AEP by INLINEFORM3 . Finally, we determine the most appropriate section for the ASP task and we denote our approach with INLINEFORM4 .
In the following, we describe the process of learning the functions INLINEFORM0 and INLINEFORM1 . We introduce features for the learning process, which encode information regarding the entity salience, relative authority and novelty in the case of AEP task. For the ASP task, we measure the overall fit of an article to the entity sections, with the entity being an input from AEP task. Additionally, considering that the entity profiles INLINEFORM2 are incomplete, in the case of a missing section we suggest and expand the entity profiles based on section templates generated from entities of the same class INLINEFORM3 (see Section UID34 ).
## Article–Entity Placement
In this step we learn the function INLINEFORM0 to correctly determine whether INLINEFORM1 should be suggested for INLINEFORM2 , basically a binary classification model (0=`non-relevant' and 1=`relevant'). Note that we are mainly interested in finding the relevant pairs in this task. For every news article, the number of disambiguated entities is around 30 (but INLINEFORM3 is suggested for only two of them on average). Therefore, the distribution of `non-relevant' and `relevant' pairs is skewed towards the earlier, and by simply choosing the `non-relevant' label we can achieve a high accuracy for INLINEFORM4 . Finding the relevant pairs is therefore a considerable challenge.
An article INLINEFORM0 is suggested to INLINEFORM1 by our function INLINEFORM2 if it fulfills the following properties. The entity INLINEFORM3 is salient in INLINEFORM4 (a central concept), therefore ensuring that INLINEFORM5 is about INLINEFORM6 and that INLINEFORM7 is important for INLINEFORM8 . Next, given the fact there might be many articles in which INLINEFORM9 is salient, we also look at the reverse property, namely whether INLINEFORM10 is important for INLINEFORM11 . We do this by comparing the authority of INLINEFORM12 (which is a measure of popularity of an entity, such as its frequency of mention in a whole corpus) with the authority of its co-occurring entities in INLINEFORM13 , leading to a feature we call relative authority. The intuition is that for an entity that has overall lower authority than its co-occurring entities, a news article is more easily of importance. Finally, if the article we are about to suggest is already covered in the entity profile INLINEFORM14 , we do not wish to suggest redundant information, hence the novelty. Therefore, the learning objective of INLINEFORM15 should fulfill the following properties. Table TABREF21 shows a summary of the computed features for INLINEFORM16 .
Salience: entity INLINEFORM0 should be a salient entity in news article INLINEFORM1
Relative Authority: the set of entities INLINEFORM0 with which INLINEFORM1 co-occurs should have higher authority than INLINEFORM2 , making INLINEFORM3 important for INLINEFORM4
Novelty: news article INLINEFORM0 should provide novel information for entity INLINEFORM1 taking into account its profile INLINEFORM2
Baseline Features. As discussed in Section SECREF2 , a variety of features that measure salience of an entity in text are available from the NLP community. We reimplemented the ones in Dunietz and Gillick BIBREF11 . This includes a variety of features, e.g. positional features, occurrence frequency and the internal POS structure of the entity and the sentence it occurs in. Table 2 in BIBREF11 gives details.
Relative Entity Frequency. Although frequency of mention and positional features play some role in baseline features, their interaction is not modeled by a single feature nor do the positional features encode more than sentence position. We therefore suggest a novel feature called relative entity frequency, INLINEFORM0 , that has three properties.: (i) It rewards entities for occurring throughout the text instead of only in some parts of the text, measured by the number of paragraphs it occurs in (ii) it rewards entities that occur more frequently in the opening paragraphs of an article as we model INLINEFORM1 as an exponential decay function. The decay corresponds to the positional index of the news paragraph. This is inspired by the news-specific discourse structure that tends to give short summaries of the most important facts and entities in the opening paragraphs. (iii) it compares entity frequency to the frequency of its co-occurring mentions as the weight of an entity appearing in a specific paragraph, normalized by the sum of the frequencies of other entities in INLINEFORM2 . DISPLAYFORM0
where, INLINEFORM0 represents a news paragraph from INLINEFORM1 , and with INLINEFORM2 we indicate the set of all paragraphs in INLINEFORM3 . The frequency of INLINEFORM4 in a paragraph INLINEFORM5 is denoted by INLINEFORM6 . With INLINEFORM7 and INLINEFORM8 we indicate the number of paragraphs in which entity INLINEFORM9 occurs, and the total number of paragraphs, respectively.
Relative Authority. In this case, we consider the comparative relevance of the news article to the different entities occurring in it. As an example, let us consider the meeting of the Sudanese bishop Elias Taban with Hillary Clinton. Both entities are salient for the meeting. However, in Taban's Wikipedia page, this meeting is discussed prominently with a corresponding news reference, whereas in Hillary Clinton's Wikipedia page it is not reported at all. We believe this is not just an omission in Clinton's page but mirrors the fact that for the lesser known Taban the meeting is big news whereas for the more famous Clinton these kind of meetings are a regular occurrence, not all of which can be reported in what is supposed to be a selection of the most important events for her. Therefore, if two entities co-occur, the news is more relevant for the entity with the lower a priori authority.
The a priori authority of an entity (denoted by INLINEFORM0 ) can be measured in several ways. We opt for two approaches: (i) probability of entity INLINEFORM1 occurring in the corpus INLINEFORM2 , and (ii) authority assessed through centrality measures like PageRank BIBREF16 . For the second case we construct the graph INLINEFORM3 consisting of entities in INLINEFORM4 and news articles in INLINEFORM5 as vertices. The edges are established between INLINEFORM6 and entities in INLINEFORM7 , that is INLINEFORM8 , and the out-links from INLINEFORM9 , that is INLINEFORM10 (arrows present the edge direction).
Starting from a priori authority, we proceed to relative authority by comparing the a priori authority of co-occurring entities in INLINEFORM0 . We define the relative authority of INLINEFORM1 as the proportion of co-occurring entities INLINEFORM2 that have a higher a priori authority than INLINEFORM3 (see Equation EQREF28 . DISPLAYFORM0
As we might run the danger of not suggesting any news articles for entities with very high a priori authority (such as Clinton) due to the strict inequality constraint, we can relax the constraint such that the authority of co-occurring entities is above a certain threshold.
News Domain Authority. The news domain authority addresses two main aspects. Firstly, if bundled together with the relative authority feature, we can ensure that dependent on the entity authority, we suggest news from authoritative sources, hence ensuring the quality of suggested articles. The second aspect is in a news streaming scenario where multiple news domains report the same event — ideally only articles coming from authoritative sources would fulfill the conditions for the news suggestion task.
The news domain authority is computed based on the number of news references in Wikipedia coming from a particular news domain INLINEFORM0 . This represents a simple prior that a news article INLINEFORM1 is from domain INLINEFORM2 in corpus INLINEFORM3 . We extract the domains by taking the base URLs from the news article URLs.
An important feature when suggesting an article INLINEFORM0 to an entity INLINEFORM1 is the novelty of INLINEFORM2 w.r.t the already existing entity profile INLINEFORM3 . Studies BIBREF17 have shown that on comparable collections to ours (TREC GOV2) the number of duplicates can go up to INLINEFORM4 . This figure is likely higher for major events concerning highly authoritative entities on which all news media will report.
Given an entity INLINEFORM0 and the already added news references INLINEFORM1 up to year INLINEFORM2 , the novelty of INLINEFORM3 at year INLINEFORM4 is measured by the KL divergence between the language model of INLINEFORM5 and articles in INLINEFORM6 . We combine this measure with the entity overlap of INLINEFORM7 and INLINEFORM8 . The novelty value of INLINEFORM9 is given by the minimal divergence value. Low scores indicate low novelty for the entity profile INLINEFORM10 .
N(n|e) = n'Nt-1{DKL((n') || (n)) + DKL((N) || (n)).
DKL((n') || (n)). (1-) jaccard((n'),(n))} where INLINEFORM0 is the KL divergence of the language models ( INLINEFORM1 and INLINEFORM2 ), whereas INLINEFORM3 is the mixing weight ( INLINEFORM4 ) between the language models INLINEFORM5 and the entity overlap in INLINEFORM6 and INLINEFORM7 .
Here we introduce the evaluation setup and analyze the results for the article–entity (AEP) placement task. We only report the evaluation metrics for the `relevant' news-entity pairs. A detailed explanation on why we focus on the `relevant' pairs is provided in Section SECREF16 .
Baselines. We consider the following baselines for this task.
B1. The first baseline uses only the salience-based features by Dunietz and Gillick BIBREF11 .
B2. The second baseline assigns the value relevant to a pair INLINEFORM0 , if and only if INLINEFORM1 appears in the title of INLINEFORM2 .
Learning Models. We use Random Forests (RF) BIBREF23 . We learn the RF on all computed features in Table TABREF21 . The optimization on RF is done by splitting the feature space into multiple trees that are considered as ensemble classifiers. Consequently, for each classifier it computes the margin function as a measure of the average count of predicting the correct class in contrast to any other class. The higher the margin score the more robust the model.
Metrics. We compute precision P, recall R and F1 score for the relevant class. For example, precision is the number of news-entity pairs we correctly labeled as relevant compared to our ground truth divided by the number of all news-entity pairs we labeled as relevant.
The following results measure the effectiveness of our approach in three main aspects: (i) overall performance of INLINEFORM0 and comparison to baselines, (ii) robustness across the years, and (iii) optimal model for the AEP placement task.
Performance. Figure FIGREF55 shows the results for the years 2009 and 2013, where we optimized the learning objective with instances from year INLINEFORM0 and evaluate on the years INLINEFORM1 (see Section SECREF46 ). The results show the precision–recall curve. The red curve shows baseline B1 BIBREF11 , and the blue one shows the performance of INLINEFORM2 . The curve shows for varying confidence scores (high to low) the precision on labeling the pair INLINEFORM3 as `relevant'. In addition, at each confidence score we can compute the corresponding recall for the `relevant' label. For high confidence scores on labeling the news-entity pairs, the baseline B1 achieves on average a precision score of P=0.50, while INLINEFORM4 has P=0.93. We note that with the drop in the confidence score the corresponding precision and recall values drop too, and the overall F1 score for B1 is around F1=0.2, in contrast we achieve an average score of F1=0.67.
It is evident from Figure FIGREF55 that for the years 2009 and 2013, INLINEFORM0 significantly outperforms the baseline B1. We measure the significance through the t-test statistic and get a p-value of INLINEFORM1 . The improvement we achieve over B1 in absolute numbers, INLINEFORM2 P=+0.5 in terms of precision for the years between 2009 and 2014, and a similar improvement in terms of F1 score. The improvement for recall is INLINEFORM3 R=+0.4. The relative improvement over B1 for P and F1 is almost 1.8 times better, while for recall we are 3.5 times better. In Table TABREF58 we show the overall scores for the evaluation metrics for B1 and INLINEFORM4 . Finally, for B2 we achieve much poorer performance, with average scores of P=0.21, R=0.20 and F1=0.21.
Robustness. In Table TABREF58 , we show the overall performance for the years between 2009 and 2013. An interesting observation we make is that we have a very robust performance and the results are stable across the years. If we consider the experimental setup, where for year INLINEFORM0 we optimize the learning objective with only 74k training instances and evaluate on the rest of the instances, it achieves a very good performance. We predict with F1=0.68 the remaining 469k instances for the years INLINEFORM1 .
The results are particularly promising considering the fact that the distribution between our two classes is highly skewed. On average the number of `relevant' pairs account for only around INLINEFORM0 of all pairs. A good indicator to support such a statement is the kappa (denoted by INLINEFORM1 ) statistic. INLINEFORM2 measures agreement between the algorithm and the gold standard on both labels while correcting for chance agreement (often expected due to extreme distributions). The INLINEFORM3 scores for B1 across the years is on average INLINEFORM4 , while for INLINEFORM5 we achieve a score of INLINEFORM6 (the maximum score for INLINEFORM7 is 1).
In Figure FIGREF60 we show the impact of the individual feature groups that contribute to the superior performance in comparison to the baselines. Relative entity frequency from the salience feature, models the entity salience as an exponentially decaying function based on the positional index of the paragraph where the entity appears. The performance of INLINEFORM0 with relative entity frequency from the salience feature group is close to that of all the features combined. The authority and novelty features account to a further improvement in terms of precision, by adding roughly a 7%-10% increase. However, if both feature groups are considered separately, they significantly outperform the baseline B1.
## Article–Section Placement
We model the ASP placement task as a successor of the AEP task. For all the `relevant' news entity pairs, the task is to determine the correct entity section. Each section in a Wikipedia entity page represents a different topic. For example, Barack Obama has the sections `Early Life', `Presidency', `Family and Personal Life' etc. However, many entity pages have an incomplete section structure. Incomplete or missing sections are due to two Wikipedia properties. First, long-tail entities miss information and sections due to their lack of popularity. Second, for all entities whether popular or not, certain sections might occur for the first time due to real world developments. As an example, the entity Germanwings did not have an `Accidents' section before this year's disaster, which was the first in the history of the airline.
Even if sections are missing for certain entities, similar sections usually occur in other entities of the same class (e.g. other airlines had disasters and therefore their pages have an accidents section). We exploit such homogeneity of section structure and construct templates that we use to expand entity profiles. The learning objective for INLINEFORM0 takes into account the following properties:
Section-templates: account for incomplete section structure for an entity profile INLINEFORM0 by constructing section templates INLINEFORM1 from an entity class INLINEFORM2
Overall fit: measures the overall fit of a news article to sections in the section templates INLINEFORM0
Given the fact that entity profiles are often incomplete, we construct section templates for every entity class. We group entities based on their class INLINEFORM0 and construct section templates INLINEFORM1 . For different entity classes, e.g. Person and Location, the section structure and the information represented in those section varies heavily. Therefore, the section templates are with respect to the individual classes in our experimental setup (see Figure FIGREF42 ). DISPLAYFORM0
Generating section templates has two main advantages. Firstly, by considering class-based profiles, we can overcome the problem of incomplete individual entity profiles and thereby are able to suggest news articles to sections that do not yet exist in a specific entity INLINEFORM0 . The second advantage is that we are able to canonicalize the sections, i.e. `Early Life' and `Early Life and Childhood' would be treated similarly.
To generate the section template INLINEFORM0 , we extract all sections from entities of a given type INLINEFORM1 at year INLINEFORM2 . Next, we cluster the entity sections, based on an extended version of k–means clustering BIBREF18 , namely x–means clustering introduced in Pelleg et al. which estimates the number of clusters efficiently BIBREF19 . As a similarity metric we use the cosine similarity computed based on the tf–idf models of the sections. Using the x–means algorithm we overcome the requirement to provide the number of clusters k beforehand. x–means extends the k–means algorithm, such that a user only specifies a range [ INLINEFORM3 , INLINEFORM4 ] that the number of clusters may reasonably lie in.
The learning objective of INLINEFORM0 is to determine the overall fit of a news article INLINEFORM1 to one of the sections in a given section template INLINEFORM2 . The template is pre-determined by the class of the entity for which the news is suggested as relevant by INLINEFORM3 . In all cases, we measure how well INLINEFORM4 fits each of the sections INLINEFORM5 as well as the specific entity section INLINEFORM6 . The section profiles in INLINEFORM7 represent the aggregated entity profiles from all entities of class INLINEFORM8 at year INLINEFORM9 .
To learn INLINEFORM0 we rely on a variety of features that consider several similarity aspects as shown in Table TABREF31 . For the sake of simplicity we do not make the distinction in Table TABREF31 between the individual entity section and class-based section similarities, INLINEFORM1 and INLINEFORM2 , respectively. Bear in mind that an entity section INLINEFORM3 might be present at year INLINEFORM4 but not at year INLINEFORM5 (see for more details the discussion on entity profile expansion in Section UID69 ).
Topic. We use topic similarities to ensure (i) that the content of INLINEFORM0 fits topic-wise with a specific section text and (ii) that it has a similar topic to previously referred news articles in that section. In a pre-processing stage we compute the topic models for the news articles, entity sections INLINEFORM1 and the aggregated class-based sections in INLINEFORM2 . The topic models are computed using LDA BIBREF20 . We only computed a single topic per article/section as we are only interested in topic term overlaps between article and sections. We distinguish two main features: the first feature measures the overlap of topic terms between INLINEFORM3 and the entity section INLINEFORM4 and INLINEFORM5 , and the second feature measures the overlap of the topic model of INLINEFORM6 against referred news articles in INLINEFORM7 at time INLINEFORM8 .
Syntactic. These features represent a mechanism for conveying the importance of a specific text snippet, solely based on the frequency of specific POS tags (i.e. NNP, CD etc.), as commonly used in text summarization tasks. Following the same intuition as in BIBREF8 , we weigh the importance of articles by the count of specific POS tags. We expect that for different sections, the importance of POS tags will vary. We measure the similarity of POS tags in a news article against the section text. Additionally, we consider bi-gram and tri-gram POS tag overlap. This exploits similarity in syntactical patterns between the news and section text.
Lexical. As lexical features, we measure the similarity of INLINEFORM0 against the entity section text INLINEFORM1 and the aggregate section text INLINEFORM2 . Further, we distinguish between the overall similarity of INLINEFORM3 and that of the different news paragraphs ( INLINEFORM4 which denotes the paragraphs of INLINEFORM5 up to the 5th paragraph). A higher similarity on the first paragraphs represents a more confident indicator that INLINEFORM6 should be suggested to a specific section INLINEFORM7 . We measure the similarity based on two metrics: (i) the KL-divergence between the computed language models and (ii) cosine similarity of the corresponding paragraph text INLINEFORM8 and section text.
Entity-based. Another feature set we consider is the overlap of named entities and their corresponding entity classes. For different entity sections, we expect to find a particular set of entity classes that will correlate with the section, e.g. `Early Life' contains mostly entities related to family, school, universities etc.
Frequency. Finally, we gather statistics about the number of entities, paragraphs, news article length, top– INLINEFORM0 entities and entity classes, and the frequency of different POS tags. Here we try to capture patterns of articles that are usually cited in specific sections.
## Evaluation Plan
In this section we outline the evaluation plan to verify the effectiveness of our learning approaches. To evaluate the news suggestion problem we are faced with two challenges.
What comprises the ground truth for such a task ?
How do we construct training and test splits given that entity pages consists of text added at different points in time ?
Consider the ground truth challenge. Evaluating if an arbitrary news article should be included in Wikipedia is both subjective and difficult for a human if she is not an expert. An invasive approach, which was proposed by Barzilay and Sauper BIBREF8 , adds content directly to Wikipedia and expects the editors or other users to redact irrelevant content over a period of time. The limitations of such an evaluation technique is that content added to long-tail entities might not be evaluated by informed users or editors in the experiment time frame. It is hard to estimate how much time the added content should be left on the entity page. A more non-invasive approach could involve crowdsourcing of entity and news article pairs in an IR style relevance assessment setup. The problem of such an approach is again finding knowledgeable users or experts for long-tail entities. Thus the notion of relevance of a news recommendation is challenging to evaluate in a crowd setup.
We take a slightly different approach by making an assumption that the news articles already present in Wikipedia entity pages are relevant. To this extent, we extract a dataset comprising of all news articles referenced in entity pages (details in Section SECREF40 ). At the expense of not evaluating the space comprising of news articles absent in Wikipedia, we succeed in (i) avoiding restrictive assumptions about the quality of human judgments, (ii) being invasive and polluting Wikipedia, and (iii) deriving a reusable test bed for quicker experimentation.
The second challenge of construction of training and test set separation is slightly easier and is addressed in Section SECREF46 .
## Datasets
The datasets we use for our experimental evaluation are directly extracted from the Wikipedia entity pages and their revision history. The generated data represents one of the contributions of our paper. The datasets are the following:
Entity Classes. We focus on a manually predetermined set of entity classes for which we expect to have news coverage. The number of analyzed entity classes is 27, including INLINEFORM0 entities with at least one news reference. The entity classes were selected from the DBpedia class ontology. Figure FIGREF42 shows the number of entities per class for the years (2009-2014).
News Articles. We extract all news references from the collected Wikipedia entity pages. The extracted news references are associated with the sections in which they appear. In total there were INLINEFORM0 news references, and after crawling we end up with INLINEFORM1 successfully crawled news articles. The details of the news article distribution, and the number of entities and sections from which they are referred are shown in Table TABREF44 .
Article-Entity Ground-truth. The dataset comprises of the news and entity pairs INLINEFORM0 . News-entity pairs are relevant if the news article is referenced in the entity page. Non-relevant pairs (i.e. negative training examples) consist of news articles that contain an entity but are not referenced in that entity's page. If a news article INLINEFORM1 is referred from INLINEFORM2 at year INLINEFORM3 , the features are computed taking into account the entity profiles at year INLINEFORM4 .
Article-Section Ground-truth. The dataset consists of the triple INLINEFORM0 , where INLINEFORM1 , where we assume that INLINEFORM2 has already been determined as relevant. We therefore have a multi-class classification problem where we need to determine the section of INLINEFORM3 where INLINEFORM4 is cited. Similar to the article-entity ground truth, here too the features compute the similarity between INLINEFORM5 , INLINEFORM6 and INLINEFORM7 .
## Data Pre-Processing
We POS-tag the news articles and entity profiles INLINEFORM0 with the Stanford tagger BIBREF21 . For entity linking the news articles, we use TagMe! BIBREF22 with a confidence score of 0.3. On a manual inspection of a random sample of 1000 disambiguated entities, the accuracy is above 0.9. On average, the number of entities per news article is approximately 30. For entity linking the entity profiles, we simply follow the anchor text that refers to Wikipedia entities.
## Train and Testing Evaluation Setup
We evaluate the generated supervised models for the two tasks, AEP and ASP, by splitting the train and testing instances. It is important to note that for the pairs INLINEFORM0 and the triple INLINEFORM1 , the news article INLINEFORM2 is referenced at time INLINEFORM3 by entity INLINEFORM4 , while the features take into account the entity profile at time INLINEFORM5 . This avoids any `overlapping' content between the news article and the entity page, which could affect the learning task of the functions INLINEFORM6 and INLINEFORM7 . Table TABREF47 shows the statistics of train and test instances. We learn the functions at year INLINEFORM8 and test on instances for the years greater than INLINEFORM9 . Please note that we do not show the performance for year 2014 as we do not have data for 2015 for evaluation.
## Article-Section Placement
Here we show the evaluation setup for ASP task and discuss the results with a focus on three main aspects, (i) the overall performance across the years, (ii) the entity class specific performance, and (iii) the impact on entity profile expansion by suggesting missing sections to entities based on the pre-computed templates.
Baselines. To the best of our knowledge, we are not aware of any comparable approach for this task. Therefore, the baselines we consider are the following:
S1: Pick the section from template INLINEFORM0 with the highest lexical similarity to INLINEFORM1 : S1 INLINEFORM2
S2: Place the news into the most frequent section in INLINEFORM0
Learning Models. We use Random Forests (RF) BIBREF23 and Support Vector Machines (SVM) BIBREF24 . The models are optimized taking into account the features in Table TABREF31 . In contrast to the AEP task, here the scale of the number of instances allows us to learn the SVM models. The SVM model is optimized using the INLINEFORM0 loss function and uses the Gaussian kernels.
Metrics. We compute precision P as the ratio of news for which we pick a section INLINEFORM0 from INLINEFORM1 and INLINEFORM2 conforms to the one in our ground-truth (see Section SECREF40 ). The definition of recall R and F1 score follows from that of precision.
Figure FIGREF66 shows the overall performance and a comparison of our approach (when INLINEFORM0 is optimized using SVM) against the best performing baseline S2. With the increase in the number of training instances for the ASP task the performance is a monotonically non-decreasing function. For the year 2009, we optimize the learning objective of INLINEFORM1 with around 8% of the total instances, and evaluate on the rest. The performance on average is around P=0.66 across all classes. Even though for many classes the performance is already stable (as we will see in the next section), for some classes we improve further. If we take into account the years between 2010 and 2012, we have an increase of INLINEFORM2 P=0.17, with around 70% of instances used for training and the remainder for evaluation. For the remaining years the total improvement is INLINEFORM3 P=0.18 in contrast to the performance at year 2009.
On the other hand, the baseline S1 has an average precision of P=0.12. The performance across the years varies slightly, with the year 2011 having the highest average precision of P=0.13. Always picking the most frequent section as in S2, as shown in Figure FIGREF66 , results in an average precision of P=0.17, with a uniform distribution across the years.
Here we show the performance of INLINEFORM0 decomposed for the different entity classes. Specifically we analyze the 27 classes in Figure FIGREF42 . In Table TABREF68 , we show the results for a range of years (we omit showing all years due to space constraints). For illustration purposes only, we group them into four main classes ( INLINEFORM1 Person, Organization, Location, Event INLINEFORM2 ) and into the specific sub-classes shown in the second column in Table TABREF68 . For instance, the entity classes OfficeHolder and Politician are aggregated into Person–Politics.
It is evident that in the first year the performance is lower in contrast to the later years. This is due to the fact that as we proceed, we can better generalize and accurately determine the correct fit of an article INLINEFORM0 into one of the sections from the pre-computed templates INLINEFORM1 . The results are already stable for the year range INLINEFORM2 . For a few Person sub-classes, e.g. Politics, Entertainment, we achieve an F1 score above 0.9. These additionally represent classes with a sufficient number of training instances for the years INLINEFORM3 . The lowest F1 score is for the Criminal and Television classes. However, this is directly correlated with the insufficient number of instances.
The baseline approaches for the ASP task perform poorly. S1, based on lexical similarity, has a varying performance for different entity classes. The best performance is achieved for the class Person – Politics, with P=0.43. This highlights the importance of our feature choice and that the ASP cannot be considered as a linear function, where the maximum similarity yields the best results. For different entity classes different features and combination of features is necessary. Considering that S2 is the overall best performing baseline, through our approach INLINEFORM0 we have a significant improvement of over INLINEFORM1 P=+0.64.
The models we learn are very robust and obtain high accuracy, fulfilling our pre-condition for accurate news suggestions into the entity sections. We measure the robustness of INLINEFORM0 through the INLINEFORM1 statistic. In this case, we have a model with roughly 10 labels (corresponding to the number of sections in a template INLINEFORM2 ). The score we achieve shows that our model predicts with high confidence with INLINEFORM3 .
The last analysis is the impact we have on expanding entity profiles INLINEFORM0 with new sections. Figure FIGREF70 shows the ratio of sections for which we correctly suggest an article INLINEFORM1 to the right section in the section template INLINEFORM2 . The ratio here corresponds to sections that are not present in the entity profile at year INLINEFORM3 , that is INLINEFORM4 . However, given the generated templates INLINEFORM5 , we can expand the entity profile INLINEFORM6 with a new section at time INLINEFORM7 . In details, in the absence of a section at time INLINEFORM8 , our model trains well on similar sections from the section template INLINEFORM9 , hence we can predict accurately the section and in this case suggest its addition to the entity profile. With time, it is obvious that the expansion rate decreases at later years as the entity profiles become more `complete'.
This is particularly interesting for expanding the entity profiles of long-tail entities as well as updating entities with real-world emerging events that are added constantly. In many cases such missing sections are present at one of the entities of the respective entity class INLINEFORM0 . An obvious case is the example taken in Section SECREF16 , where the `Accidents' is rather common for entities of type Airline. However, it is non-existent for some specific entity instances, i.e Germanwings airline.
Through our ASP approach INLINEFORM0 , we are able to expand both long-tail and trunk entities. We distinguish between the two types of entities by simply measuring their section text length. The real distribution in the ground truth (see Section SECREF40 ) is 27% and 73% are long-tail and trunk entities, respectively. We are able to expand the entity profiles for both cases and all entity classes without a significant difference, with the only exception being the class Creative Work, where we expand significantly more trunk entities.
## Conclusion and Future Work
In this work, we have proposed an automated approach for the novel task of suggesting news articles to Wikipedia entity pages to facilitate Wikipedia updating. The process consists of two stages. In the first stage, article–entity placement, we suggest news articles to entity pages by considering three main factors, such as entity salience in a news article, relative authority and novelty of news articles for an entity page. In the second stage, article–section placement, we determine the best fitting section in an entity page. Here, we remedy the problem of incomplete entity section profiles by constructing section templates for specific entity classes. This allows us to add missing sections to entity pages. We carry out an extensive experimental evaluation on 351,983 news articles and 73,734 entities coming from 27 distinct entity classes. For the first stage, we achieve an overall performance with P=0.93, R=0.514 and F1=0.676, outperforming our baseline competitors significantly. For the second stage, we show that we can learn incrementally to determine the correct section for a news article based on section templates. The overall performance across different classes is P=0.844, R=0.885 and F1=0.860.
In the future, we will enhance our work by extracting facts from the suggested news articles. Results suggest that the news content cited in entity pages comes from the first paragraphs. However, challenging task such as the canonicalization and chronological ordering of facts, still remain.
| 13
|
1704.00939
|
Fortia-FBK at SemEval-2017 Task 5: Bullish or Bearish? Inferring Sentiment towards Brands from Financial News Headlines
|
# Fortia-FBK at SemEval-2017 Task 5: Bullish or Bearish? Inferring Sentiment towards Brands from Financial News Headlines
## Abstract
In this paper, we describe a methodology to infer Bullish or Bearish sentiment towards companies/brands. More specifically, our approach leverages affective lexica and word embeddings in combination with convolutional neural networks to infer the sentiment of financial news headlines towards a target company. Such architecture was used and evaluated in the context of the SemEval 2017 challenge (task 5, subtask 2), in which it obtained the best performance.
## Introduction
Real time information is key for decision making in highly technical domains such as finance. The explosive growth of financial technology industry (Fintech) continued in 2016, partially due to the current interest in the market for Artificial Intelligence-based technologies.
Opinion-rich texts such as micro-blogging and news can have an important impact in the financial sector (e.g. raise or fall in stock value) or in the overall economy (e.g. the Greek public debt crisis). In such a context, having granular access to the opinions of an important part of the population is of key importance to any public and private actor in the field. In order to take advantage of this raw data, it is thus needed to develop machine learning methods allowing to convert unstructured text into information that can be managed and exploited.
In this paper, we address the sentiment analysis problem applied to financial headlines, where the goal is, for a given news headline and target company, to infer its polarity score i.e. how positive (or negative) the sentence is with respect to the target company. Previous research BIBREF0 has highlighted the association between news items and market fluctiations; hence, in the financial domain, sentiment analysis can be used as a proxy for bullish (i.e. positive, upwards trend) or bearish (i.e. negative, downwards trend) attitude towards a specific financial actor, allowing to identify and monitor in real-time the sentiment associated with e.g. stocks or brands.
Our contribution leverages pre-trained word embeddings (GloVe, trained on wikipedia+gigaword corpus), the DepecheMood affective lexicon, and convolutional neural networks.
## Related Works
While image and sound come with a natural high dimensional embedding, the issue of which is the best representation is still an open research problem in the context of natural language and text. It is beyond the scope of this paper to do a thorough overview of word representations, for this we refer the interest reader to the excellent review provided by BIBREF1 . Here, we will just introduce the main representations that are related to the proposed method.
## Data
The data consists of a set of financial news headlines, crawled from several online outlets such as Yahoo Finance, where each sentence contains one or more company names/brands.
Each tuple (headline, company) is annotated with a sentiment score ranging from -1 (very negative, bearish) to 1 (very positive, bullish). The training/test sets provided contain 1142 and 491 annotated sentences, respectively.
A sample instance is reported below:
Headline: “Morrisons book second consecutive quarter of sales growth”
Company name: “Morrisons”
Sentiment score: 0.43
## Method
In Figure FIGREF5 , we can see the overall architecture of our model.
## Sentence representation and preprocessing
Minimal preprocessing was adopted in our approach: we replaced the target company's name with a fixed word <company> and numbers with <number>. The sentences were then tokenized using spaces as separator and keeping punctuation symbols as separate tokens.
The words are represented as fixed length vectors INLINEFORM0 resulting from the concatenation of GloVe pre-trained embeddings and DepecheMood BIBREF19 lexicon representation. Since we cannot directly concatenate token-based embeddings (provided in GloVe) with the lemma#PoS-based representation available in DepecheMood, we proceeded to re-build the latter in token-based form, applying the exact same methodology albeit with two differences: we started from a larger dataset (51.9K news articles instead of 25.3K) and used a frequency cut-off, i.e. keeping only those tokens that appear at least 5 times in the corpus.
These word-level representation are used as the first layer of our network. During training we allow the weights of the representation to be updated. We further add the VADER score for the sentence under analysis. The complete sentence representation is presented in Algorithm UID8 .
InputInput OutputOutput The sentence embedding INLINEFORM0 INLINEFORM1
INLINEFORM0 in INLINEFORM1 INLINEFORM2 = [GloVe( INLINEFORM3 , INLINEFORM4 ), DepecheMood( INLINEFORM5 )] INLINEFORM6 Sentence representation
## Architectural Details
A 1D convolutional layer with filters of multiple sizes {2, 3, 4} is applied to the sequence of word embeddings. The filters are used to learn useful translation-invariant representations of the sequential input data. A global max-pooling is then applied across the sequence for each filter output.
We apply the concatenation layer to the output of the global max-pooling and the output of VADER.
The activation function used between layers is ReLU BIBREF24 except for the out layer where tanh is used to map the output into [-1, 1] range.
Dropout BIBREF25 was used to avoid over-fitting to the training data: it prevents the co-adaptation of the neurones and it also provides an inexpensive way to average an exponential number of networks. In addition, we averaged the output of multiple networks with the same architecture but trained independently with different random seeds in order to reduce noise.
The loss function used is the cosine distance between the predicted scores and the gold standard for each batch. Even though stochastic optimization methods like Adam BIBREF26 are usually applied to loss functions that are written as a sum of per-sample loss, which is not the case for the cosine, it converges to an acceptable solution. The loss can be written as : DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are the predicted and true sentiment scores for batch INLINEFORM2 , respectively.
The algorithm for training/testing our model is reported in Algorithm UID15 .
InputInput OutputOutput ParameterParameters A set of trained models INLINEFORM0 , and the predictions INLINEFORM1 for the test set INLINEFORM2 The number INLINEFORM3 of models to train INLINEFORM4 see sec 3.1 INLINEFORM5 in INLINEFORM6 INLINEFORM7 ) see Alg. UID8 INLINEFORM8 INLINEFORM9 see Eq. EQREF16 INLINEFORM10 INLINEFORM11 INLINEFORM12 Training/Testing algorithm. To build our model, we set N=10.
## Results
In this section, we report the results obtained by our model according to challenge official evaluation metric, which is based cosine-similarity and described in BIBREF27 . Results are reported for three diverse configurations: (i) the full system; (ii) the system without using word embeddings (i.e. Glove and DepecheMood); and (iii) the system without using pre-processing. In Table TABREF17 we show model's performances on the challenge training data, in a 5-fold cross-validation setting.
Further, the final performances obtained with our approach on the challenge test set are reported in Table TABREF18 . Consistently with the cross-validation performances shown earlier, we observe the beneficial impact of word-representations and basic pre-processing.
## Conclusions
In this paper, we presented the network architecture used for the Fortia-FBK submission to the Semeval-2017 Task 5, Subtask 2 challenge, with the goal of predicting positive (bullish) or negative (bearish) attitude towards a target brand from financial news headlines. The proposed system ranked 1st in such challenge.
Our approach is based on 1d convolutions and uses fine-tuning of unsupervised word representations and a rule based sentiment model in its inputs. We showed that the use of pre-computed word representations allows to reduce over-fitting and to achieve significantly better generalization, while some basic pre-processing was needed to further improve the performance.
| 8
|
1704.02686
|
Word Embeddings via Tensor Factorization
|
# Word Embeddings via Tensor Factorization
## Abstract
Most popular word embedding techniques involve implicit or explicit factorization of a word co-occurrence based matrix into low rank factors. In this paper, we aim to generalize this trend by using numerical methods to factor higher-order word co-occurrence based arrays, or \textit{tensors}. We present four word embeddings using tensor factorization and analyze their advantages and disadvantages. One of our main contributions is a novel joint symmetric tensor factorization technique related to the idea of coupled tensor factorization. We show that embeddings based on tensor factorization can be used to discern the various meanings of polysemous words without being explicitly trained to do so, and motivate the intuition behind why this works in a way that doesn't with existing methods. We also modify an existing word embedding evaluation metric known as Outlier Detection [Camacho-Collados and Navigli, 2016] to evaluate the quality of the order-$N$ relations that a word embedding captures, and show that tensor-based methods outperform existing matrix-based methods at this task. Experimentally, we show that all of our word embeddings either outperform or are competitive with state-of-the-art baselines commonly used today on a variety of recent datasets. Suggested applications of tensor factorization-based word embeddings are given, and all source code and pre-trained vectors are publicly available online.
## Introduction
Word embeddings have been used to improve the performance of many NLP tasks including language modelling BIBREF1 , machine translation BIBREF2 , and sentiment analysis BIBREF3 . The broad applicability of word embeddings to NLP implies that improvements to their quality will likely have widespread benefits for the field.
The word embedding problem is to learn a mapping INLINEFORM0 ( INLINEFORM1 100-300 in most applications) that encodes meaningful semantic and/or syntactic information. For instance, in many word embeddings, INLINEFORM2 car INLINEFORM3 truck INLINEFORM4 , since the words are semantically similar.
More complex relationships than similarity can also be encoded in word embeddings. For example, we can answer analogy queries of the form INLINEFORM0 ? using simple arithmetic in many state-of-the-art embeddings BIBREF4 . The answer to bed INLINEFORM1 sleep INLINEFORM2 chair INLINEFORM3 INLINEFORM4 is given by the word whose vector representation is closest to INLINEFORM5 sleep INLINEFORM6 bed INLINEFORM7 chair INLINEFORM8 ( INLINEFORM9 sit INLINEFORM10 ). Other embeddings may encode such information in a nonlinear way BIBREF5 .
BIBREF4 demonstrates the additive compositionality of their word2vec vectors: one can sum vectors produced by their embedding to compute vectors for certain phrases rather than just vectors for words. Later in this paper, we will show that our embeddings naturally give rise to a form of multiplicative compositionality that has not yet been explored in the literature.
Almost all recent word embeddings rely on the distributional hypothesis BIBREF6 , which states that a word's meaning can be inferred from the words that tend to surround it. To utilize the distributional hypothesis, many embeddings are given by a low-rank factor of a matrix derived from co-occurrences in a large unsupervised corpus, see BIBREF7 , BIBREF8 , BIBREF9 and BIBREF10 .
Approaches that rely on matrix factorization only utilize pairwise co-occurrence information in the corpus. We aim to extend this approach by creating word embeddings given by factors of tensors containing higher order co-occurrence data.
## Related work
Some common word embeddings related to co-occurrence based matrix factorization include GloVe BIBREF7 , word2vec BIBREF9 , LexVec BIBREF10 , and NNSE BIBREF8 . In contrast, our work studies word embeddings given by factorization of tensors. An overview of tensor factorization methods is given in BIBREF11 .
Our work uses factorization of symmetric nonnegative tensors, which has been studied in the past BIBREF12 , BIBREF13 . In general, factorization of tensors has been applied to NLP in BIBREF14 and factorization of nonnegative tensors BIBREF15 . Recently, factorization of symmetric tensors has been used to create a generic word embedding BIBREF16 but the idea was not explored extensively. Our work studies this idea in much greater detail, fully demonstrating the viability of tensor factorization as a technique for training word embeddings.
Composition of word vectors to create novel representations has been studied in depth, including additive, multiplicative, and tensor-based methods BIBREF17 , BIBREF18 . Typically, composition is used to create vectors that represent phrases or sentences. Our work, instead, shows that pairs of word vectors can be composed multiplicatively to create different vector representations for the various meanings of a single polysemous word.
## Notation
Throughout this paper we will write scalars in lowercase italics INLINEFORM0 , vectors in lowercase bold letters INLINEFORM1 , matrices with uppercase bold letters INLINEFORM2 , and tensors (of order INLINEFORM3 ) with Euler script notation INLINEFORM4 , as is standard in the literature.
## Pointwise Mutual Information
Pointwise mutual information (PMI) is a useful property in NLP that quantifies the likelihood that two words co-occur BIBREF9 . It is defined as: INLINEFORM0
where INLINEFORM0 is the probability that INLINEFORM1 and INLINEFORM2 occur together in a given fixed-length context window in the corpus, irrespective of order.
It is often useful to consider the positive PMI (PPMI), defined as: INLINEFORM0
since negative PMI values have little grounded interpretation BIBREF19 , BIBREF9 , BIBREF15 .
Given an indexed vocabulary INLINEFORM0 , one can construct a INLINEFORM1 PPMI matrix INLINEFORM2 where INLINEFORM3 . Many existing word embedding techniques involve factorizing this PPMI matrix BIBREF9 , BIBREF8 , BIBREF10 .
PMI can be generalized to INLINEFORM0 variables. While there are many ways to do so BIBREF20 , in this paper we use the form defined by: INLINEFORM1
where INLINEFORM0 is the probability that all of INLINEFORM1 occur together in a given fixed-length context window in the corpus, irrespective of their order.
In this paper we study 3-way PPMI tensors INLINEFORM0 , where INLINEFORM1 , as this is the natural higher-order generalization of the PPMI matrix. We leave the study of creating word embeddings with INLINEFORM2 -dimensional PPMI tensors ( INLINEFORM3 ) to future work.
## Tensor factorization
Just as the rank- INLINEFORM0 matrix decomposition is defined to be the product of two factor matrices ( INLINEFORM1 ), the canonical rank- INLINEFORM2 tensor decomposition for a third order tensor is defined to be the product of three factor matrices BIBREF11 : DISPLAYFORM0
where INLINEFORM0 is the outer product: INLINEFORM1 . This is also commonly referred to as the rank-R CP Decomposition. Elementwise, this is written as: INLINEFORM2
where INLINEFORM0 is elementwise vector multiplication and INLINEFORM1 is the INLINEFORM2 row of INLINEFORM3 . In our later section on multiplicative compositionality, we will see this formulation gives rise to a meaningful interpretation of the elementwise product between vectors in our word embeddings.
Symmetric CP Decomposition. In this paper, we will consider symmetric CP decomposition of nonnegative tensors BIBREF21 , BIBREF11 . Since our INLINEFORM0 -way PPMI is nonnegative and invariant under permutation, the PPMI tensor INLINEFORM1 is nonnegative and supersymmetric, i.e. INLINEFORM2 for any permutation INLINEFORM3 .
In the symmetric CP decomposition, instead of factorizing INLINEFORM0 , we factorize INLINEFORM1 as the triple product of a single factor matrix INLINEFORM2 such that INLINEFORM3
In this formulation, we use INLINEFORM0 to be the word embedding so the vector for INLINEFORM1 is the INLINEFORM2 row of INLINEFORM3 similar to the formulations in BIBREF9 , BIBREF8 , BIBREF7 .
It is known that the optimal rank- INLINEFORM0 CP decomposition exists for symmetric nonnegative tensors such as the PPMI tensor BIBREF21 . However, finding such a decomposition is NP hard in general BIBREF22 so we must consider approximate methods.
In this work, we only consider the symmetric CP decomposition, leaving the study of other tensor decompositions (such as the Tensor Train or HOSVD BIBREF23 , BIBREF11 ) to future work.
## Computing the Symmetric CP Decomposition
The INLINEFORM0 size of the third order PPMI tensor presents a number of computational challenges. In practice, INLINEFORM1 can vary from INLINEFORM2 to INLINEFORM3 , resulting in a tensor whose naive representation requires at least INLINEFORM4 bytes = 4 TB of floats. Even the sparse representation of the tensor takes up such a large fraction of memory that standard algorithms such as successive rank-1 approximation BIBREF12 , BIBREF24 and alternating least-squares BIBREF11 are infeasible for our uses. Thus, in this paper we will consider a stochastic online formulation similar to that of BIBREF25 .
We optimize the CP decomposition in an online fashion, using small random subsets INLINEFORM0 of the nonzero tensor entries to update the decomposition at time INLINEFORM1 . In this minibatch setting, we optimize the decomposition based on the current minibatch and the previous decomposition at time INLINEFORM2 . To update INLINEFORM3 (and thus the symmetric decomposition), we first define a decomposition loss INLINEFORM4 and minimize this loss with respect to INLINEFORM5 using Adam BIBREF26 .
At each time INLINEFORM0 , we take INLINEFORM1 to be all co-occurrence triples (weighted by PPMI) in a fixed number of sentences (around 1,000) from the corpus. We continue training until we have depleted the entire corpus.
For INLINEFORM0 to accurately model INLINEFORM1 , we also include a certain proportion of elements with zero PPMI (or “negative samples”) in INLINEFORM2 , similar to that of BIBREF10 . We use an empirically found proportion of negative samples for training, and leave discovery of the optimal negative sample proportion to future work.
## Word Embedding Proposals
CP-S. The first embedding we propose is based on symmetic CP decomposition of the PPMI tensor INLINEFORM0 as discussed in the mathematical preliminaries section. The optimal setting for the word embedding INLINEFORM1 is: INLINEFORM2
Since we cannot feasibly compute this exactly, we minimize the loss function defined as the squared error between the values in INLINEFORM0 and their predicted values: INLINEFORM1
using the techniques discussed in the previous section.
JCP-S. A potential problem with CP-S is that it is only trained on third order information. To rectify this issue, we propose a novel joint tensor factorization problem we call Joint Symmetric Rank- INLINEFORM0 CP Decomposition. In this problem, the input is the fixed rank INLINEFORM1 and a list of supersymmetric tensors INLINEFORM2 of different orders but whose axis lengths all equal INLINEFORM3 . Each tensor INLINEFORM4 is to be factorized via rank- INLINEFORM5 symmetric CP decomposition using a single INLINEFORM6 factor matrix INLINEFORM7 .
To produce a solution, we first define the loss at time INLINEFORM0 to be the sum of the reconstruction losses of each different tensor: INLINEFORM1
where INLINEFORM0 is an INLINEFORM1 -dimensional supersymmetric PPMI tensor. We then minimize the loss with respect to INLINEFORM2 . Since we are using at most third order tensors in this work, we assign our word embedding INLINEFORM3 to be: INLINEFORM4
This problem is a specific instance of Coupled Tensor Decomposition, which has been studied in the past BIBREF27 , BIBREF28 . In this problem, the goal is to factorize multiple tensors using at least one factor matrix in common. A similar formulation to our problem can be found in BIBREF29 , which studies blind source separation using the algebraic geometric aspects of jointly factorizing numerous supersymmetric tensors (to unknown rank). In contrast to our work, they outline some generic rank properties of such a decomposition rather than attacking the problem numerically. Also, in our formulation the rank is fixed and an approximate solution must be found. Exploring the connection between the theoretical aspects of joint decomposition and quality of word embeddings would be an interesting avenue for future work.
To the best of our knowledge this is the first study of Joint Symmetric Rank- INLINEFORM0 CP Decomposition.
## Shifted PMI
In the same way BIBREF9 considers factorization of positive shifted PMI matrices, we consider factorization of positive shifted PMI tensors INLINEFORM0 , where INLINEFORM1 for some constant shift INLINEFORM2 . We empirically found that different levels of shifting resulted in different qualities of word embeddings – the best shift we found for CP-S was a shift of INLINEFORM3 , whereas any nonzero shift for JCP-S resulted in a worse embedding across the board. When we discuss evaluation we report the results given by factorization of the PPMI tensors shifted by the best value we found for each specific embedding.
## Computational notes
When considering going from two dimensions to three, it is perhaps necessary to discuss the computational issues in such a problem size increase. However, it should be noted that the creation of pre-trained embeddings can be seen as a pre-processing step for many future NLP tasks, so if the training can be completed once, it can be used forever thereafter without having to take training time into account. Despite this, we found that the training of our embeddings was not considerably slower than the training of order-2 equivalents such as SGNS. Explicitly, our GPU trained CBOW vectors (using the experimental settings found below) in 3568 seconds, whereas training CP-S and JCP-S took 6786 and 8686 seconds respectively.
## Evaluation
In this section we present a quantitative evaluation comparing our embeddings to an informationless embedding and two strong baselines. Our baselines are:
For a fair comparison, we trained each model on the same corpus of 10 million sentences gathered from Wikipedia. We removed stopwords and words appearing fewer than 2,000 times (130 million tokens total) to reduce noise and uninformative words. Our word2vec and NNSE baselines were trained using the recommended hyperparameters from their original publications, and all optimizers were using using the default settings. Hyperparameters are always consistent across evaluations.
Because of the dataset size, the results shown should be considered a proof of concept rather than an objective comparison to state-of-the-art pre-trained embeddings. Due to the natural computational challenges arising from working with tensors, we leave creation of a full-scale production ready embedding based on tensor factorization to future work.
As is common in the literature BIBREF4 , BIBREF8 , we use 300-dimensional vectors for our embeddings and all word vectors are normalized to unit length prior to evaluation.
## Quantitative tasks
Outlier Detection. The Outlier Detection task BIBREF0 is to determine which word in a list INLINEFORM0 of INLINEFORM1 words is unrelated to the other INLINEFORM2 which were chosen to be related. For each INLINEFORM3 , one can compute its compactness score INLINEFORM4 , which is the compactness of INLINEFORM5 . INLINEFORM6 is explicitly computed as the mean similarity of all word pairs INLINEFORM7 . The predicted outlier is INLINEFORM8 , as the INLINEFORM9 related words should form a compact cluster with high mean similarity.
We use the WikiSem500 dataset BIBREF30 which includes sets of INLINEFORM0 words per group gathered based on semantic similarity. Thus, performance on this task is correlated with the amount of semantic information encoded in a word embedding. Performance on this dataset was shown to be well-correlated with performance at the common NLP task of sentiment analysis BIBREF30 .
The two metrics associated with this task are accuracy and Outlier Position Percentage (OPP). Accuracy is the fraction of cases in which the true outlier correctly had the highest compactness score. OPP measures how close the true outlier was to having the highest compactness score, rewarding embeddings more for predicting the outlier to be in 2nd place rather than INLINEFORM0 when sorting the words by their compactness score INLINEFORM1 .
3-way Outlier Detection. As our tensor-based embeddings encode higher order relationships between words, we introduce a new way to compute INLINEFORM0 based on groups of 3 words rather than pairs of words. We define the compactness score for a word INLINEFORM1 to be: INLINEFORM2
where INLINEFORM0 denotes similarity between a group of 3 vectors. INLINEFORM1 is defined as: INLINEFORM2
We call this evaluation method OD3.
The purpose of OD3 is to evaluate the extent to which an embedding captures 3rd order relationships between words. As we will see in the results of our quantitative experiments, our tensor methods outperform the baselines on OD3, which validates our approach.
This approach can easily be generalized to OD INLINEFORM0 INLINEFORM1 , but again we leave the study of higher order relationships to future work.
Simple supervised tasks. BIBREF5 points out that the primary application of word embeddings is transfer learning to NLP tasks. They argue that to evaluate an embedding's ability to transfer information to a relevant task, one must measure the embedding's accessibility of information for actual downstream tasks. To do so, one must cite the performance of simple supervised tasks as training set size increases, which is commonly done in transfer learning evaluation BIBREF5 . If an algorithm using a word embedding performs well with just a small amount of training data, then the information encoded in the embedding is easily accessible.
The simple supervised downstream tasks we use to evaluate the embeddings are as follows:
Supervised Analogy Recovery. We consider the task of solving queries of the form a : b :: c : ? using a simple neural network as suggested in BIBREF5 . The analogy dataset we use is from the Google analogy testbed BIBREF4 .
Sentiment analysis. We also consider sentiment analysis as described by BIBREF31 . We use the suggested Large Movie Review dataset BIBREF32 , containing 50,000 movie reviews.
All code is implemented using scikit-learn or TensorFlow and uses the suggested train/test split.
Word similarity. To standardize our evaluation methodology, we evaluate the embeddings using word similarity on the common MEN and MTurk datasets BIBREF33 , BIBREF34 . For an overview of word similarity evaluation, see BIBREF31 .
## Quantitative results
Outlier Detection results. The results are shown in Table TABREF20 . The first thing to note is that CP-S outperforms the other methods across each Outlier Detection metric. Since the WikiSem500 dataset is semantically focused, performance at this task demonstrates the quality of semantic information encoded in our embeddings.
On OD2, the baselines perform more competitively with our CP Decomposition based models, but when OD3 is considered our methods clearly excel. Since the tensor-based methods are trained directly on third order information and perform much better at OD3, we see that OD3 scores reflect the amount of third order information in a word embedding. This is a validation of OD3, as our 3rd order embeddings would naturally out perform 2nd order embeddings at a task that requires third order information. Still, the superiority of our tensor-based embeddings at OD2 demonstrates the quality of the semantic information they encode.
Supervised analogy results. The results are shown in Figure FIGREF18 . At the supervised semantic analogy task, CP-S vastly outperforms the baselines at all levels of training data, further signifying the amount of semantic information encoded by this embedding technique.
Also, when only 10% of the training data is presented, our tensor methods are the only ones that attain nonzero performance – even in such a limited data setting, use of CP-S's vectors results in nearly 40% accuracy. This phenomenon is also observed in the syntactic analogy tasks: our embeddings consistently outperform the others until 100% of the training data is presented. These two observations demonstrate the accessibility of the information encoded in our word embeddings. We can thus conclude that this relational information encoded in the tensor-based embeddings is more easily accessible than that of CBOW and NNSE. Thus, our methods would likely be better suited for transfer learning to actual NLP tasks, particularly those in data-sparse settings.
Sentiment analysis results. The results are shown in Figure FIGREF19 . In this task, JCP-S is the dominant method across all levels of training data, but the difference is more obvious when training data is limited. This again indicates that for this specific task the information encoded by our tensor-based methods is more readily available as that of the baselines. It is thus evident that exploiting both second and third order co-occurrence data leads to higher quality semantic information being encoded in the embedding. At this point it is not clear why JCP-S so vastly outperforms CP-S at this task, but its superiority to the other strong baselines demonstrates the quality of information encoded by JCP-S. This discrepancy is also illustrative of the fact that there is no single “best word embedding” BIBREF5 – different embeddings encode different types of information, and thus should be used where they shine rather than for every NLP task.
Word Similarity results.
We show the results in Table TABREF21 . As we can see, our embeddings very clearly outperform the random embedding at this task. They even outperform CBOW on both of these datasets. It is worth including these results as the word similarity task is a very common way of evaluating embedding quality in the literature. However, due to the many intrinsic problems with evaluating word embeddings using word similarity BIBREF35 , we do not discuss this further.
## Multiplicative Compositionality
We find that even though they are not explicitly trained to do so, our tensor-based embeddings capture polysemy information naturally through multiplicative compositionality. We demonstrate this property qualitatively and provide proper motivation for it, leaving automated utilization to future work.
In our tensor-based embeddings, we found that one can create a vector that represents a word INLINEFORM0 in the context of another word INLINEFORM1 by taking the elementwise product INLINEFORM2 . We call INLINEFORM3 a “meaning vector” for the polysemous word INLINEFORM4 .
For example, consider the word star, which can denote a lead performer or a celestial body. We can create a vector for star in the “lead performer” sense by taking the elementwise product INLINEFORM0 . This produces a vector that lies near vectors for words related to lead performers and far from those related to star's other senses.
To motivate why this works, recall that the values in a third order PPMI tensor INLINEFORM0 are given by: INLINEFORM1
where INLINEFORM0 is the word vector for INLINEFORM1 . If words INLINEFORM2 have a high PPMI, then INLINEFORM3 will also be high, meaning INLINEFORM4 will be close to INLINEFORM5 in the vector space by cosine similarity.
For example, even though galaxy is likely to appear in the context of the word star in in the “celestial body” sense, INLINEFORM0 PPMI(star, actor, galaxy) is low whereas INLINEFORM1 PPMI(star, actor, drama) is high. Thus , INLINEFORM2 represents the meaning of star in the “lead performer” sense.
In Table TABREF22 we present the nearest neighbors of multiplicative and additive composed vectors for a variety of polysemous words. As we can see, the words corresponding to the nearest neighbors of the composed vectors for our tensor methods are semantically related to the intended sense both for multiplicative and additive composition. In contrast, for CBOW, only additive composition yields vectors whose nearest neighbors are semantically related to the intended sense. Thus, our embeddings can produce complementary sets of polysemous word representations that are qualitatively valid whereas CBOW (seemingly) only guarantees meaningful additive compositionality. We leave automated usage of this property to future work.
## Conclusion
Our key contributions are as follows:
Tensor factorization appears to be a highly applicable and effective tool for learning word embeddings, with many areas of potential future work. Leveraging higher order data in training word embeddings is useful for encoding new types of information and semantic relationships compared to models that are trained using only pairwise data. This indicates that such techniques will prove useful for training word embeddings to be used in downstream NLP tasks.
| 14
|
1704.04452
|
Bringing Structure into Summaries: Crowdsourcing a Benchmark Corpus of Concept Maps
|
# Bringing Structure into Summaries: Crowdsourcing a Benchmark Corpus of Concept Maps
## Abstract
Concept maps can be used to concisely represent important information and bring structure into large document collections. Therefore, we study a variant of multi-document summarization that produces summaries in the form of concept maps. However, suitable evaluation datasets for this task are currently missing. To close this gap, we present a newly created corpus of concept maps that summarize heterogeneous collections of web documents on educational topics. It was created using a novel crowdsourcing approach that allows us to efficiently determine important elements in large document collections. We release the corpus along with a baseline system and proposed evaluation protocol to enable further research on this variant of summarization.
## Introduction
Multi-document summarization (MDS), the transformation of a set of documents into a short text containing their most important aspects, is a long-studied problem in NLP. Generated summaries have been shown to support humans dealing with large document collections in information seeking tasks BIBREF0 , BIBREF1 , BIBREF2 . However, when exploring a set of documents manually, humans rarely write a fully-formulated summary for themselves. Instead, user studies BIBREF3 , BIBREF4 show that they note down important keywords and phrases, try to identify relationships between them and organize them accordingly. Therefore, we believe that the study of summarization with similarly structured outputs is an important extension of the traditional task.
A representation that is more in line with observed user behavior is a concept map BIBREF5 , a labeled graph showing concepts as nodes and relationships between them as edges (Figure FIGREF2 ). Introduced in 1972 as a teaching tool BIBREF6 , concept maps have found many applications in education BIBREF7 , BIBREF8 , for writing assistance BIBREF9 or to structure information repositories BIBREF10 , BIBREF11 . For summarization, concept maps make it possible to represent a summary concisely and clearly reveal relationships. Moreover, we see a second interesting use case that goes beyond the capabilities of textual summaries: When concepts and relations are linked to corresponding locations in the documents they have been extracted from, the graph can be used to navigate in a document collection, similar to a table of contents. An implementation of this idea has been recently described by BIBREF12 .
The corresponding task that we propose is concept-map-based MDS, the summarization of a document cluster in the form of a concept map. In order to develop and evaluate methods for the task, gold-standard corpora are necessary, but no suitable corpus is available. The manual creation of such a dataset is very time-consuming, as the annotation includes many subtasks. In particular, an annotator would need to manually identify all concepts in the documents, while only a few of them will eventually end up in the summary.
To overcome these issues, we present a corpus creation method that effectively combines automatic preprocessing, scalable crowdsourcing and high-quality expert annotations. Using it, we can avoid the high effort for single annotators, allowing us to scale to document clusters that are 15 times larger than in traditional summarization corpora. We created a new corpus of 30 topics, each with around 40 source documents on educational topics and a summarizing concept map that is the consensus of many crowdworkers (see Figure FIGREF3 ).
As a crucial step of the corpus creation, we developed a new crowdsourcing scheme called low-context importance annotation. In contrast to traditional approaches, it allows us to determine important elements in a document cluster without requiring annotators to read all documents, making it feasible to crowdsource the task and overcome quality issues observed in previous work BIBREF13 . We show that the approach creates reliable data for our focused summarization scenario and, when tested on traditional summarization corpora, creates annotations that are similar to those obtained by earlier efforts.
To summarize, we make the following contributions: (1) We propose a novel task, concept-map-based MDS (§ SECREF2 ), (2) present a new crowdsourcing scheme to create reference summaries (§ SECREF4 ), (3) publish a new dataset for the proposed task (§ SECREF5 ) and (4) provide an evaluation protocol and baseline (§ SECREF7 ). We make these resources publicly available under a permissive license.
## Task
Concept-map-based MDS is defined as follows: Given a set of related documents, create a concept map that represents its most important content, satisfies a specified size limit and is connected.
We define a concept map as a labeled graph showing concepts as nodes and relationships between them as edges. Labels are arbitrary sequences of tokens taken from the documents, making the summarization task extractive. A concept can be an entity, abstract idea, event or activity, designated by its unique label. Good maps should be propositionally coherent, meaning that every relation together with the two connected concepts form a meaningful proposition.
The task is complex, consisting of several interdependent subtasks. One has to extract appropriate labels for concepts and relations and recognize different expressions that refer to the same concept across multiple documents. Further, one has to select the most important concepts and relations for the summary and finally organize them in a graph satisfying the connectedness and size constraints.
## Related Work
Some attempts have been made to automatically construct concept maps from text, working with either single documents BIBREF14 , BIBREF9 , BIBREF15 , BIBREF16 or document clusters BIBREF17 , BIBREF18 , BIBREF19 . These approaches extract concept and relation labels from syntactic structures and connect them to build a concept map. However, common task definitions and comparable evaluations are missing. In addition, only a few of them, namely Villalon.2012 and Valerio.2006, define summarization as their goal and try to compress the input to a substantially smaller size. Our newly proposed task and the created large-cluster dataset fill these gaps as they emphasize the summarization aspect of the task.
For the subtask of selecting summary-worthy concepts and relations, techniques developed for traditional summarization BIBREF20 and keyphrase extraction BIBREF21 are related and applicable. Approaches that build graphs of propositions to create a summary BIBREF22 , BIBREF23 , BIBREF24 , BIBREF25 seem to be particularly related, however, there is one important difference: While they use graphs as an intermediate representation from which a textual summary is then generated, the goal of the proposed task is to create a graph that is directly interpretable and useful for a user. In contrast, these intermediate graphs, e.g. AMR, are hardly useful for a typical, non-linguist user.
For traditional summarization, the most well-known datasets emerged out of the DUC and TAC competitions. They provide clusters of news articles with gold-standard summaries. Extending these efforts, several more specialized corpora have been created: With regard to size, Nakano.2010 present a corpus of summaries for large-scale collections of web pages. Recently, corpora with more heterogeneous documents have been suggested, e.g. BIBREF26 and BIBREF27 . The corpus we present combines these aspects, as it has large clusters of heterogeneous documents, and provides a necessary benchmark to evaluate the proposed task.
For concept map generation, one corpus with human-created summary concept maps for student essays has been created BIBREF28 . In contrast to our corpus, it only deals with single documents, requires a two orders of magnitude smaller amount of compression of the input and is not publicly available .
Other types of information representation that also model concepts and their relationships are knowledge bases, such as Freebase BIBREF29 , and ontologies. However, they both differ in important aspects: Whereas concept maps follow an open label paradigm and are meant to be interpretable by humans, knowledge bases and ontologies are usually more strictly typed and made to be machine-readable. Moreover, approaches to automatically construct them from text typically try to extract as much information as possible, while we want to summarize a document.
## Low-Context Importance Annotation
Lloret.2013 describe several experiments to crowdsource reference summaries. Workers are asked to read 10 documents and then select 10 summary sentences from them for a reward of $0.05. They discovered several challenges, including poor work quality and the subjectiveness of the annotation task, indicating that crowdsourcing is not useful for this purpose.
To overcome these issues, we introduce a new task design, low-context importance annotation, to determine summary-worthy parts of documents. Compared to Lloret et al.'s approach, it is more in line with crowdsourcing best practices, as the tasks are simple, intuitive and small BIBREF30 and workers receive reasonable payment BIBREF31 . Most importantly, it is also much more efficient and scalable, as it does not require workers to read all documents in a cluster.
## Task Design
We break down the task of importance annotation to the level of single propositions. The goal of our crowdsourcing scheme is to obtain a score for each proposition indicating its importance in a document cluster, such that a ranking according to the score would reveal what is most important and should be included in a summary. In contrast to other work, we do not show the documents to the workers at all, but provide only a description of the document cluster's topic along with the propositions. This ensures that tasks are small, simple and can be done quickly (see Figure FIGREF4 ).
In preliminary tests, we found that this design, despite the minimal context, works reasonably on our focused clusters on common educational topics. For instance, consider Figure FIGREF4 : One can easily say that P1 is more important than P2 without reading the documents.
We distinguish two task variants:
Instead of enforcing binary importance decisions, we use a 5-point Likert-scale to allow more fine-grained annotations. The obtained labels are translated into scores (5..1) and the average of all scores for a proposition is used as an estimate for its importance. This follows the idea that while single workers might find the task subjective, the consensus of multiple workers, represented in the average score, tends to be less subjective due to the “wisdom of the crowd”. We randomly group five propositions into a task.
As an alternative, we use a second task design based on pairwise comparisons. Comparisons are known to be easier to make and more consistent BIBREF32 , but also more expensive, as the number of pairs grows quadratically with the number of objects. To reduce the cost, we group five propositions into a task and ask workers to order them by importance per drag-and-drop. From the results, we derive pairwise comparisons and use TrueSkill BIBREF35 , a powerful Bayesian rank induction model BIBREF34 , to obtain importance estimates for each proposition.
## Pilot Study
To verify the proposed approach, we conducted a pilot study on Amazon Mechanical Turk using data from TAC2008 BIBREF36 . We collected importance estimates for 474 propositions extracted from the first three clusters using both task designs. Each Likert-scale task was assigned to 5 different workers and awarded $0.06. For comparison tasks, we also collected 5 labels each, paid $0.05 and sampled around 7% of all possible pairs. We submitted them in batches of 100 pairs and selected pairs for subsequent batches based on the confidence of the TrueSkill model.
Following the observations of Lloret.2013, we established several measures for quality control. First, we restricted our tasks to workers from the US with an approval rate of at least 95%. Second, we identified low quality workers by measuring the correlation of each worker's Likert-scores with the average of the other four scores. The worst workers (at most 5% of all labels) were removed.
In addition, we included trap sentences, similar as in BIBREF13 , in around 80 of the tasks. In contrast to Lloret et al.'s findings, both an obvious trap sentence (This sentence is not important) and a less obvious but unimportant one (Barack Obama graduated from Harvard Law) were consistently labeled as unimportant (1.08 and 1.14), indicating that the workers did the task properly.
For Likert-scale tasks, we follow Snow.2008 and calculate agreement as the average Pearson correlation of a worker's Likert-score with the average score of the remaining workers. This measure is less strict than exact label agreement and can account for close labels and high- or low-scoring workers. We observe a correlation of 0.81, indicating substantial agreement. For comparisons, the majority agreement is 0.73. To further examine the reliability of the collected data, we followed the approach of Kiritchenko.2016 and simply repeated the crowdsourcing for one of the three topics. Between the importance estimates calculated from the first and second run, we found a Pearson correlation of 0.82 (Spearman 0.78) for Likert-scale tasks and 0.69 (Spearman 0.66) for comparison tasks. This shows that the approach, despite the subjectiveness of the task, allows us to collect reliable annotations.
In addition to the reliability studies, we extrinsically evaluated the annotations in the task of summary evaluation. For each of the 58 peer summaries in TAC2008, we calculated a score as the sum of the importance estimates of the propositions it contains. Table TABREF13 shows how these peer scores, averaged over the three topics, correlate with the manual responsiveness scores assigned during TAC in comparison to ROUGE-2 and Pyramid scores. The results demonstrate that with both task designs, we obtain importance annotations that are similarly useful for summary evaluation as pyramid annotations or gold-standard summaries (used for ROUGE).
Based on the pilot study, we conclude that the proposed crowdsourcing scheme allows us to obtain proper importance annotations for propositions. As workers are not required to read all documents, the annotation is much more efficient and scalable as with traditional methods.
## Corpus Creation
This section presents the corpus construction process, as outlined in Figure FIGREF16 , combining automatic preprocessing, scalable crowdsourcing and high-quality expert annotations to be able to scale to the size of our document clusters. For every topic, we spent about $150 on crowdsourcing and 1.5h of expert annotations, while just a single annotator would already need over 8 hours (at 200 words per minute) to read all documents of a topic.
## Source Data
As a starting point, we used the DIP corpus BIBREF37 , a collection of 49 clusters of 100 web pages on educational topics (e.g. bullying, homeschooling, drugs) with a short description of each topic. It was created from a large web crawl using state-of-the-art information retrieval. We selected 30 of the topics for which we created the necessary concept map annotations.
## Proposition Extraction
As concept maps consist of propositions expressing the relation between concepts (see Figure FIGREF2 ), we need to impose such a structure upon the plain text in the document clusters. This could be done by manually annotating spans representing concepts and relations, however, the size of our clusters makes this a huge effort: 2288 sentences per topic (69k in total) need to be processed. Therefore, we resort to an automatic approach.
The Open Information Extraction paradigm BIBREF38 offers a representation very similar to the desired one. For instance, from
Students with bad credit history should not lose hope and apply for federal loans with the FAFSA.
Open IE systems extract tuples of two arguments and a relation phrase representing propositions:
(s. with bad credit history, should not lose, hope)
(s. with bad credit history, apply for, federal loans with the FAFSA)
While the relation phrase is similar to a relation in a concept map, many arguments in these tuples represent useful concepts. We used Open IE 4, a state-of-the-art system BIBREF39 to process all sentences. After removing duplicates, we obtained 4137 tuples per topic.
Since we want to create a gold-standard corpus, we have to ensure that we produce high-quality data. We therefore made use of the confidence assigned to every extracted tuple to filter out low quality ones. To ensure that we do not filter too aggressively (and miss important aspects in the final summary), we manually annotated 500 tuples sampled from all topics for correctness. On the first 250 of them, we tuned the filter threshold to 0.5, which keeps 98.7% of the correct extractions in the unseen second half. After filtering, a topic had on average 2850 propositions (85k in total).
## Proposition Filtering
Despite the similarity of the Open IE paradigm, not every extracted tuple is a suitable proposition for a concept map. To reduce the effort in the subsequent steps, we therefore want to filter out unsuitable ones. A tuple is suitable if it (1) is a correct extraction, (2) is meaningful without any context and (3) has arguments that represent proper concepts. We created a guideline explaining when to label a tuple as suitable for a concept map and performed a small annotation study. Three annotators independently labeled 500 randomly sampled tuples. The agreement was 82% ( INLINEFORM0 ). We found tuples to be unsuitable mostly because they had unresolvable pronouns, conflicting with (2), or arguments that were full clauses or propositions, conflicting with (3), while (1) was mostly taken care of by the confidence filtering in § SECREF21 .
Due to the high number of tuples we decided to automate the filtering step. We trained a linear SVM on the majority voted annotations. As features, we used the extraction confidence, length of arguments and relations as well as part-of-speech tags, among others. To ensure that the automatic classification does not remove suitable propositions, we tuned the classifier to avoid false negatives. In particular, we introduced class weights, improving precision on the negative class at the cost of a higher fraction of positive classifications. Additionally, we manually verified a certain number of the most uncertain negative classifications to further improve performance. When 20% of the classifications are manually verified and corrected, we found that our model trained on 350 labeled instances achieves 93% precision on negative classifications on the unseen 150 instances. We found this to be a reasonable trade-off of automation and data quality and applied the model to the full dataset.
The classifier filtered out 43% of the propositions, leaving 1622 per topic. We manually examined the 17k least confident negative classifications and corrected 955 of them. We also corrected positive classifications for certain types of tuples for which we knew the classifier to be imprecise. Finally, each topic was left with an average of 1554 propositions (47k in total).
## Importance Annotation
Given the propositions identified in the previous step, we now applied our crowdsourcing scheme as described in § SECREF4 to determine their importance. To cope with the large number of propositions, we combine the two task designs: First, we collect Likert-scores from 5 workers for each proposition, clean the data and calculate average scores. Then, using only the top 100 propositions according to these scores, we crowdsource 10% of all possible pairwise comparisons among them. Using TrueSkill, we obtain a fine-grained ranking of the 100 most important propositions.
For Likert-scores, the average agreement over all topics is 0.80, while the majority agreement for comparisons is 0.78. We repeated the data collection for three randomly selected topics and found the Pearson correlation between both runs to be 0.73 (Spearman 0.73) for Likert-scores and 0.72 (Spearman 0.71) for comparisons. These figures show that the crowdsourcing approach works on this dataset as reliably as on the TAC documents.
In total, we uploaded 53k scoring and 12k comparison tasks to Mechanical Turk, spending $4425.45 including fees. From the fine-grained ranking of the 100 most important propositions, we select the top 50 per topic to construct a summary concept map in the subsequent steps.
## Proposition Revision
Having a manageable number of propositions, an annotator then applied a few straightforward transformations that correct common errors of the Open IE system. First, we break down propositions with conjunctions in either of the arguments into separate propositions per conjunct, which the Open IE system sometimes fails to do. And second, we correct span errors that might occur in the argument or relation phrases, especially when sentences were not properly segmented. As a result, we have a set of high quality propositions for our concept map, consisting of, due to the first transformation, 56.1 propositions per topic on average.
## Concept Map Construction
In this final step, we connect the set of important propositions to form a graph. For instance, given the following two propositions
(student, may borrow, Stafford Loan)
(the student, does not have, a credit history)
one can easily see, although the first arguments differ slightly, that both labels describe the concept student, allowing us to build a concept map with the concepts student, Stafford Loan and credit history. The annotation task thus involves deciding which of the available propositions to include in the map, which of their concepts to merge and, when merging, which of the available labels to use. As these decisions highly depend upon each other and require context, we decided to use expert annotators rather than crowdsource the subtasks.
Annotators were given the topic description and the most important, ranked propositions. Using a simple annotation tool providing a visualization of the graph, they could connect the propositions step by step. They were instructed to reach a size of 25 concepts, the recommended maximum size for a concept map BIBREF6 . Further, they should prefer more important propositions and ensure connectedness. When connecting two propositions, they were asked to keep the concept label that was appropriate for both propositions. To support the annotators, the tool used ADW BIBREF40 , a state-of-the-art approach for semantic similarity, to suggest possible connections. The annotation was carried out by graduate students with a background in NLP after receiving an introduction into the guidelines and tool and annotating a first example.
If an annotator was not able to connect 25 concepts, she was allowed to create up to three synthetic relations with freely defined labels, making the maps slightly abstractive. On average, the constructed maps have 0.77 synthetic relations, mostly connecting concepts whose relation is too obvious to be explicitly stated in text (e.g. between Montessori teacher and Montessori education).
To assess the reliability of this annotation step, we had the first three maps created by two annotators. We casted the task of selecting propositions to be included in the map as a binary decision task and observed an agreement of 84% ( INLINEFORM0 ). Second, we modeled the decision which concepts to join as a binary decision on all pairs of common concepts, observing an agreement of 95% ( INLINEFORM1 ). And finally, we compared which concept labels the annotators decided to include in the final map, observing 85% ( INLINEFORM2 ) agreement. Hence, the annotation shows substantial agreement BIBREF41 .
## Corpus Analysis
In this section, we describe our newly created corpus, which, in addition to having summaries in the form of concept maps, differs from traditional summarization corpora in several aspects.
## Document Clusters
The corpus consists of document clusters for 30 different topics. Each of them contains around 40 documents with on average 2413 tokens, which leads to an average cluster size of 97,880 token. With these characteristics, the document clusters are 15 times larger than typical DUC clusters of ten documents and five times larger than the 25-document-clusters (Table TABREF26 ). In addition, the documents are also more variable in terms of length, as the (length-adjusted) standard deviation is twice as high as in the other corpora. With these properties, the corpus represents an interesting challenge towards real-world application scenarios, in which users typically have to deal with much more than ten documents.
Because we used a large web crawl as the source for our corpus, it contains documents from a variety of genres. To further analyze this property, we categorized a sample of 50 documents from the corpus. Among them, we found professionally written articles and blog posts (28%), educational material for parents and kids (26%), personal blog posts (16%), forum discussions and comments (12%), commented link collections (12%) and scientific articles (6%).
In addition to the variety of genres, the documents also differ in terms of language use. To capture this property, we follow Zopf.2016 and compute, for every topic, the average Jensen-Shannon divergence between the word distribution of one document and the word distribution in the remaining documents. The higher this value is, the more the language differs between documents. We found the average divergence over all topics to be 0.3490, whereas it is 0.3019 in DUC 2004 and 0.3188 in TAC 2008A.
## Concept Maps
As Table TABREF33 shows, each of the 30 reference concept maps has exactly 25 concepts and between 24 and 28 relations. Labels for both concepts and relations consist on average of 3.2 tokens, whereas the latter are a bit shorter in characters.
To obtain a better picture of what kind of text spans have been used as labels, we automatically tagged them with their part-of-speech and determined their head with a dependency parser. Concept labels tend to be headed by nouns (82%) or verbs (15%), while they also contain adjectives, prepositions and determiners. Relation labels, on the other hand, are almost always headed by a verb (94%) and contain prepositions, nouns and particles in addition. These distributions are very similar to those reported by Villalon.2010 for their (single-document) concept map corpus.
Analyzing the graph structure of the maps, we found that all of them are connected. They have on average 7.2 central concepts with more than one relation, while the remaining ones occur in only one proposition. We found that achieving a higher number of connections would mean compromising importance, i.e. including less important propositions, and decided against it.
## Baseline Experiments
In this section, we briefly describe a baseline and evaluation scripts that we release, with a detailed documentation, along with the corpus.
## Conclusion
In this work, we presented low-context importance annotation, a novel crowdsourcing scheme that we used to create a new benchmark corpus for concept-map-based MDS. The corpus has large-scale document clusters of heterogeneous web documents, posing a challenging summarization task. Together with the corpus, we provide implementations of a baseline method and evaluation scripts and hope that our efforts facilitate future research on this variant of summarization.
## Acknowledgments
We would like to thank Teresa Botschen, Andreas Hanselowski and Markus Zopf for their help with the annotation work and Christian Meyer for his valuable feedback. This work has been supported by the German Research Foundation as part of the Research Training Group “Adaptive Preparation of Information from Heterogeneous Sources” (AIPHES) under grant No. GRK 1994/1.
| 19
|
1704.05572
|
Answering Complex Questions Using Open Information Extraction
|
# Answering Complex Questions Using Open Information Extraction
## Abstract
While there has been substantial progress in factoid question-answering (QA), answering complex questions remains challenging, typically requiring both a large body of knowledge and inference techniques. Open Information Extraction (Open IE) provides a way to generate semi-structured knowledge for QA, but to date such knowledge has only been used to answer simple questions with retrieval-based methods. We overcome this limitation by presenting a method for reasoning with Open IE knowledge, allowing more complex questions to be handled. Using a recently proposed support graph optimization framework for QA, we develop a new inference model for Open IE, in particular one that can work effectively with multiple short facts, noise, and the relational structure of tuples. Our model significantly outperforms a state-of-the-art structured solver on complex questions of varying difficulty, while also removing the reliance on manually curated knowledge.
## Introduction
Effective question answering (QA) systems have been a long-standing quest of AI research. Structured curated KBs have been used successfully for this task BIBREF0 , BIBREF1 . However, these KBs are expensive to build and typically domain-specific. Automatically constructed open vocabulary (subject; predicate; object) style tuples have broader coverage, but have only been used for simple questions where a single tuple suffices BIBREF2 , BIBREF3 .
Our goal in this work is to develop a QA system that can perform reasoning with Open IE BIBREF4 tuples for complex multiple-choice questions that require tuples from multiple sentences. Such a system can answer complex questions in resource-poor domains where curated knowledge is unavailable. Elementary-level science exams is one such domain, requiring complex reasoning BIBREF5 . Due to the lack of a large-scale structured KB, state-of-the-art systems for this task either rely on shallow reasoning with large text corpora BIBREF6 , BIBREF7 or deeper, structured reasoning with a small amount of automatically acquired BIBREF8 or manually curated BIBREF9 knowledge.
Consider the following question from an Alaska state 4th grade science test:
Which object in our solar system reflects light and is a satellite that orbits around one planet? (A) Earth (B) Mercury (C) the Sun (D) the Moon
This question is challenging for QA systems because of its complex structure and the need for multi-fact reasoning. A natural way to answer it is by combining facts such as (Moon; is; in the solar system), (Moon; reflects; light), (Moon; is; satellite), and (Moon; orbits; around one planet).
A candidate system for such reasoning, and which we draw inspiration from, is the TableILP system of BIBREF9 . TableILP treats QA as a search for an optimal subgraph that connects terms in the question and answer via rows in a set of curated tables, and solves the optimization problem using Integer Linear Programming (ILP). We similarly want to search for an optimal subgraph. However, a large, automatically extracted tuple KB makes the reasoning context different on three fronts: (a) unlike reasoning with tables, chaining tuples is less important and reliable as join rules aren't available; (b) conjunctive evidence becomes paramount, as, unlike a long table row, a single tuple is less likely to cover the entire question; and (c) again, unlike table rows, tuples are noisy, making combining redundant evidence essential. Consequently, a table-knowledge centered inference model isn't the best fit for noisy tuples.
To address this challenge, we present a new ILP-based model of inference with tuples, implemented in a reasoner called TupleInf. We demonstrate that TupleInf significantly outperforms TableILP by 11.8% on a broad set of over 1,300 science questions, without requiring manually curated tables, using a substantially simpler ILP formulation, and generalizing well to higher grade levels. The gains persist even when both solvers are provided identical knowledge. This demonstrates for the first time how Open IE based QA can be extended from simple lookup questions to an effective system for complex questions.
## Related Work
We discuss two classes of related work: retrieval-based web question-answering (simple reasoning with large scale KB) and science question-answering (complex reasoning with small KB).
## Tuple Inference Solver
We first describe the tuples used by our solver. We define a tuple as (subject; predicate; objects) with zero or more objects. We refer to the subject, predicate, and objects as the fields of the tuple.
## Tuple KB
We use the text corpora (S) from BIBREF6 aristo2016:combining to build our tuple KB. For each test set, we use the corresponding training questions $Q_\mathit {tr}$ to retrieve domain-relevant sentences from S. Specifically, for each multiple-choice question $(q,A) \in Q_\mathit {tr}$ and each choice $a \in A$ , we use all non-stopword tokens in $q$ and $a$ as an ElasticSearch query against S. We take the top 200 hits, run Open IE v4, and aggregate the resulting tuples over all $a \in A$ and over all questions in $Q_\mathit {tr}$ to create the tuple KB (T).
## Tuple Selection
Given a multiple-choice question $qa$ with question text $q$ and answer choices A= $\lbrace a_i\rbrace $ , we select the most relevant tuples from $T$ and $S$ as follows.
Selecting from Tuple KB: We use an inverted index to find the 1,000 tuples that have the most overlapping tokens with question tokens $tok(qa).$ . We also filter out any tuples that overlap only with $tok(q)$ as they do not support any answer. We compute the normalized TF-IDF score treating the question, $q$ as a query and each tuple, $t$ as a document: $
&\textit {tf}(x, q)=1\; \textmd {if x} \in q ; \textit {idf}(x) = log(1 + N/n_x) \\
&\textit {tf-idf}(t, q)=\sum _{x \in t\cap q} idf(x)
$
where $N$ is the number of tuples in the KB and $n_x$ are the number of tuples containing $x$ . We normalize the tf-idf score by the number of tokens in $t$ and $q$ . We finally take the 50 top-scoring tuples $T_{qa}$ .
On-the-fly tuples from text: To handle questions from new domains not covered by the training set, we extract additional tuples on the fly from S (similar to BIBREF17 knowlhunting). We perform the same ElasticSearch query described earlier for building T. We ignore sentences that cover none or all answer choices as they are not discriminative. We also ignore long sentences ( $>$ 300 characters) and sentences with negation as they tend to lead to noisy inference. We then run Open IE on these sentences and re-score the resulting tuples using the Jaccard score due to the lossy nature of Open IE, and finally take the 50 top-scoring tuples $T^{\prime }_{qa}$ .
## Support Graph Search
Similar to TableILP, we view the QA task as searching for a graph that best connects the terms in the question (qterms) with an answer choice via the knowledge; see Figure 1 for a simple illustrative example. Unlike standard alignment models used for tasks such as Recognizing Textual Entailment (RTE) BIBREF18 , however, we must score alignments between a set $T_{qa} \cup T^{\prime }_{qa}$ of structured tuples and a (potentially multi-sentence) multiple-choice question $qa$ .
The qterms, answer choices, and tuples fields form the set of possible vertices, $\mathcal {V}$ , of the support graph. Edges connecting qterms to tuple fields and tuple fields to answer choices form the set of possible edges, $\mathcal {E}$ . The support graph, $G(V, E)$ , is a subgraph of $\mathcal {G}(\mathcal {V}, \mathcal {E})$ where $V$ and $E$ denote “active” nodes and edges, resp. We define the desired behavior of an optimal support graph via an ILP model as follows.
Similar to TableILP, we score the support graph based on the weight of the active nodes and edges. Each edge $e(t, h)$ is weighted based on a word-overlap score. While TableILP used WordNet BIBREF19 paths to compute the weight, this measure results in unreliable scores when faced with longer phrases found in Open IE tuples.
Compared to a curated KB, it is easy to find Open IE tuples that match irrelevant parts of the questions. To mitigate this issue, we improve the scoring of qterms in our ILP objective to focus on important terms. Since the later terms in a question tend to provide the most critical information, we scale qterm coefficients based on their position. Also, qterms that appear in almost all of the selected tuples tend not to be discriminative as any tuple would support such a qterm. Hence we scale the coefficients by the inverse frequency of the tokens in the selected tuples.
Since Open IE tuples do not come with schema and join rules, we can define a substantially simpler model compared to TableILP. This reduces the reasoning capability but also eliminates the reliance on hand-authored join rules and regular expressions used in TableILP. We discovered (see empirical evaluation) that this simple model can achieve the same score as TableILP on the Regents test (target test set used by TableILP) and generalizes better to different grade levels.
We define active vertices and edges using ILP constraints: an active edge must connect two active vertices and an active vertex must have at least one active edge. To avoid positive edge coefficients in the objective function resulting in spurious edges in the support graph, we limit the number of active edges from an active tuple, question choice, tuple fields, and qterms (first group of constraints in Table 1 ). Our model is also capable of using multiple tuples to support different parts of the question as illustrated in Figure 1 . To avoid spurious tuples that only connect with the question (or choice) or ignore the relation being expressed in the tuple, we add constraints that require each tuple to connect a qterm with an answer choice (second group of constraints in Table 1 ).
We also define new constraints based on the Open IE tuple structure. Since an Open IE tuple expresses a fact about the tuple's subject, we require the subject to be active in the support graph. To avoid issues such as (Planet; orbit; Sun) matching the sample question in the introduction (“Which object $\ldots $ orbits around a planet”), we also add an ordering constraint (third group in Table 1 ).
Its worth mentioning that TupleInf only combines parallel evidence i.e. each tuple must connect words in the question to the answer choice. For reliable multi-hop reasoning using OpenIE tuples, we can add inter-tuple connections to the support graph search, controlled by a small number of rules over the OpenIE predicates. Learning such rules for the Science domain is an open problem and potential avenue of future work.
## Experiments
Comparing our method with two state-of-the-art systems for 4th and 8th grade science exams, we demonstrate that (a) TupleInf with only automatically extracted tuples significantly outperforms TableILP with its original curated knowledge as well as with additional tuples, and (b) TupleInf's complementary approach to IR leads to an improved ensemble. Numbers in bold indicate statistical significance based on the Binomial exact test BIBREF20 at $p=0.05$ .
We consider two question sets. (1) 4th Grade set (1220 train, 1304 test) is a 10x larger superset of the NY Regents questions BIBREF6 , and includes professionally written licensed questions. (2) 8th Grade set (293 train, 282 test) contains 8th grade questions from various states.
We consider two knowledge sources. The Sentence corpus (S) consists of domain-targeted $~$ 80K sentences and 280 GB of plain text extracted from web pages used by BIBREF6 aristo2016:combining. This corpus is used by the IR solver and also used to create the tuple KB T and on-the-fly tuples $T^{\prime }_{qa}$ . Additionally, TableILP uses $\sim $ 70 Curated tables (C) designed for 4th grade NY Regents exams.
We compare TupleInf with two state-of-the-art baselines. IR is a simple yet powerful information-retrieval baseline BIBREF6 that selects the answer option with the best matching sentence in a corpus. TableILP is the state-of-the-art structured inference baseline BIBREF9 developed for science questions.
## Results
Table 2 shows that TupleInf, with no curated knowledge, outperforms TableILP on both question sets by more than 11%. The lower half of the table shows that even when both solvers are given the same knowledge (C+T), the improved selection and simplified model of TupleInf results in a statistically significant improvement. Our simple model, TupleInf(C + T), also achieves scores comparable to TableILP on the latter's target Regents questions (61.4% vs TableILP's reported 61.5%) without any specialized rules.
Table 3 shows that while TupleInf achieves similar scores as the IR solver, the approaches are complementary (structured lossy knowledge reasoning vs. lossless sentence retrieval). The two solvers, in fact, differ on 47.3% of the training questions. To exploit this complementarity, we train an ensemble system BIBREF6 which, as shown in the table, provides a substantial boost over the individual solvers. Further, IR + TupleInf is consistently better than IR + TableILP. Finally, in combination with IR and the statistical association based PMI solver (that scores 54.1% by itself) of BIBREF6 aristo2016:combining, TupleInf achieves a score of 58.2% as compared to TableILP's ensemble score of 56.7% on the 4th grade set, again attesting to TupleInf's strength.
## Error Analysis
We describe four classes of failures that we observed, and the future work they suggest.
Missing Important Words: Which material will spread out to completely fill a larger container? (A)air (B)ice (C)sand (D)water
In this question, we have tuples that support water will spread out and fill a larger container but miss the critical word “completely”. An approach capable of detecting salient question words could help avoid that.
Lossy IE: Which action is the best method to separate a mixture of salt and water? ...
The IR solver correctly answers this question by using the sentence: Separate the salt and water mixture by evaporating the water. However, TupleInf is not able to answer this question as Open IE is unable to extract tuples from this imperative sentence. While the additional structure from Open IE is useful for more robust matching, converting sentences to Open IE tuples may lose important bits of information.
Bad Alignment: Which of the following gases is necessary for humans to breathe in order to live?(A) Oxygen(B) Carbon dioxide(C) Helium(D) Water vapor
TupleInf returns “Carbon dioxide” as the answer because of the tuple (humans; breathe out; carbon dioxide). The chunk “to breathe” in the question has a high alignment score to the “breathe out” relation in the tuple even though they have completely different meanings. Improving the phrase alignment can mitigate this issue.
Out of scope: Deer live in forest for shelter. If the forest was cut down, which situation would most likely happen?...
Such questions that require modeling a state presented in the question and reasoning over the state are out of scope of our solver.
## Conclusion
We presented a new QA system, TupleInf, that can reason over a large, potentially noisy tuple KB to answer complex questions. Our results show that TupleInf is a new state-of-the-art structured solver for elementary-level science that does not rely on curated knowledge and generalizes to higher grades. Errors due to lossy IE and misalignments suggest future work in incorporating context and distributional measures.
## Appendix: ILP Model Details
To build the ILP model, we first need to get the questions terms (qterm) from the question by chunking the question using an in-house chunker based on the postagger from FACTORIE.
## Experiment Details
We use the SCIP ILP optimization engine BIBREF21 to optimize our ILP model. To get the score for each answer choice $a_i$ , we force the active variable for that choice $x_{a_i}$ to be one and use the objective function value of the ILP model as the score. For evaluations, we use a 2-core 2.5 GHz Amazon EC2 linux machine with 16 GB RAM. To evaluate TableILP and TupleInf on curated tables and tuples, we converted them into the expected format of each solver as follows.
## Using curated tables with TupleInf
For each question, we select the 7 best matching tables using the tf-idf score of the table w.r.t. the question tokens and top 20 rows from each table using the Jaccard similarity of the row with the question. (same as BIBREF9 tableilp2016). We then convert the table rows into the tuple structure using the relations defined by TableILP. For every pair of cells connected by a relation, we create a tuple with the two cells as the subject and primary object with the relation as the predicate. The other cells of the table are used as additional objects to provide context to the solver. We pick top-scoring 50 tuples using the Jaccard score.
## Using Open IE tuples with TableILP
We create an additional table in TableILP with all the tuples in $T$ . Since TableILP uses fixed-length $(subject; predicate; object)$ triples, we need to map tuples with multiple objects to this format. For each object, $O_i$ in the input Open IE tuple $(S; P; O_1; O_2 \ldots )$ , we add a triple $(S; P; O_i)$ to this table.
| 14
|
1705.00108
|
Semi-supervised sequence tagging with bidirectional language models
|
# Semi-supervised sequence tagging with bidirectional language models
## Abstract
Pre-trained word embeddings learned from unlabeled text have become a standard component of neural network architectures for NLP tasks. However, in most cases, the recurrent network that operates on word-level representations to produce context sensitive representations is trained on relatively little labeled data. In this paper, we demonstrate a general semi-supervised approach for adding pre- trained context embeddings from bidirectional language models to NLP systems and apply it to sequence labeling tasks. We evaluate our model on two standard datasets for named entity recognition (NER) and chunking, and in both cases achieve state of the art results, surpassing previous systems that use other forms of transfer or joint learning with additional labeled data and task specific gazetteers.
## Introduction
Due to their simplicity and efficacy, pre-trained word embedding have become ubiquitous in NLP systems. Many prior studies have shown that they capture useful semantic and syntactic information BIBREF0 , BIBREF1 and including them in NLP systems has been shown to be enormously helpful for a variety of downstream tasks BIBREF2 .
However, in many NLP tasks it is essential to represent not just the meaning of a word, but also the word in context. For example, in the two phrases “A Central Bank spokesman” and “The Central African Republic”, the word `Central' is used as part of both an Organization and Location. Accordingly, current state of the art sequence tagging models typically include a bidirectional recurrent neural network (RNN) that encodes token sequences into a context sensitive representation before making token specific predictions BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 .
Although the token representation is initialized with pre-trained embeddings, the parameters of the bidirectional RNN are typically learned only on labeled data. Previous work has explored methods for jointly learning the bidirectional RNN with supplemental labeled data from other tasks BIBREF7 , BIBREF3 .
In this paper, we explore an alternate semi-supervised approach which does not require additional labeled data. We use a neural language model (LM), pre-trained on a large, unlabeled corpus to compute an encoding of the context at each position in the sequence (hereafter an LM embedding) and use it in the supervised sequence tagging model. Since the LM embeddings are used to compute the probability of future words in a neural LM, they are likely to encode both the semantic and syntactic roles of words in context.
Our main contribution is to show that the context sensitive representation captured in the LM embeddings is useful in the supervised sequence tagging setting. When we include the LM embeddings in our system overall performance increases from 90.87% to 91.93% INLINEFORM0 for the CoNLL 2003 NER task, a more then 1% absolute F1 increase, and a substantial improvement over the previous state of the art. We also establish a new state of the art result (96.37% INLINEFORM1 ) for the CoNLL 2000 Chunking task.
As a secondary contribution, we show that using both forward and backward LM embeddings boosts performance over a forward only LM. We also demonstrate that domain specific pre-training is not necessary by applying a LM trained in the news domain to scientific papers.
## Overview
The main components in our language-model-augmented sequence tagger (TagLM) are illustrated in Fig. FIGREF4 . After pre-training word embeddings and a neural LM on large, unlabeled corpora (Step 1), we extract the word and LM embeddings for every token in a given input sequence (Step 2) and use them in the supervised sequence tagging model (Step 3).
## Baseline sequence tagging model
Our baseline sequence tagging model is a hierarchical neural tagging model, closely following a number of recent studies BIBREF4 , BIBREF5 , BIBREF3 , BIBREF8 (left side of Figure FIGREF5 ).
Given a sentence of tokens INLINEFORM0 it first forms a representation, INLINEFORM1 , for each token by concatenating a character based representation INLINEFORM2 with a token embedding INLINEFORM3 : DISPLAYFORM0
The character representation INLINEFORM0 captures morphological information and is either a convolutional neural network (CNN) BIBREF4 , BIBREF8 or RNN BIBREF3 , BIBREF5 . It is parameterized by INLINEFORM1 with parameters INLINEFORM2 . The token embeddings, INLINEFORM3 , are obtained as a lookup INLINEFORM4 , initialized using pre-trained word embeddings, and fine tuned during training BIBREF2 .
To learn a context sensitive representation, we employ multiple layers of bidirectional RNNs. For each token position, INLINEFORM0 , the hidden state INLINEFORM1 of RNN layer INLINEFORM2 is formed by concatenating the hidden states from the forward ( INLINEFORM3 ) and backward ( INLINEFORM4 ) RNNs. As a result, the bidirectional RNN is able to use both past and future information to make a prediction at token INLINEFORM5 . More formally, for the first RNN layer that operates on INLINEFORM6 to output INLINEFORM7 : DISPLAYFORM0
The second RNN layer is similar and uses INLINEFORM0 to output INLINEFORM1 . In this paper, we use INLINEFORM2 layers of RNNs in all experiments and parameterize INLINEFORM3 as either Gated Recurrent Units (GRU) BIBREF9 or Long Short-Term Memory units (LSTM) BIBREF10 depending on the task.
Finally, the output of the final RNN layer INLINEFORM0 is used to predict a score for each possible tag using a single dense layer. Due to the dependencies between successive tags in our sequence labeling tasks (e.g. using the BIOES labeling scheme, it is not possible for I-PER to follow B-LOC), it is beneficial to model and decode each sentence jointly instead of independently predicting the label for each token. Accordingly, we add another layer with parameters for each label bigram, computing the sentence conditional random field (CRF) loss BIBREF11 using the forward-backward algorithm at training time, and using the Viterbi algorithm to find the most likely tag sequence at test time, similar to BIBREF2 .
## Bidirectional LM
A language model computes the probability of a token sequence INLINEFORM0 INLINEFORM1
Recent state of the art neural language models BIBREF12 use a similar architecture to our baseline sequence tagger where they pass a token representation (either from a CNN over characters or as token embeddings) through multiple layers of LSTMs to embed the history INLINEFORM0 into a fixed dimensional vector INLINEFORM1 . This is the forward LM embedding of the token at position INLINEFORM2 and is the output of the top LSTM layer in the language model. Finally, the language model predicts the probability of token INLINEFORM3 using a softmax layer over words in the vocabulary.
The need to capture future context in the LM embeddings suggests it is beneficial to also consider a backward LM in additional to the traditional forward LM. A backward LM predicts the previous token given the future context. Given a sentence with INLINEFORM0 tokens, it computes INLINEFORM1
A backward LM can be implemented in an analogous way to a forward LM and produces the backward LM embedding INLINEFORM0 , for the sequence INLINEFORM1 , the output embeddings of the top layer LSTM.
In our final system, after pre-training the forward and backward LMs separately, we remove the top layer softmax and concatenate the forward and backward LM embeddings to form bidirectional LM embeddings, i.e., INLINEFORM0 . Note that in our formulation, the forward and backward LMs are independent, without any shared parameters.
## Combining LM with sequence model
Our combined system, TagLM, uses the LM embeddings as additional inputs to the sequence tagging model. In particular, we concatenate the LM embeddings INLINEFORM0 with the output from one of the bidirectional RNN layers in the sequence model. In our experiments, we found that introducing the LM embeddings at the output of the first layer performed the best. More formally, we simply replace ( EQREF6 ) with DISPLAYFORM0
There are alternate possibilities for adding the LM embeddings to the sequence model. One possibility adds a non-linear mapping after the concatenation and before the second RNN (e.g. replacing ( EQREF9 ) with INLINEFORM0 where INLINEFORM1 is a non-linear function). Another possibility introduces an attention-like mechanism that weights the all LM embeddings in a sentence before including them in the sequence model. Our initial results with the simple concatenation were encouraging so we did not explore these alternatives in this study, preferring to leave them for future work.
## Experiments
We evaluate our approach on two well benchmarked sequence tagging tasks, the CoNLL 2003 NER task BIBREF13 and the CoNLL 2000 Chunking task BIBREF14 . We report the official evaluation metric (micro-averaged INLINEFORM0 ). In both cases, we use the BIOES labeling scheme for the output tags, following previous work which showed it outperforms other options BIBREF15 . Following BIBREF8 , we use the Senna word embeddings BIBREF2 and pre-processed the text by lowercasing all tokens and replacing all digits with 0.
## Overall system results
Tables TABREF15 and TABREF16 compare results from TagLM with previously published state of the art results without additional labeled data or task specific gazetteers. Tables TABREF17 and TABREF18 compare results of TagLM to other systems that include additional labeled data or gazetteers. In both tasks, TagLM establishes a new state of the art using bidirectional LMs (the forward CNN-BIG-LSTM and the backward LSTM-2048-512).
In the CoNLL 2003 NER task, our model scores 91.93 mean INLINEFORM0 , which is a statistically significant increase over the previous best result of 91.62 INLINEFORM1 from BIBREF8 that used gazetteers (at 95%, two-sided Welch t-test, INLINEFORM2 ).
In the CoNLL 2000 Chunking task, TagLM achieves 96.37 mean INLINEFORM0 , exceeding all previously published results without additional labeled data by more then 1% absolute INLINEFORM1 . The improvement over the previous best result of 95.77 in BIBREF6 that jointly trains with Penn Treebank (PTB) POS tags is statistically significant at 95% ( INLINEFORM2 assuming standard deviation of INLINEFORM3 ).
Importantly, the LM embeddings amounts to an average absolute improvement of 1.06 and 1.37 INLINEFORM0 in the NER and Chunking tasks, respectively.
Although we do not use external labeled data or gazetteers, we found that TagLM outperforms previous state of the art results in both tasks when external resources (labeled data or task specific gazetteers) are available. Furthermore, Tables TABREF17 and TABREF18 show that, in most cases, the improvements we obtain by adding LM embeddings are larger then the improvements previously obtained by adding other forms of transfer or joint learning. For example, BIBREF3 noted an improvement of only 0.06 INLINEFORM0 in the NER task when transfer learning from both CoNLL 2000 chunks and PTB POS tags and BIBREF8 reported an increase of 0.71 INLINEFORM1 when adding gazetteers to their baseline. In the Chunking task, previous work has reported from 0.28 to 0.75 improvement in INLINEFORM2 when including supervised labels from the PTB POS tags or CoNLL 2003 entities BIBREF3 , BIBREF7 , BIBREF6 .
## Analysis
To elucidate the characteristics of our LM augmented sequence tagger, we ran a number of additional experiments on the CoNLL 2003 NER task.
In this experiment, we concatenate the LM embeddings at different locations in the baseline sequence tagger. In particular, we used the LM embeddings INLINEFORM0 to:
augment the input of the first RNN layer; i.e.,
INLINEFORM0 ,
augment the output of the first RNN layer; i.e., INLINEFORM0 , and
augment the output of the second RNN layer; i.e., INLINEFORM0 .
Table TABREF20 shows that the second alternative performs best. We speculate that the second RNN layer in the sequence tagging model is able to capture interactions between task specific context as expressed in the first RNN layer and general context as expressed in the LM embeddings in a way that improves overall system performance. These results are consistent with BIBREF7 who found that chunking performance was sensitive to the level at which additional POS supervision was added.
In this experiment, we compare six different configurations of the forward and backward language models (including the baseline model which does not use any language models). The results are reported in Table TABREF21 .
We find that adding backward LM embeddings consistently outperforms forward-only LM embeddings, with INLINEFORM0 improvements between 0.22 and 0.27%, even with the relatively small backward LSTM-2048-512 LM.
LM size is important, and replacing the forward LSTM-2048-512 with CNN-BIG-LSTM (test perplexities of 47.7 to 30.0 on 1B Word Benchmark) improves INLINEFORM0 by 0.26 - 0.31%, about as much as adding backward LM. Accordingly, we hypothesize (but have not tested) that replacing the backward LSTM-2048-512 with a backward LM analogous to the CNN-BIG-LSTM would further improve performance.
To highlight the importance of including language models trained on a large scale data, we also experimented with training a language model on just the CoNLL 2003 training and development data. Due to the much smaller size of this data set, we decreased the model size to 512 hidden units with a 256 dimension projection and normalized tokens in the same manner as input to the sequence tagging model (lower-cased, with all digits replaced with 0). The test set perplexities for the forward and backward models (measured on the CoNLL 2003 test data) were 106.9 and 104.2, respectively. Including embeddings from these language models decreased performance slightly compared to the baseline system without any LM. This result supports the hypothesis that adding language models help because they learn composition functions (i.e., the RNN parameters in the language model) from much larger data compared to the composition functions in the baseline tagger, which are only learned from labeled data.
To understand the importance of including a task specific sequence RNN we ran an experiment that removed the task specific sequence RNN and used only the LM embeddings with a dense layer and CRF to predict output tags. In this setup, performance was very low, 88.17 INLINEFORM0 , well below our baseline. This result confirms that the RNNs in the baseline tagger encode essential information which is not encoded in the LM embeddings. This is unsurprising since the RNNs in the baseline tagger are trained on labeled examples, unlike the RNN in the language model which is only trained on unlabeled examples. Note that the LM weights are fixed in this experiment.
A priori, we expect the addition of LM embeddings to be most beneficial in cases where the task specific annotated datasets are small. To test this hypothesis, we replicated the setup from BIBREF3 that samples 1% of the CoNLL 2003 training set and compared the performance of TagLM to our baseline without LM. In this scenario, test INLINEFORM0 increased 3.35% (from 67.66 to 71.01%) compared to an increase of 1.06% INLINEFORM1 for a similar comparison with the full training dataset. The analogous increases in BIBREF3 are 3.97% for cross-lingual transfer from CoNLL 2002 Spanish NER and 6.28% INLINEFORM2 for transfer from PTB POS tags. However, they found only a 0.06% INLINEFORM3 increase when using the full training data and transferring from both CoNLL 2000 chunks and PTB POS tags. Taken together, this suggests that for very small labeled training sets, transferring from other tasks yields a large improvement, but this improvement almost disappears when the training data is large. On the other hand, our approach is less dependent on the training set size and significantly improves performance even with larger training sets.
Our TagLM formulation increases the number of parameters in the second RNN layer INLINEFORM0 due to the increase in the input dimension INLINEFORM1 if all other hyperparameters are held constant. To confirm that this did not have a material impact on the results, we ran two additional experiments. In the first, we trained a system without a LM but increased the second RNN layer hidden dimension so that number of parameters was the same as in TagLM. In this case, performance decreased slightly (by 0.15% INLINEFORM2 ) compared to the baseline model, indicating that solely increasing parameters does not improve performance. In the second experiment, we decreased the hidden dimension of the second RNN layer in TagLM to give it the same number of parameters as the baseline no LM model. In this case, test INLINEFORM3 increased slightly to INLINEFORM4 indicating that the additional parameters in TagLM are slightly hurting performance.
One artifact of our evaluation framework is that both the labeled data in the chunking and NER tasks and the unlabeled text in the 1 Billion Word Benchmark used to train the bidirectional LMs are derived from news articles. To test the sensitivity to the LM training domain, we also applied TagLM with a LM trained on news articles to the SemEval 2017 Shared Task 10, ScienceIE. ScienceIE requires end-to-end joint entity and relationship extraction from scientific publications across three diverse fields (computer science, material sciences, and physics) and defines three broad entity types (Task, Material and Process). For this task, TagLM increased INLINEFORM0 on the development set by 4.12% (from 49.93 to to 54.05%) for entity extraction over our baseline without LM embeddings and it was a major component in our winning submission to ScienceIE, Scenario 1 BIBREF20 . We conclude that LM embeddings can improve the performance of a sequence tagger even when the data comes from a different domain.
## Conclusion
In this paper, we proposed a simple and general semi-supervised method using pre-trained neural language models to augment token representations in sequence tagging models. Our method significantly outperforms current state of the art models in two popular datasets for NER and Chunking. Our analysis shows that adding a backward LM in addition to traditional forward LMs consistently improves performance. The proposed method is robust even when the LM is trained on unlabeled data from a different domain, or when the baseline model is trained on a large number of labeled examples.
## Acknowledgments
We thank Chris Dyer, Julia Hockenmaier, Jayant Krishnamurthy, Matt Gardner and Oren Etzioni for comments on earlier drafts that led to substantial improvements in the final version.
| 10
|
1705.03261
|
Drug-drug Interaction Extraction via Recurrent Neural Network with Multiple Attention Layers
|
# Drug-drug Interaction Extraction via Recurrent Neural Network with Multiple Attention Layers
## Abstract
Drug-drug interaction (DDI) is a vital information when physicians and pharmacists intend to co-administer two or more drugs. Thus, several DDI databases are constructed to avoid mistakenly combined use. In recent years, automatically extracting DDIs from biomedical text has drawn researchers' attention. However, the existing work utilize either complex feature engineering or NLP tools, both of which are insufficient for sentence comprehension. Inspired by the deep learning approaches in natural language processing, we propose a recur- rent neural network model with multiple attention layers for DDI classification. We evaluate our model on 2013 SemEval DDIExtraction dataset. The experiments show that our model classifies most of the drug pairs into correct DDI categories, which outperforms the existing NLP or deep learning methods.
## Introduction
Drug-drug interaction (DDI) is a situation when one drug increases or decreases the effect of another drug BIBREF0 . Adverse drug reactions may cause severe side effect, if two or more medicines were taken and their DDI were not investigated in detail. DDI is a common cause of illness, even a cause of death BIBREF1 . Thus, DDI databases for clinical medication decisions are proposed by some researchers. These databases such as SFINX BIBREF2 , KEGG BIBREF3 , CredibleMeds BIBREF4 help physicians and pharmacists avoid most adverse drug reactions.
Traditional DDI databases are manually constructed according to clinical records, scientific research and drug specifications. For instance, The sentence “With combined use, clinicians should be aware, when phenytoin is added, of the potential for reexacerbation of pulmonary symptomatology due to lowered serum theophylline concentrations BIBREF5 ”, which is from a pharmacotherapy report, describe the side effect of phenytoin and theophylline's combined use. Then this information on specific medicines will be added to DDI databases. As drug-drug interactions have being increasingly found, manually constructing DDI database would consume a lot of manpower and resources.
There has been many efforts to automatically extract DDIs from natural language BIBREF0 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , mainly medical literature and clinical records. These works can be divided into the following categories:
To avoid complex feature engineering and NLP toolkits' usage, we employ deep learning approaches for sentence comprehension as a whole. Our model takes in a sentence from biomedical literature which contains a drug pair and outputs what kind of DDI this drug pair belongs. This assists physicians refrain from improper combined use of drugs. In addition, the word and sentence level attentions are introduced to our model for better DDI predictions.
We train our language comprehension model with labeled instances. Figure FIGREF5 shows partial records in DDI corpus BIBREF16 . We extract the sentence and drug pairs in the records. There are 3 drug pairs in this example thus we have 3 instances. The DDI corpus annotate each drug pair in the sentence with a DDI type. The DDI type, which is the most concerned information, is described in table TABREF4 . The details about how we train our model and extract the DDI type from text are described in the remaining sections.
## Related Work
In DDI extraction task, NLP methods or machine learning approaches are proposed by most of the work. Chowdhury BIBREF14 and Thomas et al. BIBREF11 proposed methods that use linguistic phenomenons and two-stage SVM to classify DDIs. FBK-irst BIBREF10 is a follow-on work which applies kernel method to the existing model and outperforms it.
Neural network based approaches have been proposed by several works. Liu et al. BIBREF9 employ CNN for DDI extraction for the first time which outperforms the traditional machine learning based methods. Limited by the convolutional kernel size, the CNN can only extracted features of continuous 3 to 5 words rather than distant words. Liu et al. BIBREF8 proposed dependency-based CNN to handle distant but relevant words. Sahu et al. BIBREF12 proposed LSTM based DDI extraction approach and outperforms CNN based approach, since LSTM handles sentence as a sequence instead of slide windows. To conclude, Neural network based approaches have advantages of 1) less reliance on extra NLP toolkits, 2) simpler preprocessing procedure, 3) better performance than text analysis and machine learning methods.
Drug-drug interaction extraction is a relation extraction task of natural language processing. Relation extraction aims to determine the relation between two given entities in a sentence. In recent years, attention mechanism and various neural networks are applied to relation extraction BIBREF17 , BIBREF18 , BIBREF19 , BIBREF20 , BIBREF21 . Convolutional deep neural network are utilized for extracting sentence level features in BIBREF19 . Then the sentence level features are concatenated with lexical level features, which are obtained by NLP toolkit WordNet BIBREF22 , followed by a multilayer perceptron (MLP) to classify the entities' relation. A fixed work is proposed by Nguyen et al. BIBREF21 . The convolutional kernel is set various size to capture more n-gram features. In addition, the word and position embedding are trained automatically instead of keeping constant as in BIBREF19 . Wang et al. BIBREF20 introduce multi-level attention mechanism to CNN in order to emphasize the keywords and ignore the non-critical words during relation detection. The attention CNN model outperforms previous state-of-the-art methods.
Besides CNN, Recurrent neural network (RNN) has been applied to relation extraction as well. Zhang et al. BIBREF18 utilize long short-term memory network (LSTM), a typical RNN model, to represent sentence. The bidirectional LSTM chronologically captures the previous and future information, after which a pooling layer and MLP have been set to extract feature and classify the relation. Attention mechanism is added to bidirectional LSTM in BIBREF17 for relation extraction. An attention layer gives each memory cell a weight so that classifier can catch the principal feature for the relation detection. The Attention based bidirectional LSTM has been proven better than previous work.
## Proposed Model
In this section, we present our bidirectional recurrent neural network with multiple attention layer model. The overview of our architecture is shown in figure FIGREF15 . For a given instance, which describes the details about two or more drugs, the model represents each word as a vector in embedding layer. Then the bidirectional RNN layer generates a sentence matrix, each column vector in which is the semantic representation of the corresponding word. The word level attention layer transforms the sentence matrix to vector representation. Then sentence level attention layer generates final representation for the instance by combining several relevant sentences in view of the fact that these sentences have the same drug pair. Followed by a softmax classifier, the model classifies the drug pair in the given instance as specific DDI.
## Preprocessing
The DDI corpus contains thousands of XML files, each of which are constructed by several records. For a sentence containing INLINEFORM0 drugs, there are INLINEFORM1 drug pairs. We replace the interested two drugs with “drug1” and “drug2” while the other drugs are replaced by “durg0”, as in BIBREF9 did. This step is called drug blinding. For example, the sentence in figure FIGREF5 generates 3 instances after drug blinding: “drug1: an increased risk of hepatitis has been reported to result from combined use of drug2 and drug0”, “drug1: an increased risk of hepatitis has been reported to result from combined use of drug0 and drug2”, “drug0: an increased risk of hepatitis has been reported to result from combined use of drug1 and drug2”. The drug blinded sentences are the instances that are fed to our model.
We put the sentences with the same drug pairs together as a set, since the sentence level attention layer (will be described in Section SECREF21 ) will use the sentences which contain the same drugs.
## Embedding Layer
Given an instance INLINEFORM0 which contains specified two drugs INLINEFORM1 , INLINEFORM2 , each word is embedded in a INLINEFORM3 dimensional space ( INLINEFORM4 , INLINEFORM5 are the dimension of word embedding and position embedding). The look up table function INLINEFORM6 maps a word or a relative position to a column vector. After embedding layer the sentence is represented by INLINEFORM7 , where DISPLAYFORM0
The INLINEFORM0 function is usually implemented with matrix-vector product. Let INLINEFORM1 , INLINEFORM2 denote the one-hot representation (column vector) of word and relative distance. INLINEFORM3 , INLINEFORM4 are word and position embedding query matrix. The look up functions are implemented by DISPLAYFORM0
Then the word sequence INLINEFORM0 is fed to the RNN layer. Note that the sentence will be filled with INLINEFORM1 if its length is less than INLINEFORM2 .
## Bidirectional RNN Encoding Layer
The words in the sequence are read by RNN's gated recurrent unit (GRU) one by one. The GRU takes the current word INLINEFORM0 and the previous GRU's hidden state INLINEFORM1 as input. The current GRU encodes INLINEFORM2 and INLINEFORM3 into a new hidden state INLINEFORM4 (its dimension is INLINEFORM5 , a hyperparameter), which can be regarded as informations the GRU remembered.
Figure FIGREF25 shows the details in GRU. The reset gate INLINEFORM0 selectively forgets informations delivered by previous GRU. Then the hidden state becomes INLINEFORM1 . The update gate INLINEFORM2 updates the informations according to INLINEFORM3 and INLINEFORM4 . The equations below describe these procedures. Note that INLINEFORM5 stands for element wise multiplication. DISPLAYFORM0 DISPLAYFORM1
The bidirectional RNN contains forward RNN and backward RNN. Forward RNN reads sentence from INLINEFORM0 to INLINEFORM1 , generating INLINEFORM2 , INLINEFORM3 , ..., INLINEFORM4 . Backward RNN reads sentence from INLINEFORM5 to INLINEFORM6 , generating INLINEFORM7 , INLINEFORM8 , ..., INLINEFORM9 . Then the encode result of this layer is DISPLAYFORM0
We apply dropout technique in RNN layer to avoid overfitting. Each GRU have a probability (denoted by INLINEFORM0 , also a hyperparameter) of being dropped. The dropped GRU has no output and will not affect the subsequent GRUs. With bidirectional RNN and dropout technique, the input INLINEFORM1 is encoded into sentence matrix INLINEFORM2 .
## Word Level Attention
The purpose of word level attention layer is to extract sentence representation (also known as feature vector) from encoded matrix. We use word level attention instead of max pooling, since attention mechanism can determine the importance of individual encoded word in each row of INLINEFORM0 . Let INLINEFORM1 denotes the attention vector (column vector), INLINEFORM2 denotes the filter that gives each element in the row of INLINEFORM3 a weight. The following equations shows the attention operation, which is also illustrated in figure FIGREF15 . DISPLAYFORM0 DISPLAYFORM1
The softmax function takes a vector INLINEFORM0 as input and outputs a vector, DISPLAYFORM0
INLINEFORM0 denotes the feature vector captured by this layer. Several approaches BIBREF12 , BIBREF17 use this vector and softmax classifier for classification. Inspired by BIBREF23 we propose the sentence level attention to combine the information of other sentences for a improved DDI classification.
## Sentence Level Attention
The previous layers captures the features only from the given sentence. However, other sentences may contains informations that contribute to the understanding of this sentence. It is reasonable to look over other relevant instances when determine two drugs' interaction from the given sentence. In our implementation, the instances that have the same drug pair are believed to be relevant. The relevant instances set is denoted by INLINEFORM0 , where INLINEFORM1 is the sentence feature vector. INLINEFORM2 stands for how well the instance INLINEFORM3 matches its DDI INLINEFORM4 (Vector representation of a specific DDI). INLINEFORM5 is a diagonal attention matrix, multiplied by which the feature vector INLINEFORM6 can concentrate on those most representative features. DISPLAYFORM0 DISPLAYFORM1
INLINEFORM0 is the softmax result of INLINEFORM1 . The final sentence representation is decided by all of the relevant sentences' feature vector, as Equation EQREF24 shows. DISPLAYFORM0
Note that the set INLINEFORM0 is gradually growing as new sentence with the same drugs pairs is found when training. An instance INLINEFORM1 is represented by INLINEFORM2 before sentence level attention. The sentence level attention layer finds the set INLINEFORM3 , instances in which have the same drug pair as in INLINEFORM4 , and put INLINEFORM5 in INLINEFORM6 . Then the final sentence representation INLINEFORM7 is calculated in this layer.
## Classification and Training
A given sentence INLINEFORM0 is finally represented by the feature vector INLINEFORM1 . Then we feed it to a softmax classifier. Let INLINEFORM2 denotes the set of all kinds of DDI. The output INLINEFORM3 is the probabilities of each class INLINEFORM4 belongs. DISPLAYFORM0
We use cross entropy cost function and INLINEFORM0 regularization as the optimization objective. For INLINEFORM1 -th instance, INLINEFORM2 denotes the one-hot representation of it's label, where the model outputs INLINEFORM3 . The cross entropy cost is: DISPLAYFORM0
For a mini-batch INLINEFORM0 , the optimization objective is: DISPLAYFORM0
All parameters in this model is: DISPLAYFORM0
We optimize the parameters of objective function INLINEFORM0 with Adam BIBREF24 , which is a variant of mini-batch stochastic gradient descent. During each train step, the gradient of INLINEFORM1 is calculated. Then INLINEFORM2 is adjusted according to the gradient. After the end of training, we have a model that is able to predict two drugs' interactions when a sentence about these drugs is given.
## DDI Prediction
The model is trained for DDI classification. The parameters in list INLINEFORM0 are tuned during the training process. Given a new sentence with two drugs, we can use this model to classify the DDI type.
The DDI prediction follows the procedure described in Section SECREF6 - SECREF26 . The given sentence is eventually represented by feature vector INLINEFORM0 . Then INLINEFORM1 is classified to a specific DDI type with a softmax classifier. In next section, we will evaluate our model's DDI prediction performance and see the advantages and shortcomings of our model.
## Datasets and Evaluation Metrics
We use the DDI corpus of the 2013 DDIExtraction challenge BIBREF16 to train and test our model. The DDIs in this corpus are classified as five types. We give the definitions of these types and their example sentences, as shown in table TABREF4 . This standard dataset is made up of training set and testing set. We use the same metrics as in other drug-drug interaction extraction literature BIBREF11 , BIBREF10 , BIBREF25 , BIBREF9 , BIBREF8 , BIBREF12 : the overall precision, recall, and F1 score on testing set. INLINEFORM0 denotes the set of {False, Mechanism, Effect, Advise, Int}. The precision and recall of each INLINEFORM1 are calculated by DISPLAYFORM0 DISPLAYFORM1
Then the overall precision, recall, and F1 score are calculated by DISPLAYFORM0
Besides, we evaluate the captured feature vectors with t-SNE BIBREF26 , a visualizing and intuitive way to map a high dimensional vector into a 2 or 3-dimensional space. If the points in a low dimensional space are easy to be split, the feature vectors are believed to be more distinguishable.
## Hyperparameter Settings and Training
We use TensorFlow BIBREF27 r0.11 to implement the proposed model. The input of each word is an ordered triple (word, relative distance from drug1, relative distance from drug2). The sentence, which is represented as a matrix, is fed to the model. The output of the model is a INLINEFORM0 -dimensional vector representing the probabilities of being corresponding DDI. It is the network, parameters, and hyperparameters which decides the output vector. The network's parameters are adjusted during training, where the hyperparameters are tuned by hand. The hyperparameters after tuning are as follows. The word embedding's dimension INLINEFORM1 , the position embedding's dimension INLINEFORM2 , the hidden state's dimension INLINEFORM3 , the probability of dropout INLINEFORM4 , other hyperparameters which are not shown here are set to TensorFlow's default values.
The word embedding is initialized by pre-trained word vectors using GloVe BIBREF28 , while other parameters are initialized randomly. During each training step, a mini-batch (the mini-batch size INLINEFORM0 in our implementation) of sentences is selected from training set. The gradient of objective function is calculated for parameters updating (See Section SECREF26 ).
Figure FIGREF32 shows the training process. The objective function INLINEFORM0 is declining as the training mini-batches continuously sent to the model. As the testing mini-batches, the INLINEFORM1 function is fluctuating while its overall trend is descending. The instances in testing set are not participated in training so that INLINEFORM2 function is not descending so fast. However, training and testing instances have similar distribution in sample space, causing that testing instances' INLINEFORM3 tends to be smaller along with the training process. INLINEFORM4 has inverse relationship with the performance measurement. The F1 score is getting fluctuating around a specific value after enough training steps. The reason why fluctuating range is considerable is that only a tiny part of the whole training or testing set has been calculated the F1 score. Testing the whole set during every step is time consuming and not necessary. We will evaluate the model on the whole testing set in Section SECREF47 .
## Experimental Results
We save our model every 100 step and predict all the DDIs of the instances in the testing set. These predictions' F1 score is shown in figure FIGREF40 . To demonstrate the sentence level attention layer is effective, we drop this layer and then directly use INLINEFORM0 for softmax classification (See figure FIGREF15 ). The result is shown with “RNN + dynamic word embedding + ATT” curve, which illustrates that the sentence level attention layer contributes to a more accurate model.
Whether a dynamic or static word embedding is better for a DDI extraction task is under consideration. Nguyen et al. BIBREF21 shows that updating word embedding at the time of other parameters being trained makes a better performance in relation extraction task. We let the embedding be static when training, while other conditions are all the same. The “RNN + static word embedding + 2ATT” curve shows this case. We can draw a conclusion that updating the initialized word embedding trains more suitable word vectors for the task, which promotes the performance.
We compare our best F1 score with other state-of-the-art approaches in table TABREF39 , which shows our model has competitive advantage in dealing with drug-drug interaction extraction. The predictions confusion matrix is shown in table TABREF46 . The DDIs other than false being classified as false makes most of the classification error. It may perform better if a classifier which can tells true and false DDI apart is trained. We leave this two-stage classifier to our future work. Another phenomenon is that the “Int” type is often classified as “Effect”. The “Int” sentence describes there exists interaction between two drugs and this information implies the two drugs' combination will have good or bed effect. That's the reason why “Int” and “Effect” are often obfuscated.
To evaluate the features our model captured, we employ scikit-learn BIBREF29 's t-SNE class to map high dimensional feature vectors to 2-dimensional vectors, which can be depicted on a plane. We depict all the features of the instances in testing set, as shown in figure FIGREF41 . The RNN model using dynamic word embedding and 2 layers of attention is the most distinguishable one. Unfortunately, the classifier can not classify all the instances into correct classes. Comparing table TABREF46 with figure UID44 , both of which are from the best performed model, we can observe some conclusions. The “Int” DDIs are often misclassified as “Effect”, for the reason that some of the “Int” points are in the “Effect” cluster. The “Effect” points are too scattered so that plenty of “Effect” DDIs are classified to other types. The “Mechanism” points are gathered around two clusters, causing that most of the “mechanism” DDIs are classified to two types: “False” and “Mechanism”. In short, the visualizability of feature mapping gives better explanations for the prediction results and the quality of captured features.
## Conclusion and Future Work
To conclude, we propose a recurrent neural network with multiple attention layers to extract DDIs from biomedical text. The sentence level attention layer, which combines other sentences containing the same drugs, has been added to our model. The experiments shows that our model outperforms the state-of-the-art DDI extraction systems. Task relevant word embedding and two attention layers improved the performance to some extent.
The imbalance of the classes and the ambiguity of semantics cause most of the misclassifications. We consider that instance generation using generative adversarial networks would cover the instance shortage in specific category. It is also reasonable to use distant supervision learning (which utilize other relevant material) for knowledge supplement and obtain a better performed DDI extraction system.
## Acknowledgment
This work is supported by the NSFC under Grant 61303191, 61303190, 61402504, 61103015.
| 15
|
1705.07368
|
Mixed Membership Word Embeddings for Computational Social Science
|
# Mixed Membership Word Embeddings for Computational Social Science
## Abstract
Word embeddings improve the performance of NLP systems by revealing the hidden structural relationships between words. Despite their success in many applications, word embeddings have seen very little use in computational social science NLP tasks, presumably due to their reliance on big data, and to a lack of interpretability. I propose a probabilistic model-based word embedding method which can recover interpretable embeddings, without big data. The key insight is to leverage mixed membership modeling, in which global representations are shared, but individual entities (i.e. dictionary words) are free to use these representations to uniquely differing degrees. I show how to train the model using a combination of state-of-the-art training techniques for word embeddings and topic models. The experimental results show an improvement in predictive language modeling of up to 63% in MRR over the skip-gram, and demonstrate that the representations are beneficial for supervised learning. I illustrate the interpretability of the models with computational social science case studies on State of the Union addresses and NIPS articles.
## Introduction
Word embedding models, which learn to encode dictionary words with vector space representations, have been shown to be valuable for a variety of natural language processing (NLP) tasks such as statistical machine translation BIBREF2 , part-of-speech tagging, chunking, and named entity recogition BIBREF3 , as they provide a more nuanced representation of words than a simple indicator vector into a dictionary. These models follow a long line of research in data-driven semantic representations of text, including latent semantic analysis BIBREF4 and its probabilistic extensions BIBREF5 , BIBREF6 . In particular, topic models BIBREF7 have found broad applications in computational social science BIBREF8 , BIBREF9 and the digital humanities BIBREF10 , where interpretable representations reveal meaningful insights. Despite widespread success at NLP tasks, word embeddings have not yet supplanted topic models as the method of choice in computational social science applications. I speculate that this is due to two primary factors: 1) a perceived reliance on big data, and 2) a lack of interpretability. In this work, I develop new models to address both of these limitations.
Word embeddings have risen in popularity for NLP applications due to the success of models designed specifically for the big data setting. In particular, BIBREF0 , BIBREF1 showed that very simple word embedding models with high-dimensional representations can scale up to massive datasets, allowing them to outperform more sophisticated neural network language models which can process fewer documents. In this work, I offer a somewhat contrarian perspective to the currently prevailing trend of big data optimism, as exemplified by the work of BIBREF0 , BIBREF1 , BIBREF3 , and others, who argue that massive datasets are sufficient to allow language models to automatically resolve many challenging NLP tasks. Note that “big” datasets are not always available, particularly in computational social science NLP applications, where the data of interest are often not obtained from large scale sources such as the internet and social media, but from sources such as press releases BIBREF11 , academic journals BIBREF10 , books BIBREF12 , and transcripts of recorded speech BIBREF13 , BIBREF14 , BIBREF15 . A standard practice in the literature is to train word embedding models on a generic large corpus such as Wikipedia, and use the embeddings for NLP tasks on the target dataset, cf. BIBREF3 , BIBREF0 , BIBREF16 , BIBREF17 . However, as we shall see here, this standard practice might not always be effective, as the size of a dataset does not correspond to its degree of relevance for a particular analysis. Even very large corpora have idiosyncrasies that can make their embeddings invalid for other domains. For instance, suppose we would like to use word embeddings to analyze scientific articles on machine learning. In Table TABREF1 , I report the most similar words to the word “learning” based on word embedding models trained on two corpora. For embeddings trained on articles from the NIPS conference, the most similar words are related to machine learning, as desired, while for embeddings trained on the massive, generic Google News corpus, the most similar words relate to learning and teaching in the classroom. Evidently, domain-specific data can be important.
Even more concerningly, BIBREF18 show that word embeddings can encode implicit sexist assumptions. This suggests that when trained on large generic corpora they could also encode the hegemonic worldview, which is inappropriate for studying, e.g., black female hip-hop artists' lyrics, or poetry by Syrian refugees, and could potentially lead to systematic bias against minorities, women, and people of color in NLP applications with real-world consequences, such as automatic essay grading and college admissions. In order to proactively combat these kinds of biases in large generic datasets, and to address computational social science tasks, there is a need for effective word embeddings for small datasets, so that the most relevant datasets can be used for training, even when they are small. To make word embeddings a viable alternative to topic models for applications in the social sciences, we further desire that the embeddings are semantically meaningful to human analysts.
In this paper, I introduce an interpretable word embedding model, and an associated topic model, which are designed to work well when trained on a small to medium-sized corpus of interest. The primary insight is to use a data-efficient parameter sharing scheme via mixed membership modeling, with inspiration from topic models. Mixed membership models provide a flexible yet efficient latent representation, in which entities are associated with shared, global representations, but to uniquely varying degrees. I identify the skip-gram word2vec model of BIBREF0 , BIBREF1 as corresponding to a certain naive Bayes topic model, which leads to mixed membership extensions, allowing the use of fewer vectors than words. I show that this leads to better modeling performance without big data, as measured by predictive performance (when the context is leveraged for prediction), as well as to interpretable latent representations that are highly valuable for computational social science applications. The interpretability of the representations arises from defining embeddings for words (and hence, documents) in terms of embeddings for topics. My experiments also shed light on the relative merits of training embeddings on generic big data corpora versus domain-specific data.
## Background
In this section, I provide the necessary background on word embeddings, as well as on topic models and mixed membership models. Traditional language models aim to predict words given the contexts that they are found in, thereby forming a joint probabilistic model for sequences of words in a language. BIBREF19 developed improved language models by using distributed representations BIBREF20 , in which words are represented by neural network synapse weights, or equivalently, vector space embeddings.
Later authors have noted that these word embeddings are useful for semantic representations of words, independently of whether a full joint probabilistic language model is learned, and that alternative training schemes can be beneficial for learning the embeddings. In particular, BIBREF0 , BIBREF1 proposed the skip-gram model, which inverts the language model prediction task and aims to predict the context given an input word. The skip-gram model is a log-bilinear discriminative probabilistic classifier parameterized by “input” word embedding vectors INLINEFORM0 for the input words INLINEFORM1 , and “output” word embedding vectors INLINEFORM2 for context words INLINEFORM3 , as shown in Table TABREF2 , top-left.
Topic models such as latent Dirichlet allocation (LDA) BIBREF7 are another class of probabilistic language models that have been used for semantic representation BIBREF6 . A straightforward way to model text corpora is via unsupervised multinomial naive Bayes, in which a latent cluster assignment for each document selects a multinomial distribution over words, referred to as a topic, with which the documents' words are assumed to be generated. LDA topic models improve over naive Bayes by using a mixed membership model, in which the assumption that all words in a document INLINEFORM0 belong to the same topic is relaxed, and replaced with a distribution over topics INLINEFORM1 . In the model's assumed generative process, for each word INLINEFORM2 in document INLINEFORM3 , a topic assignment INLINEFORM4 is drawn via INLINEFORM5 , then the word is drawn from the chosen topic INLINEFORM6 . The mixed membership formalism provides a useful compromise between model flexibility and statistical efficiency: the INLINEFORM7 topics INLINEFORM8 are shared across all documents, thereby sharing statistical strength, but each document is free to use the topics to its own unique degree. Bayesian inference further aids data efficiency, as uncertainty over INLINEFORM9 can be managed for shorter documents. Some recent papers have aimed to combine topic models and word embeddings BIBREF21 , BIBREF22 , but they do not aim to address the small data problem for computational social science, which I focus on here. I provide a more detailed discussion of related work in the supplementary.
## The Mixed Membership Skip-Gram
To design an interpretable word embedding model for small corpora, we identify novel connections between word embeddings and topic models, and adapt advances from topic modeling. Following the distributional hypothesis BIBREF23 , the skip-gram's word embeddings parameterize discrete probability distributions over words INLINEFORM0 which tend to co-occur, and tend to be semantically coherent – a property leveraged by the Gaussian LDA model of BIBREF21 . This suggests that these discrete distributions can be reinterpreted as topics INLINEFORM1 . We thus reinterpret the skip-gram as a parameterization of a certain supervised naive Bayes topic model (Table TABREF2 , top-right). In this topic model, input words INLINEFORM2 are fully observed “cluster assignments,” and the words in INLINEFORM3 's contexts are a “document.” The skip-gram differs from this supervised topic model only in the parameterization of the “topics” via word vectors which encode the distributions with a log-bilinear model. Note that although the skip-gram is discriminative, in the sense that it does not jointly model the input words INLINEFORM4 , we are here equivalently interpreting it as encoding a “conditionally generative” process for the context given the words, in order to develop probabilistic models that extend the skip-gram.
As in LDA, this model can be improved by replacing the naive Bayes assumption with a mixed membership assumption. By applying the mixed membership representation to this topic model version of the skip-gram, we obtain the model in the bottom-right of Table TABREF2 . After once again parameterizing this model with word embeddings, we obtain our final model, the mixed membership skip-gram (MMSG) (Table TABREF2 , bottom-left). In the model, each input word has a distribution over topics INLINEFORM0 . Each topic has a vector-space embedding INLINEFORM1 and each output word has a vector INLINEFORM2 (a parameter, not an embedding for INLINEFORM3 ). A topic INLINEFORM4 is drawn for each context, and the words in the context are drawn from the log-bilinear model using INLINEFORM5 : DISPLAYFORM0
We can expect that the resulting mixed membership word embeddings are beneficial in the small-to-medium data regime for the following reasons:
Of course, the model also requires some new parameters to be learned, namely the mixed membership proportions INLINEFORM0 . Based on topic modeling, I hypothesized that with care, these added parameters need not adversely affect performance in the small-medium data regime, for two reasons: 1) we can use a Bayesian approach to effectively manage uncertainty in them, and to marginalize them out, which prevents them being a bottleneck during training; and 2) at test time, using the posterior for INLINEFORM1 given the context, instead of the “prior” INLINEFORM2 , mitigates the impact of uncertainty in INLINEFORM3 due to limited training data: DISPLAYFORM0
To obtain a vector for a word type INLINEFORM0 , we can use the prior mean, INLINEFORM1 . For a word token INLINEFORM2 , we can leverage its context via the posterior mean, INLINEFORM3 . These embeddings are convex combinations of topic vectors (see Figure FIGREF23 for an example). With fewer vectors than words, some model capacity is lost, but the flexibility of the mixed membership representation allows the model to compensate. When the number of shared vectors equals the number of words, the mixed membership skip-gram is strictly more representationally powerful than the skip-gram. With more vectors than words, we can expect that the increased representational power would be beneficial in the big data regime. As this is not my goal, I leave this for future work.
## Experimental Results
The goals of our experiments were to study the relative merits of big data and domain-specific small data, to validate the proposed methods, and to study their applicability for computational social science research.
## Quantitative Experiments
I first measured the effectiveness of the embeddings at the skip-gram's training task, predicting context words INLINEFORM0 given input words INLINEFORM1 . This task measures the methods' performance for predictive language modeling. I used four datasets of sociopolitical, scientific, and literary interest: the corpus of NIPS articles from 1987 – 1999 ( INLINEFORM2 million), the U.S. presidential state of the Union addresses from 1790 – 2015 ( INLINEFORM3 ), the complete works of Shakespeare ( INLINEFORM4 ; this version did not contain the Sonnets), and the writings of black scholar and activist W.E.B. Du Bois, as digitized by Project Gutenberg ( INLINEFORM5 ). For each dataset, I held out 10,000 INLINEFORM6 pairs uniformly at random, where INLINEFORM7 , and aimed to predict INLINEFORM8 given INLINEFORM9 (and optionally, INLINEFORM10 ). Since there are a large number of classes, I treat this as a ranking problem, and report the mean reciprocal rank. The experiments were repeated and averaged over 5 train/test splits.
The results are shown in Table TABREF25 . I compared to a word frequency baseline, the skip-gram (SG), and Tomas Mikolov/Google's vectors trained on Google News, INLINEFORM0 billion, via CBOW. Simulated annealing was performed for 1,000 iterations, NCE was performed for 1 million minibatches of size 128, and 128-dimensional embeddings were used (300 for Google). I used INLINEFORM1 for NIPS, INLINEFORM2 for state of the Union, and INLINEFORM3 for the two smaller datasets. Methods were able to leverage the remainder of the context, either by adding the context's vectors, or via the posterior (Equation EQREF22 ), which helped for all methods except the naive skip-gram. We can identify several noteworthy findings. First, the generic big data vectors (Google+context) were outperformed by the skip-gram on 3 out of 4 datasets (and by the skip-gram topic model on the other), by a large margin, indicating that domain-specific embeddings are often important. Second, the mixed membership models, using posterior inference, beat or matched their naive Bayes counterparts, for both the word embedding models and the topic models. As hypothesized, posterior inference on INLINEFORM4 at test time was important for good performance. Finally, the topic models beat their corresponding word embedding models at prediction. I therefore recommend the use of our MMSG topic model variant for predictive language modeling in the small data regime.
I tested the performance of the representations as features for document categorization and regression tasks. The results are given in Table TABREF26 . For document categorization, I used three standard benchmark datasets: 20 Newsgroups (19,997 newsgroup posts), Reuters-150 newswire articles (15,500 articles and 150 classes), and Ohsumed medical abstracts on 23 cardiovascular diseases (20,000 articles). I held out 4,000 test documents for 20 Newsgroups, and used the standard train/test splits from the literature in the other corpora (e.g. for Ohsumed, 50% of documents were assigned to training and to test sets). I obtained document embeddings for the MMSG, in the same latent space as the topic embeddings, by summing the posterior mean vectors INLINEFORM0 for each token. Vector addition was similarly used to construct document vectors for the other embedding models. All vectors were normalized to unit length. I also considered a tf-idf baseline. Logistic regression models were trained on the features extracted on the training set for each method.
Across the three datasets, several clear trends emerged (Table TABREF26 ). First, the generic Google vectors were consistently and substantially outperformed in classification performance by the skipgram (SG) and MMSG vectors, highlighting the importance of corpus-specific embeddings. Second, despite the MMSG's superior performance at language modeling on small datasets, the SG features outperformed the MMSG's at the document categorization task. By encoding vectors at the topic level instead of the word level, the MMSG loses word level resolution in the embeddings, which turned out to be valuable for these particular classification tasks. We are not, however, restricted to use only one type of embedding to construct features for classification. Interestingly, when the SG and MMSG features were concatenated (SG+MMSG), this improved classification performance over these vectors individually. This suggests that the topic-level MMSG vectors and word-level SG vectors encode complementary information, and both are beneficial for performance. Finally, further concatenating the generic Google vectors' features (SG+MMSG+Google) improved performance again, despite the fact that these vectors performed poorly on their own. It should be noted that tf-idf, which is notoriously effective for document categorization, outperformed the embedding methods on these datasets.
I also analyzed the regression task of predicting the year of a state of the Union address based on its text information. I used lasso-regularized linear regression models, evaluated via a leave-one-out cross-validation experimental setup. Root-mean-square error (RMSE) results are reported in Table TABREF26 (bottom). Unlike for the other tasks, the Google big data vectors were the best individual features in this case, outperforming the domain-specific SG and MMSG embeddings individually. On the other hand, SG+MMSG+Google performed the best overall, showing that domain-specific embeddings can improve performance even when big data embeddings are successful. The tf-idf baseline was beaten by all of the embedding models on this task.
## Computational Social Science Case Studies: State of the Union and NIPS
I also performed several case studies. I obtained document embeddings, in the same latent space as the topic embeddings, by summing the posterior mean vectors INLINEFORM0 for each token, and visualized them in two dimensions using INLINEFORM1 -SNE BIBREF24 (all vectors were normalized to unit length). The state of the Union addresses (Figure FIGREF27 ) are embedded almost linearly by year, with a major jump around the New Deal (1930s), and are well separated by party at any given time period. The embedded topics (gray) allow us to interpret the space. The George W. Bush addresses are embedded near a “war on terror” topic (“weapons, war...”), and the Barack Obama addresses are embedded near a “stimulus” topic (“people, work...”).
On the NIPS corpus, for the input word “Bayesian” (Table ), the naive Bayes and skip-gram models learned a topic with words that refer to Bayesian networks, probabilistic models, and neural networks. The mixed membership models are able to separate this into more coherent and specific topics including Bayesian modeling, Bayesian training of neural networks (for which Sir David MacKay was a strong proponent, and Andreas Weigend wrote an influential early paper), and Monte Carlo methods. By performing the additive composition of word vectors, which we obtain by finding the prior mean vector for each word type INLINEFORM0 , INLINEFORM1 (and then normalizing), we obtain relevant topics INLINEFORM2 as nearest neighbors (Figure FIGREF28 ). Similarly, we find that the additive composition of topic and word vectors works correctly: INLINEFORM3 , and INLINEFORM4 .
The INLINEFORM0 -SNE visualization of NIPS documents (Figure FIGREF28 ) shows some temporal clustering patterns (blue documents are more recent, red documents are older, and gray points are topics). I provide a more detailed case study on NIPS in the supplementary material.
## Conclusion
I have proposed a model-based method for training interpretable corpus-specific word embeddings for computational social science, using mixed membership representations, Metropolis-Hastings-Walker sampling, and NCE. Experimental results for prediction, supervised learning, and case studies on state of the Union addresses and NIPS articles, indicate that high-quality embeddings and topics can be obtained using the method. The results highlight the fact that big data is not always best, as domain-specific data can be very valuable, even when it is small. I plan to use this approach for substantive social science applications, and to address algorithmic bias and fairness issues.
Acknowledgements
I thank Eric Nalisnick and Padhraic Smyth for many helpful discussions.
## Supplementary Material
]
## Related Work
In this supplementary document, we discuss related work in the literature and its relation to our proposed methods, provide a case study on NIPS articles, and derive the collapsed Gibbs sampling update for the MMSGTM, which we leverage when training the MMSG.
## Topic Modeling and Word Embeddings
The Gaussian LDA model of BIBREF21 improves the performance of topic modeling by leveraging the semantic information encoded in word embeddings. Gaussian LDA modifies the generative process of LDA such that each topic is assumed to generate the vectors via its own Gaussian distribution. Similarly to our MMSG model, in Gaussian LDA each topic is encoded with a vector, in this case the mean of the Gaussian. It takes pre-trained word embeddings as input, rather than learning the embeddings from data within the same model, and does not aim to perform word embedding.
The topical word embedding (TWE) models of BIBREF22 reverse this, as they take LDA topic assignments of words as input, and aim to use them to improve the resultant word embeddings. The authors propose three variants, each of which modifies the skip-gram training objective to use LDA topic assignments together with words. In the best performing variant, called TWE-1, a standard skip-gram word embedding model is trained independently with another skip-gram variant, which tries to predict context words given the input word's topic assignment. The skip-gram embedding and the topic embeddings are concatenated to form the final embedding.
At test time, a distribution over topics for the word given the context, INLINEFORM0 is estimated according to the topic counts over the other context words. Using this as a prior, a posterior over topics given both the input word and the context is calculated, and similarities between pairs of words (with their contexts) are averaged over this posterior, in a procedure inspired by those used by BIBREF43 , BIBREF36 . The primary similarity to our MMSG approach is the use of a training algorithm involving the prediction of context words, given a topic. Our method does this as part of an overall model-based inference procedure, and we learn mixed membership proportions INLINEFORM1 rather than using empirical counts as the prior over topics for a word token. In accordance with the skip-gram's prediction model, we are thus able to model the context words in the data likelihood term when computing the posterior probability of the topic assignment. TWE-1 requires that topic assignments are available at test time. It provides a mechanism to predict contextual similarity, but not to predict held-out context words, so we are unable to compare to it in our experiments.
Other neurally-inspired topic models include replicated softmax BIBREF34 , and its successor, DocNADE BIBREF37 . Replicated softmax extends the restricted Boltzmann machine to handle multinomial counts for document modeling. DocNADE builds on the ideas of replicated softmax, but uses the NADE architecture, where observations (i.e. words) are modeled sequentially given the previous observations.
## Multi-Prototype Embedding Models
Multi-prototype embeddings models are another relevant line of work. These models address lexical ambiguity by assigning multiple vectors to each word type, each corresponding to a different meaning of that word. BIBREF43 propose to cluster the occurrences of each word type, based on features extracted from its context. Embeddings are then learned for each cluster. BIBREF36 apply a similar approach, but they use initial single-prototype word embeddings to provide the features used for clustering. These clustering methods have some resemblance to our topic model pre-clustering step, although their clustering is applied within instances of a given word type, rather than globally across all word types, as in our methods. This results in models with more vectors than words, while we aim to find fewer vectors than words, to reduce the model's complexity for small datasets. Rather than employing an off-the-shelf clustering algorithm and then applying an unrelated embedding model to its output, our approach aims to perform model-based clustering within an overall joint model of topic/cluster assignments and word vectors.
Perhaps the most similar model to ours in the literature is the probabilistic multi-prototype embedding model of BIBREF45 , who treat the prototype assignment of a word as a latent variable, assumed drawn from a mixture over prototypes for each word. The embeddings are then trained using EM. Our MMSG model can be understood as the mixed membership version of this model, in which the prototypes (vectors) are shared across all word types, and each word type has its own mixed membership proportions across the shared prototypes. While a similar EM algorithm can be applied to the MMSG, the E-step is much more expensive, as we typically desire many more shared vectors (often in the thousands) than we would prototypes per a single word type (Tian et al. use ten in their experiments). We use the Metropolis-Hastings-Walker algorithm with the topic model reparameterization of our model in order to address this by efficiently pre-solving the E-step.
## Mixed Membership Modeling
Mixed membership modeling is a flexible alternative to traditional clustering, in which each data point is assigned to a single cluster. Instead, mixed membership models posit that individual entities are associated with multiple underlying clusters, to differing degrees, as encoded by a mixed membership vector that sums to one across the clusters BIBREF28 , BIBREF26 . These mixed membership proportions are generally used to model lower-level grouped data, such as the words inside a document. Each lower-level data point inside a group is assumed to be assigned to one of the shared, global clusters according to the group-level membership proportions. Thus, a mixed membership model consists of a mixture model for each group, which share common mixture component parameters, but with differing mixture proportions.
This formalism has lead to probabilistic models for a variety of applications, including medical diagnosis BIBREF39 , population genetics BIBREF42 , survey analysis BIBREF29 , computer vision BIBREF27 , BIBREF30 , text documents BIBREF35 , BIBREF7 , and social network analysis BIBREF25 . Nonparametric Bayesian extensions, in which the number of underlying clusters is learned from data via Bayesian inference, have also been proposed BIBREF44 . In this work, dictionary words are assigned a mixed membership distribution over a set of shared latent vector space embeddings. Each instantiation of a dictionary word (an “input” word) is assigned to one of the shared embeddings based on its dictionary word's membership vector. The words in its context (“output” words) are assumed to be drawn based on the chosen embedding.
## Case Study on NIPS
In Figure FIGREF33 , we show a zoomed in INLINEFORM0 -SNE visualization of NIPS document embeddings. We can see regions of the space corresponding to learning algorithms (bottom), data space and latent space (center), training neural networks (top), and nearest neighbors (bottom-left). We also visualized the authors' embeddings via INLINEFORM1 -SNE (Figure FIGREF34 ). We find regions of latent space for reinforcement learning authors (left: “state, action,...,” Singh, Barto,Sutton), probabilistic methods (right: “mixture, model,” “monte, carlo,” Bishop, Williams, Barber, Opper, Jordan, Ghahramani, Tresp, Smyth), and evaluation (top-right: “results, performance, experiments,...”).
## Derivation of the Collapsed Gibbs Update
Let INLINEFORM0 be the number of output words in the INLINEFORM1 th context, let INLINEFORM2 be those output words, and let INLINEFORM3 be the input words other that INLINEFORM4 (similarly, topic assignments INLINEFORM5 and output words INLINEFORM6 ). Then the collapsed Gibbs update samples from the conditional distribution INLINEFORM7
We recognize the first integral as the mean of a Dirichlet distribution which we obtain via conjugacy: INLINEFORM0
The above can also be understood as the probability of the next ball drawn from a multivariate Polya urn model, also known as the Dirichlet-compound multinomial distribution, arising from the posterior predictive distribution of a discrete likelihood with a Dirichlet prior. We will need the full form of such a distribution to analyze the second integral. Once again leveraging conjugacy, we have: INLINEFORM0
INLINEFORM0
where INLINEFORM0 is the number of times that output word INLINEFORM1 occurs in the INLINEFORM2 th context, since the final integral is over the full support of a Dirichlet distribution, which integrates to one. Eliminating terms that aren't affected by the INLINEFORM3 assignment, the above is INLINEFORM4
where we have used the fact that INLINEFORM0 for any INLINEFORM1 , and integer INLINEFORM2 . We can interpret this as the probability of drawing the context words under the multivariate Polya urn model, in which the number of “colored balls” (word counts plus prior counts) is increased by one each time a certain color (word) is selected. In other words, in each step, corresponding to the selection of each context word, we draw a ball from the urn, then put it back, along with another ball of the same color. The INLINEFORM3 and INLINEFORM4 terms reflect that the counts have been changed by adding these extra balls into the urn in each step. The second to last equation shows that this process is exchangeable: it does not matter which order the balls were drawn in when determining the probability of the sequence. Multiplying this with the term from the first integral, calculated earlier, gives us the final form of the update equation, INLINEFORM5
| 14
|
1705.09665
|
Community Identity and User Engagement in a Multi-Community Landscape
|
# Community Identity and User Engagement in a Multi-Community Landscape
## Abstract
A community's identity defines and shapes its internal dynamics. Our current understanding of this interplay is mostly limited to glimpses gathered from isolated studies of individual communities. In this work we provide a systematic exploration of the nature of this relation across a wide variety of online communities. To this end we introduce a quantitative, language-based typology reflecting two key aspects of a community's identity: how distinctive, and how temporally dynamic it is. By mapping almost 300 Reddit communities into the landscape induced by this typology, we reveal regularities in how patterns of user engagement vary with the characteristics of a community. Our results suggest that the way new and existing users engage with a community depends strongly and systematically on the nature of the collective identity it fosters, in ways that are highly consequential to community maintainers. For example, communities with distinctive and highly dynamic identities are more likely to retain their users. However, such niche communities also exhibit much larger acculturation gaps between existing users and newcomers, which potentially hinder the integration of the latter. More generally, our methodology reveals differences in how various social phenomena manifest across communities, and shows that structuring the multi-community landscape can lead to a better understanding of the systematic nature of this diversity.
## Introduction
“If each city is like a game of chess, the day when I have learned the rules, I shall finally possess my empire, even if I shall never succeed in knowing all the cities it contains.”
— Italo Calvino, Invisible Cities
A community's identity—defined through the common interests and shared experiences of its users—shapes various facets of the social dynamics within it BIBREF0 , BIBREF1 , BIBREF2 . Numerous instances of this interplay between a community's identity and social dynamics have been extensively studied in the context of individual online communities BIBREF3 , BIBREF4 , BIBREF5 . However, the sheer variety of online platforms complicates the task of generalizing insights beyond these isolated, single-community glimpses. A new way to reason about the variation across multiple communities is needed in order to systematically characterize the relationship between properties of a community and the dynamics taking place within.
One especially important component of community dynamics is user engagement. We can aim to understand why users join certain communities BIBREF6 , what factors influence user retention BIBREF7 , and how users react to innovation BIBREF5 . While striking patterns of user engagement have been uncovered in prior case studies of individual communities BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , we do not know whether these observations hold beyond these cases, or when we can draw analogies between different communities. Are there certain types of communities where we can expect similar or contrasting engagement patterns?
To address such questions quantitatively we need to provide structure to the diverse and complex space of online communities. Organizing the multi-community landscape would allow us to both characterize individual points within this space, and reason about systematic variations in patterns of user engagement across the space.
Present work: Structuring the multi-community space. In order to systematically understand the relationship between community identityand user engagement we introduce a quantitative typology of online communities. Our typology is based on two key aspects of community identity: how distinctive—or niche—a community's interests are relative to other communities, and how dynamic—or volatile—these interests are over time. These axes aim to capture the salience of a community's identity and dynamics of its temporal evolution.
Our main insight in implementing this typology automatically and at scale is that the language used within a community can simultaneously capture how distinctive and dynamic its interests are. This language-based approach draws on a wealth of literature characterizing linguistic variation in online communities and its relationship to community and user identity BIBREF16 , BIBREF5 , BIBREF17 , BIBREF18 , BIBREF19 . Basing our typology on language is also convenient since it renders our framework immediately applicable to a wide variety of online communities, where communication is primarily recorded in a textual format.
Using our framework, we map almost 300 Reddit communities onto the landscape defined by the two axes of our typology (Section SECREF2 ). We find that this mapping induces conceptually sound categorizations that effectively capture key aspects of community-level social dynamics. In particular, we quantitatively validate the effectiveness of our mapping by showing that our two-dimensional typology encodes signals that are predictive of community-level rates of user retention, complementing strong activity-based features.
Engagement and community identity. We apply our framework to understand how two important aspects of user engagement in a community—the community's propensity to retain its users (Section SECREF3 ), and its permeability to new members (Section SECREF4 )—vary according to the type of collective identity it fosters. We find that communities that are characterized by specialized, constantly-updating content have higher user retention rates, but also exhibit larger linguistic gaps that separate newcomers from established members.
More closely examining factors that could contribute to this linguistic gap, we find that especially within distinctive communities, established users have an increased propensity to engage with the community's specialized content, compared to newcomers (Section SECREF5 ). Interestingly, while established members of distinctive communities more avidly respond to temporal updates than newcomers, in more generic communities it is the outsiders who engage more with volatile content, perhaps suggesting that such content may serve as an entry-point to the community (but not necessarily a reason to stay). Such insights into the relation between collective identity and user engagement can be informative to community maintainers seeking to better understand growth patterns within their online communities.
More generally, our methodology stands as an example of how sociological questions can be addressed in a multi-community setting. In performing our analyses across a rich variety of communities, we reveal both the diversity of phenomena that can occur, as well as the systematic nature of this diversity.
## A typology of community identity
A community's identity derives from its members' common interests and shared experiences BIBREF15 , BIBREF20 . In this work, we structure the multi-community landscape along these two key dimensions of community identity: how distinctive a community's interests are, and how dynamic the community is over time.
We now proceed to outline our quantitative typology, which maps communities along these two dimensions. We start by providing an intuition through inspecting a few example communities. We then introduce a generalizable language-based methodology and use it to map a large set of Reddit communities onto the landscape defined by our typology of community identity.
## Overview and intuition
In order to illustrate the diversity within the multi-community space, and to provide an intuition for the underlying structure captured by the proposed typology, we first examine a few example communities and draw attention to some key social dynamics that occur within them.
We consider four communities from Reddit: in Seahawks, fans of the Seahawks football team gather to discuss games and players; in BabyBumps, expecting mothers trade advice and updates on their pregnancy; Cooking consists of recipe ideas and general discussion about cooking; while in pics, users share various images of random things (like eels and hornets). We note that these communities are topically contrasting and foster fairly disjoint user bases. Additionally, these communities exhibit varied patterns of user engagement. While Seahawks maintains a devoted set of users from month to month, pics is dominated by transient users who post a few times and then depart.
Discussions within these communities also span varied sets of interests. Some of these interests are more specific to the community than others: risotto, for example, is seldom a discussion point beyond Cooking. Additionally, some interests consistently recur, while others are specific to a particular time: kitchens are a consistent focus point for cooking, but mint is only in season during spring. Coupling specificity and consistency we find interests such as easter, which isn't particularly specific to BabyBumps but gains prominence in that community around Easter (see Figure FIGREF3 .A for further examples).
These specific interests provide a window into the nature of the communities' interests as a whole, and by extension their community identities. Overall, discussions in Cooking focus on topics which are highly distinctive and consistently recur (like risotto). In contrast, discussions in Seahawks are highly dynamic, rapidly shifting over time as new games occur and players are traded in and out. In the remainder of this section we formally introduce a methodology for mapping communities in this space defined by their distinctiveness and dynamicity (examples in Figure FIGREF3 .B).
## Language-based formalization
Our approach follows the intuition that a distinctive community will use language that is particularly specific, or unique, to that community. Similarly, a dynamic community will use volatile language that rapidly changes across successive windows of time. To capture this intuition automatically, we start by defining word-level measures of specificity and volatility. We then extend these word-level primitives to characterize entire comments, and the community itself.
Our characterizations of words in a community are motivated by methodology from prior literature that compares the frequency of a word in a particular setting to its frequency in some background distribution, in order to identify instances of linguistic variation BIBREF21 , BIBREF19 . Our particular framework makes this comparison by way of pointwise mutual information (PMI).
In the following, we use INLINEFORM0 to denote one community within a set INLINEFORM1 of communities, and INLINEFORM2 to denote one time period within the entire history INLINEFORM3 of INLINEFORM4 . We account for temporal as well as inter-community variation by computing word-level measures for each time period of each community's history, INLINEFORM5 . Given a word INLINEFORM6 used within a particular community INLINEFORM7 at time INLINEFORM8 , we define two word-level measures:
Specificity. We quantify the specificity INLINEFORM0 of INLINEFORM1 to INLINEFORM2 by calculating the PMI of INLINEFORM3 and INLINEFORM4 , relative to INLINEFORM5 , INLINEFORM6
where INLINEFORM0 is INLINEFORM1 's frequency in INLINEFORM2 . INLINEFORM3 is specific to INLINEFORM4 if it occurs more frequently in INLINEFORM5 than in the entire set INLINEFORM6 , hence distinguishing this community from the rest. A word INLINEFORM7 whose occurrence is decoupled from INLINEFORM8 , and thus has INLINEFORM9 close to 0, is said to be generic.
We compute values of INLINEFORM0 for each time period INLINEFORM1 in INLINEFORM2 ; in the above description we drop the time-based subscripts for clarity.
Volatility. We quantify the volatility INLINEFORM0 of INLINEFORM1 to INLINEFORM2 as the PMI of INLINEFORM3 and INLINEFORM4 relative to INLINEFORM5 , the entire history of INLINEFORM6 : INLINEFORM7
A word INLINEFORM0 is volatile at time INLINEFORM1 in INLINEFORM2 if it occurs more frequently at INLINEFORM3 than in the entire history INLINEFORM4 , behaving as a fad within a small window of time. A word that occurs with similar frequency across time, and hence has INLINEFORM5 close to 0, is said to be stable.
Extending to utterances. Using our word-level primitives, we define the specificity of an utterance INLINEFORM0 in INLINEFORM1 , INLINEFORM2 as the average specificity of each word in the utterance. The volatility of utterances is defined analogously.
## Community-level measures
Having described these word-level measures, we now proceed to establish the primary axes of our typology:
Distinctiveness. A community with a very distinctive identity will tend to have distinctive interests, expressed through specialized language. Formally, we define the distinctiveness of a community INLINEFORM0 as the average specificity of all utterances in INLINEFORM1 . We refer to a community with a less distinctive identity as being generic.
Dynamicity. A highly dynamic community constantly shifts interests from one time window to another, and these temporal variations are reflected in its use of volatile language. Formally, we define the dynamicity of a community INLINEFORM0 as the average volatility of all utterances in INLINEFORM1 . We refer to a community whose language is relatively consistent throughout time as being stable.
In our subsequent analyses, we focus mostly on examing the average distinctiveness and dynamicity of a community over time, denoted INLINEFORM0 and INLINEFORM1 .
## Applying the typology to Reddit
We now explain how our typology can be applied to the particular setting of Reddit, and describe the overall behaviour of our linguistic axes in this context.
Dataset description. Reddit is a popular website where users form and participate in discussion-based communities called subreddits. Within these communities, users post content—such as images, URLs, or questions—which often spark vibrant lengthy discussions in thread-based comment sections.
The website contains many highly active subreddits with thousands of active subscribers. These communities span an extremely rich variety of topical interests, as represented by the examples described earlier. They also vary along a rich multitude of structural dimensions, such as the number of users, the amount of conversation and social interaction, and the social norms determining which types of content become popular. The diversity and scope of Reddit's multicommunity ecosystem make it an ideal landscape in which to closely examine the relation between varying community identities and social dynamics.
Our full dataset consists of all subreddits on Reddit from January 2013 to December 2014, for which there are at least 500 words in the vocabulary used to estimate our measures, in at least 4 months of the subreddit's history. We compute our measures over the comments written by users in a community in time windows of months, for each sufficiently active month, and manually remove communities where the bulk of the contributions are in a foreign language. This results in 283 communities ( INLINEFORM0 ), for a total of 4,872 community-months ( INLINEFORM1 ).
Estimating linguistic measures. We estimate word frequencies INLINEFORM0 , and by extension each downstream measure, in a carefully controlled manner in order to ensure we capture robust and meaningful linguistic behaviour. First, we only consider top-level comments which are initial responses to a post, as the content of lower-level responses might reflect conventions of dialogue more than a community's high-level interests. Next, in order to prevent a few highly active users from dominating our frequency estimates, we count each unique word once per user, ignoring successive uses of the same word by the same user. This ensures that our word-level characterizations are not skewed by a small subset of highly active contributors.
In our subsequent analyses, we will only look at these measures computed over the nouns used in comments. In principle, our framework can be applied to any choice of vocabulary. However, in the case of Reddit using nouns provides a convenient degree of interpretability. We can easily understand the implication of a community preferentially mentioning a noun such as gamer or feminist, but interpreting the overuse of verbs or function words such as take or of is less straightforward. Additionally, in focusing on nouns we adopt the view emphasized in modern “third wave” accounts of sociolinguistic variation, that stylistic variation is inseparable from topical content BIBREF23 . In the case of online communities, the choice of what people choose to talk about serves as a primary signal of social identity. That said, a typology based on more purely stylistic differences is an interesting avenue for future work.
Accounting for rare words. One complication when using measures such as PMI, which are based off of ratios of frequencies, is that estimates for very infrequent words could be overemphasized BIBREF24 . Words that only appear a few times in a community tend to score at the extreme ends of our measures (e.g. as highly specific or highly generic), obfuscating the impact of more frequent words in the community. To address this issue, we discard the long tail of infrequent words in our analyses, using only the top 5th percentile of words, by frequency within each INLINEFORM0 , to score comments and communities.
Typology output on Reddit. The distribution of INLINEFORM0 and INLINEFORM1 across Reddit communities is shown in Figure FIGREF3 .B, along with examples of communities at the extremes of our typology. We find that interpretable groupings of communities emerge at various points within our axes. For instance, highly distinctive and dynamic communities tend to focus on rapidly-updating interests like sports teams and games, while generic and consistent communities tend to be large “link-sharing” hubs where users generally post content with no clear dominating themes. More examples of communities at the extremes of our typology are shown in Table TABREF9 .
We note that these groupings capture abstract properties of a community's content that go beyond its topic. For instance, our typology relates topically contrasting communities such as yugioh (which is about a popular trading card game) and Seahawks through the shared trait that their content is particularly distinctive. Additionally, the axes can clarify differences between topically similar communities: while startrek and thewalkingdead both focus on TV shows, startrek is less dynamic than the median community, while thewalkingdead is among the most dynamic communities, as the show was still airing during the years considered.
## Community identity and user retention
We have seen that our typology produces qualitatively satisfying groupings of communities according to the nature of their collective identity. This section shows that there is an informative and highly predictive relationship between a community's position in this typology and its user engagement patterns. We find that communities with distinctive and dynamic identities have higher rates of user engagement, and further show that a community's position in our identity-based landscape holds important predictive information that is complementary to a strong activity baseline.
In particular user retention is one of the most crucial aspects of engagement and is critical to community maintenance BIBREF2 . We quantify how successful communities are at retaining users in terms of both short and long-term commitment. Our results indicate that rates of user retention vary drastically, yet systematically according to how distinctive and dynamic a community is (Figure FIGREF3 ).
We find a strong, explanatory relationship between the temporal consistency of a community's identity and rates of user engagement: dynamic communities that continually update and renew their discussion content tend to have far higher rates of user engagement. The relationship between distinctiveness and engagement is less universal, but still highly informative: niche communities tend to engender strong, focused interest from users at one particular point in time, though this does not necessarily translate into long-term retention.
## Community-type and monthly retention
We find that dynamic communities, such as Seahawks or starcraft, have substantially higher rates of monthly user retention than more stable communities (Spearman's INLINEFORM0 = 0.70, INLINEFORM1 0.001, computed with community points averaged over months; Figure FIGREF11 .A, left). Similarly, more distinctive communities, like Cooking and Naruto, exhibit moderately higher monthly retention rates than more generic communities (Spearman's INLINEFORM2 = 0.33, INLINEFORM3 0.001; Figure FIGREF11 .A, right).
Monthly retention is formally defined as the proportion of users who contribute in month INLINEFORM0 and then return to contribute again in month INLINEFORM1 . Each monthly datapoint is treated as unique and the trends in Figure FIGREF11 show 95% bootstrapped confidence intervals, cluster-resampled at the level of subreddit BIBREF25 , to account for differences in the number of months each subreddit contributes to the data.
Importantly, we find that in the task of predicting community-level user retention our identity-based typology holds additional predictive value on top of strong baseline features based on community-size (# contributing users) and activity levels (mean # contributions per user), which are commonly used for churn prediction BIBREF7 . We compared out-of-sample predictive performance via leave-one-community-out cross validation using random forest regressors with ensembles of size 100, and otherwise default hyperparameters BIBREF26 . A model predicting average monthly retention based on a community's average distinctiveness and dynamicity achieves an average mean squared error ( INLINEFORM0 ) of INLINEFORM1 and INLINEFORM2 , while an analogous model predicting based on a community's size and average activity level (both log-transformed) achieves INLINEFORM4 and INLINEFORM5 . The difference between the two models is not statistically significant ( INLINEFORM6 , Wilcoxon signed-rank test). However, combining features from both models results in a large and statistically significant improvement over each independent model ( INLINEFORM7 , INLINEFORM8 , INLINEFORM9 Bonferroni-corrected pairwise Wilcoxon tests). These results indicate that our typology can explain variance in community-level retention rates, and provides information beyond what is present in standard activity-based features.
## Community-type and user tenure
As with monthly retention, we find a strong positive relationship between a community's dynamicity and the average number of months that a user will stay in that community (Spearman's INLINEFORM0 = 0.41, INLINEFORM1 0.001, computed over all community points; Figure FIGREF11 .B, left). This verifies that the short-term trend observed for monthly retention translates into longer-term engagement and suggests that long-term user retention might be strongly driven by the extent to which a community continually provides novel content. Interestingly, there is no significant relationship between distinctiveness and long-term engagement (Spearman's INLINEFORM2 = 0.03, INLINEFORM3 0.77; Figure FIGREF11 .B, right). Thus, while highly distinctive communities like RandomActsOfMakeup may generate focused commitment from users over a short period of time, such communities are unlikely to retain long-term users unless they also have sufficiently dynamic content.
To measure user tenures we focused on one slice of data (May, 2013) and measured how many months a user spends in each community, on average—the average number of months between a user's first and last comment in each community. We have activity data up until May 2015, so the maximum tenure is 24 months in this set-up, which is exceptionally long relative to the average community member (throughout our entire data less than INLINEFORM0 of users have tenures of more than 24 months in any community).
## Community identity and acculturation
The previous section shows that there is a strong connection between the nature of a community's identity and its basic user engagement patterns. In this section, we probe the relationship between a community's identity and how permeable, or accessible, it is to outsiders.
We measure this phenomenon using what we call the acculturation gap, which compares the extent to which engaged vs. non-engaged users employ community-specific language. While previous work has found this gap to be large and predictive of future user engagement in two beer-review communities BIBREF5 , we find that the size of the acculturation gap depends strongly on the nature of a community's identity, with the gap being most pronounced in stable, highly distinctive communities (Figure FIGREF13 ).
This finding has important implications for our understanding of online communities. Though many works have analyzed the dynamics of “linguistic belonging” in online communities BIBREF16 , BIBREF28 , BIBREF5 , BIBREF17 , our results suggest that the process of linguistically fitting in is highly contingent on the nature of a community's identity. At one extreme, in generic communities like pics or worldnews there is no distinctive, linguistic identity for users to adopt.
To measure the acculturation gap for a community, we follow Danescu-Niculescu-Mizil et al danescu-niculescu-mizilno2013 and build “snapshot language models” (SLMs) for each community, which capture the linguistic state of a community at one point of time. Using these language models we can capture how linguistically close a particular utterance is to the community by measuring the cross-entropy of this utterance relative to the SLM: DISPLAYFORM0
where INLINEFORM0 is the probability assigned to bigram INLINEFORM1 from comment INLINEFORM2 in community-month INLINEFORM3 . We build the SLMs by randomly sampling 200 active users—defined as users with at least 5 comments in the respective community and month. For each of these 200 active users we select 5 random 10-word spans from 5 unique comments. To ensure robustness and maximize data efficiency, we construct 100 SLMs for each community-month pair that has enough data, bootstrap-resampling from the set of active users.
We compute a basic measure of the acculturation gap for a community-month INLINEFORM0 as the relative difference of the cross-entropy of comments by users active in INLINEFORM1 with that of singleton comments by outsiders—i.e., users who only ever commented once in INLINEFORM2 , but who are still active in Reddit in general: DISPLAYFORM0
INLINEFORM0 denotes the distribution over singleton comments, INLINEFORM1 denotes the distribution over comments from users active in INLINEFORM2 , and INLINEFORM3 the expected values of the cross-entropy over these respective distributions. For each bootstrap-sampled SLM we compute the cross-entropy of 50 comments by active users (10 comments from 5 randomly sampled active users, who were not used to construct the SLM) and 50 comments from randomly-sampled outsiders.
Figure FIGREF13 .A shows that the acculturation gap varies substantially with how distinctive and dynamic a community is. Highly distinctive communities have far higher acculturation gaps, while dynamicity exhibits a non-linear relationship: relatively stable communities have a higher linguistic `entry barrier', as do very dynamic ones. Thus, in communities like IAmA (a general Q&A forum) that are very generic, with content that is highly, but not extremely dynamic, outsiders are at no disadvantage in matching the community's language. In contrast, the acculturation gap is large in stable, distinctive communities like Cooking that have consistent community-specific language. The gap is also large in extremely dynamic communities like Seahawks, which perhaps require more attention or interest on the part of active users to keep up-to-date with recent trends in content.
These results show that phenomena like the acculturation gap, which were previously observed in individual communities BIBREF28 , BIBREF5 , cannot be easily generalized to a larger, heterogeneous set of communities. At the same time, we see that structuring the space of possible communities enables us to observe systematic patterns in how such phenomena vary.
## Community identity and content affinity
Through the acculturation gap, we have shown that communities exhibit large yet systematic variations in their permeability to outsiders. We now turn to understanding the divide in commenting behaviour between outsiders and active community members at a finer granularity, by focusing on two particular ways in which such gaps might manifest among users: through different levels of engagement with specific content and with temporally volatile content.
Echoing previous results, we find that community type mediates the extent and nature of the divide in content affinity. While in distinctive communities active members have a higher affinity for both community-specific content and for highly volatile content, the opposite is true for generic communities, where it is the outsiders who engage more with volatile content.
We quantify these divides in content affinity by measuring differences in the language of the comments written by active users and outsiders. Concretely, for each community INLINEFORM0 , we define the specificity gap INLINEFORM1 as the relative difference between the average specificity of comments authored by active members, and by outsiders, where these measures are macroaveraged over users. Large, positive INLINEFORM2 then occur in communities where active users tend to engage with substantially more community-specific content than outsiders.
We analogously define the volatility gap INLINEFORM0 as the relative difference in volatilities of active member and outsider comments. Large, positive values of INLINEFORM1 characterize communities where active users tend to have more volatile interests than outsiders, while negative values indicate communities where active users tend to have more stable interests.
We find that in 94% of communities, INLINEFORM0 , indicating (somewhat unsurprisingly) that in almost all communities, active users tend to engage with more community-specific content than outsiders. However, the magnitude of this divide can vary greatly: for instance, in Homebrewing, which is dedicated to brewing beer, the divide is very pronounced ( INLINEFORM1 0.33) compared to funny, a large hub where users share humorous content ( INLINEFORM2 0.011).
The nature of the volatility gap is comparatively more varied. In Homebrewing ( INLINEFORM0 0.16), as in 68% of communities, active users tend to write more volatile comments than outsiders ( INLINEFORM1 0). However, communities like funny ( INLINEFORM2 -0.16), where active users contribute relatively stable comments compared to outsiders ( INLINEFORM3 0), are also well-represented on Reddit.
To understand whether these variations manifest systematically across communities, we examine the relationship between divides in content affinity and community type. In particular, following the intuition that active users have a relatively high affinity for a community's niche, we expect that the distinctiveness of a community will be a salient mediator of specificity and volatility gaps. Indeed, we find a strong correlation between a community's distinctiveness and its specificity gap (Spearman's INLINEFORM0 0.34, INLINEFORM1 0.001).
We also find a strong correlation between distinctiveness and community volatility gaps (Spearman's INLINEFORM0 0.53, INLINEFORM1 0.001). In particular, we see that among the most distinctive communities (i.e., the top third of communities by distinctiveness), active users tend to write more volatile comments than outsiders (mean INLINEFORM2 0.098), while across the most generic communities (i.e., the bottom third), active users tend to write more stable comments (mean INLINEFORM3 -0.047, Mann-Whitney U test INLINEFORM4 0.001). The relative affinity of outsiders for volatile content in these communities indicates that temporally ephemeral content might serve as an entry point into such a community, without necessarily engaging users in the long term.
## Further related work
Our language-based typology and analysis of user engagement draws on and contributes to several distinct research threads, in addition to the many foundational studies cited in the previous sections.
Multicommunity studies. Our investigation of user engagement in multicommunity settings follows prior literature which has examined differences in user and community dynamics across various online groups, such as email listservs. Such studies have primarily related variations in user behaviour to structural features such as group size and volume of content BIBREF30 , BIBREF31 , BIBREF32 , BIBREF33 . In focusing on the linguistic content of communities, we extend this research by providing a content-based framework through which user engagement can be examined.
Reddit has been a particularly useful setting for studying multiple communities in prior work. Such studies have mostly focused on characterizing how individual users engage across a multi-community platform BIBREF34 , BIBREF35 , or on specific user engagement patterns such as loyalty to particular communities BIBREF22 . We complement these studies by seeking to understand how features of communities can mediate a broad array of user engagement patterns within them.
Typologies of online communities. Prior attempts to typologize online communities have primarily been qualitative and based on hand-designed categories, making them difficult to apply at scale. These typologies often hinge on having some well-defined function the community serves, such as supporting a business or non-profit cause BIBREF36 , which can be difficult or impossible to identify in massive, anonymous multi-community settings. Other typologies emphasize differences in communication platforms and other functional requirements BIBREF37 , BIBREF38 , which are important but preclude analyzing differences between communities within the same multi-community platform. Similarly, previous computational methods of characterizing multiple communities have relied on the presence of markers such as affixes in community names BIBREF35 , or platform-specific affordances such as evaluation mechanisms BIBREF39 .
Our typology is also distinguished from community detection techniques that rely on structural or functional categorizations BIBREF40 , BIBREF41 . While the focus of those studies is to identify and characterize sub-communities within a larger social network, our typology provides a characterization of pre-defined communities based on the nature of their identity.
Broader work on collective identity. Our focus on community identity dovetails with a long line of research on collective identity and user engagement, in both online and offline communities BIBREF42 , BIBREF1 , BIBREF2 . These studies focus on individual-level psychological manifestations of collective (or social) identity, and their relationship to user engagement BIBREF42 , BIBREF43 , BIBREF44 , BIBREF0 .
In contrast, we seek to characterize community identities at an aggregate level and in an interpretable manner, with the goal of systematically organizing the diverse space of online communities. Typologies of this kind are critical to these broader, social-psychological studies of collective identity: they allow researchers to systematically analyze how the psychological manifestations and implications of collective identity vary across diverse sets of communities.
## Conclusion and future work
Our current understanding of engagement patterns in online communities is patched up from glimpses offered by several disparate studies focusing on a few individual communities. This work calls into attention the need for a method to systematically reason about similarities and differences across communities. By proposing a way to structure the multi-community space, we find not only that radically contrasting engagement patterns emerge in different parts of this space, but also that this variation can be at least partly explained by the type of identity each community fosters.
Our choice in this work is to structure the multi-community space according to a typology based on community identity, as reflected in language use. We show that this effectively explains cross-community variation of three different user engagement measures—retention, acculturation and content affinity—and complements measures based on activity and size with additional interpretable information. For example, we find that in niche communities established members are more likely to engage with volatile content than outsiders, while the opposite is true in generic communities. Such insights can be useful for community maintainers seeking to understand engagement patterns in their own communities.
One main area of future research is to examine the temporal dynamics in the multi-community landscape. By averaging our measures of distinctiveness and dynamicity across time, our present study treated community identity as a static property. However, as communities experience internal changes and respond to external events, we can expect the nature of their identity to shift as well. For instance, the relative consistency of harrypotter may be disrupted by the release of a new novel, while Seahawks may foster different identities during and between football seasons. Conversely, a community's type may also mediate the impact of new events. Moving beyond a static view of community identity could enable us to better understand how temporal phenomena such as linguistic change manifest across different communities, and also provide a more nuanced view of user engagement—for instance, are communities more welcoming to newcomers at certain points in their lifecycle?
Another important avenue of future work is to explore other ways of mapping the landscape of online communities. For example, combining structural properties of communities BIBREF40 with topical information BIBREF35 and with our identity-based measures could further characterize and explain variations in user engagement patterns. Furthermore, extending the present analyses to even more diverse communities supported by different platforms (e.g., GitHub, StackExchange, Wikipedia) could enable the characterization of more complex behavioral patterns such as collaboration and altruism, which become salient in different multicommunity landscapes.
## Acknowledgements
The authors thank Liye Fu, Jack Hessel, David Jurgens and Lillian Lee for their helpful comments. This research has been supported in part by a Discovery and Innovation Research Seed Award from the Office of the Vice Provost for Research at Cornell, NSF CNS-1010921, IIS-1149837, IIS-1514268 NIH BD2K, ARO MURI, DARPA XDATA, DARPA SIMPLEX, DARPA NGS2, Stanford Data Science Initiative, SAP Stanford Graduate Fellowship, NSERC PGS-D, Boeing, Lightspeed, and Volkswagen.
| 14
|
1706.06894
|
Stance Detection in Turkish Tweets
|
# Stance Detection in Turkish Tweets
## Abstract
Stance detection is a classification problem in natural language processing where for a text and target pair, a class result from the set {Favor, Against, Neither} is expected. It is similar to the sentiment analysis problem but instead of the sentiment of the text author, the stance expressed for a particular target is investigated in stance detection. In this paper, we present a stance detection tweet data set for Turkish comprising stance annotations of these tweets for two popular sports clubs as targets. Additionally, we provide the evaluation results of SVM classifiers for each target on this data set, where the classifiers use unigram, bigram, and hashtag features. This study is significant as it presents one of the initial stance detection data sets proposed so far and the first one for Turkish language, to the best of our knowledge. The data set and the evaluation results of the corresponding SVM-based approaches will form plausible baselines for the comparison of future studies on stance detection.
## Introduction
Stance detection (also called stance identification or stance classification) is one of the considerably recent research topics in natural language processing (NLP). It is usually defined as a classification problem where for a text and target pair, the stance of the author of the text for that target is expected as a classification output from the set: {Favor, Against, Neither} BIBREF0 .
Stance detection is usually considered as a subtask of sentiment analysis (opinion mining) BIBREF1 topic in NLP. Both are mostly performed on social media texts, particularly on tweets, hence both are important components of social media analysis. Nevertheless, in sentiment analysis, the sentiment of the author of a piece of text usually as Positive, Negative, and Neutral is explored while in stance detection, the stance of the author of the text for a particular target (an entity, event, etc.) either explicitly or implicitly referred to in the text is considered. Like sentiment analysis, stance detection systems can be valuable components of information retrieval and other text analysis systems BIBREF0 .
Previous work on stance detection include BIBREF2 where a stance classifier based on sentiment and arguing features is proposed in addition to an arguing lexicon automatically compiled. The ultimate approach performs better than distribution-based and uni-gram-based baseline systems BIBREF2 . In BIBREF3 , the authors show that the use of dialogue structure improves stance detection in on-line debates. In BIBREF4 , Hasan and Ng carry out stance detection experiments using different machine learning algorithms, training data sets, features, and inter-post constraints in on-line debates, and draw insightful conclusions based on these experiments. For instance, they find that sequence models like HMMs perform better at stance detection when compared with non-sequence models like Naive Bayes (NB) BIBREF4 . In another related study BIBREF5 , the authors conclude that topic-independent features can be exploited for disagreement detection in on-line dialogues. The employed features include agreement, cue words, denial, hedges, duration, polarity, and punctuation BIBREF5 . Stance detection on a corpus of student essays is considered in BIBREF6 . After using linguistically-motivated feature sets together with multivalued NB and SVM as the learning models, the authors conclude that they outperform two baseline approaches BIBREF6 . In BIBREF7 , the author claims that Wikipedia can be used to determine stances about controversial topics based on their previous work regarding controversy extraction on the Web.
Among more recent related work, in BIBREF8 stance detection for unseen targets is studied and bidirectional conditional encoding is employed. The authors state that their approach achieves state-of-the art performance rates BIBREF8 on SemEval 2016 Twitter Stance Detection corpus BIBREF0 . In BIBREF9 , a stance-community detection approach called SCIFNET is proposed. SCIFNET creates networks of people who are stance targets, automatically from the related document collections BIBREF9 using stance expansion and refinement techniques to arrive at stance-coherent networks. A tweet data set annotated with stance information regarding six predefined targets is proposed in BIBREF10 where this data set is annotated through crowdsourcing. The authors indicate that the data set is also annotated with sentiment information in addition to stance, so it can help reveal associations between stance and sentiment BIBREF10 . Lastly, in BIBREF0 , SemEval 2016's aforementioned shared task on Twitter Stance Detection is described. Also provided are the results of the evaluations of 19 systems participating in two subtasks (one with training data set provided and the other without an annotated data set) of the shared task BIBREF0 .
In this paper, we present a tweet data set in Turkish annotated with stance information, where the corresponding annotations are made publicly available. The domain of the tweets comprises two popular football clubs which constitute the targets of the tweets included. We also provide the evaluation results of SVM classifiers (for each target) on this data set using unigram, bigram, and hashtag features.
To the best of our knowledge, the current study is the first one to target at stance detection in Turkish tweets. Together with the provided annotated data set and the corresponding evaluations with the aforementioned SVM classifiers which can be used as baseline systems, our study will hopefully help increase social media analysis studies on Turkish content.
The rest of the paper is organized as follows: In Section SECREF2 , we describe our tweet data set annotated with the target and stance information. Section SECREF3 includes the details of our SVM-based stance classifiers and their evaluation results with discussions. Section SECREF4 includes future research topics based on the current study, and finally Section SECREF5 concludes the paper with a summary.
## A Stance Detection Data Set
We have decided to consider tweets about popular sports clubs as our domain for stance detection. Considerable amounts of tweets are being published for sports-related events at every instant. Hence we have determined our targets as Galatasaray (namely Target-1) and Fenerbahçe (namely, Target-2) which are two of the most popular football clubs in Turkey. As is the case for the sentiment analysis tools, the outputs of the stance detection systems on a stream of tweets about these clubs can facilitate the use of the opinions of the football followers by these clubs.
In a previous study on the identification of public health-related tweets, two tweet data sets in Turkish (each set containing 1 million random tweets) have been compiled where these sets belong to two different periods of 20 consecutive days BIBREF11 . We have decided to use one of these sets (corresponding to the period between August 18 and September 6, 2015) and firstly filtered the tweets using the possible names used to refer to the target clubs. Then, we have annotated the stance information in the tweets for these targets as Favor or Against. Within the course of this study, we have not considered those tweets in which the target is not explicitly mentioned, as our initial filtering process reveals.
For the purposes of the current study, we have not annotated any tweets with the Neither class. This stance class and even finer-grained classes can be considered in further annotation studies. We should also note that in a few tweets, the target of the stance was the management of the club while in some others a particular footballer of the club is praised or criticised. Still, we have considered the club as the target of the stance in all of the cases and carried out our annotations accordingly.
At the end of the annotation process, we have annotated 700 tweets, where 175 tweets are in favor of and 175 tweets are against Target-1, and similarly 175 tweets are in favor of and 175 are against Target-2. Hence, our data set is a balanced one although it is currently limited in size. The corresponding stance annotations are made publicly available at http://ceng.metu.edu.tr/ INLINEFORM0 e120329/ Turkish_Stance_Detection_Tweet_Dataset.csv in Comma Separated Values (CSV) format. The file contains three columns with the corresponding headers. The first column is the tweet id of the corresponding tweet, the second column contains the name of the stance target, and the last column includes the stance of the tweet for the target as Favor or Against.
To the best of our knowledge, this is the first publicly-available stance-annotated data set for Turkish. Hence, it is a significant resource as there is a scarcity of annotated data sets, linguistic resources, and NLP tools available for Turkish. Additionally, to the best of our knowledge, it is also significant for being the first stance-annotated data set including sports-related tweets, as previous stance detection data sets mostly include on-line texts on political/ethical issues.
## Stance Detection Experiments Using SVM Classifiers
It is emphasized in the related literature that unigram-based methods are reliable for the stance detection task BIBREF2 and similarly unigram-based models have been used as baseline models in studies such as BIBREF0 . In order to be used as a baseline and reference system for further studies on stance detection in Turkish tweets, we have trained two SVM classifiers (one for each target) using unigrams as features. Before the extraction of unigrams, we have employed automated preprocessing to filter out the stopwords in our annotated data set of 700 tweets. The stopword list used is the list presented in BIBREF12 which, in turn, is the slightly extended version of the stopword list provided in BIBREF13 .
We have used the SVM implementation available in the Weka data mining application BIBREF14 where this particular implementation employs the SMO algorithm BIBREF15 to train a classifier with a linear kernel. The 10-fold cross-validation results of the two classifiers are provided in Table TABREF1 using the metrics of precision, recall, and F-Measure.
The evaluation results are quite favorable for both targets and particularly higher for Target-1, considering the fact that they are the initial experiments on the data set. The performance of the classifiers is better for the Favor class for both targets when compared with the performance results for the Against class. This outcome may be due to the common use of some terms when expressing positive stance towards sports clubs in Turkish tweets. The same percentage of common terms may not have been observed in tweets during the expression of negative stances towards the targets. Yet, completely the opposite pattern is observed in stance detection results of baseline systems given in BIBREF0 , i.e., better F-Measure rates have been obtained for the Against class when compared with the Favor class BIBREF0 . Some of the baseline systems reported in BIBREF0 are SVM-based systems using unigrams and ngrams as features similar to our study, but their data sets include all three stance classes of Favor, Against, and Neither, while our data set comprises only tweets classified as belonging to Favor or Against classes. Another difference is that the data sets in BIBREF0 have been divided into training and test sets, while in our study we provide 10-fold cross-validation results on the whole data set. On the other hand, we should also note that SVM-based sentiment analysis systems (such as those given in BIBREF16 ) have been reported to achieve better F-Measure rates for the Positive sentiment class when compared with the results obtained for the Negative class. Therefore, our evaluation results for each stance class seem to be in line with such sentiment analysis systems. Yet, further experiments on the extended versions of our data set should be conducted and the results should again be compared with the stance detection results given in the literature.
We have also evaluated SVM classifiers which use only bigrams as features, as ngram-based classifiers have been reported to perform better for the stance detection problem BIBREF0 . However, we have observed that using bigrams as the sole features of the SVM classifiers leads to quite poor results. This observation may be due to the relatively limited size of the tweet data set employed. Still, we can conclude that unigram-based features lead to superior results compared to the results obtained using bigrams as features, based on our experiments on our data set. Yet, ngram-based features may be employed on the extended versions of the data set to verify this conclusion within the course of future work.
With an intention to exploit the contribution of hashtag use to stance detection, we have also used the existence of hashtags in tweets as an additional feature to unigrams. The corresponding evaluation results of the SVM classifiers using unigrams together the existence of hashtags as features are provided in Table TABREF2 .
When the results given in Table TABREF2 are compared with the results in Table TABREF1 , a slight decrease in F-Measure (0.5%) for Target-1 is observed, while the overall F-Measure value for Target-2 has increased by 1.8%. Although we could not derive sound conclusions mainly due to the relatively small size of our data set, the increase in the performance of the SVM classifier Target-2 is an encouraging evidence for the exploitation of hashtags in a stance detection system. We leave other ways of exploiting hashtags for stance detection as a future work.
To sum up, our evaluation results are significant as reference results to be used for comparison purposes and provides evidence for the utility of unigram-based and hashtag-related features in SVM classifiers for the stance detection problem in Turkish tweets.
## Future Prospects
Future work based on the current study includes the following:
## Conclusion
Stance detection is a considerably new research area in natural language processing and is considered within the scope of the well-studied topic of sentiment analysis. It is the detection of stance within text towards a target which may be explicitly specified in the text or not. In this study, we present a stance-annotated tweet data set in Turkish where the targets of the annotated stances are two popular sports clubs in Turkey. The corresponding annotations are made publicly-available for research purposes. To the best of our knowledge, this is the first stance detection data set for the Turkish language and also the first sports-related stance-annotated data set. Also presented in this study are SVM classifiers (one for each target) utilizing unigram and bigram features in addition to using the existence of hashtags as another feature. 10-fold cross validation results of these classifiers are presented which can be used as reference results by prospective systems. Both the annotated data set and the classifiers with evaluations are significant since they are the initial contributions to stance detection problem in Turkish tweets.
| 5
|
1706.07179
|
RelNet: End-to-End Modeling of Entities & Relations
|
# RelNet: End-to-End Modeling of Entities & Relations
## Abstract
We introduce RelNet: a new model for relational reasoning. RelNet is a memory augmented neural network which models entities as abstract memory slots and is equipped with an additional relational memory which models relations between all memory pairs. The model thus builds an abstract knowledge graph on the entities and relations present in a document which can then be used to answer questions about the document. It is trained end-to-end: only supervision to the model is in the form of correct answers to the questions. We test the model on the 20 bAbI question-answering tasks with 10k examples per task and find that it solves all the tasks with a mean error of 0.3%, achieving 0% error on 11 of the 20 tasks.
## Introduction
Reasoning about entities and their relations is an important problem for achieving general artificial intelligence. Often such problems are formulated as reasoning over graph-structured representation of knowledge. Knowledge graphs, for example, consist of entities and relations between them BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . Representation learning BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 and reasoning BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 with such structured representations is an important and active area of research.
Most previous work on knowledge representation and reasoning relies on a pipeline of natural language processing systems, often consisting of named entity extraction BIBREF12 , entity resolution and coreference BIBREF13 , relationship extraction BIBREF4 , and knowledge graph inference BIBREF14 . While this cascaded approach of using NLP systems can be effective at reasoning with knowledge bases at scale, it also leads to a problem of compounding of the error from each component sub-system. The importance of each of these sub-component on a particular downstream application is also not clear.
For the task of question-answering, we instead make an attempt at an end-to-end approach which directly models the entities and relations in the text as memory slots. While incorporating existing knowledge (from curated knowledge bases) for the purpose of question-answering BIBREF11 , BIBREF8 , BIBREF15 is an important area of research, we consider the simpler setting where all the information is contained within the text itself – which is the approach taken by many recent memory based neural network models BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 .
Recently, BIBREF17 proposed a dynamic memory based neural network for implicitly modeling the state of entities present in the text for question answering. However, this model lacks any module for relational reasoning. In response, we propose RelNet, which extends memory-augmented neural networks with a relational memory to reason about relationships between multiple entities present within the text. Our end-to-end method reads text, and writes to both memory slots and edges between them. Intuitively, the memory slots correspond to entities and the edges correspond to relationships between entities, each represented as a vector. The only supervision signal for our method comes from answering questions on the text.
We demonstrate the utility of the model through experiments on the bAbI tasks BIBREF18 and find that the model achieves smaller mean error across the tasks than the best previously published result BIBREF17 in the 10k examples regime and achieves 0% error on 11 of the 20 tasks.
## RelNet Model
We describe the RelNet model in this section. Figure 1 provides a high-level view of the model. The model is sequential in nature, consisting of the following steps: read text, process it into a dynamic relational memory and then attention conditioned on the question generates the answer. We model the dynamic memory in a fashion similar to Recurrent Entity Networks BIBREF17 and then equip it with an additional relational memory.
There are three main components to the model: 1) input encoder 2) dynamic memory, and 3) output module. We will describe these three modules in details. The input encoder and output module implementations are similar to the Entity Network BIBREF17 and main novelty lies in the dynamic memory. We describe the operations executed by the network for a single example consisting of a document with $T$ sentences, where each sentence consists of a sequence of words represented with $K$ -dimensional word embeddings $\lbrace e_1, \ldots , e_N\rbrace $ , a question on the document represented as another sequence of words and an answer to the question.
## Related Work
There is a long line of work in textual question-answering systems BIBREF21 , BIBREF22 . Recent successful approaches use memory based neural networks for question answering, for example BIBREF23 , BIBREF18 , BIBREF24 , BIBREF19 , BIBREF17 . Our model is also a memory network based model and is also related to the neural turing machine BIBREF25 . As described previously, the model is closely related to the Recurrent Entity Networks model BIBREF17 which describes an end-to-end approach to model entities in text but does not directly model relations. Other approaches to question answering use external knowledge, for instance external knowledge bases BIBREF26 , BIBREF11 , BIBREF27 , BIBREF28 , BIBREF9 or external text like Wikipedia BIBREF29 , BIBREF30 .
Very recently, and in parallel to this work, a method for relational reasoning called relation networks BIBREF31 was proposed. They demonstrated that simple neural network modules are not as effective at relational reasoning and their proposed module is similar to our model. However, relation network is not a memory-based model and there is no mechanism to read and write relevant information for each pair. Moreover, while their approach scales as the square of the number of sentences, our approach scales as the square of the number of memory slots used per QA pair. The output module in our model can be seen as a type of relation network.
Representation learning and reasoning over graph structured data is also relevant to this work. Graph based neural network models BIBREF32 , BIBREF33 , BIBREF34 have been proposed which take graph data as an input. The relational memory however does not rely on a specified graph structure and such models can potentially be used for multi-hop reasoning over the relational memory. BIBREF35 proposed a method for learning a graphical representation of the text data for question answering, however the model requires explicit supervision for the graph at every step whereas RelNet does not require explicit supervision for the graph.
## Experiments
We evaluate the model's performance on the bAbI tasks BIBREF18 , a collection of 20 question answering tasks which have become a benchmark for evaluating memory-augmented neural networks. We compare the performance with the Recurrent Entity Networks model (EntNet) BIBREF17 . Performance is measured in terms of mean percentage error on the tasks.
Training Details: We used Adam and did a grid search for the learning rate in {0.01, 0.005, 0.001} and choose a fixed learning rate of 0.005 based on performance on the validation set, and clip the gradient norm at 2. We keep all other details similar to BIBREF17 for a fair comparison. embedding dimensions were fixed to be 100, models were trained for a maximum of 250 epochs with mini-batches size of 32 for all tasks except 3 for which the batch size was 16. The document sizes were limited to most recent 70 sentences for all tasks, except for task 3 for which it was limited to 130. The RelNet models were run for 5 times with random seed on each task and the model with best validation performance was chosen as the final model. The baseline EntNet model was run for 10 times for each task BIBREF17 .
The results are shown in Table 1 . The RelNet model achieves a mean error of 0.285% across tasks which is better than the results of the EntNet model BIBREF17 . The RelNet model is able to achieve 0% test error on 11 of the tasks, whereas the EntNet model achieves 0% error on 7 of the tasks.
## Conclusion
We demonstrated an end-to-end trained neural network augmented with a structured memory representation which can reason about entities and relations for question answering. Future work will investigate the performance of these models on more real world datasets, interpreting what the models learn, and scaling these models to answer questions about entities and relations from reading massive text corpora.
| 5
|
1706.08032
|
A Deep Neural Architecture for Sentence-level Sentiment Classification in Twitter Social Networking
|
# A Deep Neural Architecture for Sentence-level Sentiment Classification in Twitter Social Networking
## Abstract
This paper introduces a novel deep learning framework including a lexicon-based approach for sentence-level prediction of sentiment label distribution. We propose to first apply semantic rules and then use a Deep Convolutional Neural Network (DeepCNN) for character-level embeddings in order to increase information for word-level embedding. After that, a Bidirectional Long Short-Term Memory Network (Bi-LSTM) produces a sentence-wide feature representation from the word-level embedding. We evaluate our approach on three Twitter sentiment classification datasets. Experimental results show that our model can improve the classification accuracy of sentence-level sentiment analysis in Twitter social networking.
## Introduction
Twitter sentiment classification have intensively researched in recent years BIBREF0 BIBREF1 . Different approaches were developed for Twitter sentiment classification by using machine learning such as Support Vector Machine (SVM) with rule-based features BIBREF2 and the combination of SVMs and Naive Bayes (NB) BIBREF3 . In addition, hybrid approaches combining lexicon-based and machine learning methods also achieved high performance described in BIBREF4 . However, a problem of traditional machine learning is how to define a feature extractor for a specific domain in order to extract important features.
Deep learning models are different from traditional machine learning methods in that a deep learning model does not depend on feature extractors because features are extracted during training progress. The use of deep learning methods becomes to achieve remarkable results for sentiment analysis BIBREF5 BIBREF6 BIBREF7 . Some researchers used Convolutional Neural Network (CNN) for sentiment classification. CNN models have been shown to be effective for NLP. For example, BIBREF6 proposed various kinds of CNN to learn sentiment-bearing sentence vectors, BIBREF5 adopted two CNNs in character-level to sentence-level representation for sentiment analysis. BIBREF7 constructs experiments on a character-level CNN for several large-scale datasets. In addition, Long Short-Term Memory (LSTM) is another state-of-the-art semantic composition model for sentiment classification with many variants described in BIBREF8 . The studies reveal that using a CNN is useful in extracting information and finding feature detectors from texts. In addition, a LSTM can be good in maintaining word order and the context of words. However, in some important aspects, the use of CNN or LSTM separately may not capture enough information.
Inspired by the models above, the goal of this research is using a Deep Convolutional Neural Network (DeepCNN) to exploit the information of characters of words in order to support word-level embedding. A Bi-LSTM produces a sentence-wide feature representation based on these embeddings. The Bi-LSTM is a version of BIBREF9 with Full Gradient described in BIBREF10 . In addition, the rules-based approach also effects classification accuracy by focusing on important sub-sentences expressing the main sentiment of a tweet while removing unnecessary parts of a tweet. The paper makes the following contributions:
The organization of the present paper is as follows: In section 2, we describe the model architecture which introduces the structure of the model. We explain the basic idea of model and the way of constructing the model. Section 3 show results and analysis and section 4 summarize this paper.
## Basic idea
Our proposed model consists of a deep learning classifier and a tweet processor. The deep learning classifier is a combination of DeepCNN and Bi-LSTM. The tweet processor standardizes tweets and then applies semantic rules on datasets. We construct a framework that treats the deep learning classifier and the tweet processor as two distinct components. We believe that standardizing data is an important step to achieve high accuracy. To formulate our problem in increasing the accuracy of the classifier, we illustrate our model in Figure. FIGREF4 as follows:
Tweets are firstly considered via a processor based on preprocessing steps BIBREF0 and the semantic rules-based method BIBREF11 in order to standardize tweets and capture only important information containing the main sentiment of a tweet.
We use DeepCNN with Wide convolution for character-level embeddings. A wide convolution can learn to recognize specific n-grams at every position in a word that allows features to be extracted independently of these positions in the word. These features maintain the order and relative positions of characters. A DeepCNN is constructed by two wide convolution layers and the need of multiple wide convolution layers is widely accepted that a model constructing by multiple processing layers have the ability to learn representations of data with higher levels of abstraction BIBREF12 . Therefore, we use DeepCNN for character-level embeddings to support morphological and shape information for a word. The DeepCNN produces INLINEFORM0 global fixed-sized feature vectors for INLINEFORM1 words.
A combination of the global fixed-size feature vectors and word-level embedding is fed into Bi-LSTM. The Bi-LSTM produces a sentence-level representation by maintaining the order of words.
Our work is philosophically similar to BIBREF5 . However, our model is distinguished with their approaches in two aspects:
Using DeepCNN with two wide convolution layers to increase representation with multiple levels of abstraction.
Integrating global character fixed-sized feature vectors with word-level embedding to extract a sentence-wide feature set via Bi-LSTM. This deals with three main problems: (i) Sentences have any different size; (ii) The semantic and the syntactic of words in a sentence are captured in order to increase information for a word; (iii) Important information of characters that can appear at any position in a word are extracted.
In sub-section B, we introduce various kinds of dataset. The modules of our model are constructed in other sub-sections.
## Data Preparation
Stanford - Twitter Sentiment Corpus (STS Corpus): STS Corpus contains 1,600K training tweets collected by a crawler from BIBREF0 . BIBREF0 constructed a test set manually with 177 negative and 182 positive tweets. The Stanford test set is small. However, it has been widely used in different evaluation tasks BIBREF0 BIBREF5 BIBREF13 .
Sanders - Twitter Sentiment Corpus: This dataset consists of hand-classified tweets collected by using search terms: INLINEFORM0 , #google, #microsoft and #twitter. We construct the dataset as BIBREF14 for binary classification.
Health Care Reform (HCR): This dataset was constructed by crawling tweets containing the hashtag #hcr BIBREF15 . Task is to predict positive/negative tweets BIBREF14 .
## Preprocessing
We firstly take unique properties of Twitter in order to reduce the feature space such as Username, Usage of links, None, URLs and Repeated Letters. We then process retweets, stop words, links, URLs, mentions, punctuation and accentuation. For emoticons, BIBREF0 revealed that the training process makes the use of emoticons as noisy labels and they stripped the emoticons out from their training dataset because BIBREF0 believed that if we consider the emoticons, there is a negative impact on the accuracies of classifiers. In addition, removing emoticons makes the classifiers learns from other features (e.g. unigrams and bi-grams) presented in tweets and the classifiers only use these non-emoticon features to predict the sentiment of tweets. However, there is a problem is that if the test set contains emoticons, they do not influence the classifiers because emoticon features do not contain in its training data. This is a limitation of BIBREF0 , because the emoticon features would be useful when classifying test data. Therefore, we keep emoticon features in the datasets because deep learning models can capture more information from emoticon features for increasing classification accuracy.
## Semantic Rules (SR)
In Twitter social networking, people express their opinions containing sub-sentences. These sub-sentences using specific PoS particles (Conjunction and Conjunctive adverbs), like "but, while, however, despite, however" have different polarities. However, the overall sentiment of tweets often focus on certain sub-sentences. For example:
@lonedog bwahahah...you are amazing! However, it was quite the letdown.
@kirstiealley my dentist is great but she's expensive...=(
In two tweets above, the overall sentiment is negative. However, the main sentiment is only in the sub-sentences following but and however. This inspires a processing step to remove unessential parts in a tweet. Rule-based approach can assists these problems in handling negation and dealing with specific PoS particles led to effectively affect the final output of classification BIBREF11 BIBREF16 . BIBREF11 summarized a full presentation of their semantic rules approach and devised ten semantic rules in their hybrid approach based on the presentation of BIBREF16 . We use five rules in the semantic rules set because other five rules are only used to compute polarity of words after POS tagging or Parsing steps. We follow the same naming convention for rules utilized by BIBREF11 to represent the rules utilized in our proposed method. The rules utilized in the proposed method are displayed in Table TABREF15 in which is included examples from STS Corpus and output after using the rules. Table TABREF16 illustrates the number of processed sentences on each dataset.
## Representation Levels
To construct embedding inputs for our model, we use a fixed-sized word vocabulary INLINEFORM0 and a fixed-sized character vocabulary INLINEFORM1 . Given a word INLINEFORM2 is composed from characters INLINEFORM3 , the character-level embeddings are encoded by column vectors INLINEFORM4 in the embedding matrix INLINEFORM5 , where INLINEFORM6 is the size of the character vocabulary. For word-level embedding INLINEFORM7 , we use a pre-trained word-level embedding with dimension 200 or 300. A pre-trained word-level embedding can capture the syntactic and semantic information of words BIBREF17 . We build every word INLINEFORM8 into an embedding INLINEFORM9 which is constructed by two sub-vectors: the word-level embedding INLINEFORM10 and the character fixed-size feature vector INLINEFORM11 of INLINEFORM12 where INLINEFORM13 is the length of the filter of wide convolutions. We have INLINEFORM14 character fixed-size feature vectors corresponding to word-level embedding in a sentence.
## Deep Learning Module
DeepCNN in the deep learning module is illustrated in Figure. FIGREF22 . The DeepCNN has two wide convolution layers. The first layer extract local features around each character windows of the given word and using a max pooling over character windows to produce a global fixed-sized feature vector for the word. The second layer retrieves important context characters and transforms the representation at previous level into a representation at higher abstract level. We have INLINEFORM0 global character fixed-sized feature vectors for INLINEFORM1 words.
In the next step of Figure. FIGREF4 , we construct the vector INLINEFORM0 by concatenating the word-level embedding with the global character fixed-size feature vectors. The input of Bi-LSTM is a sequence of embeddings INLINEFORM1 . The use of the global character fixed-size feature vectors increases the relationship of words in the word-level embedding. The purpose of this Bi-LSTM is to capture the context of words in a sentence and maintain the order of words toward to extract sentence-level representation. The top of the model is a softmax function to predict sentiment label. We describe in detail the kinds of CNN and LSTM that we use in next sub-part 1 and 2.
The one-dimensional convolution called time-delay neural net has a filter vector INLINEFORM0 and take the dot product of filter INLINEFORM1 with each m-grams in the sequence of characters INLINEFORM2 of a word in order to obtain a sequence INLINEFORM3 : DISPLAYFORM0
Based on Equation 1, we have two types of convolutions that depend on the range of the index INLINEFORM0 . The narrow type of convolution requires that INLINEFORM1 and produce a sequence INLINEFORM2 . The wide type of convolution does not require on INLINEFORM3 or INLINEFORM4 and produce a sequence INLINEFORM5 . Out-of-range input values INLINEFORM6 where INLINEFORM7 or INLINEFORM8 are taken to be zero. We use wide convolution for our model.
Given a word INLINEFORM0 composed of INLINEFORM1 characters INLINEFORM2 , we take a character embedding INLINEFORM3 for each character INLINEFORM4 and construct a character matrix INLINEFORM5 as following Equation. 2: DISPLAYFORM0
The values of the embeddings INLINEFORM0 are parameters that are optimized during training. The trained weights in the filter INLINEFORM1 correspond to a feature detector which learns to recognize a specific class of n-grams. The n-grams have size INLINEFORM2 . The use of a wide convolution has some advantages more than a narrow convolution because a wide convolution ensures that all weights of filter reach the whole characters of a word at the margins. The resulting matrix has dimension INLINEFORM3 .
Long Short-Term Memory networks usually called LSTMs are a improved version of RNN. The core idea behind LSTMs is the cell state which can maintain its state over time, and non-linear gating units which regulate the information flow into and out of the cell. The LSTM architecture that we used in our proposed model is described in BIBREF9 . A single LSTM memory cell is implemented by the following composite function: DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 is the logistic sigmoid function, INLINEFORM1 and INLINEFORM2 are the input gate, forget gate, output gate, cell and cell input activation vectors respectively. All of them have a same size as the hidden vector INLINEFORM3 . INLINEFORM4 is the hidden-input gate matrix, INLINEFORM5 is the input-output gate matrix. The bias terms which are added to INLINEFORM6 and INLINEFORM7 have been omitted for clarity. In addition, we also use the full gradient for calculating with full backpropagation through time (BPTT) described in BIBREF10 . A LSTM gradients using finite differences could be checked and making practical implementations more reliable.
## Regularization
For regularization, we use a constraint on INLINEFORM0 of the weight vectors BIBREF18 .
## Experimental setups
For the Stanford Twitter Sentiment Corpus, we use the number of samples as BIBREF5 . The training data is selected 80K tweets for a training data and 16K tweets for the development set randomly from the training data of BIBREF0 . We conduct a binary prediction for STS Corpus.
For Sander dataset, we use standard 10-fold cross validation as BIBREF14 . We construct the development set by selecting 10% randomly from 9-fold training data.
In Health Care Reform Corpus, we also select 10% randomly for the development set in a training set and construct as BIBREF14 for comparison. We describe the summary of datasets in Table III.
for all datasets, the filter window size ( INLINEFORM0 ) is 7 with 6 feature maps each for the first wide convolution layer, the second wide convolution layer has a filter window size of 5 with 14 feature maps each. Dropout rate ( INLINEFORM1 ) is 0.5, INLINEFORM2 constraint, learning rate is 0.1 and momentum of 0.9. Mini-batch size for STS Corpus is 100 and others are 4. In addition, training is done through stochastic gradient descent over shuffled mini-batches with Adadelta update rule BIBREF19 .
we use the publicly available Word2Vec trained from 100 billion words from Google and TwitterGlove of Stanford is performed on aggregated global word-word co-occurrence statistics from a corpus. Word2Vec has dimensionality of 300 and Twitter Glove have dimensionality of 200. Words that do not present in the set of pre-train words are initialized randomly.
## Experimental results
Table IV shows the result of our model for sentiment classification against other models. We compare our model performance with the approaches of BIBREF0 BIBREF5 on STS Corpus. BIBREF0 reported the results of Maximum Entropy (MaxEnt), NB, SVM on STS Corpus having good performance in previous time. The model of BIBREF5 is a state-of-the-art so far by using a CharSCNN. As can be seen, 86.63 is the best prediction accuracy of our model so far for the STS Corpus.
For Sanders and HCR datasets, we compare results with the model of BIBREF14 that used a ensemble of multiple base classifiers (ENS) such as NB, Random Forest (RF), SVM and Logistic Regression (LR). The ENS model is combined with bag-of-words (BoW), feature hashing (FH) and lexicons. The model of BIBREF14 is a state-of-the-art on Sanders and HCR datasets. Our models outperform the model of BIBREF14 for the Sanders dataset and HCR dataset.
## Analysis
As can be seen, the models with SR outperforms the model with no SR. Semantic rules is effective in order to increase classification accuracy. We evaluate the efficiency of SR for the model in Table V of our full paper . We also conduct two experiments on two separate models: DeepCNN and Bi-LSTM in order to show the effectiveness of combination of DeepCNN and Bi-LSTM. In addition, the model using TwitterGlove outperform the model using GoogleW2V because TwitterGlove captures more information in Twitter than GoogleW2V. These results show that the character-level information and SR have a great impact on Twitter Data. The pre-train word vectors are good, universal feature extractors. The difference between our model and other approaches is the ability of our model to capture important features by using SR and combine these features at high benefit. The use of DeepCNN can learn a representation of words in higher abstract level. The combination of global character fixed-sized feature vectors and a word embedding helps the model to find important detectors for particles such as 'not' that negate sentiment and potentiate sentiment such as 'too', 'so' standing beside expected features. The model not only learns to recognize single n-grams, but also patterns in n-grams lead to form a structure significance of a sentence.
## Conclusions
In the present work, we have pointed out that the use of character embeddings through a DeepCNN to enhance information for word embeddings built on top of Word2Vec or TwitterGlove improves classification accuracy in Tweet sentiment classification. Our results add to the well-establish evidence that character vectors are an important ingredient for word-level in deep learning for NLP. In addition, semantic rules contribute handling non-essential sub-tweets in order to improve classification accuracy.
| 12
|
1707.05236
|
Artificial Error Generation with Machine Translation and Syntactic Patterns
|
# Artificial Error Generation with Machine Translation and Syntactic Patterns
## Abstract
Shortage of available training data is holding back progress in the area of automated error detection. This paper investigates two alternative methods for artificially generating writing errors, in order to create additional resources. We propose treating error generation as a machine translation task, where grammatically correct text is translated to contain errors. In addition, we explore a system for extracting textual patterns from an annotated corpus, which can then be used to insert errors into grammatically correct sentences. Our experiments show that the inclusion of artificially generated errors significantly improves error detection accuracy on both FCE and CoNLL 2014 datasets.
## Introduction
Writing errors can occur in many different forms – from relatively simple punctuation and determiner errors, to mistakes including word tense and form, incorrect collocations and erroneous idioms. Automatically identifying all of these errors is a challenging task, especially as the amount of available annotated data is very limited. Rei2016 showed that while some error detection algorithms perform better than others, it is additional training data that has the biggest impact on improving performance.
Being able to generate realistic artificial data would allow for any grammatically correct text to be transformed into annotated examples containing writing errors, producing large amounts of additional training examples. Supervised error generation systems would also provide an efficient method for anonymising the source corpus – error statistics from a private corpus can be aggregated and applied to a different target text, obscuring sensitive information in the original examination scripts. However, the task of creating incorrect data is somewhat more difficult than might initially appear – naive methods for error generation can create data that does not resemble natural errors, thereby making downstream systems learn misleading or uninformative patterns.
Previous work on artificial error generation (AEG) has focused on specific error types, such as prepositions and determiners BIBREF0 , BIBREF1 , or noun number errors BIBREF2 . Felice2014a investigated the use of linguistic information when generating artificial data for error correction, but also restricting the approach to only five error types. There has been very limited research on generating artificial data for all types, which is important for general-purpose error detection systems. For example, the error types investigated by Felice2014a cover only 35.74% of all errors present in the CoNLL 2014 training dataset, providing no additional information for the majority of errors.
In this paper, we investigate two supervised approaches for generating all types of artificial errors. We propose a framework for generating errors based on statistical machine translation (SMT), training a model to translate from correct into incorrect sentences. In addition, we describe a method for learning error patterns from an annotated corpus and transplanting them into error-free text. We evaluate the effect of introducing artificial data on two error detection benchmarks. Our results show that each method provides significant improvements over using only the available training set, and a combination of both gives an absolute improvement of 4.3% in INLINEFORM0 , without requiring any additional annotated data.
## Error Generation Methods
We investigate two alternative methods for AEG. The models receive grammatically correct text as input and modify certain tokens to produce incorrect sequences. The alternative versions of each sentence are aligned using Levenshtein distance, allowing us to identify specific words that need to be marked as errors. While these alignments are not always perfect, we found them to be sufficient for practical purposes, since alternative alignments of similar sentences often result in the same binary labeling. Future work could explore more advanced alignment methods, such as proposed by felice-bryant-briscoe.
In Section SECREF4 , this automatically labeled data is then used for training error detection models.
## Machine Translation
We treat AEG as a translation task – given a correct sentence as input, the system would learn to translate it to contain likely errors, based on a training corpus of parallel data. Existing SMT approaches are already optimised for identifying context patterns that correspond to specific output sequences, which is also required for generating human-like errors. The reverse of this idea, translating from incorrect to correct sentences, has been shown to work well for error correction tasks BIBREF2 , BIBREF3 , and round-trip translation has also been shown to be promising for correcting grammatical errors BIBREF4 .
Following previous work BIBREF2 , BIBREF5 , we build a phrase-based SMT error generation system. During training, error-corrected sentences in the training data are treated as the source, and the original sentences written by language learners as the target. Pialign BIBREF6 is used to create a phrase translation table directly from model probabilities. In addition to default features, we add character-level Levenshtein distance to each mapping in the phrase table, as proposed by Felice:2014-CoNLL. Decoding is performed using Moses BIBREF7 and the language model used during decoding is built from the original erroneous sentences in the learner corpus. The IRSTLM Toolkit BIBREF8 is used for building a 5-gram language model with modified Kneser-Ney smoothing BIBREF9 .
## Pattern Extraction
We also describe a method for AEG using patterns over words and part-of-speech (POS) tags, extracting known incorrect sequences from a corpus of annotated corrections. This approach is based on the best method identified by Felice2014a, using error type distributions; while they covered only 5 error types, we relax this restriction and learn patterns for generating all types of errors.
The original and corrected sentences in the corpus are aligned and used to identify short transformation patterns in the form of (incorrect phrase, correct phrase). The length of each pattern is the affected phrase, plus up to one token of context on both sides. If a word form changes between the incorrect and correct text, it is fully saved in the pattern, otherwise the POS tags are used for matching.
For example, the original sentence `We went shop on Saturday' and the corrected version `We went shopping on Saturday' would produce the following pattern:
(VVD shop_VV0 II, VVD shopping_VVG II)
After collecting statistics from the background corpus, errors can be inserted into error-free text. The learned patterns are now reversed, looking for the correct side of the tuple in the input sentence. We only use patterns with frequency INLINEFORM0 , which yields a total of 35,625 patterns from our training data. For each input sentence, we first decide how many errors will be generated (using probabilities from the background corpus) and attempt to create them by sampling from the collection of applicable patterns. This process is repeated until all the required errors have been generated or the sentence is exhausted. During generation, we try to balance the distribution of error types as well as keeping the same proportion of incorrect and correct sentences as in the background corpus BIBREF10 . The required POS tags were generated with RASP BIBREF11 , using the CLAWS2 tagset.
## Error Detection Model
We construct a neural sequence labeling model for error detection, following the previous work BIBREF12 , BIBREF13 . The model receives a sequence of tokens as input and outputs a prediction for each position, indicating whether the token is correct or incorrect in the current context. The tokens are first mapped to a distributed vector space, resulting in a sequence of word embeddings. Next, the embeddings are given as input to a bidirectional LSTM BIBREF14 , in order to create context-dependent representations for every token. The hidden states from forward- and backward-LSTMs are concatenated for each word position, resulting in representations that are conditioned on the whole sequence. This concatenated vector is then passed through an additional feedforward layer, and a softmax over the two possible labels (correct and incorrect) is used to output a probability distribution for each token. The model is optimised by minimising categorical cross-entropy with respect to the correct labels. We use AdaDelta BIBREF15 for calculating an adaptive learning rate during training, which accounts for a higher baseline performance compared to previous results.
## Evaluation
We trained our error generation models on the public FCE training set BIBREF16 and used them to generate additional artificial training data. Grammatically correct text is needed as the starting point for inserting artificial errors, and we used two different sources: 1) the corrected version of the same FCE training set on which the system is trained (450K tokens), and 2) example sentences extracted from the English Vocabulary Profile (270K tokens).. While there are other text corpora that could be used (e.g., Wikipedia and news articles), our development experiments showed that keeping the writing style and vocabulary close to the target domain gives better results compared to simply including more data.
We evaluated our detection models on three benchmarks: the FCE test data (41K tokens) and the two alternative annotations of the CoNLL 2014 Shared Task dataset (30K tokens) BIBREF3 . Each artificial error generation system was used to generate 3 different versions of the artificial data, which were then combined with the original annotated dataset and used for training an error detection system. Table TABREF1 contains example sentences from the error generation systems, highlighting each of the edits that are marked as errors.
The error detection results can be seen in Table TABREF4 . We use INLINEFORM0 as the main evaluation measure, which was established as the preferred measure for error correction and detection by the CoNLL-14 shared task BIBREF3 . INLINEFORM1 calculates a weighted harmonic mean of precision and recall, which assigns twice as much importance to precision – this is motivated by practical applications, where accurate predictions from an error detection system are more important compared to coverage. For comparison, we also report the performance of the error detection system by Rei2016, trained using the same FCE dataset.
The results show that error detection performance is substantially improved by making use of artificially generated data, created by any of the described methods. When comparing the error generation system by Felice2014a (FY14) with our pattern-based (PAT) and machine translation (MT) approaches, we see that the latter methods covering all error types consistently improve performance. While the added error types tend to be less frequent and more complicated to capture, the added coverage is indeed beneficial for error detection. Combining the pattern-based approach with the machine translation system (Ann+PAT+MT) gave the best overall performance on all datasets. The two frameworks learn to generate different types of errors, and taking advantage of both leads to substantial improvements in error detection.
We used the Approximate Randomisation Test BIBREF17 , BIBREF18 to calculate statistical significance and found that the improvement for each of the systems using artificial data was significant over using only manual annotation. In addition, the final combination system is also significantly better compared to the Felice2014a system, on all three datasets. While Rei2016 also report separate experiments that achieve even higher performance, these models were trained on a considerably larger proprietary corpus. In this paper we compare error detection frameworks trained on the same publicly available FCE dataset, thereby removing the confounding factor of dataset size and only focusing on the model architectures.
The error generation methods can generate alternative versions of the same input text – the pattern-based method randomly samples the error locations, and the SMT system can provide an n-best list of alternative translations. Therefore, we also investigated the combination of multiple error-generated versions of the input files when training error detection models. Figure FIGREF6 shows the INLINEFORM0 score on the development set, as the training data is increased by using more translations from the n-best list of the SMT system. These results reveal that allowing the model to see multiple alternative versions of the same file gives a distinct improvement – showing the model both correct and incorrect variations of the same sentences likely assists in learning a discriminative model.
## Related Work
Our work builds on prior research into AEG. Brockett2006 constructed regular expressions for transforming correct sentences to contain noun number errors. Rozovskaya2010a learned confusion sets from an annotated corpus in order to generate preposition errors. Foster2009 devised a tool for generating errors for different types using patterns provided by the user or collected automatically from an annotated corpus. However, their method uses a limited number of edit operations and is thus unable to generate complex errors. Cahill2013 compared different training methodologies and showed that artificial errors helped correct prepositions. Felice2014a learned error type distributions for generating five types of errors, and the system in Section SECREF3 is an extension of this model. While previous work focused on generating a specific subset of error types, we explored two holistic approaches to AEG and showed that they are able to significantly improve error detection performance.
## Conclusion
This paper investigated two AEG methods, in order to create additional training data for error detection. First, we explored a method using textual patterns learned from an annotated corpus, which are used for inserting errors into correct input text. In addition, we proposed formulating error generation as an MT framework, learning to translate from grammatically correct to incorrect sentences.
The addition of artificial data to the training process was evaluated on three error detection annotations, using the FCE and CoNLL 2014 datasets. Making use of artificial data provided improvements for all data generation methods. By relaxing the type restrictions and generating all types of errors, our pattern-based method consistently outperformed the system by Felice2014a. The combination of the pattern-based method with the machine translation approach gave further substantial improvements and the best performance on all datasets.
| 8
|
1707.06806
|
Shallow reading with Deep Learning: Predicting popularity of online content using only its title
|
# Shallow reading with Deep Learning: Predicting popularity of online content using only its title
## Abstract
With the ever decreasing attention span of contemporary Internet users, the title of online content (such as a news article or video) can be a major factor in determining its popularity. To take advantage of this phenomenon, we propose a new method based on a bidirectional Long Short-Term Memory (LSTM) neural network designed to predict the popularity of online content using only its title. We evaluate the proposed architecture on two distinct datasets of news articles and news videos distributed in social media that contain over 40,000 samples in total. On those datasets, our approach improves the performance over traditional shallow approaches by a margin of 15%. Additionally, we show that using pre-trained word vectors in the embedding layer improves the results of LSTM models, especially when the training set is small. To our knowledge, this is the first attempt of applying popularity prediction using only textual information from the title.
## Introduction
The distribution of textual content is typically very fast and catches user attention for only a short period of time BIBREF0 . For this reason, proper wording of the article title may play a significant role in determining the future popularity of the article. The reflection of this phenomenon is the proliferation of click-baits - short snippets of text whose main purpose is to encourage viewers to click on the link embedded in the snippet. Although detection of click-baits is a separate research topic BIBREF1 , in this paper we address a more general problem of predicting popularity of online content based solely on its title.
Predicting popularity in the Internet is a challenging and non-trivial task due to a multitude of factors impacting the distribution of the information: external context, social network of the publishing party, relevance of the video to the final user, etc. This topic has therefore attracted a lot of attention from the research community BIBREF2 , BIBREF3 , BIBREF0 , BIBREF4 .
In this paper we propose a method for online content popularity prediction based on a bidirectional recurrent neural network called BiLSTM. This work is inspired by recent successful applications of deep neural networks in many natural language processing problems BIBREF5 , BIBREF6 . Our method attempts to model complex relationships between the title of an article and its popularity using novel deep network architecture that, in contrast to the previous approaches, gives highly interpretable results. Last but not least, the proposed BiLSTM method provides a significant performance boost in terms of prediction accuracy over the standard shallow approach, while outperforming the current state-of-the-art on two distinct datasets with over 40,000 samples.
To summarize, the contributions presented in this paper are the following:
The remainder of this paper is organized in the following manner: first, we review the relevant literature and compare our approach to existing work. Next, we formulate the problem of popularity prediction and propose a model that takes advantage of BiLSTM architecture to address it. Then, we evaluate our model on two datasets using several pre-trained word embeddings and compare it to benchmark models. We conclude this work with discussion on future research paths.
## Related Work
The ever increasing popularity of the Internet as a virtual space to share content inspired research community to analyze different aspects of online information distribution. Various types of content were analyzed, ranging from textual data, such as Twitter posts BIBREF0 or Digg stories BIBREF2 to images BIBREF7 to videos BIBREF8 , BIBREF3 , BIBREF9 . Although several similarities were observed across content domains, e.g. log-normal distribution of data popularity BIBREF10 , in this work we focus only on textual content and, more precisely, on the popularity of news articles and its relation to the article's title.
Forecasting popularity of news articles was especially well studied in the context of Twitter - a social media platform designed specifically for sharing textual data BIBREF11 , BIBREF12 . Not only did the previous works focus on the prediction part, but also on modeling message propagation within the network BIBREF13 . However, most of the works were focused on analyzing the social interactions between the users and the characteristics of so-called social graph of users' connections, rather than on the textual features. Contrary to those approaches, in this paper we base our predictions using only textual features of the article title. We also validate our proposed method on one dataset collected using a different social media platform, namely Facebook, and another one created from various news articles BIBREF4 .
Recently, several works have touched on the topic of popularity prediction of news article from a multimodal perspective BIBREF4 , BIBREF14 . Although in BIBREF4 the authors analyze news articles on a per-modality basis, they do not approach the problem of popularity prediction in a holistic way. To address this shortcoming, BIBREF14 have proposed a multimodal approach to predicting popularity of short videos shares in social media platform Vine using a model that fuses features related to different modalities. In our work, we focus only on textual features of the article title for the purpose of popularity prediction, as our goal is to empower the journalists to quantitatively assess the quality of the headlines they create before the publication. Nevertheless, we believe that in future research we will extend our method towards multimodal popularity prediction.
## Method
In this section we present the bidirectional LSTM model for popularity prediction. We start by formulating the problem and follow up with the description of word embeddings used in our approach. We then present the Long Short-Term Memory network that serves as a backbone for our bidirectional LSTM architecture. We conclude this section with our interpretation of hidden bidirectional states and describe how they can be employed for title introspection.
## Problem Formulation
We cast the problem of popularity prediction as a binary classification task. We assume our data points contain a string of characters representing article title and a popularity metric, such as number of comments or views. The input of our classification is the character string, while the output is the binary label corresponding to popular or unpopular class. To enable the comparison of the methods on datasets containing content published on different websites and with different audience sizes, we determine that a video is popular if its popularity metric exceeds the median value of the corresponding metric for other points in the set, otherwise - it is labeled as unpopular. The details of the labeling procedure are discussed separately in the Datasets section.
## Text Representation
Since the input of our method is textual data, we follow the approach of BIBREF15 and map the text into a fixed-size vector representation. To this end, we use word embeddings that were successfully applied in other domains. We follow BIBREF5 and use pre-trained GloVe word vectors BIBREF16 to initialize the embedding layer (also known as look-up table). Section SECREF18 discusses the embedding layer in more details.
## Bidirectional Long Short-Term Memory Network
Our method for popularity prediction using article's title is inspired by a bidirectional LSTM architecture. The overview of the model can be seen in Fig. FIGREF8 .
Let INLINEFORM0 be INLINEFORM1 -dimensional word vector corresponding to the INLINEFORM2 -the word in the headline, then a variable length sequence: INLINEFORM3 represents a headline. A recurrent neural network (RNN) processes this sequence by recursively applying a transformation function to the current element of sequence INLINEFORM4 and its previous hidden internal state INLINEFORM5 (optionally outputting INLINEFORM6 ). At each time step INLINEFORM7 , the hidden state is updated by: DISPLAYFORM0
where INLINEFORM0 is a non-linear activation function. LSTM network BIBREF17 updates its internal state differently, at each step INLINEFORM1 it calculates: DISPLAYFORM0
where INLINEFORM0 is the sigmoid activation function, tanh is the hyperbolic tangent function and INLINEFORM1 denotes component-wise multiplication. In our experiments we used 128, 256 for the dimensionality of hidden layer in both LSTM and BiLSTM. The term in equation EQREF10 INLINEFORM2 , is called the input gate and it uses the input word and the past hidden state to determine whether the input is worth remembering or not. The amount of information that is being discarded is controlled by forget gate INLINEFORM3 , while INLINEFORM4 is the output gate that controls the amount of information that leaks from memory cell INLINEFORM5 to the hidden state INLINEFORM6 . In the context of classification, we typically treat the output of the hidden state at the last time step of LSTM as the document representation and feed it to sigmoid layer to perform classification BIBREF18 .
Due to its sequential nature, a recurrent neural network puts more emphasis on the recent elements. To circumvent this problem BIBREF19 introduced a bidirectional RNN in which each training sequence is presented forwards and backwards to two separate recurrent nets, both of which are connected to the same output layer. Therefore, at any time-step we have the whole information about the sequence. This is shown by the following equation: DISPLAYFORM0
In our method, we use the bidirectional LSTM architecture for content popularity prediction using only textual cues. We have to therefore map the neural network outputs from a set of hidden states INLINEFORM0 to classification labels. We evaluated several approaches to this problem, such as max or mean pooling. The initial experiments showed that the highest performance was achieved using late fusion approach, that is by concatenating the last hidden state in forward and backward sequence. The intuition behind this design choice is that the importance of the first few words of the headline is relatively high, as the information contained in INLINEFORM1 , i.e. the last item in the backward sequence, is mostly taken from the first word.
## Hidden State Interpretation
One interesting property of bidirectional RNNs is the fact, that the concatenation of hidden states INLINEFORM0 and INLINEFORM1 can be interpreted as a context-dependent vector representation of word INLINEFORM2 . This allows us to introspect a given title and approximate the contribution of each word to the estimated popularity. To that end one can process the headline representation INLINEFORM3 through the bidirectional recurrent network and then retrieve pairs of forward and backwards hidden state INLINEFORM4 for each word INLINEFORM5 . Then, the output of the last fully-connected layer INLINEFORM6 could be interpreted as context-depended popularity of a word INLINEFORM7 .
## Training
In our experiments we minimize the binary cross-entropy loss using Stochastic Gradient Descent on randomly shuffled mini-batches with the Adam optimization algorithm BIBREF20 . We reduce the learning rate by a factor of 0.2 once learning plateaus. We also employ early stopping strategy, i.e. stopping the training algorithm before convergence based on the values of loss function on the validation set.
## Evaluation
In this section, we evaluate our method and compare its performance against the competitive approaches. We use INLINEFORM0 -fold evaluation protocol with INLINEFORM1 with random dataset split. We measure the performance using standard accuracy metric which we define as a ratio between correctly classified data samples from test dataset and all test samples.
## Datasets
In this section we present two datasets used in our experiments: The NowThisNews dataset, collected for the purpose of this paper, and The BreakingNews dataset BIBREF4 , publicly available dataset of news articles.
contains 4090 posts with associated videos from NowThisNews Facebook page collected between 07/2015 and 07/2016. For each post we collected its title and the number of views of the corresponding video, which we consider our popularity metric. Due to a fairly lengthy data collection process, we decided to normalize our data by first grouping posts according to their publication month and then labeling the posts for which the popularity metric exceeds the median monthly value as popular, the remaining part as unpopular.
BIBREF4 contains a variety of news-related information such as images, captions, geo-location information and comments which could be used as a proxy for article popularity. The articles in this dataset were collected between January and December 2014. Although we tried to retrieve the entire dataset, we were able to download only 38,182 articles due to the dead links published in the dataset. The retrieved articles were published in main news channels, such as Yahoo News, The Guardian or The Washington Post. Similarly, to The NowThisNews dataset we normalize the data by grouping articles per publisher, and classifying them as popular, when the number of comments exceeds the median value for given publisher.
## Baselines
As a first baseline we use Bag-of-Words, a well-known and robust text representations used in various domains BIBREF21 , combined with a standard shallow classifier, namely, a Support Vector Machine with linear kernel. We used LIBSVM implementation of SVM.
Our second baseline is a deep Convectional Neural Network applied on word embeddings. This baseline represents state-of-the-art method presented in BIBREF4 with minor adjustments to the binary classification task. The architecture of the CNN benchmark we use is the following: the embedding layer transforms one-hot encoded words to their dense vector representations, followed by the convolution layer of 256 filters with width equal to 5 followed by max pooling layer (repeated three times), fully-connected layer with dropout and INLINEFORM0 regularization and finally, sigmoid activation layer. For fair comparison, both baselines were trained using the same training procedure as our method.
## Embeddings
As a text embedding in our experiments, we use publicly available GloVe word vectors BIBREF16 pre-trained on two datasets: Wikipedia 2014 with Gigaword5 (W+G5) and Common Crawl (CC). Since their output dimensionality can be modified, we show the results for varying dimensionality sizes. On top of that, we evaluate two training approaches: using static word vectors and fine-tuning them during training phase.
## Results
The results of our experiments can be seen in Tab. TABREF21 and TABREF22 . Our proposed BiLSTM approach outperforms the competing methods consistently across both datasets. The performance improvement is especially visible for The NowThisNews dataset and reaches over 15% with respect to the shallow architecture in terms of of accuracy. Although the improvement with respect to the other methods based on deep neural network is less evident, the recurrent nature of our method provides much more intuitive interpretation of the results and allow for parsing the contribution of each single word to the overall score.
To present how our model works in practice, we show in Tab. TABREF23 a list of 3 headlines from NowThisNews dataset that are scored with the highest probability of belonging to a popular class, as well as 3 headlines with the lowest score. As can be seen, our model correctly detected videos that become viral at the same time assigning low score to content that underperformed. We believe that BiLSTM could be successfully applied in real-life scenarios.
## Conclusions
In this paper we present a novel approach to the problem of online article popularity prediction. To our knowledge, this is the first attempt of predicting the performance of content on social media using only textual information from its title. We show that our method consistently outperforms benchmark models. Additionally, the proposed method could not only be used to compare competing titles with regard to their estimated probability, but also to gain insights about what constitutes a good title. Future work includes modeling popularity prediction problem with multiple data modalities, such as images or videos. Furthermore, all of the evaluated models function at the word level, which could be problematic due to idiosyncratic nature of social media and Internet content. It is, therefore, worth investigating, whether combining models that operate at the character level to learn and generate vector representation of titles with visual features could improve the overall performance.
| 14
|
1708.00111
|
A Continuous Relaxation of Beam Search for End-to-end Training of Neural Sequence Models
|
# A Continuous Relaxation of Beam Search for End-to-end Training of Neural Sequence Models
## Abstract
Beam search is a desirable choice of test-time decoding algorithm for neural sequence models because it potentially avoids search errors made by simpler greedy methods. However, typical cross entropy training procedures for these models do not directly consider the behaviour of the final decoding method. As a result, for cross-entropy trained models, beam decoding can sometimes yield reduced test performance when compared with greedy decoding. In order to train models that can more effectively make use of beam search, we propose a new training procedure that focuses on the final loss metric (e.g. Hamming loss) evaluated on the output of beam search. While well-defined, this"direct loss"objective is itself discontinuous and thus difficult to optimize. Hence, in our approach, we form a sub-differentiable surrogate objective by introducing a novel continuous approximation of the beam search decoding procedure. In experiments, we show that optimizing this new training objective yields substantially better results on two sequence tasks (Named Entity Recognition and CCG Supertagging) when compared with both cross entropy trained greedy decoding and cross entropy trained beam decoding baselines.
## Introduction
[t] Standard Beam Search [1] INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 t = 0 to T i = 1 to k INLINEFORM4 INLINEFORM5 INLINEFORM6 is the local output scoring function INLINEFORM7 top-k-max INLINEFORM8 Top k values of the input matrix INLINEFORM9 top-k-argmax INLINEFORM10 Top INLINEFORM11 argmax index pairs of the input matrix i = 1 to k INLINEFORM12 embedding( INLINEFORM13 ) INLINEFORM14 INLINEFORM15 is a nonlinear recurrent function that returns state at next step INLINEFORM16 INLINEFORM17 follow-backpointer( INLINEFORM18 ) INLINEFORM19 Sequence-to-sequence (seq2seq) models have been successfully used for many sequential decision tasks such as machine translation BIBREF0 , BIBREF1 , parsing BIBREF2 , BIBREF3 , summarization BIBREF4 , dialog generation BIBREF5 , and image captioning BIBREF6 . Beam search is a desirable choice of test-time decoding algorithm for such models because it potentially avoids search errors made by simpler greedy methods. However, the typical approach to training neural sequence models is to use a locally normalized maximum likelihood objective (cross-entropy training) BIBREF0 . This objective does not directly reason about the behaviour of the final decoding method. As a result, for cross-entropy trained models, beam decoding can sometimes yield reduced test performance when compared with greedy decoding BIBREF7 , BIBREF8 , BIBREF9 . These negative results are not unexpected. The training procedure was not search-aware: it was not able to consider the effect that changing the model's scores might have on the ease of search while using a beam decoding, greedy decoding, or otherwise.
We hypothesize that the under-performance of beam search in certain scenarios can be resolved by using a better designed training objective. Because beam search potentially offers more accurate search when compared to greedy decoding, we hope that appropriately trained models should be able to leverage beam search to improve performance. In order to train models that can more effectively make use of beam search, we propose a new training procedure that focuses on the final loss metric (e.g. Hamming loss) evaluated on the output of beam search. While well-defined and a valid training criterion, this “direct loss” objective is discontinuous and thus difficult to optimize. Hence, in our approach, we form a sub-differentiable surrogate objective by introducing a novel continuous approximation of the beam search decoding procedure. In experiments, we show that optimizing this new training objective yields substantially better results on two sequence tasks (Named Entity Recognition and CCG Supertagging) when compared with both cross-entropy trained greedy decoding and cross-entropy trained beam decoding baselines.
Several related methods, including reinforcement learning BIBREF10 , BIBREF11 , imitation learning BIBREF12 , BIBREF13 , BIBREF14 , and discrete search based methods BIBREF15 , BIBREF16 , BIBREF17 , BIBREF18 , have also been proposed to make training search-aware. These methods include approaches that forgo direct optimization of a global training objective, instead incorporating credit assignment for search errors by using methods like early updates BIBREF19 that explicitly track the reachability of the gold target sequence during the search procedure. While addressing a related problem – credit assignment for search errors during training – in this paper, we propose an approach with a novel property: we directly optimize a continuous and global training objective using backpropagation. As a result, in our approach, credit assignment is handled directly via gradient optimization in an end-to-end computation graph. The most closely related work to our own approach was proposed by Goyal et al. BIBREF20 . They do not consider beam search, but develop a continuous approximation of greedy decoding for scheduled sampling objectives. Other related work involves training a generator with a Gumbel reparamterized sampling module to more reliably find the MAP sequences at decode-time BIBREF21 , and constructing surrogate loss functions BIBREF22 that are close to task losses.
## Model
We denote the seq2seq model parameterized by INLINEFORM0 as INLINEFORM1 . We denote the input sequence as INLINEFORM2 , the gold output sequence as INLINEFORM3 and the result of beam search over INLINEFORM4 as INLINEFORM5 . Ideally, we would like to directly minimize a final evaluation loss, INLINEFORM6 , evaluated on the result of running beam search with input INLINEFORM7 and model INLINEFORM8 . Throughout this paper we assume that the evaluation loss decomposes over time steps INLINEFORM9 as: INLINEFORM10 . We refer to this idealized training objective that directly evaluates prediction loss as the “direct loss” objective and define it as: DISPLAYFORM0
Unfortunately, optimizing this objective using gradient methods is difficult because the objective is discontinuous. The two sources of discontinuity are:
We introduce a surrogate training objective that avoids these problems and as a result is fully continuous. In order to accomplish this, we propose a continuous relaxation to the composition of our final loss metric, INLINEFORM0 , and our decoder function, INLINEFORM1 : INLINEFORM2
Specifically, we form a continuous function softLB that seeks to approximate the result of running our decoder on input INLINEFORM0 and then evaluating the result against INLINEFORM1 using INLINEFORM2 . By introducing this new module, we are now able to construct our surrogate training objective: DISPLAYFORM0
Specified in more detail in Section SECREF9 , our surrogate objective in Equation 2 will additionally take a hyperparameter INLINEFORM0 that trades approximation quality for smoothness of the objective. Under certain conditions, Equation 2 converges to the objective in Equation 1 as INLINEFORM1 is increased. We first describe the standard discontinuous beam search procedure and then our training approach (Equation 2) involving a continuous relaxation of beam search.
## Discontinuity in Beam Search
[t] continuous-top-k-argmax [1] INLINEFORM0 INLINEFORM1 , s.t. INLINEFORM2 INLINEFORM3 INLINEFORM4 = 1 to k peaked-softmax will be dominated by scores closer to INLINEFORM5 INLINEFORM6 The square operation is element-wise Formally, beam search is a procedure with hyperparameter INLINEFORM7 that maintains a beam of INLINEFORM8 elements at each time step and expands each of the INLINEFORM9 elements to find the INLINEFORM10 -best candidates for the next time step. The procedure finds an approximate argmax of a scoring function defined on output sequences.
We describe beam search in the context of seq2seq models in Algorithm SECREF1 – more specifically, for an encoder-decoder BIBREF0 model with a nonlinear auto-regressive decoder (e.g. an LSTM BIBREF23 ). We define the global model score of a sequence INLINEFORM0 with length INLINEFORM1 to be the sum of local output scores at each time step of the seq2seq model: INLINEFORM2 . In neural models, the function INLINEFORM3 is implemented as a differentiable mapping, INLINEFORM4 , which yields scores for vocabulary elements using the recurrent hidden states at corresponding time steps. In our notation, INLINEFORM5 is the hidden state of the decoder at time step INLINEFORM6 for beam element INLINEFORM7 , INLINEFORM8 is the embedding of the output symbol at time-step INLINEFORM9 for beam element INLINEFORM10 , and INLINEFORM11 is the cumulative model score at step INLINEFORM12 for beam element INLINEFORM13 . In Algorithm SECREF1 , we denote by INLINEFORM14 the cumulative candidate score matrix which represents the model score of each successor candidate in the vocabulary for each beam element. This score is obtained by adding the local output score (computed as INLINEFORM15 ) to the running total of the score for the candidate. The function INLINEFORM16 in Algorithms SECREF1 and SECREF7 yields successive hidden states in recurrent neural models like RNNs, LSTMs etc. The INLINEFORM17 operation maps a word in the vocabulary INLINEFORM18 , to a continuous embedding vector. Finally, backpointers at each time step to the beam elements at the previous time step are also stored for identifying the best sequence INLINEFORM19 , at the conclusion of the search procedure. A backpointer at time step INLINEFORM20 for a beam element INLINEFORM21 is denoted by INLINEFORM22 which points to one of the INLINEFORM23 elements at the previous beam. We denote a vector of backpointers for all the beam elements by INLINEFORM24 . The INLINEFORM25 operation takes as input backpointers ( INLINEFORM26 ) and candidates ( INLINEFORM27 ) for all the beam elements at each time step and traverses the sequence in reverse (from time-step INLINEFORM28 through 1) following backpointers at each time step and identifying candidate words associated with each backpointer that results in a sequence INLINEFORM29 , of length INLINEFORM30 .
The procedure described in Algorithm SECREF1 is discontinuous because of the top-k-argmax procedure that returns a pair of vectors corresponding to the INLINEFORM0 highest-scoring indices for backpointers and vocabulary items from the score matrix INLINEFORM1 . This index selection results in hard backpointers at each time step which restrict the gradient flow during backpropagation. In the next section, we describe a continuous relaxation to the top-k-argmax procedure which forms the crux of our approach.
## Continuous Approximation to top-k-argmax
[t] Continuous relaxation to beam search [1] INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 , INLINEFORM4 t = 0 to T INLINEFORM5 i=1 to k INLINEFORM6 INLINEFORM7 is a local output scoring function INLINEFORM8 INLINEFORM9 is used to compute INLINEFORM10 INLINEFORM11 Call Algorithm 2 i = 1 to k INLINEFORM12 Soft back pointer computation INLINEFORM13 Contribution from vocabulary items INLINEFORM14 Peaked distribution over the candidates to compute INLINEFORM15 INLINEFORM16 INLINEFORM17 INLINEFORM18 j = 1 to k Get contributions from soft backpointers for each beam element INLINEFORM19 INLINEFORM20 INLINEFORM21 INLINEFORM22 is a nonlinear recurrent function that returns state at next step INLINEFORM23 Pick the loss for the sequence with highest model score on the beam in a soft manner.
The key property that we use in our approximation is that for a real valued vector INLINEFORM0 , the argmax with respect to a vector of scores, INLINEFORM1 , can be approximated by a temperature controlled softmax operation. The argmax operation can be represented as: INLINEFORM2
which can be relaxed by replacing the indicator function with a peaked-softmax operation with hyperparameter INLINEFORM0 : INLINEFORM1
As INLINEFORM0 , INLINEFORM1 so long as there is only one maximum value in the vector INLINEFORM2 . This peaked-softmax operation has been shown to be effective in recent work BIBREF24 , BIBREF25 , BIBREF20 involving continuous relaxation to the argmax operation, although to our knowledge, this is the first work to apply it to approximate the beam search procedure.
Using this peaked-softmax operation, we propose an iterative algorithm for computing a continuous relaxation to the top-k-argmax procedure in Algorithm SECREF6 which takes as input a score matrix of size INLINEFORM0 and returns INLINEFORM1 peaked matrices INLINEFORM2 of size INLINEFORM3 . Each matrix INLINEFORM4 represents the index of INLINEFORM5 -th max. For example, INLINEFORM6 will have most of its mass concentrated on the index in the matrix that corresponds to the argmax, while INLINEFORM7 will have most of its mass concentrated on the index of the 2nd-highest scoring element. Specifically, we obtain matrix INLINEFORM8 by computing the squared difference between the INLINEFORM9 -highest score and all the scores in the matrix and then using the peaked-softmax operation over the negative squared differences. This results in scores closer to the INLINEFORM10 -highest score to have a higher mass than scores far away from the INLINEFORM11 -highest score.
Hence, the continuous relaxation to top-k-argmax operation can be simply implemented by iteratively using the max operation which is continuous and allows for gradient flow during backpropagation. As INLINEFORM0 , each INLINEFORM1 vector converges to hard index pairs representing hard backpointers and successor candidates described in Algorithm SECREF1 . For finite INLINEFORM2 , we introduce a notion of a soft backpointer, represented as a vector INLINEFORM3 in the INLINEFORM4 -probability simplex, which represents the contribution of each beam element from the previous time step to a beam element at current time step. This is obtained by a row-wise sum over INLINEFORM5 to get INLINEFORM6 values representing soft backpointers.
## Training with Continuous Relaxation of Beam Search
We describe our approach in detail in Algorithm 3 and illustrate the soft beam recurrence step in Figure 1. For composing the loss function and the beam search function for our optimization as proposed in Equation 2, we make use of decomposability of the loss function across time-steps. Thus for a sequence y, the total loss is: INLINEFORM0 . In our experiments, INLINEFORM1 is the Hamming loss which can be easily computed at each time-step by simply comparing gold INLINEFORM2 with INLINEFORM3 . While exact computation of INLINEFORM4 will vary according to the loss, our proposed procedure will be applicable as long as the total loss is decomposable across time-steps. While decomposability of loss is a strong assumption, existing literature on structured prediction BIBREF26 , BIBREF27 has made due with this assumption, often using decomposable losses as surrogates for non-decomposable ones. We detail the continuous relaxation to beam search in Algorithm SECREF7 with INLINEFORM5 being the cumulative loss of beam element INLINEFORM6 at time step INLINEFORM7 and INLINEFORM8 being the embedding matrix of the target vocabulary which is of size INLINEFORM9 where INLINEFORM10 is the size of the embedding vector.
In Algorithm SECREF7 , all the discrete selection functions have been replaced by their soft, continuous counterparts which can be backpropagated through. This results in all the operations being matrix and vector operations which is ideal for a GPU implementation. An important aspect of this algorithm is that we no longer rely on exactly identifying a discrete search prediction INLINEFORM0 since we are only interested in a continuous approximation to the direct loss INLINEFORM1 (line 18 of Algorithm SECREF7 ), and all the computation is expressed via the soft beam search formulation which eliminates all the sources of discontinuities associated with the training objective in Equation 1. The computational complexity of our approach for training scales linearly with the beam size and hence is roughly INLINEFORM2 times slower than standard CE training for beam size INLINEFORM3 . Since we have established the pointwise convergence of peaked-softmax to argmax as INLINEFORM4 for all vectors that have a unique maximum value, we can establish pointwise convergence of objective in Equation 2 to objective in Equation 1 as INLINEFORM5 , as long as there are no ties among the top-k scores of the beam expansion candidates at any time step. We posit that absolute ties are unlikely due to random initialization of weights and the domain of the scores being INLINEFORM6 . Empirically, we did not observe any noticeable impact of potential ties on the training procedure and our approach performed well on the tasks as discussed in Section SECREF4 . DISPLAYFORM0
We experimented with different annealing schedules for INLINEFORM0 starting with non-peaked softmax moving toward peaked-softmax across epochs so that learning is stable with informative gradients. This is important because cost functions like Hamming distance with very high INLINEFORM1 tend to be non-smooth and are generally flat in regions far away from changepoints and have a very large gradient near the changepoints which makes optimization difficult.
## Decoding
The motivation behind our approach is to make the optimization aware of beam search decoding while maintaining the continuity of the objective. However, since our approach doesn't introduce any new model parameters and optimization is agnostic to the architecture of the seq2seq model, we were able to experiment with various decoding schemes like locally normalized greedy decoding, and hard beam search, once the model has been trained.
However, to reduce the gap between the training procedure and test procedure, we also experimented with soft beam search decoding. This decoding approach closely follows Algorithm SECREF7 , but along with soft back pointers, we also compute hard back pointers at each time step. After computing all the relevant quantities like model score, loss etc., we follow the hard backpointers to obtain the best sequence INLINEFORM0 . This is very different from hard beam decoding because at each time step, the selection decisions are made via our soft continuous relaxation which influences the scores, LSTM hidden states and input embeddings at subsequent time-steps. The hard backpointers are essentially the MAP estimate of the soft backpointers at each step. With small, finite INLINEFORM1 , we observe differences between soft beam search and hard beam search decoding in our experiments.
## Comparison with Max-Margin Objectives
Max-margin based objectives are typically motivated as another kind of surrogate training objective which avoid the discontinuities associated with direct loss optimization. Hinge loss for structured prediction typically takes the form: INLINEFORM0
where INLINEFORM0 is the input sequence, INLINEFORM1 is the gold target sequence, INLINEFORM2 is the output search space and INLINEFORM3 is the discontinuous cost function which we assume is decomposable across the time-steps of a sequence. Finding the cost augmented maximum score is generally difficult in large structured models and often involves searching over the output space and computing the approximate cost augmented maximal output sequence and the score associated with it via beam search. This procedure introduces discontinuities in the training procedure of structured max-margin objectives and renders it non amenable to training via backpropagation. Related work BIBREF15 on incorporating beam search into the training of neural sequence models does involve cost-augmented max-margin loss but it relies on discontinuous beam search forward passes and an explicit mechanism to ensure that the gold sequence stays in the beam during training, and hence does not involve back propagation through the beam search procedure itself.
Our continuous approximation to beam search can very easily be modified to compute an approximation to the structured hinge loss so that it can be trained via backpropagation if the cost function is decomposable across time-steps. In Algorithm SECREF7 , we only need to modify line 5 as: INLINEFORM0
and instead of computing INLINEFORM0 in Algorithm SECREF7 , we first compute the cost augmented maximum score as: INLINEFORM1
and also compute the target score INLINEFORM0 by simply running the forward pass of the LSTM decoder over the gold target sequence. The continuous approximation to the hinge loss to be optimized is then: INLINEFORM1 . We empirically compare this approach with the proposed approach to optimize direct loss in experiments.
## Experimental Setup
Since our goal is to investigate the efficacy of our approach for training generic seq2seq models, we perform experiments on two NLP tagging tasks with very different characteristics and output search spaces: Named Entity Recognition (NER) and CCG supertagging. While seq2seq models are appropriate for CCG supertagging task because of the long-range correlations between the sequential output elements and a large search space, they are not ideal for NER which has a considerably smaller search space and weaker correlations between predictions at subsequent time steps. In our experiments, we observe improvements from our approach on both of the tasks. We use a seq2seq model with a bi-directional LSTM encoder (1 layer with tanh activation function) for the input sequence INLINEFORM0 , and an LSTM decoder (1 layer with tanh activation function) with a fixed attention mechanism that deterministically attends to the INLINEFORM1 -th input token when decoding the INLINEFORM2 -th output, and hence does not involve learning of any attention parameters. Since, computational complexity of our approach for optimization scales linearly with beam size for each instance, it is impractical to use very large beam sizes for training. Hence, beam size for all the beam search based experiments was set to 3 which resulted in improvements on both the tasks as discussed in the results. For both tasks, the direct loss function was the Hamming distance cost which aims to maximize word level accuracy.
## Named Entity Recognition
For named entity recognition, we use the CONLL 2003 shared task data BIBREF28 for German language and use the provided data splits. We perform no preprocessing on the data. The output vocabulary length (label space) is 10. A peculiar characteristic of this problem is that the training data is naturally skewed toward one default label (`O') because sentences typically do not contain many named entities and the evaluation focuses on the performance recognizing entities. Therefore, we modify the Hamming cost such that incorrect prediction of `O' is doubly penalized compared to other incorrect predictions. We use the hidden layers of size 64 and label embeddings of size 8. As mentioned earlier, seq2seq models are not an ideal choice for NER (tag-level correlations are short-ranged in NER – the unnecessary expressivity of full seq2seq models over simple encoder-classifier neural models makes training harder). However, we wanted to evaluate the effectiveness of our approach on different instantiations of seq2seq models.
## CCG Supertagging
We used the standard splits of CCG bank BIBREF29 for training, development, and testing. The label space of supertags is 1,284 which is much larger than NER. The distribution of supertags in the training data exhibits a long tail because these supertags encode specific syntactic information about the words' usage. The supertag labels are correlated with each other and many tags encode similar information about the syntax. Moreover, this task is sensitive to the long range sequential decisions and search effects because of how it holistically encodes the syntax of the entire sentence. We perform minor preprocessing on the data similar to the preprocessing in BIBREF30 . For this task, we used hidden layers of size 512 and the supertag label embeddings were also of size 512. The standard evaluation metric for this task is the word level label accuracy which directly corresponds to Hamming loss.
## Hyperparameter tuning
For tuning all the hyperparameters related to optimization we trained our models for 50 epochs and picked the models with the best performance on the development set. We also ran multiple random restarts for all the systems evaluated to account for performance variance across randomly started runs. We pretrained all our models with standard cross entropy training which was important for stable optimization of the non convex neural objective with a large parameter search space. This warm starting is a common practice in prior work on complex neural models BIBREF10 , BIBREF4 , BIBREF14 .
## Comparison
We report performance on validation and test sets for both the tasks in Tables 1 and 2. The baseline model is a cross entropy trained seq2seq model (Baseline CE) which is also used to warm start the the proposed optimization procedures in this paper. This baseline has been compared against the approximate direct loss training objective (Section SECREF9 ), referred to as INLINEFORM0 in the tables, and the approximate max-margin training objective (Section SECREF12 ), referred to as INLINEFORM1 in the tables. Results are reported for models when trained with annealing INLINEFORM2 , and also with a constant setting of INLINEFORM3 which is a very smooth but inaccurate approximation of the original direct loss that we aim to optimize. Comparisons have been made on the basis of performance of the models under different decoding paradigms (represented as different column in the tables): locally normalized decoding (CE greedy), hard beam search decoding and soft beam search decoding described in Section SECREF11 .
## Results
As shown in Tables 1 and 2, our approach INLINEFORM0 shows significant improvements over the locally normalized CE baseline with greedy decoding for both the tasks (+5.5 accuracy points gain for supertagging and +1.5 F1 points for NER). The improvement is more pronounced on the supertagging task, which is not surprising because: (i) the evaluation metric is tag-level accuracy which is congruent with the Hamming loss that INLINEFORM1 directly optimizes and (ii) the supertagging task itself is very sensitive to the search procedure because tags across time-steps tend to exhibit long range dependencies as they encode specialized syntactic information about word usage in the sentence.
Another common trend to observe is that annealing INLINEFORM0 always results in better performance than training with a constant INLINEFORM1 for both INLINEFORM2 (Section SECREF9 ) and INLINEFORM3 (Section SECREF12 ). This shows that a stable training scheme that smoothly approaches minimizing the actual direct loss is important for our proposed approach. Additionally, we did not observe a large difference when our soft approximation is used for decoding (Section SECREF11 ) compared to hard beam search decoding, which suggests that our approximation to the hard beam search is as effective as its discrete counterpart.
For supertagging, we observe that the baseline cross entropy trained model improves its predictions with beam search decoding compared to greedy decoding by 2 accuracy points, which suggests that beam search is already helpful for this task, even without search-aware training. Both the optimization schemes proposed in this paper improve upon the baseline with soft direct loss optimization ( INLINEFORM0 ), performing better than the approximate max-margin approach.
For NER, we observe that optimizing INLINEFORM0 outperforms all the other approaches but we also observe interesting behaviour of beam search decoding and the approximate max-margin objective for this task. The pretrained CE baseline model yields worse performance when beam search is done instead of greedy locally normalized decoding. This is because the training data is heavily skewed toward the `O' label and hence the absolute score resolution between different tags at each time-step during decoding isn't enough to avoid leading beam search toward a wrong hypothesis path. We observed in our experiments that hard beam search resulted in predicting more `O's which also hurt the prediction of tags at future time steps and hurt precision as well as recall. Encouragingly, INLINEFORM1 optimization, even though warm started with a CE trained model that performs worse with beam search, led to the NER model becoming more search aware, which resulted in superior performance. However, we also observe that the approximate max-margin approach ( INLINEFORM2 ) performs poorly here. We attribute this to a deficiency in the max-margin objective when coupled with approximate search methods like beam search that do not provide guarantees on finding the supremum: one way to drive this objective down is to learn model scores such that the search for the best hypothesis is difficult, so that the value of the loss augmented decode is low, while the gold sequence maintains higher model score. Because we also warm started with a pre-trained model that results in a worse performance with beam search decode than with greedy decode, we observe the adverse effect of this deficiency. The result is a model that scores the gold hypothesis highly, but yields poor decoding outputs. This observation indicates that using max-margin based objectives with beam search during training actually may achieve the opposite of our original intent: the objective can be driven down by introducing search errors.
The observation that our optimization method led to improvements on both the tasks–even on NER for which hard beam search during decoding on a CE trained model hurt the performance–by making the optimization more search aware, indicates the effectiveness of our approach for training seq2seq models.
## Conclusion
While beam search is a method of choice for performing search in neural sequence models, as our experiments confirm, it is not necessarily guaranteed to improve accuracy when applied to cross-entropy-trained models. In this paper, we propose a novel method for optimizing model parameters that directly takes into account the process of beam search itself through a continuous, end-to-end sub-differentiable relaxation of beam search composed with the final evaluation loss. Experiments demonstrate that our method is able to improve overall test-time results for models using beam search as a test-time inference method, leading to substantial improvements in accuracy.
| 14
|
1709.01256
|
Semantic Document Distance Measures and Unsupervised Document Revision Detection
|
# Semantic Document Distance Measures and Unsupervised Document Revision Detection
## Abstract
In this paper, we model the document revision detection problem as a minimum cost branching problem that relies on computing document distances. Furthermore, we propose two new document distance measures, word vector-based Dynamic Time Warping (wDTW) and word vector-based Tree Edit Distance (wTED). Our revision detection system is designed for a large scale corpus and implemented in Apache Spark. We demonstrate that our system can more precisely detect revisions than state-of-the-art methods by utilizing the Wikipedia revision dumps https://snap.stanford.edu/data/wiki-meta.html and simulated data sets.
## Introduction
It is a common habit for people to keep several versions of documents, which creates duplicate data. A scholarly article is normally revised several times before being published. An academic paper may be listed on personal websites, digital conference libraries, Google Scholar, etc. In major corporations, a document typically goes through several revisions involving multiple editors and authors. Users would benefit from visualizing the entire history of a document. It is worthwhile to develop a system that is able to intelligently identify, manage and represent revisions. Given a collection of text documents, our study identifies revision relationships in a completely unsupervised way. For each document in a corpus we only use its content and the last modified timestamp. We assume that a document can be revised by many users, but that the documents are not merged together. We consider collaborative editing as revising documents one by one.
The two research problems that are most relevant to document revision detection are plagiarism detection and revision provenance. In a plagiarism detection system, every incoming document is compared with all registered non-plagiarized documents BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . The system returns true if an original copy is found in the database; otherwise, the system returns false and adds the document to the database. Thus, it is a 1-to-n problem. Revision provenance is a 1-to-1 problem as it keeps track of detailed updates of one document BIBREF4 , BIBREF5 . Real-world applications include GitHub, version control in Microsoft Word and Wikipedia version trees BIBREF6 . In contrast, our system solves an n-to-n problem on a large scale. Our potential target data sources, such as the entire web or internal corpora in corporations, contain numerous original documents and their revisions. The aim is to find all revision document pairs within a reasonable time.
Document revision detection, plagiarism detection and revision provenance all rely on comparing the content of two documents and assessing a distance/similarity score. The classic document similarity measure, especially for plagiarism detection, is fingerprinting BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 . Fixed-length fingerprints are created using hash functions to represent document features and are then used to measure document similarities. However, the main purpose of fingerprinting is to reduce computation instead of improving accuracy, and it cannot capture word semantics. Another widely used approach is computing the sentence-to-sentence Levenshtein distance and assigning an overall score for every document pair BIBREF13 . Nevertheless, due to the large number of existing documents, as well as the large number of sentences in each document, the Levenshtein distance is not computation-friendly. Although alternatives such as the vector space model (VSM) can largely reduce the computation time, their effectiveness is low. More importantly, none of the above approaches can capture semantic meanings of words, which heavily limits the performances of these approaches. For instance, from a semantic perspective, “I went to the bank" is expected to be similar to “I withdrew some money" rather than “I went hiking." Our document distance measures are inspired by the weaknesses of current document distance/similarity measures and recently proposed models for word representations such as word2vec BIBREF14 and Paragraph Vector (PV) BIBREF15 . Replacing words with distributed vector embeddings makes it feasible to measure semantic distances using advanced algorithms, e.g., Dynamic Time Warping (DTW) BIBREF16 , BIBREF17 , BIBREF18 and Tree Edit Distance (TED) BIBREF19 , BIBREF20 , BIBREF21 , BIBREF22 , BIBREF23 , BIBREF24 , BIBREF25 , BIBREF26 . Although calculating text distance using DTW BIBREF27 , TED BIBREF28 or Word Mover's Distance (WMV) BIBREF29 has been attempted in the past, these measures are not ideal for large-scale document distance calculation. The first two algorithms were designed for sentence distances instead of document distances. The third measure computes the distance of two documents by solving a transshipment problem between words in the two documents and uses word2vec embeddings to calculate semantic distances of words. The biggest limitation of WMV is its long computation time. We show in Section SECREF54 that our wDTW and wTED measures yield more precise distance scores with much shorter running time than WMV.
We recast the problem of detecting document revisions as a network optimization problem (see Section SECREF2 ) and consequently as a set of document distance problems (see Section SECREF4 ). We use trained word vectors to represent words, concatenate the word vectors to represent documents and combine word2vec with DTW or TED. Meanwhile, in order to guarantee reasonable computation time in large data sets, we calculate document distances at the paragraph level with Apache Spark. A distance score is computed by feeding paragraph representations to DTW or TED. Our code and data are publicly available.
The primary contributions of this work are as follows.
The rest of this paper is organized in five parts. In Section 2, we clarify related terms and explain the methodology for document revision detection. In Section 3, we provide a brief background on existing document similarity measures and present our wDTW and wTED algorithms as well as the overall process flow. In Section 4, we demonstrate our revision detection results on Wikipedia revision dumps and six simulated data sets. Finally, in Section 5, we summarize some concluding remarks and discuss avenues for future work and improvements.
## Revision Network
The two requirements for a document INLINEFORM0 being a revision of another document INLINEFORM1 are that INLINEFORM2 has been created later than INLINEFORM3 and that the content of INLINEFORM4 is similar to (has been modified from) that of INLINEFORM5 . More specifically, given a corpus INLINEFORM6 , for any two documents INLINEFORM7 , we want to find out the yes/no revision relationship of INLINEFORM8 and INLINEFORM9 , and then output all such revision pairs.
We assume that each document has a creation date (the last modified timestamp) which is readily available from the meta data of the document. In this section we also assume that we have a INLINEFORM0 method and a cut-off threshold INLINEFORM1 . We represent a corpus as network INLINEFORM2 , for example Figure FIGREF5 , in which a vertex corresponds to a document. There is an arc INLINEFORM3 if and only if INLINEFORM4 and the creation date of INLINEFORM5 is before the creation date of INLINEFORM6 . In other words, INLINEFORM7 is a revision candidate for INLINEFORM8 . By construction, INLINEFORM9 is acyclic. For instance, INLINEFORM10 is a revision candidate for INLINEFORM11 and INLINEFORM12 . Note that we allow one document to be the original document of several revised documents. As we only need to focus on revision candidates, we reduce INLINEFORM13 to INLINEFORM14 , shown in Figure FIGREF5 , by removing isolated vertices. We define the weight of an arc as the distance score between the two vertices. Recall the assumption that a document can be a revision of at most one document. In other words, documents cannot be merged. Due to this assumption, all revision pairs form a branching in INLINEFORM15 . (A branching is a subgraph where each vertex has an in-degree of at most 1.) The document revision problem is to find a minimum cost branching in INLINEFORM16 (see Fig FIGREF5 ).
The minimum branching problem was earlier solved by BIBREF30 edmonds1967optimum and BIBREF31 velardi2013ontolearn. The details of his algorithm are as follows.
In our case, INLINEFORM0 is acyclic and, therefore, the second step never occurs. For this reason, Algorithm SECREF2 solves the document revision problem.
Find minimum branching INLINEFORM0 for network INLINEFORM1
[1]
Input: INLINEFORM0 INLINEFORM1
every vertex INLINEFORM0 Set INLINEFORM1 to correspond to all arcs with head INLINEFORM2 Select INLINEFORM3 such that INLINEFORM4 is minimum INLINEFORM5
Output: INLINEFORM0
The essential part of determining the minimum branching INLINEFORM0 is extracting arcs with the lowest distance scores. This is equivalent to finding the most similar document from the revision candidates for every original document.
## Distance/similarity Measures
In this section, we first introduce the classic VSM model, the word2vec model, DTW and TED. We next demonstrate how to combine the above components to construct our semantic document distance measures: wDTW and wTED. We also discuss the implementation of our revision detection system.
## Background
VSM represents a set of documents as vectors of identifiers. The identifier of a word used in this work is the tf-idf weight. We represent documents as tf-idf vectors, and thus the similarity of two documents can be described by the cosine distance between their vectors. VSM has low algorithm complexity but cannot represent the semantics of words since it is based on the bag-of-words assumption.
Word2vec produces semantic embeddings for words using a two-layer neural network. Specifically, word2vec relies on a skip-gram model that uses the current word to predict context words in a surrounding window to maximize the average log probability. Words with similar meanings tend to have similar embeddings.
DTW was developed originally for speech recognition in time series analysis and has been widely used to measure the distance between two sequences of vectors.
Given two sequences of feature vectors: INLINEFORM0 and INLINEFORM1 , DTW finds the optimal alignment for INLINEFORM2 and INLINEFORM3 by first constructing an INLINEFORM4 matrix in which the INLINEFORM5 element is the alignment cost of INLINEFORM6 and INLINEFORM7 , and then retrieving the path from one corner to the diagonal one through the matrix that has the minimal cumulative distance. This algorithm is described by the following formula. DISPLAYFORM0
TED was initially defined to calculate the minimal cost of node edit operations for transforming one labeled tree into another. The node edit operations are defined as follows.
Deletion Delete a node and connect its children to its parent maintaining the order.
Insertion Insert a node between an existing node and a subsequence of consecutive children of this node.
Substitution Rename the label of a node.
Let INLINEFORM0 and INLINEFORM1 be two labeled trees, and INLINEFORM2 be the INLINEFORM3 node in INLINEFORM4 . INLINEFORM5 corresponds to a mapping from INLINEFORM6 to INLINEFORM7 . TED finds mapping INLINEFORM8 with the minimal edit cost based on INLINEFORM9
where INLINEFORM0 means transferring INLINEFORM1 to INLINEFORM2 based on INLINEFORM3 , and INLINEFORM4 represents an empty node.
## Semantic Distance between Paragraphs
According to the description of DTW in Section UID14 , the distance between two documents can be calculated using DTW by replacing each element in the feature vectors INLINEFORM0 and INLINEFORM1 with a word vector. However, computing the DTW distance between two documents at the word level is basically as expensive as calculating the Levenshtein distance. Thus in this section we propose an improved algorithm that is more appropriate for document distance calculation.
In order to receive semantic representations for documents and maintain a reasonable algorithm complexity, we use word2vec to train word vectors and represent each paragraph as a sequence of vectors. Note that in both wDTW and wTED we take document titles and section titles as paragraphs. Although a more recently proposed model PV can directly train vector representations for short texts such as movie reviews BIBREF15 , our experiments in Section SECREF54 show that PV is not appropriate for standard paragraphs in general documents. Therefore, we use word2vec in our work. Algorithm SECREF20 describes how we compute the distance between two paragraphs based on DTW and word vectors. The distance between one paragraph in a document and one paragraph in another document can be pre-calculated in parallel using Spark to provide faster computation for wDTW and wTED.
DistPara
[h] Replace the words in paragraphs INLINEFORM0 and INLINEFORM1 with word2vec embeddings: INLINEFORM2 and INLINEFORM3 Input: INLINEFORM4 and INLINEFORM5 Initialize the first row and the first column of INLINEFORM6 matrix INLINEFORM7 INLINEFORM8 INLINEFORM9 INLINEFORM10 in range INLINEFORM11 INLINEFORM12 in range INLINEFORM13 INLINEFORM14 calculate INLINEFORM15 Return: INLINEFORM16
## Word Vector-based Dynamic Time Warping
As a document can be considered as a sequence of paragraphs, wDTW returns the distance between two documents by applying another DTW on top of paragraphs. The cost function is exactly the DistPara distance of two paragraphs given in Algorithm SECREF20 . Algorithm SECREF21 and Figure FIGREF22 describe our wDTW measure. wDTW observes semantic information from word vectors, which is fundamentally different from the word distance calculated from hierarchies among words in the algorithm proposed by BIBREF27 liu2007sentence. The shortcomings of their work are that it is difficult to learn semantic taxonomy of all words and that their DTW algorithm can only be applied to sentences not documents.
wDTW
[h] Represent documents INLINEFORM0 and INLINEFORM1 with vectors of paragraphs: INLINEFORM2 and INLINEFORM3 Input: INLINEFORM4 and INLINEFORM5 Initialize the first row and the first column of INLINEFORM6 matrix INLINEFORM7 INLINEFORM8 INLINEFORM9 INLINEFORM10 in range INLINEFORM11 INLINEFORM12 in range INLINEFORM13 INLINEFORM14 DistPara INLINEFORM15 calculate INLINEFORM16 Return: INLINEFORM17
## Word Vector-based Tree Edit Distance
TED is reasonable for measuring document distances as documents can be easily transformed to tree structures visualized in Figure FIGREF24 . The document tree concept was originally proposed by BIBREF0 si1997check. A document can be viewed at multiple abstraction levels that include the document title, its sections, subsections, etc. Thus for each document we can build a tree-like structure with title INLINEFORM0 sections INLINEFORM1 subsections INLINEFORM2 ... INLINEFORM3 paragraphs being paths from the root to leaves. Child nodes are ordered from left to right as they appear in the document.
We represent labels in a document tree as the vector sequences of titles, sections, subsections and paragraphs with word2vec embeddings. wTED converts documents to tree structures and then uses DistPara distances. More formally, the distance between two nodes is computed as follows.
The cost of substitution is the DistPara value of the two nodes.
The cost of insertion is the DistPara value of an empty sequence and the label of the inserted node. This essentially means that the cost is the sum of the L2-norms of the word vectors in that node.
The cost of deletion is the same as the cost of insertion.
Compared to the algorithm proposed by BIBREF28 sidorov2015computing, wTED provides different edit cost functions and uses document tree structures instead of syntactic n-grams, and thus wTED yields more meaningful distance scores for long documents. Algorithm SECREF23 and Figure FIGREF28 describe how we calculate the edit cost between two document trees.
wTED
[1] Convert documents INLINEFORM0 and INLINEFORM1 to trees INLINEFORM2 and INLINEFORM3 Input: INLINEFORM4 and INLINEFORM5
Initialize tree edit distance INLINEFORM0 node label INLINEFORM1 node label INLINEFORM2 Update TED mapping cost INLINEFORM3 using INLINEFORM4 DistPara INLINEFORM5 INLINEFORM6 DistPara INLINEFORM7 INLINEFORM8 DistPara INLINEFORM9
Return: INLINEFORM0
## Process Flow
Our system is a boosting learner that is composed of four modules: weak filter, strong filter, revision network and optimal subnetwork. First of all, we sort all documents by timestamps and pair up documents so that we only compare each document with documents that have been created earlier. In the first module, we calculate the VSM similarity scores for all pairs and eliminate those with scores that are lower than an empirical threshold ( INLINEFORM0 ). This is what we call the weak filter. After that, we apply the strong filter wDTW or wTED on the available pairs and filter out document pairs having distances higher than a threshold INLINEFORM1 . For VSM in Section SECREF32 , we directly filter out document pairs having similarity scores lower than a threshold INLINEFORM2 . The cut-off threshold estimation is explained in Section SECREF30 . The remaining document pairs from the strong filter are then sent to the revision network module. In the end, we output the optimal revision pairs following the minimum branching strategy.
## Estimating the Cut-off Threshold
Hyperprameter INLINEFORM0 is calibrated by calculating the absolute extreme based on an initial set of documents, i.e., all processed documents since the moment the system was put in use. Based on this set, we calculate all distance/similarity scores and create a histogram, see Figure FIGREF31 . The figure shows the correlation between the number of document pairs and the similarity scores in the training process of one simulated corpus using VSM. The optimal INLINEFORM1 in this example is around 0.6 where the number of document pairs noticeably drops.
As the system continues running, new documents become available and INLINEFORM0 can be periodically updated by using the same method.
## Numerical Experiments
This section reports the results of the experiments conducted on two data sets for evaluating the performances of wDTW and wTED against other baseline methods.
## Distance/Similarity Measures
We denote the following distance/similarity measures.
wDTW: Our semantic distance measure explained in Section SECREF21 .
wTED: Our semantic distance measure explained in Section SECREF23 .
WMD: The Word Mover's Distance introduced in Section SECREF1 . WMD adapts the earth mover's distance to the space of documents.
VSM: The similarity measure introduced in Section UID12 .
PV-DTW: PV-DTW is the same as Algorithm SECREF21 except that the distance between two paragraphs is not based on Algorithm SECREF20 but rather computed as INLINEFORM0 where INLINEFORM1 is the PV embedding of paragraph INLINEFORM2 .
PV-TED: PV-TED is the same as Algorithm SECREF23 except that the distance between two paragraphs is not based on Algorithm SECREF20 but rather computed as INLINEFORM0 .
Our experiments were conducted on an Apache Spark cluster with 32 cores and 320 GB total memory. We implemented wDTW, wTED, WMD, VSM, PV-DTW and PV-TED in Java Spark. The paragraph vectors for PV-DTW and PV-TED were trained by gensim.
## Data Sets
In this section, we introduce the two data sets we used for our revision detection experiments: Wikipedia revision dumps and a document revision data set generated by a computer simulation. The two data sets differ in that the Wikipedia revision dumps only contain linear revision chains, while the simulated data sets also contains tree-structured revision chains, which can be very common in real-world data.
The Wikipedia revision dumps that were previously introduced by Leskovec et al. leskovec2010governance contain eight GB (compressed size) revision edits with meta data.
We pre-processed the Wikipedia revision dumps using the JWPL Revision Machine BIBREF32 and produced a data set that contains 62,234 documents with 46,354 revisions. As we noticed that short documents just contributed to noise (graffiti) in the data, we eliminated documents that have fewer than three paragraphs and fewer than 300 words. We removed empty lines in the documents and trained word2vec embeddings on the entire corpus. We used the documents occurring in the first INLINEFORM0 of the revision period for INLINEFORM1 calibration, and the remaining documents for test.
The generation process of the simulated data sets is designed to mimic the real world. Users open some existing documents in a file system, make some changes (e.g. addition, deletion or replacement), and save them as separate documents. These documents become revisions of the original documents. We started from an initial corpus that did not have revisions, and kept adding new documents and revising existing documents. Similar to a file system, at any moment new documents could be added and/or some of the current documents could be revised. The revision operations we used were deletion, addition and replacement of words, sentences, paragraphs, section names and document titles. The addition of words, ..., section names, and new documents were pulled from the Wikipedia abstracts. This corpus generation process had five time periods INLINEFORM0 . Figure FIGREF42 illustrates this simulation. We set a Poisson distribution with rate INLINEFORM1 (the number of documents in the initial corpus) to control the number of new documents added in each time period, and a Poisson distribution with rate INLINEFORM2 to control the number of documents revised in each time period.
We generated six data sets using different random seeds, and each data set contained six corpora (Corpus 0 - 5). Table TABREF48 summarizes the first data set. In each data set, we name the initial corpus Corpus 0, and define INLINEFORM0 as the timestamp when we started this simulation process. We set INLINEFORM1 , INLINEFORM2 . Corpus INLINEFORM3 corresponds to documents generated before timestamp INLINEFORM4 . We extracted document revisions from Corpus INLINEFORM5 and compared the revisions generated in (Corpus INLINEFORM6 - Corpus INLINEFORM7 ) with the ground truths in Table TABREF48 . Hence, we ran four experiments on this data set in total. In every experiment, INLINEFORM8 is calibrated based on Corpus INLINEFORM9 . For instance, the training set of the first experiment was Corpus 1. We trained INLINEFORM10 from Corpus 1. We extracted all revisions in Corpus 2, and compared revisions generated in the test set (Corpus 2 - Corpus 1) with the ground truth: 258 revised documents. The word2vec model shared in the four experiments was trained on Corpus 5.
## Results
We use precision, recall and F-measure to evaluate the detected revisions. A true positive case is a correctly identified revision. A false positive case is an incorrectly identified revision. A false negative case is a missed revision record.
We illustrate the performances of wDTW, wTED, WMD, VSM, PV-DTW and PV-TED on the Wikipedia revision dumps in Figure FIGREF43 . wDTW and wTED have the highest F-measure scores compared to the rest of four measures, and wDTW also have the highest precision and recall scores. Figure FIGREF49 shows the average evaluation results on the simulated data sets. From left to right, the corpus size increases and the revision chains become longer, thus it becomes more challenging to detect document revisions. Overall, wDTW consistently performs the best. WMD is slightly better than wTED. In particular, when the corpus size increases, the performances of WMD, VSM, PV-DTW and PV-TED drop faster than wDTW and wTED. Because the revision operations were randomly selected in each corpus, it is possible that there are non-monotone points in the series.
wDTW and wTED perform better than WMD especially when the corpus is large, because they use dynamic programming to find the global optimal alignment for documents. In contrast, WMD relies on a greedy algorithm that sums up the minimal cost for every word. wDTW and wTED perform better than PV-DTW and PV-TED, which indicates that our DistPara distance in Algorithm SECREF20 is more accurate than the Euclidian distance between paragraph vectors trained by PV.
We show in Table TABREF53 the average running time of the six distance/similarity measures. In all the experiments, VSM is the fastest, wTED is faster than wDTW, and WMD is the slowest. Running WMD is extremely expensive because WMD needs to solve an INLINEFORM0 sequential transshipment problem for every two documents where INLINEFORM1 is the average number of words in a document. In contrast, by splitting this heavy computation into several smaller problems (finding the distance between any two paragraphs), which can be run in parallel, wDTW and wTED scale much better.
Combining Figure FIGREF43 , Figure FIGREF49 and Table TABREF53 we conclude that wDTW yields the most accurate results using marginally more time than VSM, PV-TED and PV-DTW, but much less running time than WMD. wTED returns satisfactory results using shorter time than wDTW.
## Conclusion
This paper has explored how DTW and TED can be extended with word2vec to construct semantic document distance measures: wDTW and wTED. By representing paragraphs with concatenations of word vectors, wDTW and wTED are able to capture the semantics of the words and thus give more accurate distance scores. In order to detect revisions, we have used minimum branching on an appropriately developed network with document distance scores serving as arc weights. We have also assessed the efficiency of the method of retrieving an optimal revision subnetwork by finding the minimum branching.
Furthermore, we have compared wDTW and wTED with several distance measures for revision detection tasks. Our results demonstrate the effectiveness and robustness of wDTW and wTED in the Wikipedia revision dumps and our simulated data sets. In order to reduce the computation time, we have computed document distances at the paragraph level and implemented a boosting learning system using Apache Spark. Although we have demonstrated the superiority of our semantic measures only in the revision detection experiments, wDTW and wTED can also be used as semantic distance measures in many clustering, classification tasks.
Our revision detection system can be enhanced with richer features such as author information and writing styles, and exact changes in revision pairs. Another interesting aspect we would like to explore in the future is reducing the complexities of calculating the distance between two paragraphs.
## Acknowledgments
This work was supported in part by Intel Corporation, Semiconductor Research Corporation (SRC).
| 15
|
1709.02271
|
Leveraging Discourse Information Effectively for Authorship Attribution
|
# Leveraging Discourse Information Effectively for Authorship Attribution
## Abstract
We explore techniques to maximize the effectiveness of discourse information in the task of authorship attribution. We present a novel method to embed discourse features in a Convolutional Neural Network text classifier, which achieves a state-of-the-art result by a substantial margin. We empirically investigate several featurization methods to understand the conditions under which discourse features contribute non-trivial performance gains, and analyze discourse embeddings.
## Introduction
Authorship attribution (AA) is the task of identifying the author of a text, given a set of author-labeled training texts. This task typically makes use of stylometric cues at the surface lexical and syntactic level BIBREF0 , although BIBREF1 and BIBREF2 go beyond the sentence level, showing that discourse information can help. However, they achieve limited performance gains and lack an in-depth analysis of discourse featurization techniques. More recently, convolutional neural networks (CNNs) have demonstrated considerable success on AA relying only on character-level INLINEFORM0 -grams BIBREF3 , BIBREF4 . The strength of these models is evidenced by findings that traditional stylometric features such as word INLINEFORM1 -grams and POS-tags do not improve, and can sometimes even hurt performance BIBREF3 , BIBREF5 . However, none of these CNN models make use of discourse.
Our work builds upon these prior studies by exploring an effective method to (i) featurize the discourse information, and (ii) integrate discourse features into the best text classifier (i.e., CNN-based models), in the expectation of achieving state-of-the-art results in AA.
BIBREF1 (henceforth F&H14) made the first comprehensive attempt at using discourse information for AA. They employ an entity-grid model, an approach introduced by BIBREF6 for the task of ordering sentences. This model tracks how the grammatical relations of salient entities (e.g., subj, obj, etc.) change between pairs of sentences in a document, thus capturing a form of discourse coherence. The grid is summarized into a vector of transition probabilities. However, because the model only records the transition between two consecutive sentences at a time, the coherence is local. BIBREF2 (henceforth F15) further extends the entity-grid model by replacing grammatical relations with discourse relations from Rhetorical Structure Theory BIBREF7 . Their study uses a linear-kernel SVM to perform pairwise author classifications, where a non-discourse model captures lexical and syntactic features. They find that adding the entity-grid with grammatical relations enhances the non-discourse model by almost 1% in accuracy, and using RST relations provides an improvement of 3%. The study, however, works with only one small dataset and their models produce overall unremarkable performance ( INLINEFORM0 85%). BIBREF8 propose an advanced Recursive Neural Network (RecNN) architecture to work with RST in the more general area of text categorization and present impressive results. However, we suspect that the massive number of parameters of RecNNs would likely cause overfitting when working with smaller datasets, as is often the case in AA tasks.
In our paper, we opt for a state-of-the-art character bigram CNN classifier BIBREF4 , and investigate various ways in which the discourse information can be featurized and integrated into the CNN. Specifically,
We explore these questions using two approaches to represent salient entities: grammatical relations, and RST discourse relations. We apply these models to datasets of varying sizes and genres, and find that adding any discourse information improves AA consistently on longer documents, but has mixed results on shorter documents. Further, embedding the discourse features in a parallel CNN at the input end yields better performance than concatenating them to the output layer as a feature vector (Section SECREF3 ). The global featurization is more effective than the local one. We also show that SVMs, which can only use discourse probability vectors, neither produce a competitive performance (even with fine-tuning), nor generalize in using the discourse information effectively.
## Background
Entity-grid model. Typical lexical features for AA are relatively superficial and restricted to within the same sentence. F&H14 hypothesize that discourse features beyond the sentence level also help authorship attribution. In particular, they propose an author has a particular style for representing entities across a discourse. Their work is based on the entity-grid model of BIBREF6 (henceforth B&L).
The entity-grid model tracks the grammatical relation (subj, obj, etc.) that salient entities take on throughout a document as a way to capture local coherence . A salient entity is defined as a noun phrase that co-occurs at least twice in a document. Extensive literature has shown that subject and object relations are a strong signal for salience and it follows from the Centering Theory that you want to avoid rough shifts in the center BIBREF9 , BIBREF10 . B&L thus focus on whether a salient entity is a subject (s), object (o), other (x), or is not present (-) in a given sentence, as illustrated in Table TABREF1 . Every sentence in a document is encoded with the grammatical relation of all the salient entities, resulting in a grid similar to Table TABREF6 .
The local coherence of a document is then defined on the basis of local entity transitions. A local entity transition is the sequence of grammatical relations that an entity can assume across INLINEFORM0 consecutive sentences, resulting in {s,o,x,-} INLINEFORM1 possible transitions. Following B&L, F&H14 consider sequences of length INLINEFORM2 =2, that is, transitions between two consecutive sentences, resulting in INLINEFORM3 =16 possible transitions. The probability for each transition is then calculated as the frequency of the transition divided by the total number of transitions. This step results in a single probability vector for every document, as illustrated in Table TABREF2 .
B&L apply this model to a sentence ordering task, where the more coherent option, as evidenced by its transition probabilities, was chosen. In authorship attribution, texts are however assumed to already be coherent. F&H14 instead hypothesize that an author unconsciously employs the same methods for describing entities as the discourse unfolds, resulting in discernible transition probability patterns across multiple of their texts. Indeed, F&H14 find that adding the B&L vectors increases the accuracy of AA by almost 1% over a baseline lexico-syntactic model.
RST discourse relations. F15 extends the notion of tracking salient entities to RST. Instead of using grammatical relations in the grid, RST discourse relations are specified. An RST discourse relation defines the relationship between two or more elementary discourse units (EDUs), which are spans of text that typically correspond to syntactic clauses. In a relation, an EDU can function as a nucleus (e.g., result.N) or as a satellite (e.g., summary.S). All the relations in a document then form a tree as in Figure FIGREF8 .
F15 finds that RST relations are more effective for AA than grammatical relations. In our paper, we populate the entity-grid in the same way as F15's “Shallow RST-style” encoding, but use fine-grained instead of coarse-grained RST relations, and do not distinguish between intra-sentential and multi-sentential RST relations, or salient and non-salient entities. We explore various featurization techniques using the coding scheme.
CNN model. shrestha2017 propose a convolutional neural network formulation for AA tasks (detailed in Section SECREF3 ). They report state-of-the-art performance on a corpus of Twitter data BIBREF11 , and compare their models with alternative architectures proposed in the literature: (i) SCH: an SVM that also uses character n-grams, among other stylometric features BIBREF11 ; (ii) LSTM-2: an LSTM trained on bigrams BIBREF12 ; (iii) CHAR: a Logistic Regression model that takes character n-grams BIBREF13 ; (iv) CNN-W: a CNN trained on word embeddings BIBREF14 . The authors show that the model CNN2 produces the best performance overall. Ruder:16 apply character INLINEFORM0 -gram CNNs to a wide range of datasets, providing strong empirical evidence that the architecture generalizes well. Further, they find that including word INLINEFORM1 -grams in addition to character INLINEFORM2 -grams reduces performance, which is in agreement with BIBREF5 's findings.
## Models
Building on shrestha2017's work, we employ their character-bigram CNN (CNN2), and propose two extensions which utilize discourse information: (i) CNN2 enhanced with relation probability vectors (CNN2-PV), and (ii) CNN2 enhanced with discourse embeddings (CNN2-DE). The CNN2-PV allows us to conduct a comparison with F&H14 and F15, which also use relation probability vectors.
CNN2. CNN2 is the baseline model with no discourse features. Illustrated in Figure FIGREF10 (center), it consists of (i) an embedding layer, (ii) a convolution layer, (iii) a max-pooling layer, and (iv) a softmax layer. We briefly sketch the processing procedure and refer the reader to BIBREF4 for mathematical details.
The network takes a sequence of character bigrams INLINEFORM0 as input, and outputs a multinomial INLINEFORM1 over class labels as the prediction. The model first looks up the embedding matrix to produce a sequence of embeddings for INLINEFORM2 (i.e., the matrix INLINEFORM3 ), then pushes the embedding sequence through convolutional filters of three bigram-window sizes INLINEFORM4 , each yielding INLINEFORM5 feature maps. We then apply the max-over-time pooling BIBREF15 to the feature maps from each filter, and concatenate the resulting vectors to obtain a single vector INLINEFORM6 , which then goes through the softmax layer to produce predictions.
CNN2-PV. This model (Figure FIGREF10 , left+center) featurizes discourse information into a vector of relation probabilities. In order to derive the discourse features, an entity grid is constructed by feeding the document through an NLP pipeline to identify salient entities. Two flavors of discourse features are created by populating the entity grid with either (i) grammatical relations (GR) or (ii) RST discourse relations (RST). The GR features are represented as grammatical relation transitions derived from the entity grid, e.g., INLINEFORM0 . The RST features are represented as RST discourse relations with their nuclearity, e.g., INLINEFORM1 . The probability vectors are then distributions over relation types. For GR, the vector is a distribution over all the entity role transitions, i.e., INLINEFORM2 (see Table TABREF2 ). For RST, the vector is a distribution over all the RST discourse relations, i.e., INLINEFORM3 Denoting a feature as such with INLINEFORM4 , we construct the pooling vector INLINEFORM5 for the char-bigrams, and concatenate INLINEFORM6 to INLINEFORM7 before feeding the resulting vector to the softmax layer.
CNN2-DE. In this model (Figure FIGREF10 , center+right), we embed discourse features in high-dimensional space (similar to char-bigram embeddings). Let INLINEFORM0 be a sequence of discourse features, we treat it in a similar fashion to the char-bigram sequence INLINEFORM1 , i.e. feeding it through a “parallel” convolutional net (Figure FIGREF10 right). The operation results in a pooling vector INLINEFORM2 . We concatenate INLINEFORM3 to the pooling vector INLINEFORM4 (which is constructed from INLINEFORM5 ) then feed INLINEFORM6 to the softmax layer for the final prediction.
## Experiments and Results
We begin by introducing the datasets (Section SECREF15 ), followed by detailing the featurization methods (Section SECREF17 ), the experiments (Section SECREF22 ), and finally reporting results (Section SECREF26 ).
## Datasets
The statistics for the three datasets used in the experiments are summarized in Table TABREF16 .
novel-9. This dataset was compiled by F&H14: a collection of 19 novels by 9 nineteenth century British and American authors in the Project Gutenberg. To compare to F&H14, we apply the same resampling method (F&H14, Section 4.2) to correct the imbalance in authors by oversampling the texts of less-represented authors.
novel-50. This dataset extends novel-9, compiling the works of 50 randomly selected authors of the same period. For each author, we randomly select 5 novels for a total 250 novels.
IMDB62. IMDB62 consists of 62K movie reviews from 62 users (1,000 each) from the Internet Movie dataset, compiled by Seroussi:11. Unlike the novel datasets, the reviews are considerably shorter, with a mean of 349 words per text.
## Featurization
As described in Section SECREF2 , in both the GR and RST variants, from each input entry we start by obtaining an entity grid.
CNN2-PV. We collect the probabilities of entity role transitions (in GR) or discourse relations (in RST) for the entries. Each entry corresponds to a probability distribution vector.
CNN2-DE. We employ two schema for creating discourse feature sequences from an entity grid. While we always read the grid by column (by a salient entity), we vary whether we track the entity across a number of sentences (n rows at a time) or across the entire document (one entire column at a time), denoted as local and global reading respectively.
For the GR discourse features, in the case of local reading, we process the entity roles one sentence pair at a time (Figure FIGREF18 , left). For example, in processing the pair INLINEFORM0 , we find the first non-empty role INLINEFORM1 for entity INLINEFORM2 in INLINEFORM3 . If INLINEFORM4 also has a non-empty role INLINEFORM5 in the INLINEFORM6 , we collect the entity role transition INLINEFORM7 . We then proceed to the following entity INLINEFORM8 , until we process all the entities in the grid and move to the next sentence pair. For the global reading, we instead read the entity roles by traversing one column of the entire document at a time (Figure FIGREF18 , right). The entity roles in all the sentences are read for one entity: we collect transitions for all the non-empty roles (e.g., INLINEFORM9 , but not INLINEFORM10 ).
For the RST discourse features, we process non-empty discourse relations also through either local or global reading. In the local reading, we read all the discourse relations in a sentence (a row) then move on to the next sentence. In the global reading, we read in discourse relations for one entity at a time. This results in sequences of discourse relations for the input entries.
## Experiments
Baseline-dataset experiments. All the baseline-dataset experiments are evaluated on novel-9. As a comparison to previous work (F15), we evaluate our models using a pairwise classification task with GR discourse features. In her model, novels are partitioned into 1000-word chunks, and the model is evaluated with accuracy. Surpassing F15's SVM model by a large margin, we then further evaluate the more difficult multi-class task, i.e., all-class prediction simultaneously, with both GR and RST discourse features and the more robust F1 evaluation. In this multi-class task, we implement two SVMs to extend F15's SVM models: (i) SVM2: a linear-kernel SVM which takes char-bigrams as input, as our CNNs, and (ii) SVM2-PV: an updated SVM2 which takes also probability vector features.
Further, we are interested in finding a performance threshold on the minimally-required input text length for discourse information to “kick in”. To this end, we chunk the novels into different sizes: 200-2000 words, at 200-word intervals, and evaluate our CNNs in the multi-class condition.
Generalization-dataset experiments. To confirm that our models generalize, we pick the best models from the baseline-dataset experiments and evaluate on the novel-50 and IMDB62 datasets. For novel-50, the chunking size applied is 2000-word as per the baseline-dataset experiment results, and for IMDB62, texts are not chunked (i.e., we feed the models with the original reviews directly). For model comparison, we also run the SVMs (i.e., SVM2 and SVM2-PV) used in the baseline-dataset experiment. All the experiments conducted here are multi-class classification with macro-averaged F1 evaluation.
Model configurations. Following F15, we perform 5-fold cross-validation. The embedding sizes are tuned on novel-9 (multi-class condition): 50 for char-bigrams; 20 for discourse features. The learning rate is 0.001 using the Adam Optimizer BIBREF18 . For all models, we apply dropout regularization of 0.75 BIBREF19 , and run 50 epochs (batch size 32). The SVMs in the baseline-dataset experiments use default settings, following F15. For the SVMs in the generalization-dataset experiments, we tuned the hyperparameters on novel-9 with a grid search, and found the optimal setting as: stopping condition tol is 1e-5, at a max-iteration of 1,500.
## Results
Baseline-dataset experiments. The results of the baseline-dataset experiments are reported in Table TABREF24 , TABREF25 and Figure FIGREF27 . In Table TABREF24 , Baseline denotes the dumb baseline model which always predicts the more-represented author of the pair. Both SVMs are from F15, and we report her results. SVM (LexSyn) takes character and word bi/trigrams and POS tags. SVM (LexSyn-PV) additionally includes probability vectors, similar to our CNN2-PV. In this part of the experiment, while the CNNs clear a large margin over SVMs, adding discourse in CNN2-PV brings only a small performance gain.
Table TABREF25 reports the results from the multi-class classification task, the more difficult task. Here, probability vector features (i.e., PV) again fail to contribute much. The discourse embedding features, on the other hand, manage to increase the F1 score by a noticeable amount, with the maximal improvement seen in the CNN2-DE (global) model with RST features (by 2.6 points). In contrast, the discourse-enhanced SVM2-PVs increase F1 by about 1 point, with overall much lower scores in comparison to the CNNs. In general, RST features work better than GR features.
The results of the varying-sizes experiments are plotted in Figure FIGREF27 . Again, we observe the overall pattern that discourse features improve the F1 score, and RST features procure superior performance. Crucially, however, we note there is no performance boost below the chunk size of 1000 for GR features, and below 600 for RST features. Where discourse features do help, the GR-based models achieve, on average, 1 extra point on F1, and the RST-based models around 2.
Generalization-dataset experiments. Table TABREF28 summarizes the results of the generalization-dataset experiments. On novel-50, most discourse-enhanced models improve the performance of the baseline non-discourse CNN2 to varying degrees. The clear pattern again emerges that RST features work better, with the best F1 score evidenced in the CNN2-DE (global) model (3.5 improvement in F1). On IMDB62, as expected with short text inputs (mean=349 words/review), the discourse features in general do not add further contribution. Even the best model CNN2-DE brings only marginal improvement, confirming our findings from varying the chunk size on novel-9, where discourse features did not help at this input size. Equipped with discourse features, SVM2-PV performs slightly better than SVM2 on novel-50 (by 0.4 with GR, 0.9 with RST features). On IMDB62, the same pattern persists for the SVMs: discourse features do not make noticeable improvements (by 0.0 and 0.5 with GR and RST respectively).
## Analysis
General analysis. Overall, we have shown that discourse information can improve authorship attribution, but only when properly encoded. This result is critical in demonstrating the particular value of discourse information, because typical stylometric features such as word INLINEFORM0 -grams and POS tags do not add additional performance improvements BIBREF3 , BIBREF5 .
In addition, the type of discourse information and the way in which it is featurized are tantamount to this performance improvement: RST features provide overall stronger improvement, and the global reading scheme for discourse embedding works better than the local one. The discourse embedding proves to be a superior featurization technique, as evidenced by the generally higher performance of CNN2-DE models over CNN2-PV models. With an SVM, where the option is not available, we are only able to use relation probability vectors to obtain a very modest performance improvement.
Further, we found an input-length threshold for the discourse features to help (Section SECREF26 ). Not surprisingly, discourse does not contribute on shorter texts. Many of the feature grids are empty for these shorter texts– either there are no coreference chains or they are not correctly resolved. Currently we only have empirical results on short novel chunks and movie reviews, but believe the finding would generalize to Twitter or blog posts.
Discourse embeddings. It does not come as a surprise that discourse embedding-based models perform better than their relation probability-based peers. The former (i) leverages the weight learning of the entire computational graph of the CNN rather than only the softmax layer, as the PV models do, and (ii) provides a more fine-grained featurization of the discourse information. Rather than merely taking a probability over grammatical relation transitions (in GR) or discourse relation types (in RST), in DE-based models we learn the dependency between grammatical relation transitions/discourse relations through the INLINEFORM0 -sized filter sweeps.
To further study the information encoded in the discourse embeddings, we perform t-SNE clustering BIBREF20 on them, using the best performing model CNN2-DE (global). We examine the closest neighbors of each embedding, and observe that similar discourse relations tend to go together (e.g., explanation and interpretation; consequence and result). Some examples are given in Table TABREF29 . However, it is unclear how this pattern helps improve classification performance. We intend to investigate this question in future work.
Global vs. Local featurization. As described in Section SECREF17 , the global reading processes all the discourse features for one entity at a time, while the local approach reads one sentence (or one sentence pair) at a time. In all the relevant experiments, global featurization showed a clear performance advantage (on average 1 point gain in F1). Recall that the creation of the grids (both GR and RST) depend on coreference chains of entities (Section SECREF2 ), and only the global reading scheme takes advantage of the coreference pattern whereas the local reading breaks the chains. To find out whether coreference pattern is the key to the performance difference, we further ran a probe experiment where we read RST discourse relations in the order in which EDUs are arranged in the RST tree (i.e., left-to-right), and evaluated this model on novel-50 and IMDB62 with the same hyperparameter setting. The F1 scores turned out to be very close to the CNN2-DE (local) model, at 97.5 and 90.9. Based on this finding, we tentatively confirm the importance of the coreference pattern, and intend to further investigate how exactly it matters for the classification performance.
GR vs. RST. RST features in general effect higher performance gains than GR features (Table TABREF28 ). The RST parser produces a tree of discourse relations for the input text, thus introducing a “global view.” The GR features, on the other hand, are more restricted to a “local view” on entities between consecutive sentences. While a deeper empirical investigation is needed, one can intuitively imagine that identifying authorship by focusing on the local transitions between grammatical relations (as in GR) is more difficult than observing how the entire text is organized (as in RST).
## Conclusion
We have conducted an in-depth investigation of techniques that (i) featurize discourse information, and (ii) effectively integrate discourse features into the state-of-the-art character-bigram CNN classifier for AA. Beyond confirming the overall superiority of RST features over GR features in larger and more difficult datasets, we present a discourse embedding technique that is unavailable for previously proposed discourse-enhanced models. The new technique enabled us to push the envelope of the current performance ceiling by a large margin.
Admittedly, in using the RST features with entity-grids, we lose the valuable RST tree structure. In future work, we intend to adopt more sophisticated methods such as RecNN, as per Ji:17, to retain more information from the RST trees while reducing the parameter size. Further, we aim to understand how discourse embeddings contribute to AA tasks, and find alternatives to coreference chains for shorter texts.
| 10
|
1709.05413
|
"How May I Help You?": Modeling Twitter Customer Service Conversations Using Fine-Grained Dialogue Acts
|
# "How May I Help You?": Modeling Twitter Customer Service Conversations Using Fine-Grained Dialogue Acts
## Abstract
Given the increasing popularity of customer service dialogue on Twitter, analysis of conversation data is essential to understand trends in customer and agent behavior for the purpose of automating customer service interactions. In this work, we develop a novel taxonomy of fine-grained"dialogue acts"frequently observed in customer service, showcasing acts that are more suited to the domain than the more generic existing taxonomies. Using a sequential SVM-HMM model, we model conversation flow, predicting the dialogue act of a given turn in real-time. We characterize differences between customer and agent behavior in Twitter customer service conversations, and investigate the effect of testing our system on different customer service industries. Finally, we use a data-driven approach to predict important conversation outcomes: customer satisfaction, customer frustration, and overall problem resolution. We show that the type and location of certain dialogue acts in a conversation have a significant effect on the probability of desirable and undesirable outcomes, and present actionable rules based on our findings. The patterns and rules we derive can be used as guidelines for outcome-driven automated customer service platforms.
## Introduction
The need for real-time, efficient, and reliable customer service has grown in recent years. Twitter has emerged as a popular medium for customer service dialogue, allowing customers to make inquiries and receive instant live support in the public domain. In order to provide useful information to customers, agents must first understand the requirements of the conversation, and offer customers the appropriate feedback. While this may be feasible at the level of a single conversation for a human agent, automatic analysis of conversations is essential for data-driven approaches towards the design of automated customer support agents and systems.
Analyzing the dialogic structure of a conversation in terms of the "dialogue acts" used, such as statements or questions, can give important meta-information about conversation flow and content, and can be used as a first step to developing automated agents. Traditional dialogue act taxonomies used to label turns in a conversation are very generic, in order to allow for broad coverage of the majority of dialogue acts possible in a conversation BIBREF0 , BIBREF1 , BIBREF2 . However, for the purpose of understanding and analyzing customer service conversations, generic taxonomies fall short. Table TABREF1 shows a sample customer service conversation between a human agent and customer on Twitter, where the customer and agent take alternating "turns" to discuss the problem. As shown from the dialogue acts used at each turn, simply knowing that a turn is a Statement or Request, as is possible with generic taxonomies, is not enough information to allow for automated handling or response to a problem. We need more fine-grained dialogue acts, such as Informative Statement, Complaint, or Request for Information to capture the speaker's intent, and act accordingly. Likewise, turns often include multiple overlapping dialogue acts, such that a multi-label approach to classification is often more informative than a single-label approach.
Dialogue act prediction can be used to guide automatic response generation, and to develop diagnostic tools for the fine-tuning of automatic agents. For example, in Table TABREF1 , the customer's first turn (Turn 1) is categorized as a Complaint, Negative Expressive Statement, and Sarcasm, and the agent's response (Turn 2) is tagged as a Request for Information, Yes-No Question, and Apology. Prediction of these dialogue acts in a real-time setting can be leveraged to generate appropriate automated agent responses to similar situations.
Additionally, important patterns can emerge from analysis of the fine-grained acts in a dialogue in a post-prediction setting. For example, if an agent does not follow-up with certain actions in response to a customer's question dialogue act, this could be found to be a violation of a best practice pattern. By analyzing large numbers of dialogue act sequences correlated with specific outcomes, various rules can be derived, i.e. "Continuing to request information late in a conversation often leads to customer dissatisfaction." This can then be codified into a best practice pattern rules for automated systems, such as "A request for information act should be issued early in a conversation, followed by an answer, informative statement, or apology towards the end of the conversation."
In this work, we are motivated to predict the dialogue acts in conversations with the intent of identifying problem spots that can be addressed in real-time, and to allow for post-conversation analysis to derive rules about conversation outcomes indicating successful/unsuccessful interactions, namely, customer satisfaction, customer frustration, and problem resolution. We focus on analysis of the dialogue acts used in customer service conversations as a first step to fully automating the interaction. We address various different challenges: dialogue act annotated data is not available for customer service on Twitter, the task of dialogue act annotation is subjective, existing taxonomies do not capture the fine-grained information we believe is valuable to our task, and tweets, although concise in nature, often consist of overlapping dialogue acts to characterize their full intent. The novelty of our work comes from the development of our fine-grained dialogue act taxonomy and multi-label approach for act prediction, as well as our analysis of the customer service domain on Twitter. Our goal is to offer useful analytics to improve outcome-oriented conversational systems.
We first expand upon previous work and generic dialogue act taxonomies, developing a fine-grained set of dialogue acts for customer service, and conducting a systematic user study to identify these acts in a dataset of 800 conversations from four Twitter customer service accounts (i.e. four different companies in the telecommunication, electronics, and insurance industries). We then aim to understand the conversation flow between customers and agents using our taxonomy, so we develop a real-time sequential SVM-HMM model to predict our fine-grained dialogue acts while a conversation is in progress, using a novel multi-label scheme to classify each turn. Finally, using our dialogue act predictions, we classify conversations based on the outcomes of customer satisfaction, frustration, and overall problem resolution, then provide actionable guidelines for the development of automated customer service systems and intelligent agents aimed at desired customer outcomes BIBREF3 , BIBREF4 .
We begin with a discussion of related work, followed by an overview of our methodology. Next, we describe our conversation modeling framework, and explain our outcome analysis experiments, to show how we derive useful patterns for designing automated customer service agents. Finally, we present conclusions and directions for future work.
## Related Work
Developing computational speech and dialogue act models has long been a topic of interest BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , with researchers from many different backgrounds studying human conversations and developing theories around conversational analysis and interpretation on intent. Modern intelligent conversational BIBREF3 , BIBREF4 and dialogue systems draw principles from many disciplines, including philosophy, linguistics, computer science, and sociology. In this section, we describe relevant previous work on speech and dialogue act modeling, general conversation modeling on Twitter, and speech and dialogue act modeling of customer service in other data sources.
Previous work has explored speech act modeling in different domains (as a predecessor to dialogue act modeling). Zhang et al. present work on recognition of speech acts on Twitter, following up with a study on scalable speech act recognition given the difficulty of obtaining labeled training data BIBREF9 . They use a simple taxonomy of four main speech acts (Statement, Question, Suggestion, Comment, and a Miscellaneous category). More recently, Vosoughi et al. develop BIBREF10 a speech act classifier for Twitter, using a modification of the taxonomy defined by Searle in 1975, including six acts they observe to commonly occur on Twitter: Assertion, Recommendation Expression, Question, Request, again plus a Miscellaneous category. They describe good features for speech act classification and the application of such a system to detect stories on social media BIBREF11 . In this work, we are interested in the dialogic characteristics of Twitter conversations, rather than speech acts in stand-alone tweets.
Different dialogue act taxonomies have been developed to characterize conversational acts. Core and Allen present the Dialogue Act Marking in Several Layers (DAMSL), a standard for discourse annotation that was developed in 1997 BIBREF0 . The taxonomy contains a total of 220 tags, divided into four main categories: communicative status, information level, forward-looking function, and backward-looking function. Jurafsky, Shriberg, and Biasca develop a less fine-grained taxonomy of 42 tags based on DAMSL BIBREF1 . Stolcke et al. employ a similar set for general conversation BIBREF2 , citing that "content- and task-related distinctions will always play an important role in effective DA [Dialogue Act] labeling." Many researchers have tackled the task of developing different speech and dialogue act taxonomies and coding schemes BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 . For the purposes of our own research, we require a set of dialogue acts that is more closely representative of customer service domain interactions - thus we expand upon previously defined taxonomies and develop a more fine-grained set.
Modeling general conversation on Twitter has also been a topic of interest in previous work. Honeycutt and Herring study conversation and collaboration on Twitter using individual tweets containing "@" mentions BIBREF16 . Ritter et al. explore unsupervised modeling of Twitter conversations, using clustering methods on a corpus of 1.3 million Twitter conversations to define a model of transitional flow between in a general Twitter dialogue BIBREF17 . While these approaches are relevant to understanding the nature of interactions on Twitter, we find that the customer service domain presents its own interesting characteristics that are worth exploring further.
The most related previous work has explored speech and dialogue act modeling in customer service, however, no previous work has focused on Twitter as a data source. In 2005, Ivanovic uses an abridged set of 12 course-grained dialogue acts (detailed in the Taxonomy section) to describe interactions between customers and agents in instant messaging chats BIBREF18 , BIBREF19 , leading to a proposal on response suggestion using the proposed dialogue acts BIBREF20 . Follow-up work using the taxonomy selected by Ivanovic comes from Kim et al., where they focus on classifying dialogue acts in both one-on-one and multi-party live instant messaging chats BIBREF21 , BIBREF22 . These works are similar to ours in the nature of the problem addressed, but we use a much more fine-grained taxonomy to define the interactions possible in the customer service domain, and focus on Twitter conversations, which are unique in their brevity and the nature of the public interactions.
The most similar work to our own is that of Herzig et al. on classifying emotions in customer support dialogues on Twitter BIBREF23 . They explore how agent responses should be tailored to the detected emotional response in customers, in order to improve the quality of service agents can provide. Rather than focusing on emotional response, we seek to model the dialogic structure and intents of the speakers using dialogue acts, with emotion included as features in our model, to characterize the emotional intent within each act.
## Methodology
The underlying goal of this work is to show how a well-defined taxonomy of dialogue acts can be used to summarize semantic information in real-time about the flow of a conversation to derive meaningful insights into the success/failure of the interaction, and then to develop actionable rules to be used in automating customer service interactions. We focus on the customer service domain on Twitter, which has not previously been explored in the context of dialogue act classification. In this new domain, we can provide meaningful recommendations about good communicative practices, based on real data. Our methodology pipeline is shown in Figure FIGREF2 .
## Taxonomy Definition
As described in the related work, the taxonomy of 12 acts to classify dialogue acts in an instant-messaging scenario, developed by Ivanovic in 2005, has been used by previous work when approaching the task of dialogue act classification for customer service BIBREF18 , BIBREF20 , BIBREF19 , BIBREF21 , BIBREF22 . The dataset used consisted of eight conversations from chat logs in the MSN Shopping Service (around 550 turns spanning around 4,500 words) BIBREF19 . The conversations were gathered by asking five volunteers to use the platform to inquire for help regarding various hypothetical situations (i.e. buying an item for someone) BIBREF19 . The process of selection of tags to develop the taxonomy, beginning with the 42 tags from the DAMSL set BIBREF0 , involved removing tags inappropriate for written text, and collapsing sets of tags into a more coarse-grained label BIBREF18 . The final taxonomy consists of the following 12 dialogue acts (sorted by frequency in the dataset): Statement (36%), Thanking (14.7%), Yes-No Question (13.9%), Response-Acknowledgement (7.2%), Request (5.9%), Open-Question (5.3%), Yes-Answer (5.1%), Conventional-Closing (2.9%), No-Answer (2.5%), Conventional-Opening (2.3%), Expressive (2.3%) and Downplayer (1.9%).
For the purposes of our own research, focused on customer service on Twitter, we found that the course-grained nature of the taxonomy presented a natural shortcoming in terms of what information could be learned by performing classification at this level. We observe that while having a smaller set of dialogue acts may be helpful for achieving good agreement between annotators (Ivanovic cites kappas of 0.87 between the three expert annotators using this tag set on his data BIBREF18 ), it is unable to offer deeper semantic insight into the specific intent behind each act for many of the categories. For example, the Statement act, which comprises the largest percentage (36% of turns), is an extremely broad category that fails to provide useful information from an analytical perspective. Likewise, the Request category also does not specify any intent behind the act, and leaves much room for improvement.
For this reason, and motivated by previous work seeking to develop dialogue act taxonomies appropriate for different domains BIBREF19 , BIBREF21 , we convert the list of dialogue acts presented by the literature into a hierarchical taxonomy, shown in Figure FIGREF6 .
We first organize the taxonomy into six high-level dialogue acts: Greeting, Statement, Request, Question, Answer, and Social Act. Then, we update the taxonomy using two main steps: restructuring and adding additional fine-grained acts.
We base our changes upon the taxonomy used by Ivanovic and Kim et al. in their work on instant messaging chat dialogues BIBREF19 , BIBREF21 , but also on general dialogue acts observed in the customer service domain, including complaints and suggestions. Our taxonomy does not make any specific restrictions on which party in the dialogue may perform each act, but we do observe that some acts are far more frequent (and sometimes non-existent) in usage, depending on whether the customer or agent is the speaker (for example, the Statement Complaint category never shows up in Agent turns).
In order to account for gaps in available act selections for annotators, we include an Other act in the broadest categories. While our taxonomy fills in many gaps from previous work in our domain, we do not claim to have handled coverage of all possible acts in this domain. Our taxonomy allows us to more closely specify the intent and motivation behind each turn, and ultimately how to address different situations.
## Data Collection
Given our taxonomy of fine-grained dialogue acts that expands upon previous work, we set out to gather annotations for Twitter customer service conversations.
For our data collection phase, we begin with conversations from the Twitter customer service pages of four different companies, from the electronics, telecommunications, and insurance industries. We perform several forms of pre-processing to the conversations. We filter out conversations if they contain more than one customer or agent speaker, do not have alternating customer/agent speaking turns (single turn per speaker), have less than 5 or more than 10 turns, have less than 70 words in total, and if any turn in the conversation ends in an ellipses followed by a link (indicating that the turn has been cut off due to length, and spans another tweet). Additionally, we remove any references to the company names (substituting with "Agent"), any references to customer usernames (substituting with "Customer"), and replacing and links or image references with INLINEFORM0 link INLINEFORM1 and INLINEFORM2 img INLINEFORM3 tokens.
Using these filters as pre-processing methods, we end up with a set of 800 conversations, spanning 5,327 turns. We conduct our annotation study on Amazon Mechanical Turk, presenting Turkers with Human Intelligence Tasks (henceforth, HITs) consisting of a single conversation between a customer and an agent. In each HIT, we present Turkers with a definition of each dialogue act, as well as a sample annotated dialogue for reference. For each turn in the conversation, we allow Turkers to select as many labels from our taxonomy as required to fully characterize the intent of the turn. Additionally, annotators are asked three questions at the end of each conversation HIT, to which they could respond that they agreed, disagreed, or could not tell:
We ask 5 Turkers to annotate each conversation HIT, and pay $0.20 per HIT. We find the list of "majority dialogue acts" for each tweet by finding any acts that have received majority-vote labels (at least 3 out of 5 judgements).
It is important to note at this point that we make an important choice as to how we will handle dialogue act tagging for each turn. We note that each turn may contain more than one dialogue act vital to carry its full meaning. Thus, we choose not to carry out a specific segmentation task on our tweets, contrary to previous work BIBREF24 , BIBREF25 , opting to characterize each tweet as a single unit composed of different, often overlapping, dialogue acts. Table TABREF16 shows examples of tweets that receive majority vote on more than one label, where the act boundaries are overlapping and not necessarily distinguishable.
It is clear that the lines differentiating these acts are not very well defined, and that segmentation would not necessarily aid in clearly separating out each intent. For these reasons, and due to the overall brevity of tweets in general, we choose to avoid the overhead of requiring annotators to provide segment boundaries, and instead ask for all appropriate dialogue acts.
## Annotation Results
Figure FIGREF17 shows the distribution of the number of times each dialogue act in our taxonomy is selected a majority act by the annotators (recall that each turn is annotated by 5 annotators). From the distribution, we see that the largest class is Statement Info which is part of the majority vote list for 2,152 of the 5,327 total turns, followed by Request Info, which appears in 1,088 of the total turns. Although Statement Informative comprises the largest set of majority labels in the data (as did Statement in Ivanovic's distribution), we do observe that other fine-grained categories of Statement occur in the most frequent labels as well, including Statement Complaint, Statement Expressive Negative, and Statement Suggestion – giving more useful information as to what form of statement is most frequently occurring. We find that 147 tweets receive no majority label (i.e. no single act received 3 or more votes out of 5). At the tail of the distribution, we see less frequent acts, such as Statement Sarcasm, Social Act Downplayer, Statement Promise, Greeting Closing, and Request Other. It is also interesting to note that both opening and closing greetings occur infrequently in the data – which is understandable given the nature of Twitter conversation, where formal greeting is not generally required.
Table TABREF19 shows a more detailed summary of the distribution of our top 12 dialogue acts according to the annotation experiments, as presented by Ivanovic BIBREF18 . Since each turn has an overlapping set of labels, the column % of Turns (5,327) represents what fraction of the total 5,327 turns contain that dialogue act label (these values do not sum to 1, since there is overlap). To give a better sense of the percentage appearance of each dialogue act class in terms of the total number of annotated labels given, we also present column % of Annotations (10,343) (these values are percentages). We measure agreement in our annotations using a few different techniques. Since each item in our annotation experiments allows for multiple labels, we first design an agreement measure that accounts for how frequently each annotator selects the acts that agree with the majority-selected labels for the turns they annotated. To calculate this for each annotator, we find the number of majority-selected acts for each conversation they annotated (call this MAJ), and the number of subset those acts that they selected (call this SUBS), and find the ratio (SUBS/MAJ). We use this ratio to systematically fine-tune our set of annotators by running our annotation in four batches, restricting our pool of annotators to those that have above a 0.60 ratio of agreement with the majority from the previous batch, as a sort of quality assurance test. We also measure Fleiss' Kappa BIBREF26 agreement between annotators in two ways: first by normalizing our annotation results into binary-valued items indicating annotators' votes for each label contain within each turn. We find an average Fleiss- INLINEFORM0 for the full dataset, including all turn-and-label items, representing moderate agreement on the 24-label problem.
We also calculate the Fleiss- INLINEFORM0 values for each label, and use the categories defined by Landis and Koch to bin our speech acts based on agreement BIBREF27 . As shown in Table TABREF18 , we find that the per-label agreement varies from "almost perfect" agreement of INLINEFORM1 for lexically defined categories such as Apology and Thanks, with only slight agreement of INLINEFORM2 for less clearly-defined categories, such as Statement (Other), Answer Response Acknowledgement and Request (Other). For the conversation-level questions, we calculate the agreement across the "Agree" label for all annotators, finding an average Fleiss- INLINEFORM3 , with question-level results of INLINEFORM4 for customer satisfaction, INLINEFORM5 for problem resolution, and INLINEFORM6 for customer frustration. These results suggest room for improvement for further development of the taxonomy, to address problem areas for annotators and remedy areas of lower agreement.
## Motivation for Multi-Label Classification
We test our hypothesis that tweet turns are often characterized by more than one distinct dialogue act label by measuring the percentage overlap between frequent pairs of labels. Of the 5,327 turns annotated, across the 800 conversations, we find that 3,593 of those turns (67.4%) contained more than one majority-act label. Table TABREF22 shows the distribution percentage of the most frequent pairs.
For example, we observe that answering with informative statements is the most frequent pair, followed by complaints coupled with negative sentiment or informative statements. We also observe that requests are usually formed as questions, but also co-occur frequently with apologies. This experiment validates our intuition that the majority of turns do contain more than a single label, and motivates our use of a multi-label classification method for characterizing each turn in the conversation modeling experiments we present in the next section.
## Conversation Modeling
In this section, we describe the setup and results of our conversational modeling experiments on the data we collected using our fine-grained taxonomy of customer service dialogue acts. We begin with an overview of the features and classes used, followed by our experimental setup and results for each experiment performed.
## Features
The following list describes the set of features used for our dialogue act classification tasks:
Word/Punctuation: binary bag-of-word unigrams, binary existence of a question mark, binary existence of an exclamation mark in a turn
Temporal: response time of a turn (time in seconds elapsed between the posting time of the previous turn and that of the current turn)
Second-Person Reference: existence of an explicit second-person reference in the turn (you, your, you're)
Emotion: count of words in each of the 8 emotion classes from the NRC emotion lexicon BIBREF28 (anger, anticipation, disgust, fear, joy, negative, positive, sadness, surprise, and trust)
Dialogue: lexical indicators in the turn: opening greetings (hi, hello, greetings, etc), closing greetings (bye, goodbye), yes-no questions (turns with questions starting with do, did, can, could, etc), wh- questions (turns with questions starting with who, what, where, etc), thanking (thank*), apology (sorry, apolog*), yes-answer, and no-answer
## Classes
Table TABREF30 shows the division of classes we use for each of our experiments. We select our classes using the distribution of annotations we observe in our data collection phase (see Table TABREF19 ), selecting the top 12 classes as candidates.
While iteratively selecting the most frequently-occurring classes helps to ensure that classes with the most data are represented in our experiments, it also introduces the problem of including classes that are very well-defined lexically, and may not require learning for classification, such as Social Act Apology and Social Act Thanking in the first 10-Class set. For this reason, we call this set 10-Class (Easy), and also experiment using a 10-Class (Hard) set, where we add in the next two less-defined and more semantically rich labels, such as Statement Offer and Question Open. When using each set of classes, a turn is either classified as one of the classes in the set, or it is classified as "other" (i.e. any of the other classes). We discuss our experiments in more detail and comment on performance differences in the experiment section.
## Experiments
Following previous work on conversation modeling BIBREF23 , we use a sequential SVM-HMM (using the INLINEFORM0 toolkit BIBREF29 ) for our conversation modeling experiments. We hypothesize that a sequential model is most suited to our dialogic data, and that we will be able to concisely capture conversational attributes such as the order in which dialogue acts often occur (i.e. some Answer act after Question a question act, or Apology acts after Complaints).
We note that with default settings for a sequence of length INLINEFORM0 , an SVM-HMM model will be able to refine its answers for any turn INLINEFORM1 as information becomes available for turns INLINEFORM2 . However, we opt to design our classifier under a real-time setting, where turn-by-turn classification is required without future knowledge or adaptation of prediction at any given stage. In our setup, turns are predicted in a real-time setting to fairly model conversation available to an intelligent agent in a conversational system. At any point, a turn INLINEFORM3 is predicted using information from turns INLINEFORM4 , and where a prediction is not changed when new information is available.
We test our hypothesis by comparing our real-time sequential SVM-HMM model to non-sequential baselines from the NLTK BIBREF30 and Scikit-Learn BIBREF31 toolkits. We use our selected feature set (described above) to be generic enough to apply to both our sequential and non-sequential models, in order to allow us to fairly compare performance. We shuffle and divide our data into 70% for training and development (560 conversations, using 10-fold cross-validation for parameter tuning), and hold out 30% of the data (240 conversations) for test.
Motivated by the prevalent overlap of dialogue acts, we conduct our learning experiments using a multi-label setup. For each of the sets of classes, we conduct binary classification task for each label: for each INLINEFORM0 -class classification task, a turn is labeled as either belonging to the current label, or not (i.e. "other"). In this setup, each turn is assigned a binary value for each label (i.e. for the 6-class experiment, each turn receives a value of 0/1 for each indicating whether the classifier predicts it to be relevant to the each of the 6 labels). Thus, for each INLINEFORM1 -class experiment, we end up with INLINEFORM2 binary labels, for example, whether the turn is a Statement Informative or Other, Request Information or Other, etc. We aggregate the INLINEFORM3 binary predictions for each turn, then compare the resultant prediction matrix for all turns to our majority-vote ground-truth labels, where at least 3 out of 5 annotators have selected a label to be true for a given turn. The difficulty of the task increases as the number of classes INLINEFORM4 increases, as there are more classifications done for each turn (i.e., for the 6-class problem, there are 6 classification tasks per turn, while for the 8-class problem, there are 8, etc). Due to the inherent imbalance of label-distribution in the data (shown in Figure FIGREF17 ), we use weighted F-macro to calculate our final scores for each feature set (which finds the average of the metrics for each label, weighted by the number of true instances for that label) BIBREF31 .
Our first experiment sets out to compare the use of a non-sequential classification algorithm versus a sequential model for dialogue act classification on our dataset. We experiment with the default Naive Bayes (NB) and Linear SVC algorithms from Scikit-Learn BIBREF31 , comparing with our sequential SVM-HMM model. We test each classifier on each of our four class sets, reporting weighted F-macro for each experiment. Figure FIGREF33 shows the results of the experiments.
From this experiment, we observe that our sequential SVM-HMM outperforms each non-sequential baseline, for each of the four class sets. We select the sequential SVM-HMM model for our preferred model for subsequent experiments. We observe that while performance may be expected to drop as the number of classes increases, we instead get a spike in performance for the 10-Class (Easy) setting. This increase occurs due to the addition of the lexically well-defined classes of Statement Apology and Statement Thanks, which are much simpler for our model to predict. Their addition results in a performance boost, comparable to that of the simpler 6-Class problem. When we remove the two well-defined classes and add in the next two broader dialogue act classes of Statement Offer and Question Open (as defined by the 10-Class (Hard) set), we observe a drop in performance, and an overall result comparable to our 8-Class problem. This result is still strong, since the number of classes has increased, but the overall performance does not drop.
We also observe that while NB and LinearSVC have the same performance trend for the smaller number of classes, Linear SVC rapidly improves in performance as the number of classes increases, following the same trend as SVM-HMM. The smallest margin of difference between SVM-HMM and Linear SVC also occurs at the 10-Class (Easy) setting, where the addition of highly-lexical classes makes for a more differentiable set of turns.
Our next experiment tests the differences in performance when training and testing our real-time sequential SVM-HMM model using only a single type of speaker's turns (i.e. only Customer or only Agent turns). Figure FIGREF35 shows the relative performance of using only speaker-specific turns, versus our standard results using all turns.
We observe that using Customer-only turns gives us lower prediction performance than using both speakers' turns, but that Agent-only turns actually gives us higher performance. Since agents are put through training on how to interact with customers (often using templates), agent behavior is significantly more predictable than customer behavior, and it is easier to predict agent turns even without utilizing any customer turn information (which is more varied, and thus more difficult to predict).
We again observe a boost in performance at out 10-Class (Easy) set, due to the inclusion of lexically well-defined classes. Notably, we achieve best performance for the 10-Class (Easy) set using only agent turns, where the use of the Apology and Thanks classes are both prevalent and predictable.
In our final experiment, we explore the changes in performance we get by splitting the training and test data based on company domain. We compare this performance with our standard setup for SVM-HMM from our baseline experiments (Figure FIGREF33 ), where our train-test data splitting is company-independent (i.e. all conversations are randomized, and no information is used to differentiate different companies or domains). To recap, our data consists of conversations from four companies from three different industrial domains (one from the telecommunication domain, two from the electronics domain, and one from the insurance domain). We create four different versions of our 6-class real-time sequential SVM-HMM, where we train on the data from three of the companies, and test on the remaining company. We present our findings in Table TABREF37 .
From the table, we see that our real-time model achieves best prediction results when we use one of the electronics companies in the test fold, even though the number of training samples is smallest in these cases. On the other hand, when we assign insurance company in the test fold, our model's prediction performance is comparatively low. Upon further investigation, we find that customer-agent conversations in the telecommunication and electronics domains are more similar than those in the insurance domain. Our findings show that our model is robust to different domains as our test set size increases, and that our more generic, company-independent experiment gives us better performance than any domain-specific experiments.
## Conversation Outcome Analysis
Given our observation that Agent turns are more predictable, and that we achieve best performance in a company-independent setting, we question whether the training that agents receive is actually reliable in terms of resulting in overall "satisfied customers", regardless of company domain. Ultimately, our goal is to discover whether we can use the insight we derive from our predicted dialogue acts to better inform conversational systems aimed at offering customer support. Our next set of experiments aims to show the utility of our real-time dialogue act classification as a method for summarizing semantic intent in a conversation into rules that can be used to guide automated systems.
## Classifying Problem Outcomes
We conduct three supervised classification experiments to better understand full conversation outcome, using the default Linear SVC classifier in Scikit-Learn BIBREF31 (which gave us our best baseline for the dialogue classification task). Each classification experiments centers around one of three problem outcomes: customer satisfaction, problem resolution, and customer frustration. For each outcome, we remove any conversation that did not receive majority consensus for a label, or received majority vote of "can't tell". Our final conversation sets consist of 216 satisfied and 500 unsatisfied customer conversations, 271 resolved and 425 unresolved problem conversations, and 534 frustrated and 229 not frustrated customer conversations. We retain the inherent imbalance in the data to match the natural distribution observed. The clear excess of consensus of responses that indicate negative outcomes further motivates us to understand what sorts of dialogic patterns results in such outcomes.
We run the experiment for each conversation outcome using 10-fold cross-validation, under each of our four class settings: 6-Class, 8-Class, 10-Class (Easy), and 10-Class (Hard). The first feature set we use is Best_Features (from the original dialogue act classification experiments), which we run as a baseline.
Our second feature set is our Dialogue_Acts predictions for each turn – we choose the most probable dialogue act prediction for each turn using our dialogue act classification framework to avoid sparsity. In this way, for each class size INLINEFORM0 , each conversation is converted into a vector of INLINEFORM1 (up to 10) features that describe the most strongly associated dialogue act from the dialogue act classification experiments for each turn, and the corresponding turn number. For example, a conversation feature vector may look as follows: INLINEFORM2
Thus, our classifier can then learn patterns based on these features (for example, that specific acts appearing at the end of a conversation are strong indicators of customer satisfaction) that allow us to derive rules about successful/unsuccessful interactions.
Figure FIGREF38 shows the results of our binary classification experiments for each outcome. For each experiment, the Best_Features set is constant over each class size, while the Dialogue_Act features are affected by class size (since the predicted act for each turn will change based on the set of acts available for that class size). Our first observation is that we achieve high performance on the binary classification task, reaching F-measures of 0.70, 0.65, and 0.83 for the satisfaction, resolution, and frustration outcomes, respectively. Also, we observe that the performance of our predicted dialogue act features is comparable to that of the much larger set of best features for each label (almost identical in the case of frustration).
In more detail, we note interesting differences comparing the performance of the small set of dialogue act features that "summarize" the large, sparse set of best features for each label, as a form of data-driven feature selection. For satisfaction, we see that the best feature set outperforms the dialogue acts for each class set except for 10-Class (Easy), where the dialogue acts are more effective. The existence of the very lexically well-defined Social Act Thanking and Social Act Apology classes makes the dialogue acts ideal for summarization. In the case of problem resolution, we see that the performance of the dialogue acts approaches that of the best feature set as the number of classes increases, showing that the dialogue features are able to express the full intent of the turns well, even at more difficult class settings. Finally, for the frustration experiment, we observe negligible different between the best features and dialogue act features, and very high classification results overall.
## Actionable Rules for Automated Customer Support
While these experiments highlight how we can use dialogue act predictions as a means to greatly reduce feature sparsity and predict conversation outcome, our main aim is to gain good insight from the use of the dialogue acts to inform and automate customer service interactions. We conduct deeper analysis by taking a closer look at the most informative dialogue act features in each experiment.
Table TABREF44 shows the most informative features and weights for each of our three conversation outcomes. To help guide our analysis, we divide the features into positions based on where they occur in the conversation: start (turns 1-3), middle (turns 4-6), and end (turns 7-10). Desirable outcomes (customers that are satisfied/not frustrated and resolved problems) are shown at the top rows of the table, and undesirable outcomes (unsatisfied/frustrated customers and unresolved problems) are shown at the bottom rows.
Our analysis helps zone in on how the use of certain dialogue acts may be likely to result in different outcomes. The weights we observe vary in the amount of insight provided: for example, offering extra help at the end of a conversation, or thanking the customer yields more satisfied customers, and more resolved problems (with ratios of above 6:1). However, some outcomes are much more subtle: for example, asking yes-no questions early-on in a conversation is highly associated with problem resolution (ratio 3:1), but asking them at the end of a conversation has as similarly strong association with unsatisfied customers. Giving elaborate answers that are not a simple affirmative, negative, or response acknowledgement (i.e. Answer (Other)) towards the middle of a conversation leads to satisfied customers that are not frustrated. Likewise, requesting information towards the end of a conversation (implying that more information is still necessary at the termination of the dialogue) leads to unsatisfied and unresolved customers, with ratios of at least 4:1.
By using the feature weights we derive from using our predicted dialogue acts in our outcome classification experiments, we can thus derive data-driven patterns that offer useful insight into good/bad practices. Our goal is to then use these rules as guidelines, serving as a basis for automated response planning in the customer service domain. For example, these rules can be used to recommend certain dialogue act responses given the position in a conversation, and based previous turns. This information, derived from correlation with conversation outcomes, gives a valuable addition to conversational flow for intelligent agents, and is more useful than canned responses.
## Conclusions
In this paper, we explore how we can analyze dialogic trends in customer service conversations on Twitter to offer insight into good/bad practices with respect to conversation outcomes. We design a novel taxonomy of fine-grained dialogue acts, tailored for the customer service domain, and gather annotations for 800 Twitter conversations. We show that dialogue acts are often semantically overlapping, and conduct multi-label supervised learning experiments to predict multiple appropriate dialogue act labels for each turn in real-time, under varying class sizes. We show that our sequential SVM-HMM model outperforms all non-sequential baselines, and plan to continue our exploration of other sequential models including Conditional Random Fields (CRF) BIBREF32 and Long Short-Term Memory (LSTM) BIBREF33 , as well as of dialogue modeling using different Markov Decision Process (MDP) BIBREF34 models such as the Partially-Observed MDP (POMDP) BIBREF35 .
We establish that agents are more predictable than customers in terms of the dialogue acts they utilize, and set out to understand whether the conversation strategies agents employ are well-correlated with desirable conversation outcomes. We conduct binary classification experiments to analyze how our predicted dialogue acts can be used to classify conversations as ending in customer satisfaction, customer frustration, and problem resolution. We observe interesting correlations between the dialogue acts agents use and the outcomes, offering insights into good/bad practices that are more useful for creating context-aware automated customer service systems than generating canned response templates.
Future directions for this work revolve around the integration of the insights derived in the design of automated customer service systems. To this end, we aim to improve the taxonomy and annotation design by consulting domain-experts and using annotator feedback and agreement information, derive more powerful features for dialogue act prediction, and automate ranking and selection of best-practice rules based on domain requirements for automated customer service system design.
| 15
|
1709.10217
|
The First Evaluation of Chinese Human-Computer Dialogue Technology
|
# The First Evaluation of Chinese Human-Computer Dialogue Technology
## Abstract
In this paper, we introduce the first evaluation of Chinese human-computer dialogue technology. We detail the evaluation scheme, tasks, metrics and how to collect and annotate the data for training, developing and test. The evaluation includes two tasks, namely user intent classification and online testing of task-oriented dialogue. To consider the different sources of the data for training and developing, the first task can also be divided into two sub tasks. Both the two tasks are coming from the real problems when using the applications developed by industry. The evaluation data is provided by the iFLYTEK Corporation. Meanwhile, in this paper, we publish the evaluation results to present the current performance of the participants in the two tasks of Chinese human-computer dialogue technology. Moreover, we analyze the existing problems of human-computer dialogue as well as the evaluation scheme itself.
## Introduction
Recently, human-computer dialogue has been emerged as a hot topic, which has attracted the attention of both academia and industry. In research, the natural language understanding (NLU), dialogue management (DM) and natural language generation (NLG) have been promoted by the technologies of big data and deep learning BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 . Following the development of machine reading comprehension BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , the NLU technology has made great progress. The development of DM technology is from rule-based approach and supervised learning based approach to reinforcement learning based approach BIBREF15 . The NLG technology is through pattern-based approach, sentence planning approach and end-to-end deep learning approach BIBREF16 , BIBREF17 , BIBREF18 . In application, there are massive products that are based on the technology of human-computer dialogue, such as Apple Siri, Amazon Echo, Microsoft Cortana, Facebook Messenger and Google Allo etc.
Although the blooming of human-computer dialogue technology in both academia and industry, how to evaluate a dialogue system, especially an open domain chit-chat system, is still an open question. Figure FIGREF6 presents a brief comparison of the open domain chit-chat system and the task-oriented dialogue system.
From Figure FIGREF6 , we can see that it is quite different between the open domain chit-chat system and the task-oriented dialogue system. For the open domain chit-chat system, as it has no exact goal in a conversation, given an input message, the responses can be various. For example, for the input message “How is it going today?”, the responses can be “I'm fine!”, “Not bad.”, “I feel so depressed!”, “What a bad day!”, etc. There may be infinite number of responses for an open domain messages. Hence, it is difficult to construct a gold standard (usually a reference set) to evaluate a response which is generated by an open domain chit-chat system. For the task-oriented system, although there are some objective evaluation metrics, such as the number of turns in a dialogue, the ratio of task completion, etc., there is no gold standard for automatically evaluating two (or more) dialogue systems when considering the satisfaction of the human and the fluency of the generated dialogue.
To promote the development of the evaluation technology for dialogue systems, especially considering the language characteristics of Chinese, we organize the first evaluation of Chinese human-computer dialogue technology. In this paper, we will present the evaluation scheme and the released corpus in detail.
The rest of this paper is as follows. In Section 2, we will briefly introduce the first evaluation of Chinese human-computer dialogue technology, which includes the descriptions and the evaluation metrics of the two tasks. We then present the evaluation data and final results in Section 3 and 4 respectively, following the conclusion and acknowledgements in the last two sections.
## The First Evaluation of Chinese Human-Computer Dialogue Technology
The First Evaluation of Chinese Human-Computer Dialogue Technology includes two tasks, namely user intent classification and online testing of task-oriented dialogue.
## Task 1: User Intent Classification
In using of human-computer dialogue based applications, human may have various intent, for example, chit-chatting, asking questions, booking air tickets, inquiring weather, etc. Therefore, after receiving an input message (text or ASR result) from a user, the first step is to classify the user intent into a specific domain for further processing. Table TABREF7 shows an example of user intent with category information.
In task 1, there are two top categories, namely, chit-chat and task-oriented dialogue. The task-oriented dialogue also includes 30 sub categories. In this evaluation, we only consider to classify the user intent in single utterance.
It is worth noting that besides the released data for training and developing, we also allow to collect external data for training and developing. To considering that, the task 1 is indeed includes two sub tasks. One is a closed evaluation, in which only the released data can be used for training and developing. The other is an open evaluation that allow to collect external data for training and developing. For task 1, we use F1-score as evaluation metric.
## Task 2: Online Testing of Task-oriented Dialogue
For the task-oriented dialogue systems, the best way for evaluation is to use the online human-computer dialogue. After finishing an online human-computer dialogue with a dialogue system, the human then manually evaluate the system by using the metrics of user satisfaction degree, dialogue fluency, etc. Therefore, in the task 2, we use an online testing of task-oriented dialogue for dialogue systems. For a human tester, we will give a complete intent with an initial sentence, which is used to start the online human-computer dialogue. Table TABREF12 shows an example of the task-oriented human-computer dialogue. Here “U” and “R” denote user and robot respectively. The complete intent is as following:
“æ¥è¯¢æ天ä»åå°æ»¨å°å京çæé´è½¯å§ç«è½¦ç¥¨ï¼ä¸ä¸éºåå¯ã
Inquire the soft berth ticket at tomorrow evening, from Harbin to Beijing, either upper or lower berth is okay.”
In task 2, there are three categories. They are “air tickets”, “train tickets” and “hotel”. Correspondingly, there are three type of tasks. All the tasks are in the scope of the three categories. However, a complete user intent may include more than one task. For example, a user may first inquiring the air tickets. However, due to the high price, the user decide to buy a train tickets. Furthermore, the user may also need to book a hotel room at the destination.
We use manual evaluation for task 2. For each system and each complete user intent, the initial sentence, which is used to start the dialogue, is the same. The tester then begin to converse to each system. A dialogue is finished if the system successfully returns the information which the user inquires or the number of dialogue turns is larger than 30 for a single task. For building the dialogue systems of participants, we release an example set of complete user intent and three data files of flight, train and hotel in JSON format. There are five evaluation metrics for task 2 as following.
Task completion ratio: The number of completed tasks divided by the number of total tasks.
User satisfaction degree: There are five scores -2, -1, 0, 1, 2, which denote very dissatisfied, dissatisfied, neutral, satisfied and very satisfied, respectively.
Response fluency: There are three scores -1, 0, 1, which indicate nonfluency, neutral, fluency.
Number of dialogue turns: The number of utterances in a task-completed dialogue.
Guidance ability for out of scope input: There are two scores 0, 1, which represent able to guide or unable to guide.
For the number of dialogue turns, we have a penalty rule that for a dialogue task, if the system cannot return the result (or accomplish the task) in 30 turns, the dialogue task is end by force. Meanwhile, if a system cannot accomplish a task in less than 30 dialogue turns, the number of dialogue turns is set to 30.
## Evaluation Data
In the evaluation, all the data for training, developing and test is provided by the iFLYTEK Corporation.
For task 1, as the descriptions in Section SECREF10 , the two top categories are chit-chat (chat in Table TABREF13 ) and task-oriented dialogue. Meanwhile, the task-oriented dialogue also includes 30 sub categories. Actually, the task 1 is a 31 categories classification task. In task 1, besides the data we released for training and developing, we also allow the participants to extend the training and developing corpus. Hence, there are two sub tasks for the task 1. One is closed test, which means the participants can only use the released data for training and developing. The other is open test, which allows the participants to explore external corpus for training and developing. Note that there is a same test set for both the closed test and the open test.
For task 2, we release 11 examples of the complete user intent and 3 data file, which includes about one month of flight, hotel and train information, for participants to build their dialogue systems. The current date for online test is set to April 18, 2017. If the tester says “today”, the systems developed by the participants should understand that he/she indicates the date of April 18, 2017.
## Evaluation Results
There are 74 participants who are signing up the evaluation. The final number of participants is 28 and the number of submitted systems is 43. Table TABREF14 and TABREF15 show the evaluation results of the closed test and open test of the task 1 respectively. Due to the space limitation, we only present the top 5 results of task 1. We will add the complete lists of the evaluation results in the version of full paper.
Note that for task 2, there are 7 submitted systems. However, only 4 systems can provide correct results or be connected in a right way at the test phase. Therefore, Table TABREF16 shows the complete results of the task 2.
## Conclusion
In this paper, we introduce the first evaluation of Chinese human-computer dialogue technology. In detail, we first present the two tasks of the evaluation as well as the evaluation metrics. We then describe the released data for evaluation. Finally, we also show the evaluation results of the two tasks. As the evaluation data is provided by the iFLYTEK Corporation from their real online applications, we believe that the released data will further promote the research of human-computer dialogue and fill the blank of the data on the two tasks.
## Acknowledgements
We would like to thank the Social Media Processing (SMP) committee of Chinese Information Processing Society of China. We thank all the participants of the first evaluation of Chinese human-computer dialogue technology. We also thank the testers from the voice resource department of the iFLYTEK Corporation for their effort to the online real-time human-computer dialogue test and offline dialogue evaluation. We thank Lingzhi Li, Yangzi Zhang, Jiaqi Zhu and Xiaoming Shi from the research center for social computing and information retrieval for their support on the data annotation, establishing the system testing environment and the communication to the participants and help connect their systems to the testing environment.
| 8
|
1710.07395
|
Detecting Online Hate Speech Using Context Aware Models
|
# Detecting Online Hate Speech Using Context Aware Models
## Abstract
In the wake of a polarizing election, the cyber world is laden with hate speech. Context accompanying a hate speech text is useful for identifying hate speech, which however has been largely overlooked in existing datasets and hate speech detection models. In this paper, we provide an annotated corpus of hate speech with context information well kept. Then we propose two types of hate speech detection models that incorporate context information, a logistic regression model with context features and a neural network model with learning components for context. Our evaluation shows that both models outperform a strong baseline by around 3% to 4% in F1 score and combining these two models further improve the performance by another 7% in F1 score.
## Introduction
Following a turbulent election season, 2016's cyber world is awash with hate speech. Automatic detection of hate speech has become an urgent need since human supervision is unable to deal with large quantities of emerging texts.
Context information, by our definition, is the text, symbols or any other kind of information related to the original text. While intuitively, context accompanying hate speech is useful for detecting hate speech, context information of hate speech has been overlooked in existing datasets and automatic detection models.
Online hate speech tends to be subtle and creative, which makes context especially important for automatic hate speech detection. For instance,
(1) barryswallows: Merkel would never say NO
This comment is posted for the News titled by "German lawmakers approve 'no means no' rape law after Cologne assaults". With context, it becomes clear that this comment is a vicious insult towards female politician. However, almost all the publicly available hate speech annotated datasets do not contain context information. BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 .
We have created a new dataset consisting of 1528 Fox News user comments, which were taken from 10 complete discussion threads for 10 widely read Fox News articles. It is different from previous datasets from the following two perspectives. First, it preserves rich context information for each comment, including its user screen name, all comments in the same thread and the news article the comment is written for. Second, there is no biased data selection and all comments in each news comment thread were annotated.
In this paper, we explored two types of models, feature-based logistic regression models and neural network models, in order to incorporate context information in automatic hate speech detection. First, logistic regression models have been used in several prior hate speech detection studies BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF0 , BIBREF2 , BIBREF9 and various features have been tried including character-level and word-level n-gram features, syntactic features, linguistic features, and comment embedding features. However, all the features were derived from the to-be-classified text itself. In contrast, we experiment with logistic regression models using features extracted from context text as well. Second, neural network models BIBREF10 , BIBREF11 , BIBREF12 have the potential to capture compositional meanings of text, but they have not been well explored for online hate speech detection until recently BIBREF13 . We experiment with neural net models containing separate learning components that model compositional meanings of context information. Furthermore, recognizing unique strengths of each type of models, we build ensemble models of the two types of models. Evaluation shows that context-aware logistic regression models and neural net models outperform their counterparts that are blind with context information. Especially, the final ensemble models outperform a strong baseline system by around 10% in F1-score.
## Related Works
Recently, a few datasets with human labeled hate speech have been created, however, most of existing datasets do not contain context information. Due to the sparsity of hate speech in everyday posts, researchers tend to sample candidates from bootstrapping instead of random sampling, in order to increase the chance of seeing hate speech. Therefore, the collected data instances are likely to be from distinct contexts.
For instance, in the Primary Data Set described in BIBREF14 and later used by BIBREF9 , 10% of the dataset is randomly selected while the remaining consists of comments tagged by users and editors. BIBREF15 built a balanced data set of 24.5k tweets by selecting from Twitter accounts that claimed to be racist or were deemed racist using their followed news sources. BIBREF5 collected hateful tweets related to the murder of Drummer Lee Rigby in 2013. BIBREF0 provided a corpus of 16k annotated tweets in which 3.3k are labeled as sexist and 1.9k are labeled as racist. They created this corpus by bootstrapping from certain key words ,specific hashtags and certain prolific users. BIBREF16 created a dataset of 9000 human labeled paragraphs that were collected using regular expression matching in order to find hate speech targeting Judaism and Israel. BIBREF7 extracted data instances from instagram that were associated with certain user accounts. BIBREF2 presented a very large corpus containing over 115k Wikipedia comments that include around 37k randomly sampled comments and the remaining 78k comments were selected from Wikipedia blocked comments.
Most of existing hate speech detection models are feature-based and use features derived from the target text itself. BIBREF5 experimented with different classification methods including Bayesian Logistic Regression, Random Forest Decision Trees and SVMs, using features such as n-grams, reduced n-grams, dependency paths, and hateful terms. BIBREF0 proposed a logistic regression model using character n-gram features. BIBREF14 used the paragraph2vec for joint modeling of comments and words, then the generated embeddings were used as feature in a logistic regression model. BIBREF9 experimented with various syntactic, linguistic and distributional semantic features including word length, sentence length, part of speech tags, and embedding features, in order to improve performance of logistic regression classifiers. Recently, BIBREF17 surveyed current approaches for hate speech detection, which interestingly also called to attention on modeling context information for resolving difficult hate speech instances.
## Corpus Overview
The Fox News User Comments corpus consists of 1528 annotated comments (435 labeled as hateful) that were posted by 678 different users in 10 complete news discussion threads in the Fox News website. The 10 threads were manually selected and represent popular discussion threads during August 2016. All of the comments included in these 10 threads were annotated. The number of comments in each of the 10 threads is roughly equal. Rich context information was kept for each comment, including its user screen name, the comments and their nested structure and the original news article. The data corpus along with annotation guidelines is posted on github.
## Annotation Guidelines
Our annotation guidelines are similar to the guidelines used by BIBREF9 . We define hateful speech to be the language which explicitly or implicitly threatens or demeans a person or a group based upon a facet of their identity such as gender, ethnicity, or sexual orientation. The labeling of hateful speech in our corpus is binary. A comment will be labeled as hateful or non-hateful.
## Annotation Procedure
We identified two native English speakers for annotating online user comments. The two annotators first discussed and practices before they started annotation. They achieved a surprisingly high Kappa score BIBREF18 of 0.98 on 648 comments from 4 threads. We think that thorough discussions in the training stage is the key for achieving this high inter-agreement. For those comments which annotators disagreed on, we label them as hateful as long as one annotator labeled them as hateful. Then one annotator continued to annotate the remaining 880 comments from the remaining six discussion threads.
## Characteristics in Fox News User Comments corpus
Hateful comments in the Fox News User Comments Corpus is often subtle, creative and implicit. Therefore, context information is necessary in order to accurately identify such hate speech.
The hatefulness of many comments depended on understanding their contexts. For instance,
(3) mastersundholm: Just remember no trabjo no cervesa
This comment is posted for the news "States moving to restore work requirements for food stamp recipients". This comment implies that Latino immigrants abuse the usage of food stamp policy, which is clearly a stereotyping.
Many hateful comments use implicit and subtle language, which contain no clear hate indicating word or phrase. In order to recognize such hard cases, we hypothesize that neural net models are more suitable by capturing overall composite meanings of a comment. For instance, the following comment is a typical implicit stereotyping against women.
(4) MarineAssassin: Hey Brianne - get in the kitchen and make me a samich. Chop Chop
11% of our annotated comments have more than 50 words each. In such long comments, the hateful indicators usually appear in a small region of a comment while the majority of the comment is neutral. For example,
(5) TMmckay: I thought ...115 words... Too many blacks winning, must be racist and needs affirmative action to make whites equally win!
Certain user screen names indicate hatefulness, which imply that comments posted by these users are likely to contain hate speech. In the following example, commie is a slur for communists.
(6)nocommie11: Blah blah blah. Israel is the only civilized nation in the region to keep the unwashed masses at bay.
## Logistic Regression Models
In logistic regression models, we extract four types of features, word-level and character-level n-gram features as well as two types of lexicon derived features. We extract these four types of features from the target comment first. Then we extract these features from two sources of context texts, specifically the title of the news article that the comment was posted for and the screen name of the user who posted the comment.
For logistic regression model implementation, we use l2 loss. We adopt the balanced class weight as described in Scikit learn. Logistic regression model with character-level n-gram features is presented as a strong baseline for comparison since it was shown very effective. BIBREF0 , BIBREF9
For character level n-grams, we extract character level bigrams, tri-grams and four-grams. For word level n-grams, we extract unigrams and bigrams.
Linguistic Inquiry and Word Count, also called LIWC, has been proven useful for text analysis and classification BIBREF19 . In the LIWC dictionary, each word is labeled with several semantic labels. In our experiment, we use the LIWC 2015 dictionary which contain 125 semantic categories. Each word is converted into a 125 dimension LIWC vector, one dimension per semantic category. The LIWC feature vector for a comment or its context is a 125 dimension vector as well, which is the sum of all its words' LIWC vectors.
NRC emotion lexicon contains a list of English words that were labeled with eight basic emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust) and sentiment polarities (negative and positive) BIBREF20 . We use NRC emotion lexicon to capture emotion clues in text. Each word is converted into a 10 dimension emotion vector, corresponding to eight emotion types and two polarity labels. The emotion vector for a comment or its context is a 10 dimension vector as well, which is the sum of all its words' emotion vectors.
As shown in table TABREF20 , given comment as the only input content, the combination of character n-grams, word n-grams, LIWC feature and NRC feature achieves the best performance. It shows that in addition to character level features, adding more features can improve hate speech detection performance. However, the improvement is limited. Compared with baseline model, the F1 score only improves 1.3%.
In contrast, when context information was taken into account, the performance greatly improved. Specifically, after incorporating features extracted from the news title and username, the model performance was improved by around 4% in both F1 score and AUC score. This shows that using additional context based features in logistic regression models is useful for hate speech detection.
## Neural Network Models
Our neural network model mainly consists of three parallel LSTM BIBREF21 layers. It has three different inputs, including the target comment, its news title and its username. Comment and news title are encoded into a sequence of word embeddings. We use pre-trained word embeddings in word2vec. Username is encoded into a sequence of characters. We use one-hot encoding of characters.
Comment is sent into a bi-directional LSTM with attention mechanism. BIBREF22 . News title and username are sent into a bi-directional LSTM. Note that we did not apply attention mechanism to the neural network models for username and news title because both types of context are relatively short and attention mechanism tends to be useful when text input is long. The three LSTM output layers are concatenated, then connected to a sigmoid layer, which outputs predictions.
The number of hidden units in each LSTM used in our model is set to be 100. The recurrent dropout rate of LSTMs is set to 0.2. In addition, we use binary cross entropy as the loss function and a batch size of 128. The neural network models are trained for 30 epochs.
As shown in table TABREF21 , given comment as the only input content, the bi-directional LSTM model with attention mechanism achieves the best performance. Note that the attention mechanism significantly improves the hate speech detection performance of the bi-directional LSTM model. We hypothesize that this is because hate indicator phrases are often concentrated in a small region of a comment, which is especially the case for long comments.
## Ensemble Models
To study the difference of logistic regression model and neural network model and potentially get performance improvement, we will build and evaluate ensemble models.
As shown in table TABREF24 , both ensemble models significantly improved hate speech detection performance. Figure FIGREF28 shows the system prediction results of comments that were labeled as hateful in the dataset. It can be seen that the two models perform differently. We further examined predicted comments and find that both types of models have unique strengths in identifying certain types of hateful comments.
The feature-based logistic regression models are capable of making good use of character-level n-gram features, which are powerful in identifying hateful comments that contains OOV words, capitalized words or misspelled words. We provide two examples from the hateful comments that were only labeled by the logistic regression model:
(7)kmawhmf:FBLM.
Here FBLM means fuck Black Lives Matter. This hateful comment contains only character information which can exactly be made use of by our logistic regression model.
(8)SFgunrmn: what a efen loon, but most femanazis are.
This comment deliberately misspelled feminazi for femanazis, which is a derogatory term for feminists. It shows that logistic regression model is capable in dealing with misspelling.
The LSTM with attention mechanism are suitable for identifying specific small regions indicating hatefulness in long comments. In addition, the neural net models are powerful in capturing implicit hateful language as well. The following are two hateful comment examples that were only identified by the neural net model:
(9)freedomscout: @LarJass Many religions are poisonous to logic and truth, that much is true...and human beings still remain fallen human beings even they are Redeemed by the Sacrifice of Jesus Christ. So there's that. But the fallacies of thinking cannot be limited or attributed to religion but to error inherent in human motivation, the motivation to utter self-centeredness as fallen sinful human beings. Nearly all of the world's many religions are expressions of that utter sinful nature...Christianity and Judaism being the sole exceptions.
This comment is expressing the stereotyping against religions which are not Christian or Judaism. The hatefulness is concentrated within the two bolded segments.
(10)mamahattheridge: blacks Love being victims.
In this comment, the four words themselves are not hateful at all. But when combined together, it is clearly hateful against black people.
## Evaluation
We evaluate our model by 10 fold cross validation using our newly created Fox News User Comments Corpus. Both types of models use the exact same 10 folds of training data and test data. We report experimental results using multiple metrics, including accuracy, precision/recall/F1-score, and accuracy area under curve (AUC).
## Experimental Results
Table TABREF20 shows the performance of logistic regression models. The first section of table TABREF20 shows the performance of logistic regression models using features extracted from a target comment only. The result shows that the logistic regression model was improved in every metric after adding both word-level n-gram features and lexicon derived features. However, the improvements are moderate.
The second section shows the performance of logistic regression models using the four types of features extracted from both a target comment and its contextsThe result shows that the logistic regression model using features extracted from a comment and both types of context achieved the best performance and obtained improvements of 2.8% and 2.5% in AUC score and F1-score respectively.
Table TABREF21 shows the performance of neural network models. The first section of table TABREF21 shows the performance of several neural network models that use comments as the only input. The model names are self-explanatory. We can see that the attention mechanism coupled with the bi-directional LSTM neural net greatly improved the online hate speech detection, by 5.7% in AUC score.
The second section of table TABREF21 shows performance of the best neural net model (bi-directional LSTM with attention) after adding additional learning components that take context as input. The results show that adding username and news title can both improve model performance. Using news title gives the best F1 score while using both news title and username gives the best AUC score.
Table TABREF24 shows performance of ensemble models by combining prediction results of the best context-aware logistic regression model and the best context-aware neural network model. We used two strategies in combining prediction results of two types of models. Specifically, the Max Score Ensemble model made the final decisions based on the maximum of two scores assigned by the two separate models; instead, the Average Score Ensemble model used the average score to make final decisions.
We can see that both ensemble models further improved hate speech detection performance compared with using one model only and achieved the best classification performance. Compared with the logistic regression baseline, the Max Score Ensemble model improved the recall by more than 20% with a comparable precision and improved the F1 score by around 10%, in addition, the Average Score Ensemble model improved the AUC score by around 7%.
## Conclusion
We demonstrated the importance of utilizing context information for online hate speech detection. We first presented a corpus of hateful speech consisting of full threads of online discussion posts. In addition, we presented two types of models, feature-based logistic regression models and neural network models, in order to incorporate context information for improving hate speech detection performance. Furthermore, we show that ensemble models leveraging strengths of both types of models achieve the best performance for automatic online hate speech detection.
| 12
|
1712.00991
|
Mining Supervisor Evaluation and Peer Feedback in Performance Appraisals
|
# Mining Supervisor Evaluation and Peer Feedback in Performance Appraisals
## Abstract
Performance appraisal (PA) is an important HR process to periodically measure and evaluate every employee's performance vis-a-vis the goals established by the organization. A PA process involves purposeful multi-step multi-modal communication between employees, their supervisors and their peers, such as self-appraisal, supervisor assessment and peer feedback. Analysis of the structured data and text produced in PA is crucial for measuring the quality of appraisals and tracking actual improvements. In this paper, we apply text mining techniques to produce insights from PA text. First, we perform sentence classification to identify strengths, weaknesses and suggestions of improvements found in the supervisor assessments and then use clustering to discover broad categories among them. Next we use multi-class multi-label classification techniques to match supervisor assessments to predefined broad perspectives on performance. Finally, we propose a short-text summarization technique to produce a summary of peer feedback comments for a given employee and compare it with manual summaries. All techniques are illustrated using a real-life dataset of supervisor assessment and peer feedback text produced during the PA of 4528 employees in a large multi-national IT company.
## Introduction
Performance appraisal (PA) is an important HR process, particularly for modern organizations that crucially depend on the skills and expertise of their workforce. The PA process enables an organization to periodically measure and evaluate every employee's performance. It also provides a mechanism to link the goals established by the organization to its each employee's day-to-day activities and performance. Design and analysis of PA processes is a lively area of research within the HR community BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 .
The PA process in any modern organization is nowadays implemented and tracked through an IT system (the PA system) that records the interactions that happen in various steps. Availability of this data in a computer-readable database opens up opportunities to analyze it using automated statistical, data-mining and text-mining techniques, to generate novel and actionable insights / patterns and to help in improving the quality and effectiveness of the PA process BIBREF4 , BIBREF5 , BIBREF6 . Automated analysis of large-scale PA data is now facilitated by technological and algorithmic advances, and is becoming essential for large organizations containing thousands of geographically distributed employees handling a wide variety of roles and tasks.
A typical PA process involves purposeful multi-step multi-modal communication between employees, their supervisors and their peers. In most PA processes, the communication includes the following steps: (i) in self-appraisal, an employee records his/her achievements, activities, tasks handled etc.; (ii) in supervisor assessment, the supervisor provides the criticism, evaluation and suggestions for improvement of performance etc.; and (iii) in peer feedback (aka INLINEFORM0 view), the peers of the employee provide their feedback. There are several business questions that managers are interested in. Examples:
In this paper, we develop text mining techniques that can automatically produce answers to these questions. Since the intended users are HR executives, ideally, the techniques should work with minimum training data and experimentation with parameter setting. These techniques have been implemented and are being used in a PA system in a large multi-national IT company.
The rest of the paper is organized as follows. Section SECREF2 summarizes related work. Section SECREF3 summarizes the PA dataset used in this paper. Section SECREF4 applies sentence classification algorithms to automatically discover three important classes of sentences in the PA corpus viz., sentences that discuss strengths, weaknesses of employees and contain suggestions for improving her performance. Section SECREF5 considers the problem of mapping the actual targets mentioned in strengths, weaknesses and suggestions to a fixed set of attributes. In Section SECREF6 , we discuss how the feedback from peers for a particular employee can be summarized. In Section SECREF7 we draw conclusions and identify some further work.
## Related Work
We first review some work related to sentence classification. Semantically classifying sentences (based on the sentence's purpose) is a much harder task, and is gaining increasing attention from linguists and NLP researchers. McKnight and Srinivasan BIBREF7 and Yamamoto and Takagi BIBREF8 used SVM to classify sentences in biomedical abstracts into classes such as INTRODUCTION, BACKGROUND, PURPOSE, METHOD, RESULT, CONCLUSION. Cohen et al. BIBREF9 applied SVM and other techniques to learn classifiers for sentences in emails into classes, which are speech acts defined by a verb-noun pair, with verbs such as request, propose, amend, commit, deliver and nouns such as meeting, document, committee; see also BIBREF10 . Khoo et al. BIBREF11 uses various classifiers to classify sentences in emails into classes such as APOLOGY, INSTRUCTION, QUESTION, REQUEST, SALUTATION, STATEMENT, SUGGESTION, THANKING etc. Qadir and Riloff BIBREF12 proposes several filters and classifiers to classify sentences on message boards (community QA systems) into 4 speech acts: COMMISSIVE (speaker commits to a future action), DIRECTIVE (speaker expects listener to take some action), EXPRESSIVE (speaker expresses his or her psychological state to the listener), REPRESENTATIVE (represents the speaker's belief of something). Hachey and Grover BIBREF13 used SVM and maximum entropy classifiers to classify sentences in legal documents into classes such as FACT, PROCEEDINGS, BACKGROUND, FRAMING, DISPOSAL; see also BIBREF14 . Deshpande et al. BIBREF15 proposes unsupervised linguistic patterns to classify sentences into classes SUGGESTION, COMPLAINT.
There is much work on a closely related problem viz., classifying sentences in dialogues through dialogue-specific categories called dialogue acts BIBREF16 , which we will not review here. Just as one example, Cotterill BIBREF17 classifies questions in emails into the dialogue acts of YES_NO_QUESTION, WH_QUESTION, ACTION_REQUEST, RHETORICAL, MULTIPLE_CHOICE etc.
We could not find much work related to mining of performance appraisals data. Pawar et al. BIBREF18 uses kernel-based classification to classify sentences in both performance appraisal text and product reviews into classes SUGGESTION, APPRECIATION, COMPLAINT. Apte et al. BIBREF6 provides two algorithms for matching the descriptions of goals or tasks assigned to employees to a standard template of model goals. One algorithm is based on the co-training framework and uses goal descriptions and self-appraisal comments as two separate perspectives. The second approach uses semantic similarity under a weak supervision framework. Ramrakhiyani et al. BIBREF5 proposes label propagation algorithms to discover aspects in supervisor assessments in performance appraisals, where an aspect is modelled as a verb-noun pair (e.g. conduct training, improve coding).
## Dataset
In this paper, we used the supervisor assessment and peer feedback text produced during the performance appraisal of 4528 employees in a large multi-national IT company. The corpus of supervisor assessment has 26972 sentences. The summary statistics about the number of words in a sentence is: min:4 max:217 average:15.5 STDEV:9.2 Q1:9 Q2:14 Q3:19.
## Sentence Classification
The PA corpus contains several classes of sentences that are of interest. In this paper, we focus on three important classes of sentences viz., sentences that discuss strengths (class STRENGTH), weaknesses of employees (class WEAKNESS) and suggestions for improving her performance (class SUGGESTION). The strengths or weaknesses are mostly about the performance in work carried out, but sometimes they can be about the working style or other personal qualities. The classes WEAKNESS and SUGGESTION are somewhat overlapping; e.g., a suggestion may address a perceived weakness. Following are two example sentences in each class.
STRENGTH:
WEAKNESS:
SUGGESTION:
Several linguistic aspects of these classes of sentences are apparent. The subject is implicit in many sentences. The strengths are often mentioned as either noun phrases (NP) with positive adjectives (Excellent technology leadership) or positive nouns (engineering strength) or through verbs with positive polarity (dedicated) or as verb phrases containing positive adjectives (delivers innovative solutions). Similarly for weaknesses, where negation is more frequently used (presentations are not his forte), or alternatively, the polarities of verbs (avoid) or adjectives (poor) tend to be negative. However, sometimes the form of both the strengths and weaknesses is the same, typically a stand-alone sentiment-neutral NP, making it difficult to distinguish between them; e.g., adherence to timing or timely closure. Suggestions often have an imperative mood and contain secondary verbs such as need to, should, has to. Suggestions are sometimes expressed using comparatives (better process compliance). We built a simple set of patterns for each of the 3 classes on the POS-tagged form of the sentences. We use each set of these patterns as an unsupervised sentence classifier for that class. If a particular sentence matched with patterns for multiple classes, then we have simple tie-breaking rules for picking the final class. The pattern for the STRENGTH class looks for the presence of positive words / phrases like takes ownership, excellent, hard working, commitment, etc. Similarly, the pattern for the WEAKNESS class looks for the presence of negative words / phrases like lacking, diffident, slow learner, less focused, etc. The SUGGESTION pattern not only looks for keywords like should, needs to but also for POS based pattern like “a verb in the base form (VB) in the beginning of a sentence”.
We randomly selected 2000 sentences from the supervisor assessment corpus and manually tagged them (dataset D1). This labelled dataset contained 705, 103, 822 and 370 sentences having the class labels STRENGTH, WEAKNESS, SUGGESTION or OTHER respectively. We trained several multi-class classifiers on this dataset. Table TABREF10 shows the results of 5-fold cross-validation experiments on dataset D1. For the first 5 classifiers, we used their implementation from the SciKit Learn library in Python (scikit-learn.org). The features used for these classifiers were simply the sentence words along with their frequencies. For the last 2 classifiers (in Table TABREF10 ), we used our own implementation. The overall accuracy for a classifier is defined as INLINEFORM0 , where the denominator is 2000 for dataset D1. Note that the pattern-based approach is unsupervised i.e., it did not use any training data. Hence, the results shown for it are for the entire dataset and not based on cross-validation.
## Comparison with Sentiment Analyzer
We also explored whether a sentiment analyzer can be used as a baseline for identifying the class labels STRENGTH and WEAKNESS. We used an implementation of sentiment analyzer from TextBlob to get a polarity score for each sentence. Table TABREF13 shows the distribution of positive, negative and neutral sentiments across the 3 class labels STRENGTH, WEAKNESS and SUGGESTION. It can be observed that distribution of positive and negative sentiments is almost similar in STRENGTH as well as SUGGESTION sentences, hence we can conclude that the information about sentiments is not much useful for our classification problem.
## Discovering Clusters within Sentence Classes
After identifying sentences in each class, we can now answer question (1) in Section SECREF1 . From 12742 sentences predicted to have label STRENGTH, we extract nouns that indicate the actual strength, and cluster them using a simple clustering algorithm which uses the cosine similarity between word embeddings of these nouns. We repeat this for the 9160 sentences with predicted label WEAKNESS or SUGGESTION as a single class. Tables TABREF15 and TABREF16 show a few representative clusters in strengths and in weaknesses, respectively. We also explored clustering 12742 STRENGTH sentences directly using CLUTO BIBREF19 and Carrot2 Lingo BIBREF20 clustering algorithms. Carrot2 Lingo discovered 167 clusters and also assigned labels to these clusters. We then generated 167 clusters using CLUTO as well. CLUTO does not generate cluster labels automatically, hence we used 5 most frequent words within the cluster as its labels. Table TABREF19 shows the largest 5 clusters by both the algorithms. It was observed that the clusters created by CLUTO were more meaningful and informative as compared to those by Carrot2 Lingo. Also, it was observed that there is some correspondence between noun clusters and sentence clusters. E.g. the nouns cluster motivation expertise knowledge talent skill (Table TABREF15 ) corresponds to the CLUTO sentence cluster skill customer management knowledge team (Table TABREF19 ). But overall, users found the nouns clusters to be more meaningful than the sentence clusters.
## PA along Attributes
In many organizations, PA is done from a predefined set of perspectives, which we call attributes. Each attribute covers one specific aspect of the work done by the employees. This has the advantage that we can easily compare the performance of any two employees (or groups of employees) along any given attribute. We can correlate various performance attributes and find dependencies among them. We can also cluster employees in the workforce using their supervisor ratings for each attribute to discover interesting insights into the workforce. The HR managers in the organization considered in this paper have defined 15 attributes (Table TABREF20 ). Each attribute is essentially a work item or work category described at an abstract level. For example, FUNCTIONAL_EXCELLENCE covers any tasks, goals or activities related to the software engineering life-cycle (e.g., requirements analysis, design, coding, testing etc.) as well as technologies such as databases, web services and GUI.
In the example in Section SECREF4 , the first sentence (which has class STRENGTH) can be mapped to two attributes: FUNCTIONAL_EXCELLENCE and BUILDING_EFFECTIVE_TEAMS. Similarly, the third sentence (which has class WEAKNESS) can be mapped to the attribute INTERPERSONAL_EFFECTIVENESS and so forth. Thus, in order to answer the second question in Section SECREF1 , we need to map each sentence in each of the 3 classes to zero, one, two or more attributes, which is a multi-class multi-label classification problem.
We manually tagged the same 2000 sentences in Dataset D1 with attributes, where each sentence may get 0, 1, 2, etc. up to 15 class labels (this is dataset D2). This labelled dataset contained 749, 206, 289, 207, 91, 223, 191, 144, 103, 80, 82, 42, 29, 15, 24 sentences having the class labels listed in Table TABREF20 in the same order. The number of sentences having 0, 1, 2, or more than 2 attributes are: 321, 1070, 470 and 139 respectively. We trained several multi-class multi-label classifiers on this dataset. Table TABREF21 shows the results of 5-fold cross-validation experiments on dataset D2.
Precision, Recall and F-measure for this multi-label classification are computed using a strategy similar to the one described in BIBREF21 . Let INLINEFORM0 be the set of predicted labels and INLINEFORM1 be the set of actual labels for the INLINEFORM2 instance. Precision and recall for this instance are computed as follows: INLINEFORM3
It can be observed that INLINEFORM0 would be undefined if INLINEFORM1 is empty and similarly INLINEFORM2 would be undefined when INLINEFORM3 is empty. Hence, overall precision and recall are computed by averaging over all the instances except where they are undefined. Instance-level F-measure can not be computed for instances where either precision or recall are undefined. Therefore, overall F-measure is computed using the overall precision and recall.
## Summarization of Peer Feedback using ILP
The PA system includes a set of peer feedback comments for each employee. To answer the third question in Section SECREF1 , we need to create a summary of all the peer feedback comments about a given employee. As an example, following are the feedback comments from 5 peers of an employee.
The individual sentences in the comments written by each peer are first identified and then POS tags are assigned to each sentence. We hypothesize that a good summary of these multiple comments can be constructed by identifying a set of important text fragments or phrases. Initially, a set of candidate phrases is extracted from these comments and a subset of these candidate phrases is chosen as the final summary, using Integer Linear Programming (ILP). The details of the ILP formulation are shown in Table TABREF36 . As an example, following is the summary generated for the above 5 peer comments.
humble nature, effective communication, technical expertise, always supportive, vast knowledge
Following rules are used to identify candidate phrases:
Various parameters are used to evaluate a candidate phrase for its importance. A candidate phrase is more important:
A complete list of parameters is described in detail in Table TABREF36 .
There is a trivial constraint INLINEFORM0 which makes sure that only INLINEFORM1 out of INLINEFORM2 candidate phrases are chosen. A suitable value of INLINEFORM3 is used for each employee depending on number of candidate phrases identified across all peers (see Algorithm SECREF6 ). Another set of constraints ( INLINEFORM4 to INLINEFORM5 ) make sure that at least one phrase is selected for each of the leadership attributes. The constraint INLINEFORM6 makes sure that multiple phrases sharing the same headword are not chosen at a time. Also, single word candidate phrases are chosen only if they are adjectives or nouns with lexical category noun.attribute. This is imposed by the constraint INLINEFORM7 . It is important to note that all the constraints except INLINEFORM8 are soft constraints, i.e. there may be feasible solutions which do not satisfy some of these constraints. But each constraint which is not satisfied, results in a penalty through the use of slack variables. These constraints are described in detail in Table TABREF36 .
The objective function maximizes the total importance score of the selected candidate phrases. At the same time, it also minimizes the sum of all slack variables so that the minimum number of constraints are broken.
INLINEFORM0 : No. of candidate phrases INLINEFORM1 : No. of phrases to select as part of summary
INLINEFORM0 INLINEFORM1 INLINEFORM2 INLINEFORM3 INLINEFORM4 INLINEFORM5 INLINEFORM6 INLINEFORM7 INLINEFORM8
INLINEFORM0 and INLINEFORM1 INLINEFORM2 INLINEFORM3 INLINEFORM4 INLINEFORM5 INLINEFORM6
INLINEFORM0 (For determining number of phrases to select to include in summary)
## Evaluation of auto-generated summaries
We considered a dataset of 100 employees, where for each employee multiple peer comments were recorded. Also, for each employee, a manual summary was generated by an HR personnel. The summaries generated by our ILP-based approach were compared with the corresponding manual summaries using the ROUGE BIBREF22 unigram score. For comparing performance of our ILP-based summarization algorithm, we explored a few summarization algorithms provided by the Sumy package. A common parameter which is required by all these algorithms is number of sentences keep in the final summary. ILP-based summarization requires a similar parameter K, which is automatically decided based on number of total candidate phrases. Assuming a sentence is equivalent to roughly 3 phrases, for Sumy algorithms, we set number of sentences parameter to the ceiling of K/3. Table TABREF51 shows average and standard deviation of ROUGE unigram f1 scores for each algorithm, over the 100 summaries. The performance of ILP-based summarization is comparable with the other algorithms, as the two sample t-test does not show statistically significant difference. Also, human evaluators preferred phrase-based summary generated by our approach to the other sentence-based summaries.
## Conclusions and Further Work
In this paper, we presented an analysis of the text generated in Performance Appraisal (PA) process in a large multi-national IT company. We performed sentence classification to identify strengths, weaknesses and suggestions for improvements found in the supervisor assessments and then used clustering to discover broad categories among them. As this is non-topical classification, we found that SVM with ADWS kernel BIBREF18 produced the best results. We also used multi-class multi-label classification techniques to match supervisor assessments to predefined broad perspectives on performance. Logistic Regression classifier was observed to produce the best results for this topical classification. Finally, we proposed an ILP-based summarization technique to produce a summary of peer feedback comments for a given employee and compared it with manual summaries.
The PA process also generates much structured data, such as supervisor ratings. It is an interesting problem to compare and combine the insights from discovered from structured data and unstructured text. Also, we are planning to automatically discover any additional performance attributes to the list of 15 attributes currently used by HR.
| 10
|
1712.05999
|
Characterizing Political Fake News in Twitter by its Meta-Data
|
# Characterizing Political Fake News in Twitter by its Meta-Data
## Abstract
This article presents a preliminary approach towards characterizing political fake news on Twitter through the analysis of their meta-data. In particular, we focus on more than 1.5M tweets collected on the day of the election of Donald Trump as 45th president of the United States of America. We use the meta-data embedded within those tweets in order to look for differences between tweets containing fake news and tweets not containing them. Specifically, we perform our analysis only on tweets that went viral, by studying proxies for users' exposure to the tweets, by characterizing accounts spreading fake news, and by looking at their polarization. We found significant differences on the distribution of followers, the number of URLs on tweets, and the verification of the users.
10pt
1.10pt
[ Characterizing Political Fake News in Twitter by its Meta-DataJulio Amador Díaz LópezAxel Oehmichen Miguel Molina-Solana( j.amador, axelfrancois.oehmichen11, mmolinas@imperial.ac.uk ) Imperial College London This article presents a preliminary approach towards characterizing political fake news on Twitter through the analysis of their meta-data. In particular, we focus on more than 1.5M tweets collected on the day of the election of Donald Trump as 45th president of the United States of America. We use the meta-data embedded within those tweets in order to look for differences between tweets containing fake news and tweets not containing them. Specifically, we perform our analysis only on tweets that went viral, by studying proxies for users' exposure to the tweets, by characterizing accounts spreading fake news, and by looking at their polarization. We found significant differences on the distribution of followers, the number of URLs on tweets, and the verification of the users.
]
## Introduction
While fake news, understood as deliberately misleading pieces of information, have existed since long ago (e.g. it is not unusual to receive news falsely claiming the death of a celebrity), the term reached the mainstream, particularly so in politics, during the 2016 presidential election in the United States BIBREF0 . Since then, governments and corporations alike (e.g. Google BIBREF1 and Facebook BIBREF2 ) have begun efforts to tackle fake news as they can affect political decisions BIBREF3 . Yet, the ability to define, identify and stop fake news from spreading is limited.
Since the Obama campaign in 2008, social media has been pervasive in the political arena in the United States. Studies report that up to 62% of American adults receive their news from social media BIBREF4 . The wide use of platforms such as Twitter and Facebook has facilitated the diffusion of fake news by simplifying the process of receiving content with no significant third party filtering, fact-checking or editorial judgement. Such characteristics make these platforms suitable means for sharing news that, disguised as legit ones, try to confuse readers.
Such use and their prominent rise has been confirmed by Craig Silverman, a Canadian journalist who is a prominent figure on fake news BIBREF5 : “In the final three months of the US presidential campaign, the top-performing fake election news stories on Facebook generated more engagement than the top stories from major news outlet”.
Our current research hence departs from the assumption that social media is a conduit for fake news and asks the question of whether fake news (as spam was some years ago) can be identified, modelled and eventually blocked. In order to do so, we use a sample of more that 1.5M tweets collected on November 8th 2016 —election day in the United States— with the goal of identifying features that tweets containing fake news are likely to have. As such, our paper aims to provide a preliminary characterization of fake news in Twitter by looking into meta-data embedded in tweets. Considering meta-data as a relevant factor of analysis is in line with findings reported by Morris et al. BIBREF6 . We argue that understanding differences between tweets containing fake news and regular tweets will allow researchers to design mechanisms to block fake news in Twitter.
Specifically, our goals are: 1) compare the characteristics of tweets labelled as containing fake news to tweets labelled as not containing them, 2) characterize, through their meta-data, viral tweets containing fake news and the accounts from which they originated, and 3) determine the extent to which tweets containing fake news expressed polarized political views.
For our study, we used the number of retweets to single-out those that went viral within our sample. Tweets within that subset (viral tweets hereafter) are varied and relate to different topics. We consider that a tweet contains fake news if its text falls within any of the following categories described by Rubin et al. BIBREF7 (see next section for the details of such categories): serious fabrication, large-scale hoaxes, jokes taken at face value, slanted reporting of real facts and stories where the truth is contentious. The dataset BIBREF8 , manually labelled by an expert, has been publicly released and is available to researchers and interested parties.
From our results, the following main observations can be made:
Our findings resonate with similar work done on fake news such as the one from Allcot and Gentzkow BIBREF9 . Therefore, even if our study is a preliminary attempt at characterizing fake news on Twitter using only their meta-data, our results provide external validity to previous research. Moreover, our work not only stresses the importance of using meta-data, but also underscores which parameters may be useful to identify fake news on Twitter.
The rest of the paper is organized as follows. The next section briefly discusses where this work is located within the literature on fake news and contextualizes the type of fake news we are studying. Then, we present our hypotheses, the data, and the methodology we follow. Finally, we present our findings, conclusions of this study, and future lines of work.
## Defining Fake news
Our research is connected to different strands of academic knowledge related to the phenomenon of fake news. In relation to Computer Science, a recent survey by Conroy and colleagues BIBREF10 identifies two popular approaches to single-out fake news. On the one hand, the authors pointed to linguistic approaches consisting in using text, its linguistic characteristics and machine learning techniques to automatically flag fake news. On the other, these researchers underscored the use of network approaches, which make use of network characteristics and meta-data, to identify fake news.
With respect to social sciences, efforts from psychology, political science and sociology, have been dedicated to understand why people consume and/or believe misinformation BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 . Most of these studies consistently reported that psychological biases such as priming effects and confirmation bias play an important role in people ability to discern misinformation.
In relation to the production and distribution of fake news, a recent paper in the field of Economics BIBREF9 found that most fake news sites use names that resemble those of legitimate organizations, and that sites supplying fake news tend to be short-lived. These authors also noticed that fake news items are more likely shared than legitimate articles coming from trusted sources, and they tend to exhibit a larger level of polarization.
The conceptual issue of how to define fake news is a serious and unresolved issue. As the focus of our work is not attempting to offer light on this, we will rely on work by other authors to describe what we consider as fake news. In particular, we use the categorization provided by Rubin et al. BIBREF7 . The five categories they described, together with illustrative examples from our dataset, are as follows:
## Research Hypotheses
Previous works on the area (presented in the section above) suggest that there may be important determinants for the adoption and diffusion of fake news. Our hypotheses builds on them and identifies three important dimensions that may help distinguishing fake news from legit information:
Taking those three dimensions into account, we propose the following hypotheses about the features that we believe can help to identify tweets containing fake news from those not containing them. They will be later tested over our collected dataset.
Exposure.
Characterization.
Polarization.
## Data and Methodology
For this study, we collected publicly available tweets using Twitter's public API. Given the nature of the data, it is important to emphasize that such tweets are subject to Twitter's terms and conditions which indicate that users consent to the collection, transfer, manipulation, storage, and disclosure of data. Therefore, we do not expect ethical, legal, or social implications from the usage of the tweets. Our data was collected using search terms related to the presidential election held in the United States on November 8th 2016. Particularly, we queried Twitter's streaming API, more precisely the filter endpoint of the streaming API, using the following hashtags and user handles: #MyVote2016, #ElectionDay, #electionnight, @realDonaldTrump and @HillaryClinton. The data collection ran for just one day (Nov 8th 2016).
One straightforward way of sharing information on Twitter is by using the retweet functionality, which enables a user to share a exact copy of a tweet with his followers. Among the reasons for retweeting, Body et al. BIBREF15 reported the will to: 1) spread tweets to a new audience, 2) to show one’s role as a listener, and 3) to agree with someone or validate the thoughts of others. As indicated, our initial interest is to characterize tweets containing fake news that went viral (as they are the most harmful ones, as they reach a wider audience), and understand how it differs from other viral tweets (that do not contain fake news). For our study, we consider that a tweet went viral if it was retweeted more than 1000 times.
Once we have the dataset of viral tweets, we eliminated duplicates (some of the tweets were collected several times because they had several handles) and an expert manually inspected the text field within the tweets to label them as containing fake news, or not containing them (according to the characterization presented before). This annotated dataset BIBREF8 is publicly available and can be freely reused.
Finally, we use the following fields within tweets (from the ones returned by Twitter's API) to compare their distributions and look for differences between viral tweets containing fake news and viral tweets not containing fake news:
In the following section, we provide graphical descriptions of the distribution of each of the identified attributes for the two sets of tweets (those labelled as containing fake news and those labelled as not containing them). Where appropriate, we normalized and/or took logarithms of the data for better representation. To gain a better understanding of the significance of those differences, we use the Kolmogorov-Smirnov test with the null hypothesis that both distributions are equal.
## Results
The sample collected consisted on 1 785 855 tweets published by 848 196 different users. Within our sample, we identified 1327 tweets that went viral (retweeted more than 1000 times by the 8th of November 2016) produced by 643 users. Such small subset of viral tweets were retweeted on 290 841 occasions in the observed time-window.
The 1327 `viral' tweets were manually annotated as containing fake news or not. The annotation was carried out by a single person in order to obtain a consistent annotation throughout the dataset. Out of those 1327 tweets, we identified 136 as potentially containing fake news (according to the categories previously described), and the rest were classified as `non containing fake news'. Note that the categorization is far from being perfect given the ambiguity of fake news themselves and human judgement involved in the process of categorization. Because of this, we do not claim that this dataset can be considered a ground truth.
The following results detail characteristics of these tweets along the previously mentioned dimensions. Table TABREF23 reports the actual differences (together with their associated p-values) of the distributions of viral tweets containing fake news and viral tweets not containing them for every variable considered.
## Exposure
Figure FIGREF24 shows that, in contrast to other kinds of viral tweets, those containing fake news were created more recently. As such, Twitter users were exposed to fake news related to the election for a shorter period of time.
However, in terms of retweets, Figure FIGREF25 shows no apparent difference between containing fake news or not containing them. That is confirmed by the Kolmogorov-Smirnoff test, which does not discard the hypothesis that the associated distributions are equal.
In relation to the number of favourites, users that generated at least a viral tweet containing fake news appear to have, on average, less favourites than users that do not generate them. Figure FIGREF26 shows the distribution of favourites. Despite the apparent visual differences, the difference are not statistically significant.
Finally, the number of hashtags used in viral fake news appears to be larger than those in other viral tweets. Figure FIGREF27 shows the density distribution of the number of hashtags used. However, once again, we were not able to find any statistical difference between the average number of hashtags in a viral tweet and the average number of hashtags in viral fake news.
## Characterization
We found that 82 users within our sample were spreading fake news (i.e. they produced at least one tweet which was labelled as fake news). Out of those, 34 had verified accounts, and the rest were unverified. From the 48 unverified accounts, 6 have been suspended by Twitter at the date of writing, 3 tried to imitate legitimate accounts of others, and 4 accounts have been already deleted. Figure FIGREF28 shows the proportion of verified accounts to unverified accounts for viral tweets (containing fake news vs. not containing fake news). From the chart, it is clear that there is a higher chance of fake news coming from unverified accounts.
Turning to friends, accounts distributing fake news appear to have, on average, the same number of friends than those distributing tweets with no fake news. However, the density distribution of friends from the accounts (Figure FIGREF29 ) shows that there is indeed a statistically significant difference in their distributions.
If we take into consideration the number of followers, accounts generating viral tweets with fake news do have a very different distribution on this dimension, compared to those accounts generating viral tweets with no fake news (see Figure FIGREF30 ). In fact, such differences are statistically significant.
A useful representation for friends and followers is the ratio between friends/followers. Figures FIGREF31 and FIGREF32 show this representation. Notice that accounts spreading viral tweets with fake news have, on average, a larger ratio of friends/followers. The distribution of those accounts not generating fake news is more evenly distributed.
With respect to the number of mentions, Figure FIGREF33 shows that viral tweets labelled as containing fake news appear to use mentions to other users less frequently than viral tweets not containing fake news. In other words, tweets containing fake news mostly contain 1 mention, whereas other tweets tend to have two). Such differences are statistically significant.
The analysis (Figure FIGREF34 ) of the presence of media in the tweets in our dataset shows that tweets labelled as not containing fake news appear to present more media elements than those labelled as fake news. However, the difference is not statistically significant.
On the other hand, Figure FIGREF35 shows that viral tweets containing fake news appear to include more URLs to other sites than viral tweets that do not contain fake news. In fact, the difference between the two distributions is statistically significant (assuming INLINEFORM0 ).
## Polarization
Finally, manual inspection of the text field of those viral tweets labelled as containing fake news shows that 117 of such tweets expressed support for Donald Trump, while only 8 supported Hillary Clinton. The remaining tweets contained fake news related to other topics, not expressing support for any of the candidates.
## Discussion
As a summary, and constrained by our existing dataset, we made the following observations regarding differences between viral tweets labelled as containing fake news and viral tweets labelled as not containing them:
These findings (related to our initial hypothesis in Table TABREF44 ) clearly suggest that there are specific pieces of meta-data about tweets that may allow the identification of fake news. One such parameter is the time of exposure. Viral tweets containing fake news are shorter-lived than those containing other type of content. This notion seems to resonate with our findings showing that a number of accounts spreading fake news have already been deleted or suspended by Twitter by the time of writing. If one considers that researchers using different data have found similar results BIBREF9 , it appears that the lifetime of accounts, together with the age of the questioned viral content could be useful to identify fake news. In the light of this finding, accounts newly created should probably put under higher scrutiny than older ones. This in fact, would be a nice a-priori bias for a Bayesian classifier.
Accounts spreading fake news appear to have a larger proportion of friends/followers (i.e. they have, on average, the same number of friends but a smaller number of followers) than those spreading viral content only. Together with the fact that, on average, tweets containing fake news have more URLs than those spreading viral content, it is possible to hypothesize that, both, the ratio of friends/followers of the account producing a viral tweet and number of URLs contained in such a tweet could be useful to single-out fake news in Twitter. Not only that, but our finding related to the number of URLs is in line with intuitions behind the incentives to create fake news commonly found in the literature BIBREF9 (in particular that of obtaining revenue through click-through advertising).
Finally, it is interesting to notice that the content of viral fake news was highly polarized. This finding is also in line with those of Alcott et al. BIBREF9 . This feature suggests that textual sentiment analysis of the content of tweets (as most researchers do), together with the above mentioned parameters from meta-data, may prove useful for identifying fake news.
## Conclusions
With the election of Donald Trump as President of the United States, the concept of fake news has become a broadly-known phenomenon that is getting tremendous attention from governments and media companies. We have presented a preliminary study on the meta-data of a publicly available dataset of tweets that became viral during the day of the 2016 US presidential election. Our aim is to advance the understanding of which features might be characteristic of viral tweets containing fake news in comparison with viral tweets without fake news.
We believe that the only way to automatically identify those deceitful tweets (i.e. containing fake news) is by actually understanding and modelling them. Only then, the automation of the processes of tagging and blocking these tweets can be successfully performed. In the same way that spam was fought, we anticipate fake news will suffer a similar evolution, with social platforms implementing tools to deal with them. With most works so far focusing on the actual content of the tweets, ours is a novel attempt from a different, but also complementary, angle.
Within the used dataset, we found there are differences around exposure, characteristics of accounts spreading fake news and the tone of the content. Those findings suggest that it is indeed possible to model and automatically detect fake news. We plan to replicate and validate our experiments in an extended sample of tweets (until 4 months after the US election), and tests the predictive power of the features we found relevant within our sample.
## Author Disclosure Statement
No competing financial interest exist.
| 12
|
1802.05574
|
Open Information Extraction on Scientific Text: An Evaluation
|
# Open Information Extraction on Scientific Text: An Evaluation
## Abstract
Open Information Extraction (OIE) is the task of the unsupervised creation of structured information from text. OIE is often used as a starting point for a number of downstream tasks including knowledge base construction, relation extraction, and question answering. While OIE methods are targeted at being domain independent, they have been evaluated primarily on newspaper, encyclopedic or general web text. In this article, we evaluate the performance of OIE on scientific texts originating from 10 different disciplines. To do so, we use two state-of-the-art OIE systems applying a crowd-sourcing approach. We find that OIE systems perform significantly worse on scientific text than encyclopedic text. We also provide an error analysis and suggest areas of work to reduce errors. Our corpus of sentences and judgments are made available.
## Introduction
This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/
The scientific literature is growing at a rapid rate BIBREF0 . To make sense of this flood of literature, for example, to extract cancer pathways BIBREF1 or find geological features BIBREF2 , increasingly requires the application of natural language processing. Given the diversity of information and its constant flux, the use of unsupervised or distantly supervised techniques are of interest BIBREF3 . In this paper, we investigate one such unsupervised method, namely, Open Information Extraction (OIE) BIBREF4 . OIE is the task of the unsupervised creation of structured information from text. OIE is often used as a starting point for a number of downstream tasks including knowledge base construction, relation extraction, and question answering BIBREF5 .
While OIE has been applied to the scientific literature before BIBREF6 , we have not found a systematic evaluation of OIE as applied to scientific publications. The most recent evaluations of OIE extraction tools BIBREF7 , BIBREF8 have instead looked at the performance of these tools on traditional NLP information sources (i.e. encyclopedic and news-wire text). Indeed, as BIBREF8 noted, there is little work on the evaluation of OIE systems. Thus, the goal of this paper is to evaluate the performance of the state of the art in OIE systems on scientific text.
Specifically, we aim to test two hypotheses:
Additionally, we seek to gain insight into the value of unsupervised approaches to information extraction and also provide information useful to implementors of these systems. We note that our evaluation differs from existing OIE evaluations in that we use crowd-sourcing annotations instead of expert annotators. This allows for a larger number of annotators to be used. All of our data, annotations and analyses are made openly available.
The rest of the paper is organized as follows. We begin with a discussion of existing evaluation approaches and then describe the OIE systems that we evaluated. We then proceed to describe the datasets used in the evaluation and the annotation process that was employed. This is followed by the results of the evaluation including an error analysis. Finally, we conclude.
## Existing Evaluation Approaches
OIE systems analyze sentences and emit relations between one predicate and two or more arguments (e.g. Washington :: was :: president). The arguments and predicates are not fixed to a given domain. (Note, that throughout this paper we use the word `triple” to refer interchangeably to binary relations.) Existing evaluation approaches for OIE systems have primarily taken a ground truth-based approach. Human annotators analyze sentences and determine correct relations to be extracted. Systems are then evaluated with respect to the overlap or similarity of their extractions to the ground truth annotations, allowing the standard metrics of precision and recall to be reported.
This seems sensible but is actually problematic because of different but equivalent representations of the information in an article. For example, consider the sentence “The patient was treated with Emtricitabine, Etravirine, and Darunavir”. One possible extraction is:
(The patient :: was treated with :: Emtricitabine, Etravirine, and Darunavir)
Another possible extraction is:
(The patient :: was treated with :: Emtricitabine)
(The patient :: was treated with :: Etravirine)
(The patient :: was treated with :: Darunavir)
Neither of these is wrong, but by choosing one approach or the other a pre-constructed gold set will falsely penalize a system that uses the other approach.
From such evaluations and their own cross dataset evaluation, BIBREF8 list the following common errors committed by OIE systems:
In our evaluation, we take a different approach. We do not define ground truth relation extractions from the sentences in advance. Instead, we manually judge the correctness of each extraction after the fact. We feel that this is the crux of the information extraction challenge. Is what is being extracted correct or not? This approach enables us to consider many more relations through the use of a crowd-sourced annotation process. Our evaluation approach is similar to the qualitative analysis performed in BIBREF8 and the evaluation performed in BIBREF7 . However, our evaluation is able to use more judges (5 instead of 2) because we apply crowd sourcing. For our labelling instructions, we adapted those used by BIBREF7 to the crowd sourcing setting.
As previously noted existing evaluations have also only looked at encyclopedic or newspaper corpora. Several systems (e.g. BIBREF4 , BIBREF9 ) have looked at text from the web as well, however, as far as we know, none have specifically looked at evaluation for scientific and medical text.
## Systems
We evaluate two OIE systems (i.e. extractors). The first, OpenIE 4 BIBREF5 , descends from two popular OIE systems OLLIE BIBREF10 and Reverb BIBREF10 . We view this as a baseline system. The second was MinIE BIBREF7 , which is reported as performing better than OLLIE, ClauseIE BIBREF9 and Stanford OIE BIBREF9 . MinIE focuses on the notion of minimization - producing compact extractions from sentences. In our experience using OIE on scientific text, we have found that these systems often produce overly specific extractions that do not provide the redundancy useful for downstream tasks. Hence, we thought this was a useful package to explore.
We note that both OpenIE 4 and MiniIE support relation extractions that go beyond binary tuples, supporting the extraction of n-ary relations. We note that the most recent version of Open IE (version 5) is focused on n-ary relations. For ease of judgement, we focused on binary relations. Additionally, both systems support the detection of negative relations.
In terms of settings, we used the off the shelf settings for OpenIE 4. For MinIE, we used their “safe mode" option, which uses slightly more aggressive minimization than the standard setting. In the recent evaluation of MiniIE, this setting performed roughly on par with the default options BIBREF7 . Driver code showing how we ran each system is available.
## Datasets
We used two different data sources in our evaluation. The first dataset (WIKI) was the same set of 200 sentences from Wikipedia used in BIBREF7 . These sentences were randomly selected by the creators of the dataset. This choice allows for a rough comparison between our results and theirs.
The second dataset (SCI) was a set of 220 sentences from the scientific literature. We sourced the sentences from the OA-STM corpus. This corpus is derived from the 10 most published in disciplines. It includes 11 articles each from the following domains: agriculture, astronomy, biology, chemistry, computer science, earth science, engineering, materials science, math, and medicine. The article text is made freely available and the corpus provides both an XML and a simple text version of each article.
We randomly selected 2 sentences with more than two words from each paper using the simple text version of the paper. We maintained the id of the source article and the line number for each sentence.
## Annotation Process
We employed the following annotation process. Each OIE extractor was applied to both datasets with the settings described above. This resulted in the generation of triples for 199 of the 200 WIKI sentences and 206 of the 220 SCI sentences. That is there were some sentences in which no triples were extracted. We discuss later the sentences in which no triples were extracted. In total 2247 triples were extracted.
The sentences and their corresponding triples were then divided. Each task contained 10 sentences and all of their unique corresponding triples from a particular OIE systems. Half of the ten sentences were randomly selected from SCI and the other half were randomly selected from WIKI. Crowd workers were asked to mark whether a triple was correct, namely, did the triple reflect the consequence of the sentence. Examples of correct and incorrect triples were provided. Complete labelling instructions and the presentation of the HITS can be found with the dataset. All triples were labelled by at least 5 workers.
Note, to ensure the every HIT had 10 sentences, some sentences were duplicated. Furthermore, we did not mandate that all workers complete all HITS.
We followed recommended practices for the use of crowd sourcing in linguistics BIBREF11 . We used Amazon Mechanical Turk as a means to present the sentences and their corresponding triples to a crowd for annotation. Within Mechanical Turk tasks are called Human Intelligence Tasks (HITs). To begin, we collected a small set of sentences and triples with known correct answers. We did this by creating a series of internal HITs and loaded them the Mechanical Turk development environment called the Mechanical Turk Sandbox. The HITs were visible to a trusted group of colleagues who were asked to complete the HITs.
Having an internal team of workers attempt HITs provides us with two valuable aspects of the eventual production HITs. First, internal users are able to provide feedback related to usability and clarity of the task. They were asked to read the instructions and let us know if there was anything that was unclear. After taking the HITs, they are able to ask questions about anomalies or confusing situations they encounter and allow us to determine if specific types of HITs are either not appropriate for the task or might need further explanation in the instructions. In addition to the internal users direct feedback, we were also able to use the Mechanical Turk Requester functionality to monitor how long (in minutes and seconds) it took each worker to complete each HIT. This would come into factor how we decided on how much to pay each Worker per HIT after they were made available to the public.
The second significant outcome from the internal annotations was the generation of a set of `expected' correct triples. Having a this set of annotations is an integral part of two aspects of our crowdsourcing process. First, it allows us to create a qualification HIT. A qualification HIT is a HIT that is made available to the public with the understanding the Workers will be evaluated based on how closely they matched the annotations of the internal annotators. Based upon this, the Workers with the most matches would be invited to work on additional tasks. Second, we are able to add the internal set of triples randomly amongst the other relations we were seeking to have annotated. This allows us to monitor quality of the individual Workers over the course of the project. Note, none of this data was used in the actual evaluation. It was only for the purposes of qualifying Workers.
We are sensitive to issues that other researchers have in regards to Mechanical Turk Workers earning fair payment in exchange for their contributions to the HITs BIBREF12 . We used the time estimates from our internal annotation to price the task in order to be above US minimum wage. All workers were qualified before being issued tasks. Overall, we employed 10 crowd workers. On average it took 30 minutes for a worker to complete a HIT. In line with BIBREF13 , we monitored for potential non-performance or spam by looking for long response times and consecutive submitted results. We saw no indicators of low quality responses.
## Judgement Data and Inter-Annotator Agreement
In total, 11262 judgements were obtained after running the annotation process. Every triple had at least 5 judgements from different annotators. All judgement data is made available. The proportion of overall agreement between annotators is 0.76 with a standard deviation of 0.25 on whether a triple is consequence of the given sentence. We also calculated inter-annotator agreement statistics. Using Krippendorff's alpha inter-annotator agreement was 0.44. This calculation was performed over all data and annotators as Krippendorff's alpha is designed to account for missing data and work across more than two annotators. Additionally, Fleiss' Kappa and Scott's pi were calculated pairwise between all annotators where there were overlapping ratings (i.e. raters had rated at least one triple in common). The average Fleiss's Kappa was 0.41 and the average of Scott's pi was 0.37. Using BIBREF14 as a guide, we interpret these statistics as suggesting there is moderate agreement between annotators and that agreement is above random chance. This moderate level of agreement is to be expected as the task itself can be difficult and requires judgement from the annotators at the margin.
Table 1 shows examples of triples that were associated with higher disagreement between annotators. One can see for example, in the third example, that annotators might be confused by the use of a pronoun (him). Another example is in the last sentence of the table, where one can see that there might be disagreement on whether the subsequent prepositional phrase behind light microscope analysis should be included as part of the extracted triple.
We take the variability of judgements into account when using this data to compute the performance of the two extraction tools. Hence, to make assessments as to whether a triple correctly reflects the content from which it is extracted, we rely on the unanimous positive agreement between crowd workers. That is to say that if we have 100% inter-annotator agreement that a triple was correctly extracted we label it as correct.
## Experimental Results
Table 2 show the results for the combinations of systems and data sources. The Correct Triples column contains the number of triples that are labelled as being correct by all annotators. Total Triples are the total number of triples extracted by the given systems over the specified data. Precision is calculated as typical where Correct Triples are treated as true positives. On average, 3.1 triples were extracted per sentence.
Figure 1 shows the performance of extractors in terms of precision as inter-annotator agreement decreases. In this figure, we look only at agreement on triples where the majority agree that the triple is correct. Furthermore, to ease comparison, we only consider triples with 5 judgements this excludes 9 triples. We indicate not only the pair-wise inter-annotator agreement but also the number of annotators who have judged a triple to be correct. For example, at the 40% agreement level at least 3 annotators have agreed that a triple is true. The figure separates the results by extractor and by data source.
We see that as expected the amount of triples agreed to as correct grows larger as we relax the requirement for agreement. For example, analyzing Open IE's results, at the 100% agreement level we see a precision of 0.56 whereas at the 40% agreement level we see a precision of 0.78. Table 3 shows the total number of correct extractions at the three agreement levels.
## Testing H1: Comparing the Performance of OIE on Scientific vs. Encyclopedic Text
From the data, we see that extractors perform better on sentences from Wikipedia (0.54 P) than scientific text (0.34 P). Additionally, we see that there is higher annotator agreement on whether triples extracted from Wikipedia and scientific text are correct or incorrect: 0.80 - SD 0.24 (WIKI) vs. 0.72 - SD 0.25 (SCI). A similar difference in agreement is observed when only looking at triples that are considered to be correct by the majority of annotators: 0.87 - SD 0.21 (WIKI) vs. 0.78 - SD 0.25 (SCI) . In both cases, the difference is significant with p-values $<$ 0.01 using Welch's t-test. The differences between data sources are also seen when looking at the individual extraction tools. For instance, for Open IE 4 the precision is 0.19 higher for wikipedia extractions over those from scientific text. With this evidence, we reject our first hypothesis that the performance of these extractors are similar across data sources.
## Testing H2: Comparing the Performance of Systems
We also compare the output of the two extractors. In terms precision, Open IE 4 performs much better across the two datasets (0.56P vs 0.39P). Looking at triples considered to be correct by the majority of annotators, we see that Open IE 4 has higher inter-annotator agreement 0.87 - SD 0.22 (Open IE) vs 0.81 - SD 0.24 (MinIE). Focusing on scientific and medical text (SCI), again where the triples are majority annotated as being correct, Open IE has higher inter-annotator agreement (Open IE: 0.83 - SD 0.24 vs MiniIE: 0.76 - SD 0.25). In both cases, the difference is significant with p-values $<$ 0.01 using Welch's t-test. This leads us to conclude that Open IE produces triples that annotators are more likely to agree as being correct.
MinIE provides many more correct extractions than OpenIE 4 (935 more across both datasets). The true recall numbers of the two systems can not be calculated with the data available, but the 40% difference in the numbers of correct extractions is strong evidence that the two systems do not have equivalent behavior.
A third indication of differences in their outputs comes from examining the complexity of the extracted relations. Open IE 4 generates longer triples on average (11.5 words) vs. 8.5 words for MinIE across all argument positions. However, Open IE 4 generates shorter relation types than MinIE (Open IE - 3.7 words; MiniIE 6.27 words) and the standard deviation in terms of word length is much more compact for Open IE 4 - 1 word vs 3 words for MinIE. Overall, our conclusion is that Open IE 4 performs better than MinIE both in terms of precision and compactness of relation types, while not matching MinIE's recall, and thus we reject our second hypothesis.
## Other Observations
The amount of triples extracted from the scientific text is slightly larger than that extracted from the Wikipedia text. This follows from the fact that the scientific sentences are on average roughly 7 words longer than encyclopedic text.
The results of our experiment also confirm the notion that an unsupervised approach to extracting relations is important. We have identified 698 unique relation types that are part of triples agreed to be correct by all annotators. This number of relation types is derived from only 400 sentences. While not every relation type is essential for downstream tasks, it is clear that building specific extractors for each relation type in a supervised setting would be difficult.
## Error Analysis
We now look more closely at the various errors that were generated by the two extractors.
Table 4 shows the sentences in which neither extractor produced triples. We see 3 distinct groups. The first are phrases that are incomplete sentences usually originating from headings (e.g. Materials and methods). The next group are descriptive headings potentially coming from paper titles or figure captions. We also see a group with more complex prepositional phrases. In general, these errors could be avoided by being more selective of the sentences used for random selection. Additionally, these systems could look at potentially just extracting noun phrases with variable relation types, hence, expressing a cooccurrence relation.
We also looked at where there was complete agreement by all annotators that a triple extraction was incorrect. In total there were 138 of these triples originating from 76 unique sentences. There were several patterns that appeared in these sentences.
We also see similar errors to those pointed out by BIBREF8 , namely, uninformative extractions, the difficulty in handling n-ary relations that are latent in the text, difficulties handling negations, and very large argument lengths. In general, these errors together point to several areas for further improvement including:
## Conclusion
The pace of change in the scientific literature means that interconnections and facts in the form of relations between entities are constantly being created. Open information extraction provides an important tool to keep up with that pace of change. We have provided evidence that unsupervised techniques are needed to be able to deal with the variety of relations present in text. The work presented here provides an independent evaluation of these tools in their use on scientific text. Past evaluations have focused on encyclopedic or news corpora which often have simpler structures. We have shown that existing OIE systems perform worse on scientific and medical content than on general audience content.
There are a range of avenues for future work. First, the application of Crowd Truth framework BIBREF15 in the analysis of these results might prove to be useful as we believe that the use of unanimous agreement tends to negatively impact the perceived performance of the OIE tools. Second, we think the application to n-ary relations and a deeper analysis of negative relations would be of interest. To do this kind of evaluation, an important area of future work is the development of guidelines and tasks for more complex analysis of sentences in a crowd sourcing environment. The ability, for example, to indicate argument boundaries or correct sentences can be expected of expert annotators but needs to implemented in a manner that is efficient and easy for the general crowd worker. Third, we would like to expand the evaluation dataset to an even larger numbers of sentences. Lastly, there are a number of core natural language processing components that might be useful for OIE in this setting, for example, the use of syntactic features as suggested by BIBREF16 . Furthermore, we think that coreference is a crucial missing component and we are actively investigating improved coreference resolution for scientific texts.
To conclude, we hope that this evaluation provides further insights for implementors of these extraction tools to deal with the complexity of scientific and medical text.
| 12
|
1802.06024
|
Towards a Continuous Knowledge Learning Engine for Chatbots
|
# Towards a Continuous Knowledge Learning Engine for Chatbots
## Abstract
Although chatbots have been very popular in recent years, they still have some serious weaknesses which limit the scope of their applications. One major weakness is that they cannot learn new knowledge during the conversation process, i.e., their knowledge is fixed beforehand and cannot be expanded or updated during conversation. In this paper, we propose to build a general knowledge learning engine for chatbots to enable them to continuously and interactively learn new knowledge during conversations. As time goes by, they become more and more knowledgeable and better and better at learning and conversation. We model the task as an open-world knowledge base completion problem and propose a novel technique called lifelong interactive learning and inference (LiLi) to solve it. LiLi works by imitating how humans acquire knowledge and perform inference during an interactive conversation. Our experimental results show LiLi is highly promising.
## Introduction
Chatbots such as dialog and question-answering systems have a long history in AI and natural language processing. Early such systems were mostly built using markup languages such as AIML, handcrafted conversation generation rules, and/or information retrieval techniques BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . Recent neural conversation models BIBREF4 , BIBREF5 , BIBREF6 are even able to perform open-ended conversations. However, since they do not use explicit knowledge bases and do not perform inference, they often suffer from generic and dull responses BIBREF5 , BIBREF7 . More recently, BIBREF8 and BIBREF9 proposed to use knowledge bases (KBs) to help generate responses for knowledge-grounded conversation. However, one major weakness of all existing chat systems is that they do not explicitly or implicitly learn new knowledge in the conversation process. This seriously limits the scope of their applications. In contrast, we humans constantly learn new knowledge in our conversations. Even if some existing systems can use very large knowledge bases either harvested from a large data source such as the Web or built manually, these KBs still miss a large number of facts (knowledge) BIBREF10 . It is thus important for a chatbot to continuously learn new knowledge in the conversation process to expand its KB and to improve its conversation ability.
In recent years, researchers have studied the problem of KB completion, i.e., inferring new facts (knowledge) automatically from existing facts in a KB. KB completion (KBC) is defined as a binary classification problem: Given a query triple, ( INLINEFORM0 , INLINEFORM1 , INLINEFORM2 ), we want to predict whether the source entity INLINEFORM3 and target entity INLINEFORM4 can be linked by the relation INLINEFORM5 . However, existing approaches BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 solve this problem under the closed-world assumption, i.e., INLINEFORM6 , INLINEFORM7 and INLINEFORM8 are all known to exist in the KB. This is a major weakness because it means that no new knowledge or facts may contain unknown entities or relations. Due to this limitation, KBC is clearly not sufficient for knowledge learning in conversations because in a conversation, the user can say anything, which may contain entities and relations that are not already in the KB.
In this paper, we remove this assumption of KBC, and allow all INLINEFORM0 , INLINEFORM1 and INLINEFORM2 to be unknown. We call the new problem open-world knowledge base completion (OKBC). OKBC generalizes KBC. Below, we show that solving OKBC naturally provides the ground for knowledge learning and inference in conversations. In essence, we formulate an abstract problem of knowledge learning and inference in conversations as a well-defined OKBC problem in the interactive setting.
From the perspective of knowledge learning in conversations, essentially we can extract two key types of information, true facts and queries, from the user utterances. Queries are facts whose truth values need to be determined. Note that we do not study fact or relation extraction in this paper as there is an extensive work on the topic. (1) For a true fact, we will incorporate it into the KB. Here we need to make sure that it is not already in the KB, which involves relation resolution and entity linking. After a fact is added to the KB, we may predict that some related facts involving some existing relations in the KB may also be true (not logical implications as they can be automatically inferred). For example, if the user says “Obama was born in USA,” the system may guess that (Obama, CitizenOf, USA) (meaning that Obama is a citizen of USA) could also be true based on the current KB. To verify this fact, it needs to solve a KBC problem by treating (Obama, CitizenOf, USA) as a query. This is a KBC problem because the fact (Obama, BornIn, USA) extracted from the original sentence has been added to the KB. Then Obama and USA are in the KB. If the KBC problem is solved, it learns a new fact (Obama, CitizenOf, USA) in addition to the extracted fact (Obama, BornIn, USA). (2) For a query fact, e.g., (Obama, BornIn, USA) extracted from the user question “Was Obama born in USA?” we need to solve the OKBC problem if any of “Obama, “BornIn”, or “USA" is not already in the KB.
We can see that OKBC is the core of a knowledge learning engine for conversation. Thus, in this paper, we focus on solving it. We assume that other tasks such as fact/relation extraction and resolution and guessing of related facts of an extracted fact are solved by other sub-systems.
We solve the OKBC problem by mimicking how humans acquire knowledge and perform reasoning in an interactive conversation. Whenever we encounter an unknown concept or relation while answering a query, we perform inference using our existing knowledge. If our knowledge does not allow us to draw a conclusion, we typically ask questions to others to acquire related knowledge and use it in inference. The process typically involves an inference strategy (a sequence of actions), which interleaves a sequence of processing and interactive actions. A processing action can be the selection of related facts, deriving inference chain, etc., that advances the inference process. An interactive action can be deciding what to ask, formulating a suitable question, etc., that enable us to interact. The process helps grow the knowledge over time and the gained knowledge enables us to communicate better in the future. We call this lifelong interactive learning and inference (LiLi). Lifelong learning is reflected by the facts that the newly acquired facts are retained in the KB and used in inference for future queries, and that the accumulated knowledge in addition to the updated KB including past inference performances are leveraged to guide future interaction and learning. LiLi should have the following capabilities:
This setting is ideal for many NLP applications like dialog and question-answering systems that naturally provide the scope for human interaction and demand real-time inference.
LiLi starts with the closed-world KBC approach path-ranking (PR) BIBREF11 , BIBREF17 and extends KBC in a major way to open-world knowledge base completion (OKBC). For a relation INLINEFORM0 , PR works by enumerating paths (except single-link path INLINEFORM1 ) between entity-pairs linked by INLINEFORM2 in the KB and use them as features to train a binary classifier to predict whether a query INLINEFORM3 should be in the KB. Here, a path between two entities is a sequence of relations linking them. In our work, we adopt the latest PR method, C-PR BIBREF16 and extend it to make it work in the open-world setting. C-PR enumerates paths by performing bidirectional random walks over the KB graph while leveraging the context of the source-target entity-pair. We also adopt and extend the compositional vector space model BIBREF20 , BIBREF21 with continual learning capability for prediction.
Given an OKBC query ( INLINEFORM0 , INLINEFORM1 , INLINEFORM2 ) (e.g., “(Obama, CitizenOf, USA), which means whether Obama a citizen of USA), LiLi interacts with the user (if needed) by dynamically formulating questions (see the interaction example in Figure 1, which will be further explained in §3) and leverages the interactively acquired knowledge (supporting facts (SFs) in the figure) for continued inference. To do so, LiLi formulates a query-specific inference strategy and executes it. We design LiLi in a Reinforcement Learning (RL) setting that performs sub-tasks like formulating and executing strategy, training a prediction model for inference, and knowledge retention for future use. To the best of our knowledge, our work is the first to address the OKBC problem and to propose an interactive learning mechanism to solve it in a continuous or lifelong manner. We empirically verify the effectiveness of LiLi on two standard real-world KBs: Freebase and WordNet. Experimental results show that LiLi is highly effective in terms of its predictive performance and strategy formulation ability.
## Related Work
To the best of our knowledge, we are not aware of any knowledge learning system that can learn new knowledge in the conversation process. This section thus discusses other related work.
Among existing KB completion approaches, BIBREF20 extended the vector space model for zero-shot KB inference. However, the model cannot handle unknown entities and can only work on fixed set of unknown relations with known embeddings. Recently, BIBREF22 proposed a method using external text corpus to perform inference on unknown entities. However, the method cannot handle unknown relations. Thus, these methods are not suitable for our open-world setting. None of the existing KB inference methods perform interactive knowledge learning like LiLi. NELL BIBREF23 continuously updates its KB using facts extracted from the Web. Our task is very different as we do not do Web fact extraction (which is also useful). We focus on user interactions in this paper. Our work is related to interactive language learning (ILL) BIBREF24 , BIBREF25 , but these are not about KB completion. The work in BIBREF26 allows a learner to ask questions in dialogue. However, this work used RL to learn about whether to ask the user or not. The “what to ask aspect" was manually designed by modeling synthetic tasks. LiLi formulates query-specific inference strategies which embed interaction behaviors. Also, no existing dialogue systems BIBREF4 , BIBREF27 , BIBREF28 , BIBREF29 , BIBREF30 employ lifelong learning to train prediction models by using information/knowledge retained in the past.
Our work is related to general lifelong learning in BIBREF31 , BIBREF32 , BIBREF33 , BIBREF34 , BIBREF35 , BIBREF36 . However, they learn only one type of tasks, e.g., supervised, topic modeling or reinforcement learning (RL) tasks. None of them is suitable for our setting, which involves interleaving of RL, supervised and interactive learning. More details about lifelong learning can be found in the book BIBREF31 .
## Interactive Knowledge Learning (LiLi)
We design LiLi as a combination of two interconnected models: (1) a RL model that learns to formulate a query-specific inference strategy for performing the OKBC task, and (2) a lifelong prediction model to predict whether a triple should be in the KB, which is invoked by an action while executing the inference strategy and is learned for each relation as in C-PR. The framework improves its performance over time through user interaction and knowledge retention. Compared to the existing KB inference methods, LiLi overcomes the following three challenges for OKBC:
1. Mapping open-world to close-world. Being a closed-world method, C-PR cannot extract path features and learn a prediction model when any of INLINEFORM0 , INLINEFORM1 or INLINEFORM2 is unknown. LiLi solves this problem through interactive knowledge acquisition. If INLINEFORM3 is unknown, LiLi asks the user to provide a clue (an example of INLINEFORM4 ). And if INLINEFORM5 or INLINEFORM6 is unknown, LiLi asks the user to provide a link (relation) to connect the unknown entity with an existing entity (automatically selected) in the KB. We refer to such a query as a connecting link query (CLQ). The acquired knowledge reduces OKBC to KBC and makes the inference task feasible.
2. Spareseness of KB. A main issue of all PR methods like C-PR is the connectivity of the KB graph. If there is no path connecting INLINEFORM0 and INLINEFORM1 in the graph, path enumeration of C-PR gets stuck and inference becomes infeasible. In such cases, LiLi uses a template relation (“@-?-@") as the missing link marker to connect entity-pairs and continues feature extraction. A path containing “@-?-@" is called an incomplete path. Thus, the extracted feature set contains both complete (no missing link) and incomplete paths. Next, LiLi selects an incomplete path from the feature set and asks the user to provide a link for path completion. We refer to such a query as missing link query (MLQ).
3. Limitation in user knowledge. If the user is unable to respond to MLQs or CLQs, LiLi uses a guessing mechanism (discussed later) to fill the gap. This enables LiLi to continue its inference even if the user cannot answer a system question.
## Components of LiLi
As lifelong learning needs to retain knowledge learned from past tasks and use it to help future learning BIBREF31 , LiLi uses a Knowledge Store (KS) for knowledge retention. KS has four components: (i) Knowledge Graph ( INLINEFORM0 ): INLINEFORM1 (the KB) is initialized with base KB triples (see §4) and gets updated over time with the acquired knowledge. (ii) Relation-Entity Matrix ( INLINEFORM2 ): INLINEFORM3 is a sparse matrix, with rows as relations and columns as entity-pairs and is used by the prediction model. Given a triple ( INLINEFORM4 , INLINEFORM5 , INLINEFORM6 ) INLINEFORM7 , we set INLINEFORM8 [ INLINEFORM9 , ( INLINEFORM10 , INLINEFORM11 )] = 1 indicating INLINEFORM12 occurs for pair ( INLINEFORM13 , INLINEFORM14 ). (iii) Task Experience Store ( INLINEFORM15 ): INLINEFORM16 stores the predictive performance of LiLi on past learned tasks in terms of Matthews correlation coefficient (MCC) that measures the quality of binary classification. So, for two tasks INLINEFORM17 and INLINEFORM18 (each relation is a task), if INLINEFORM19 [ INLINEFORM20 ] INLINEFORM21 INLINEFORM22 [ INLINEFORM23 ] [where INLINEFORM24 [ INLINEFORM25 ]=MCC( INLINEFORM26 )], we say C-PR has learned INLINEFORM27 well compared to INLINEFORM28 . (iv) Incomplete Feature DB ( INLINEFORM29 ): INLINEFORM30 stores the frequency of an incomplete path INLINEFORM31 in the form of a tuple ( INLINEFORM32 , INLINEFORM33 , INLINEFORM34 ) and is used in formulating MLQs. INLINEFORM35 [( INLINEFORM36 , INLINEFORM37 , INLINEFORM38 )] = INLINEFORM39 implies LiLi has extracted incomplete path INLINEFORM40 INLINEFORM41 times involving entity-pair INLINEFORM42 [( INLINEFORM43 , INLINEFORM44 )] for query relation INLINEFORM45 .
The RL model learns even after training whenever it encounters an unseen state (in testing) and thus, gets updated over time. KS is updated continuously over time as a result of the execution of LiLi and takes part in future learning. The prediction model uses lifelong learning (LL), where we transfer knowledge (parameter values) from the model for a past most similar task to help learn for the current task. Similar tasks are identified by factorizing INLINEFORM0 and computing a task similarity matrix INLINEFORM1 . Besides LL, LiLi uses INLINEFORM2 to identify poorly learned past tasks and acquire more clues for them to improve its skillset over time.
LiLi also uses a stack, called Inference Stack ( INLINEFORM0 ) to hold query and its state information for RL. LiLi always processes stack top ( INLINEFORM1 [top]). The clues from the user get stored in INLINEFORM2 on top of the query during strategy execution and processed first. Thus, the prediction model for INLINEFORM3 is learned before performing inference on query, transforming OKBC to a KBC problem. Table 1 shows the parameters of LiLi used in the following sections.
## Working of LiLi
Given an OKBC query ( INLINEFORM0 , INLINEFORM1 , INLINEFORM2 ), we represent it as a data instance INLINEFORM3 . INLINEFORM4 consists of INLINEFORM5 (the query triple), INLINEFORM6 (interaction limit set for INLINEFORM7 ), INLINEFORM8 (experience list storing the transition history of MDP for INLINEFORM9 in RL) and INLINEFORM10 (mode of INLINEFORM11 ) denoting if INLINEFORM12 is ` INLINEFORM13 ' (training), ` INLINEFORM14 ' (validation), ` INLINEFORM15 ' (evaluation) or ` INLINEFORM16 ' (clue) instance and INLINEFORM17 (feature set). We denote INLINEFORM18 ( INLINEFORM19 ) as the set of all complete (incomplete) path features in INLINEFORM20 . Given a data instance INLINEFORM21 , LiLi starts its initialization as follows: it sets the state as INLINEFORM22 (based on INLINEFORM23 , explained later), pushes the query tuple ( INLINEFORM24 , INLINEFORM25 ) into INLINEFORM26 and feeds INLINEFORM27 [top] to the RL-model for strategy formulation from INLINEFORM28 .
Inference Strategy Formulation. We view solving the strategy formulation problem as learning to play an inference game, where the goal is to formulate a strategy that "makes the inference task possible". Considering PR methods, inference is possible, iff (1) INLINEFORM0 becomes known to its KB (by acquiring clues when INLINEFORM1 is unknown) and (2) path features are extracted between INLINEFORM2 and INLINEFORM3 (which inturn requires INLINEFORM4 and INLINEFORM5 to be known to KB). If these conditions are met at the end of an episode (when strategy formulation finishes for a given query) of the game, LiLi wins and thus, it trains the prediction model for INLINEFORM6 and uses it for inference.
LiLi's strategy formulation is modeled as a Markov Decision Process (MDP) with finite state ( INLINEFORM0 ) and action ( INLINEFORM1 ) spaces. A state INLINEFORM2 consists of 10 binary state variables (Table 2), each of which keeps track of results of an action INLINEFORM3 taken by LiLi and thus, records the progress in inference process made so far. INLINEFORM4 is the initial state with all state bits set as 0. If the data instance (query) is a clue [ INLINEFORM5 ], INLINEFORM6 [CLUE] is set as 1. INLINEFORM7 consists of 6 actions (Table 3). INLINEFORM8 , INLINEFORM9 , INLINEFORM10 are processing actions and INLINEFORM11 , INLINEFORM12 , INLINEFORM13 are interactive actions. Whenever INLINEFORM14 is executed, the MDP reaches the terminal state. Given an action INLINEFORM15 in state INLINEFORM16 , if INLINEFORM17 is invalid in INLINEFORM21 or the objective of INLINEFORM22 is unsatisfied (* marked the condition in INLINEFORM23 ), RL receives a negative reward (empirically set); else receives a positive reward.. We use Q-learning BIBREF38 with INLINEFORM24 -greedy strategy to learn the optimal policy for training the RL model. Note that, the inference strategy is independent of KB type and correctness of prediction. Thus, the RL-model is trained only once from scratch (reused thereafter for other KBs) and also, independently of the prediction model.
Sometimes the training dataset may not be enough to learn optimal policy for all INLINEFORM0 . Thus, encountering an unseen state during test can make RL-model clueless about the action. Given a state INLINEFORM1 , whenever an invalid INLINEFORM2 is chosen, LiLi remains in INLINEFORM3 . For INLINEFORM4 , LiLi remains in INLINEFORM5 untill INLINEFORM6 (see Table 1 for INLINEFORM7 ). So, if the state remains the same for ( INLINEFORM8 +1) times, it implies LiLi has encountered a fault (an unseen state). RL-model instantly switches to the training mode and randomly explores INLINEFORM9 to learn the optimal action (fault-tolerant learning). While exploring INLINEFORM10 , the model chooses INLINEFORM11 only when it has tried all other INLINEFORM12 to avoid abrupt end of episode.
Execution of Actions. At any given point in time, let ( INLINEFORM0 , INLINEFORM1 ) be the current INLINEFORM2 [top], INLINEFORM3 is the chosen action and the current version of KS components are INLINEFORM4 , INLINEFORM5 , INLINEFORM6 and INLINEFORM7 . Then, if INLINEFORM8 is invalid in INLINEFORM9 , LiLi only updates INLINEFORM10 [top] with ( INLINEFORM11 , INLINEFORM12 ) and returns INLINEFORM13 [top] to RL-model. In this process, LiLi adds experience ( INLINEFORM14 , INLINEFORM15 , INLINEFORM16 , INLINEFORM17 ) in INLINEFORM18 and then, replaces INLINEFORM19 [top] with ( INLINEFORM20 , INLINEFORM21 ). If INLINEFORM22 is valid in INLINEFORM23 , LiLi first sets the next state INLINEFORM24 and performs a sequence of operations INLINEFORM25 based on INLINEFORM26 (discussed below). Unless specified, in INLINEFORM27 , LiLi always monitors INLINEFORM28 and if INLINEFORM29 becomes 0, LiLi sets INLINEFORM30 . Also, whenever LiLi asks the user a query, INLINEFORM31 is decremented by 1. Once INLINEFORM32 ends, LiLi updates INLINEFORM33 [top] with ( INLINEFORM34 , INLINEFORM35 ) and returns INLINEFORM36 [top] to RL-model for choosing the next action.
In INLINEFORM0 , LiLi searches INLINEFORM1 , INLINEFORM2 , INLINEFORM3 in INLINEFORM4 and sets appropriate bits in INLINEFORM5 (see Table 2). If INLINEFORM6 was unknown before and is just added to INLINEFORM7 or is in the bottom INLINEFORM8 % (see Table 1 for INLINEFORM9 ) of INLINEFORM10 , LiLi randomly sets INLINEFORM14 with probability INLINEFORM15 . If INLINEFORM16 is a clue and INLINEFORM17 , LiLi updates KS with triple INLINEFORM18 , where ( INLINEFORM19 , INLINEFORM20 , INLINEFORM21 ) and ( INLINEFORM22 , INLINEFORM23 , INLINEFORM24 ) gets added to INLINEFORM25 and INLINEFORM26 , INLINEFORM27 are set as 1.
In INLINEFORM0 , LiLi asks the user to provide a clue (+ve instance) for INLINEFORM1 and corrupts INLINEFORM2 and INLINEFORM3 of the clue once at a time, to generate -ve instances by sampling nodes from INLINEFORM4 . These instances help in training prediction model for INLINEFORM5 while executing INLINEFORM6 .
In INLINEFORM0 , LiLi selects an incomplete path INLINEFORM1 from INLINEFORM2 to formulate MLQ, such that INLINEFORM3 is most frequently observed for INLINEFORM4 and INLINEFORM5 is high, given by INLINEFORM6 . Here, INLINEFORM7 denotes the contextual similarity BIBREF16 of entity-pair INLINEFORM8 . If INLINEFORM9 is high, INLINEFORM10 is more likely to possess a relation between them and so, is a good candidate for formulating MLQ. When the user does not respond to MLQ (or CLQ in INLINEFORM11 ), the guessing mechanism is used, which works as follows: Since contextual similarity of entity-pairs is highly correlated with their class labels BIBREF16 , LiLi divides the similarity range [-1, 1] into three segments, using a low ( INLINEFORM12 ) and high ( INLINEFORM13 ) similarity threshold and replaces the missing link with INLINEFORM14 in INLINEFORM15 to make it complete as follows: If INLINEFORM16 , INLINEFORM17 = “@-LooselyRelatedTo-@"; else if INLINEFORM18 , INLINEFORM19 =“@-NotRelatedTo-@"; Otherwise, INLINEFORM20 =“@-RelatedTo-@".
In INLINEFORM0 , LiLi asks CLQs for connecting unknown entities INLINEFORM1 and/or INLINEFORM2 with INLINEFORM3 by selecting the most contextually relevant node (wrt INLINEFORM4 , INLINEFORM5 ) from INLINEFORM6 , given by link INLINEFORM7 . We adopt the contextual relevance idea in BIBREF16 which is computed using word embedding BIBREF39
In INLINEFORM0 , LiLi extracts path features INLINEFORM1 between ( INLINEFORM2 , INLINEFORM3 ) and updates INLINEFORM4 with incomplete features from INLINEFORM5 . LiLi always trains the prediction model with complete features INLINEFORM6 and once INLINEFORM7 or INLINEFORM8 , LiLi stops asking MLQs. Thus, in both INLINEFORM9 and INLINEFORM10 , LiLi always monitors INLINEFORM11 to check for the said requirements and sets INLINEFORM12 to control interactions.
In INLINEFORM0 , if LiLi wins the episode, it adds INLINEFORM1 in one of data buffers INLINEFORM2 based on its mode INLINEFORM3 . E.g., if INLINEFORM4 or INLINEFORM5 , INLINEFORM6 is used for training and added to INLINEFORM7 . Similarly validation buffer INLINEFORM8 and evaluation buffer INLINEFORM9 are populated. If INLINEFORM10 , LiLi invokes the prediction model for INLINEFORM11 .
Lifelong Relation Prediction. Given a relation INLINEFORM0 , LiLi uses INLINEFORM1 and INLINEFORM2 (see INLINEFORM3 ) to train a prediction model (say, INLINEFORM4 ) with parameters INLINEFORM5 . For a unknown INLINEFORM6 , the clue instances get stored in INLINEFORM7 and INLINEFORM8 . Thus, LiLi populates INLINEFORM9 by taking 10% (see §4) of the instances from INLINEFORM10 and starts the training. For INLINEFORM11 , LiLi uses a LSTM BIBREF40 to compose the vector representation of each feature INLINEFORM12 as INLINEFORM13 and vector representation of INLINEFORM14 as INLINEFORM15 . Next, LiLi computes the prediction value, INLINEFORM16 as sigmoid of the mean cosine similarity of all features and INLINEFORM17 , given by INLINEFORM18 ) and maximize the log-likelihood of INLINEFORM19 for training. Once INLINEFORM20 is trained, LiLi updates INLINEFORM21 [ INLINEFORM22 ] using INLINEFORM23 . We also train an inverse model for INLINEFORM24 , INLINEFORM25 by reversing the path features in INLINEFORM26 and INLINEFORM27 which help in lifelong learning (discussed below). Unlike BIBREF20 , BIBREF21 , while predicting the label for INLINEFORM28 , we compute a relation-specific prediction threshold INLINEFORM29 corresponding to INLINEFORM30 using INLINEFORM31 as: INLINEFORM32 and infer INLINEFORM33 as +ve if INLINEFORM34 and -ve otherwise. Here, INLINEFORM35 ( INLINEFORM36 ) is the mean prediction value for all +ve (-ve) examples in INLINEFORM37 .
Models trained on a few examples (e.g., clues acquired for unknown INLINEFORM0 ) with randomly initialized weights often perform poorly due to underfitting. Thus, we transfer knowledge (weights) from the past most similar (wrt INLINEFORM1 ) task in a lifelong learning manner BIBREF31 . LiLi uses INLINEFORM2 to find the past most similar task for INLINEFORM3 as follows: LiLi computes trancated SVD of INLINEFORM4 as INLINEFORM5 and then, the similarity matrix INLINEFORM6 . INLINEFORM7 provides the similarity between relations INLINEFORM8 and INLINEFORM9 in INLINEFORM10 . Thus, LiLi chooses a source relation INLINEFORM11 to transfer weights. Here, INLINEFORM12 is the set of all INLINEFORM13 and INLINEFORM14 for which LiLi has already learned a prediction model. Now, if INLINEFORM15 or INLINEFORM16 , LiLi randomly initializes the weights INLINEFORM17 for INLINEFORM18 and proceeds with the training. Otherwise, LiLi uses INLINEFORM19 as initial weights and fine-tunes INLINEFORM20 with a low learning rate.
A Running Example. Considering the example shown in Figure 1, LiLi works as follows: first, LiLi executes INLINEFORM0 and detects that the source entity “Obama" and query relation “CitizenOf" are unknown. Thus, LiLi executes INLINEFORM1 to acquire clue (SF1) for “CitizenOf" and pushes the clue (+ve example) and two generated -ve examples into INLINEFORM2 . Once the clues are processed and a prediction model is trained for “CitizenOf" by formulating separate strategies for them, LiLi becomes aware of “CitizenOf". Now, as the clues have already been popped from INLINEFORM3 , the query becomes INLINEFORM4 and the strategy formulation process for the query resumes. Next, LiLi asks user to provide a connecting link for “Obama" by performing INLINEFORM5 . Now, the query entities and relation being known, LiLi enumerates paths between “Obama" and “USA" by performing INLINEFORM6 . Let an extracted path be “ INLINEFORM7 " with missing link between ( INLINEFORM8 , INLINEFORM9 ). LiLi asks the user to fill the link by performing INLINEFORM10 and then, extracts the complete feature “ INLINEFORM11 ". The feature set is then fed to the prediction model and inference is made as a result of INLINEFORM12 . Thus, the formulated inference strategy is: “ INLINEFORM13 ".
## Experiments
We now evaluate LiLi in terms of its predictive performance and strategy formulation abilities.
Data: We use two standard datasets (see Table 4): (1) Freebase FB15k, and (2) WordNet INLINEFORM0 . Using each dataset, we build a fairly large graph and use it as the original KB ( INLINEFORM1 ) for evaluation. We also augment INLINEFORM2 with inverse triples ( INLINEFORM3 , INLINEFORM4 , INLINEFORM5 ) for each ( INLINEFORM6 , INLINEFORM7 , INLINEFORM8 ) following existing KBC methods.
Parameter Settings. Unless specified, the empirically set parameters (see Table 1) of LiLi are: INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 , INLINEFORM4 , INLINEFORM5 , INLINEFORM6 , INLINEFORM7 , INLINEFORM8 , INLINEFORM9 , INLINEFORM10 . For training RL-model with INLINEFORM11 -greedy strategy, we use INLINEFORM12 , INLINEFORM13 , pre-training steps=50000. We used Keras deep learning library to implement and train the prediction model. We set batch-size as 128, max. training epoch as 150, dropout as 0.2, hidden units and embedding size as 300 and learning rate as 5e-3 which is reduced gradually on plateau with factor 0.5 and patience 5. Adam optimizer and early stopping were used in training. We also shuffle INLINEFORM14 in each epoch and adjust class weights inversely proportional to class frequencies in INLINEFORM15 .
Labeled Dataset Generation and Simulated User Creation. We create a simulated user for each KB to evaluate LiLi. We create the labeled datasets, the simulated user’s knowledge base ( INLINEFORM0 ), and the base KB ( INLINEFORM1 ) from INLINEFORM2 . INLINEFORM3 used as the initial KB graph ( INLINEFORM4 ) of LiLi.
We followed BIBREF16 for labeled dataset generation. For Freebase, we found 86 relations with INLINEFORM0 triples and randomly selected 50 from various domains. We randomly shuffle the list of 50 relations, select 25% of them as unknown relations and consider the rest (75%) as known relations. For each known relation INLINEFORM1 , we randomly shuffle the list of distinct triples for INLINEFORM2 , choose 1000 triples and split them into 60% training, 10% validation and 20% test. Rest 10% along with the leftover (not included in the list of 1000) triples are added to INLINEFORM3 . For each unknown relation INLINEFORM4 , we remove all triples of INLINEFORM5 from INLINEFORM6 and add them to INLINEFORM7 . In this process, we also randomly choose 20% triples as test instances for unknown INLINEFORM8 which are excluded from INLINEFORM9 . Note that, now INLINEFORM10 has at least 10% of chosen triples for each INLINEFORM11 (known and unknown) and so, user is always able to provide clues for both cases. For each labeled dataset, we randomly choose 10% of the entities present in dataset triples, remove triples involving those entities from INLINEFORM12 and add to INLINEFORM13 . At this point, INLINEFORM14 gets reduced to INLINEFORM15 and is used as INLINEFORM16 for LiLi. The dataset stats in Table 4 shows that the base KB (60% triples of INLINEFORM17 ) is highly sparse (compared to original KB) which makes the inference task much harder. WordNet dataset being small, we select all 18 relations for evaluation and create labeled dataset, INLINEFORM18 and INLINEFORM19 following Freebase. Although the user may provide clues 100% of the time, it often cannot respond to MLQs and CLQs (due to lack of required triples/facts). Thus, we further enrich INLINEFORM20 with external KB triples.
Given a relation INLINEFORM0 and an observed triple ( INLINEFORM1 , INLINEFORM2 , INLINEFORM3 ) in training or testing, the pair ( INLINEFORM4 , INLINEFORM5 ) is regarded as a +ve instance for INLINEFORM6 . Following BIBREF18 , for each +ve instance ( INLINEFORM7 , INLINEFORM8 ), we generate two negative ones, one by randomly corrupting the source INLINEFORM9 , and the other by corrupting the target INLINEFORM10 . Note that, the test triples are not in INLINEFORM11 or INLINEFORM12 and none of the -ve instances overlap with the +ve ones.
Baselines. As none of the existing KBC methods can solve the OKBC problem, we choose various versions of LiLi as baselines.
Single: Version of LiLi where we train a single prediction model INLINEFORM0 for all test relations.
Sep: We do not transfer (past learned) weights for initializing INLINEFORM0 , i.e., we disable LL.
F-th): Here, we use a fixed prediction threshold 0.5 instead of relation-specific threshold INLINEFORM0 .
BG: The missing or connecting links (when the user does not respond) are filled with “@-RelatedTo-@" blindly, no guessing mechanism.
w/o PTS: LiLi does not ask for additional clues via past task selection for skillset improvement.
Evaluation Metrics. To evaluate the strategy formulation ability, we introduce a measure called Coverage( INLINEFORM0 ), defined as the fraction of total query data instances, for which LiLi has successfully formulated strategies that lead to winning. If LiLi wins on all episodes for a given dataset, INLINEFORM1 is 1.0. To evaluate the predictive performance, we use Avg. MCC and avg. +ve F1 score.
## Results and Analysis
Evaluation-I: Strategy Formulation Ability. Table 5 shows the list of inference strategies formulated by LiLi for various INLINEFORM0 and INLINEFORM1 , which control the strategy formulation of LiLi. When INLINEFORM2 , LiLi cannot interact with user and works like a closed-world method. Thus, INLINEFORM3 drops significantly (0.47). When INLINEFORM4 , i.e. with only one interaction per query, LiLi acquires knowledge well for instances where either of the entities or relation is unknown. However, as one unknown entity may appear in multiple test triples, once the entity becomes known, LiLi doesn’t need to ask for it again and can perform inference on future triples causing significant increase in INLINEFORM5 (0.97). When INLINEFORM6 , LiLi is able to perform inference on all instances and INLINEFORM7 becomes 1. For INLINEFORM8 , LiLi uses INLINEFORM9 only once (as only one MLQ satisfies INLINEFORM10 ) compared to INLINEFORM11 . In summary, LiLi’s RL-model can effectively formulate query-specific inference strategies (based on specified parameter values). Evaluation-II: Predictive Performance. Table 6 shows the comparative performance of LiLi with baselines. To judge the overall improvements, we performed paired t-test considering +ve F1 scores on each relation as paired data. Considering both KBs and all relation types, LiLi outperforms Sep with INLINEFORM12 . If we set INLINEFORM13 (training with very few clues), LiLi outperforms Sep with INLINEFORM14 on Freebase considering MCC. Thus, the lifelong learning mechanism is effective in transferring helpful knowledge. Single model performs better than Sep for unknown relations due to the sharing of knowledge (weights) across tasks. However, for known relations, performance drops because, as a new relation arrives to the system, old weights get corrupted and catastrophic forgetting occurs. For unknown relations, as the relations are evaluated just after training, there is no chance for catastrophic forgetting. The performance improvement ( INLINEFORM15 ) of LiLi over F-th on Freebase signifies that the relation-specific threshold INLINEFORM16 works better than fixed threshold 0.5 because, if all prediction values for test instances lie above (or below) 0.5, F-th predicts all instances as +ve (-ve) which degrades its performance. Due to the utilization of contextual similarity (highly correlated with class labels) of entity-pairs, LiLi’s guessing mechanism works better ( INLINEFORM17 ) than blind guessing (BG). The past task selection mechanism of LiLi also improves its performance over w/o PTS, as it acquires more clues during testing for poorly performed tasks (evaluated on validation set). For Freebase, due to a large number of past tasks [9 (25% of 38)], the performance difference is more significant ( INLINEFORM18 ). For WordNet, the number is relatively small [3 (25% of 14)] and hence, the difference is not significant.
Evaluation-III: User Interaction vs. Performance. Table 7 shows the results of LiLi by varying clue acquisition rate ( INLINEFORM0 ). We use Freebase for tuning INLINEFORM1 due to its higher number of unknown test relations compared to WordNet. LiLi’s performance improves significantly as it acquires more clues from the user. The results on INLINEFORM2 outperforms ( INLINEFORM3 ) that on INLINEFORM4 . Table 8 shows the results of LiLi on user responses to MLQ’s and CLQ’s. Answering MLQ’s and CLQ’s is very hard for simulated users (unlike crowd-sourcing) as often INLINEFORM5 lacks the required triple. Thus, we attempt to analyze how the performance is effected if the user does not respond at all. The results show a clear trend in overall performance improvement when the user responds. However, the improvement is not significant as the simulated user’s query satisfaction rate (1% MLQs and 10% CLQs) is very small. But, the analysis shows the effectiveness of LiLi’s guessing mechanism and continual learning ability that help in achieving avg. +ve F1 of 0.57 and 0.62 on FB and WN respectively with minimal participation of the user.
## Conclusion
In this paper, we are interested in building a generic engine for continuous knowledge learning in human-machine conversations. We first showed that the problem underlying the engine can be formulated as an open-world knowledge base completion (OKBC) problem. We then proposed an lifelong interactive learning and inference (LiLi) approach to solving the OKBC problem. OKBC is a generalization of KBC. LiLi solves the OKBC problem by first formulating a query-specific inference strategy using RL and then executing it to solve the problem by interacting with the user in a lifelong learning manner. Experimental results showed the effectiveness of LiLi in terms of both predictive quality and strategy formulation ability. We believe that a system with the LiLi approach can serve as a knowledge learning engine for conversations. Our future work will improve LiLi to make more accurate.
## Acknowledgments
This work was supported in part by National Science Foundation (NSF) under grant no. IIS-1407927 and IIS-1650900, and a gift from Huawei Technologies Co Ltd.
| 9
|
1802.07862
|
Multimodal Named Entity Recognition for Short Social Media Posts
|
# Multimodal Named Entity Recognition for Short Social Media Posts
## Abstract
We introduce a new task called Multimodal Named Entity Recognition (MNER) for noisy user-generated data such as tweets or Snapchat captions, which comprise short text with accompanying images. These social media posts often come in inconsistent or incomplete syntax and lexical notations with very limited surrounding textual contexts, bringing significant challenges for NER. To this end, we create a new dataset for MNER called SnapCaptions (Snapchat image-caption pairs submitted to public and crowd-sourced stories with fully annotated named entities). We then build upon the state-of-the-art Bi-LSTM word/character based NER models with 1) a deep image network which incorporates relevant visual context to augment textual information, and 2) a generic modality-attention module which learns to attenuate irrelevant modalities while amplifying the most informative ones to extract contexts from, adaptive to each sample and token. The proposed MNER model with modality attention significantly outperforms the state-of-the-art text-only NER models by successfully leveraging provided visual contexts, opening up potential applications of MNER on myriads of social media platforms.
## Introduction
Social media with abundant user-generated posts provide a rich platform for understanding events, opinions and preferences of groups and individuals. These insights are primarily hidden in unstructured forms of social media posts, such as in free-form text or images without tags. Named entity recognition (NER), the task of recognizing named entities from free-form text, is thus a critical step for building structural information, allowing for its use in personalized assistance, recommendations, advertisement, etc.
While many previous approaches BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 on NER have shown success for well-formed text in recognizing named entities via word context resolution (e.g. LSTM with word embeddings) combined with character-level features (e.g. CharLSTM/CNN), several additional challenges remain for recognizing named entities from extremely short and coarse text found in social media posts. For instance, short social media posts often do not provide enough textual contexts to resolve polysemous entities (e.g. “monopoly is da best ", where `monopoly' may refer to a board game (named entity) or a term in economics). In addition, noisy text includes a huge number of unknown tokens due to inconsistent lexical notations and frequent mentions of various newly trending entities (e.g. “xoxo Marshmelloooo ", where `Marshmelloooo' is a mis-spelling of a known entity `Marshmello', a music producer), making word embeddings based neural networks NER models vulnerable.
To address the challenges above for social media posts, we build upon the state-of-the-art neural architecture for NER with the following two novel approaches (Figure FIGREF1 ). First, we propose to leverage auxiliary modalities for additional context resolution of entities. For example, many popular social media platforms now provide ways to compose a post in multiple modalities - specifically image and text (e.g. Snapchat captions, Twitter posts with image URLs), from which we can obtain additional context for understanding posts. While “monopoly" in the previous example is ambiguous in its textual form, an accompanying snap image of a board game can help disambiguate among polysemous entities, thereby correctly recognizing it as a named entity.
Second, we also propose a general modality attention module which chooses per decoding step the most informative modality among available ones (in our case, word embeddings, character embeddings, or visual features) to extract context from. For example, the modality attention module lets the decoder attenuate the word-level signals for unknown word tokens (“Marshmellooooo" with trailing `o's) and amplifies character-level features intsead (capitalized first letter, lexical similarity to other known named entity token `Marshmello', etc.), thereby suppressing noise information (“UNK" token embedding) in decoding steps. Note that most of the previous literature in NER or other NLP tasks combine word and character-level information with naive concatenation, which is vulnerable to noisy social media posts. When an auxiliary image is available, the modality attention module determines to amplify this visual context in disambiguating polysemous entities, or to attenuate visual contexts when they are irrelevant to target named entities, selfies, etc. Note that the proposed modality attention module is distinct from how attention is used in other sequence-to-sequence literature (e.g. attending to a specific token within an input sequence). Section SECREF2 provides the detailed literature review.
Our contributions are three-fold: we propose (1) an LSTM-CNN hybrid multimodal NER network that takes as input both image and text for recognition of a named entity in text input. To the best of our knowledge, our approach is the first work to incorporate visual contexts for named entity recognition tasks. (2) We propose a general modality attention module that selectively chooses modalities to extract primary context from, maximizing information gain and suppressing irrelevant contexts from each modality (we treat words, characters, and images as separate modalities). (3) We show that the proposed approaches outperform the state-of-the-art NER models (both with and without using additional visual contexts) on our new MNER dataset SnapCaptions, a large collection of informal and extremely short social media posts paired with unique images.
## Related Work
Neural models for NER have been recently proposed, producing state-of-the-art performance on standard NER tasks. For example, some of the end-to-end NER systems BIBREF4 , BIBREF2 , BIBREF3 , BIBREF0 , BIBREF1 use a recurrent neural network usually with a CRF BIBREF5 , BIBREF6 for sequence labeling, accompanied with feature extractors for words and characters (CNN, LSTMs, etc.), and achieve the state-of-the-art performance mostly without any use of gazetteers information. Note that most of these work aggregate textual contexts via concatenation of word embeddings and character embeddings. Recently, several work have addressed the NER task specifically on noisy short text segments such as Tweets, etc. BIBREF7 , BIBREF8 . They report performance gains from leveraging external sources of information such as lexical information (POS tags, etc.) and/or from several preprocessing steps (token substitution, etc.). Our model builds upon these state-of-the-art neural models for NER tasks, and improves the model in two critical ways: (1) incorporation of visual contexts to provide auxiliary information for short media posts, and (2) addition of the modality attention module, which better incorporates word embeddings and character embeddings, especially when there are many missing tokens in the given word embedding matrix. Note that we do not explore the use of gazetteers information or other auxiliary information (POS tags, etc.) BIBREF9 as it is not the focus of our study.
Attention modules are widely applied in several deep learning tasks BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 . For example, they use an attention module to attend to a subset within a single input (a part/region of an image, a specific token in an input sequence of tokens, etc.) at each decoding step in an encoder-decoder framework for image captioning tasks, etc. BIBREF14 explore various attention mechanisms in NLP tasks, but do not incorporate visual components or investigate the impact of such models on noisy social media data. BIBREF15 propose to use attention for a subset of discrete source samples in transfer learning settings. Our modality attention differs from the previous approaches in that we attenuate or amplifies each modality input as a whole among multiple available modalities, and that we use the attention mechanism essentially to map heterogeneous modalities in a single joint embedding space. Our approach also allows for re-use of the same model for predicting labels even when some of the modalities are missing in input, as other modalities would still preserve the same semantics in the embeddings space.
Multimodal learning is studied in various domains and applications, aimed at building a joint model that extracts contextual information from multiple modalities (views) of parallel datasets.
The most relevant task to our multimodal NER system is the task of multimodal machine translation BIBREF16 , BIBREF17 , which aims at building a better machine translation system by taking as input a sentence in a source language as well as a corresponding image. Several standard sequence-to-sequence architectures are explored (a target-language LSTM decoder that takes as input an image first). Other previous literature include study of Canonical Correlation Analysis (CCA) BIBREF18 to learn feature correlations among multiple modalities, which is widely used in many applications. Other applications include image captioning BIBREF10 , audio-visual recognition BIBREF19 , visual question answering systems BIBREF20 , etc.
To the best of our knowledge, our approach is the first work to incorporate visual contexts for named entity recognition tasks.
## Proposed Methods
Figure FIGREF2 illustrates the proposed multimodal NER (MNER) model. First, we obtain word embeddings, character embeddings, and visual features (Section SECREF3 ). A Bi-LSTM-CRF model then takes as input a sequence of tokens, each of which comprises a word token, a character sequence, and an image, in their respective representation (Section SECREF4 ). At each decoding step, representations from each modality are combined via the modality attention module to produce an entity label for each token ( SECREF5 ). We formulate each component of the model in the following subsections.
Notations: Let INLINEFORM0 a sequence of input tokens with length INLINEFORM1 , with a corresponding label sequence INLINEFORM2 indicating named entities (e.g. in standard BIO formats). Each input token is composed of three modalities: INLINEFORM3 for word embeddings, character embeddings, and visual embeddings representations, respectively.
## Features
Similar to the state-of-the-art NER approaches BIBREF0 , BIBREF1 , BIBREF8 , BIBREF4 , BIBREF2 , BIBREF3 , we use both word embeddings and character embeddings.
Word embeddings are obtained from an unsupervised learning model that learns co-occurrence statistics of words from a large external corpus, yielding word embeddings as distributional semantics BIBREF21 . Specifically, we use pre-trained embeddings from GloVE BIBREF22 .
Character embeddings are obtained from a Bi-LSTM which takes as input a sequence of characters of each token, similarly to BIBREF0 . An alternative approach for obtaining character embeddings is using a convolutional neural network as in BIBREF1 , but we find that Bi-LSTM representation of characters yields empirically better results in our experiments.
Visual embeddings: To extract features from an image, we take the final hidden layer representation of a modified version of the convolutional network model called Inception (GoogLeNet) BIBREF23 , BIBREF24 trained on the ImageNet dataset BIBREF25 to classify multiple objects in the scene. Our implementation of the Inception model has deep 22 layers, training of which is made possible via “network in network" principles and several dimension reduction techniques to improve computing resource utilization. The final layer representation encodes discriminative information describing what objects are shown in an image, which provide auxiliary contexts for understanding textual tokens and entities in accompanying captions.
Incorporating this visual information onto the traditional NER system is an open challenge, and multiple approaches can be considered. For instance, one may provide visual contexts only as an initial input to decoder as in some encoder-decoder image captioning systems BIBREF26 . However, we empirically observe that an NER decoder which takes as input the visual embeddings at every decoding step (Section SECREF4 ), combined with the modality attention module (Section SECREF5 ), yields better results.
Lastly, we add a transform layer for each feature INLINEFORM0 before it is fed to the NER entity LSTM.
## Bi-LSTM + CRF for Multimodal NER
Our MNER model is built on a Bi-LSTM and CRF hybrid model. We use the following implementation for the entity Bi-LSTM.
it = (Wxiht-1 + Wcict-1)
ct = (1-it) ct-1
+ it tanh(Wxcxt + Whcht-1)
ot = (Wxoxt + Whoht-1 + Wcoct)
ht = LSTM(xt)
= ot tanh(ct)
where INLINEFORM0 is a weighted average of three modalities INLINEFORM1 via the modality attention module, which will be defined in Section SECREF5 . Bias terms for gates are omitted here for simplicity of notation.
We then obtain bi-directional entity token representations INLINEFORM0 by concatenating its left and right context representations. To enforce structural correlations between labels in sequence decoding, INLINEFORM1 is then passed to a conditional random field (CRF) to produce a label for each token maximizing the following objective. y* = y p(y|h; WCRF)
p(y|h; WCRF) = t t (yt-1,yt;h) y' t t (y't-1,y't;h)
where INLINEFORM0 is a potential function, INLINEFORM1 is a set of parameters that defines the potential functions and weight vectors for label pairs ( INLINEFORM2 ). Bias terms are omitted for brevity of formulation.
The model can be trained via log-likelihood maximization for the training set INLINEFORM0 :
L(WCRF) = i p(y|h; W)
## Modality Attention
The modality attention module learns a unified representation space for multiple available modalities (words, characters, images, etc.), and produces a single vector representation with aggregated knowledge among multiple modalities, based on their weighted importance. We motivate this module from the following observations.
A majority of the previous literature combine the word and character-level contexts by simply concatenating the word and character embeddings at each decoding step, e.g. INLINEFORM0 in Eq. SECREF4 . However, this naive concatenation of two modalities (word and characters) results in inaccurate decoding, specifically for unknown word token embeddings (an all-zero vector INLINEFORM1 or a random vector INLINEFORM2 is assigned for any unknown token INLINEFORM3 , thus INLINEFORM4 or INLINEFORM5 ). While this concatenation approach does not cause significant errors for well-formatted text, we observe that it induces performance degradation for our social media post datasets which contain a significant number of missing tokens.
Similarly, naive merging of textual and visual information ( INLINEFORM0 ) yields suboptimal results as each modality is treated equally informative, whereas in our datasets some of the images may contain irrelevant contexts to textual modalities. Hence, ideally there needs a mechanism in which the model can effectively turn the switch on and off the modalities adaptive to each sample.
To this end, we propose a general modality attention module, which adaptively attenuates or emphasizes each modality as a whole at each decoding step INLINEFORM0 , and produces a soft-attended context vector INLINEFORM1 as an input token for the entity LSTM. [at(w),at(c),at(v)] = (Wm[xt(w); xt(c); xt(v)] + bm )
t(m) = (at(m))m'{w,c,v}(at(m')) m {w,c,v}
xt = m{w,c,v} t(m)xt(m)
where INLINEFORM0 is an attention vector at each decoding step INLINEFORM1 , and INLINEFORM2 is a final context vector at INLINEFORM3 that maximizes information gain for INLINEFORM4 . Note that the optimization of the objective function (Eq. SECREF4 ) with modality attention (Eq. SECREF5 ) requires each modality to have the same dimension ( INLINEFORM5 ), and that the transformation via INLINEFORM6 essentially enforces each modality to be mapped into the same unified subspace, where the weighted average of which encodes discrimitive features for recognition of named entities.
When visual context is not provided with each token (as in the traditional NER task), we can define the modality attention for word and character embeddings only in a similar way: [at(w),at(c)] = (Wm[xt(w); xt(c)] + bm )
t(m) = (at(m))m'{w,c}(at(m')) m {w,c}
xt = m{w,c} t(m)xt(m)
Note that while we apply this modality attention module to the Bi-LSTM+CRF architecture (Section SECREF4 ) for its empirical superiority, the module itself is flexible and thus can work with other NER architectures or for other multimodal applications.
## SnapCaptions Dataset
The SnapCaptions dataset is composed of 10K user-generated image (snap) and textual caption pairs where named entities in captions are manually labeled by expert human annotators (entity types: PER, LOC, ORG, MISC). These captions are collected exclusively from snaps submitted to public and crowd-sourced stories (aka Snapchat Live Stories or Our Stories). Examples of such public crowd-sourced stories are “New York Story” or “Thanksgiving Story”, which comprise snaps that are aggregated for various public events, venues, etc. All snaps were posted between year 2016 and 2017, and do not contain raw images or other associated information (only textual captions and obfuscated visual descriptor features extracted from the pre-trained InceptionNet are available). We split the dataset into train (70%), validation (15%), and test sets (15%). The captions data have average length of 30.7 characters (5.81 words) with vocabulary size 15,733, where 6,612 are considered unknown tokens from Stanford GloVE embeddings BIBREF22 . Named entities annotated in the SnapCaptions dataset include many of new and emerging entities, and they are found in various surface forms (various nicknames, typos, etc.) To the best of our knowledge, SnapCaptions is the only dataset that contains natural image-caption pairs with expert-annotated named entities.
## Baselines
Task: given a caption and a paired image (if used), the goal is to label every token in a caption in BIO scheme (B: beginning, I: inside, O: outside) BIBREF27 . We report the performance of the following state-of-the-art NER models as baselines, as well as several configurations of our proposed approach to examine contributions of each component (W: word, C: char, V: visual).
Bi-LSTM/CRF (W only): only takes word token embeddings (Stanford GloVE) as input. The rest of the architecture is kept the same.
Bi-LSTM/CRF + Bi-CharLSTM (C only): only takes a character sequence of each word token as input. (No word embeddings)
Bi-LSTM/CRF + Bi-CharLSTM (W+C) BIBREF0 : takes as input both word embeddings and character embeddings extracted from a Bi-CharLSTM. Entity LSTM takes concatenated vectors of word and character embeddings as input tokens.
Bi-LSTM/CRF + CharCNN (W+C) BIBREF1 : uses character embeddings extracted from a CNN instead.
Bi-LSTM/CRF + CharCNN (W+C) + Multi-task BIBREF8 : trains the model to perform both recognition (into multiple entity types) as well as segmentation (binary) tasks.
(proposed) Bi-LSTM/CRF + Bi-CharLSTM with modality attention (W+C): uses the modality attention to merge word and character embeddings.
(proposed) Bi-LSTM/CRF + Bi-CharLSTM + Inception (W+C+V): takes as input visual contexts extracted from InceptionNet as well, concatenated with word and char vectors.
(proposed) Bi-LSTM/CRF + Bi-CharLSTM + Inception with modality attention (W+C+V): uses the modality attention to merge word, character, and visual embeddings as input to entity LSTM.
## Results: SnapCaptions Dataset
Table TABREF6 shows the NER performance on the Snap Captions dataset. We report both entity types recognition (PER, LOC, ORG, MISC) and named entity segmentation (named entity or not) results.
Parameters: We tune the parameters of each model with the following search space (bold indicate the choice for our final model): character embeddings dimension: {25, 50, 100, 150, 200, 300}, word embeddings size: {25, 50, 100, 150, 200, 300}, LSTM hidden states: {25, 50, 100, 150, 200, 300}, and INLINEFORM0 dimension: {25, 50, 100, 150, 200, 300}. We optimize the parameters with Adagrad BIBREF28 with batch size 10, learning rate 0.02, epsilon INLINEFORM1 , and decay 0.0.
Main Results: When visual context is available (W+C+V), we see that the model performance greatly improves over the textual models (W+C), showing that visual contexts are complimentary to textual information in named entity recognition tasks. In addition, it can be seen that the modality attention module further improves the entity type recognition performance for (W+C+V). This result indicates that the modality attention is able to focus on the most effective modality (visual, words, or characters) adaptive to each sample to maximize information gain. Note that our text-only model (W+C) with the modality attention module also significantly outperform the state-of-the-art baselines BIBREF8 , BIBREF1 , BIBREF0 that use the same textual modalities (W+C), showing the effectiveness of the modality attention module for textual models as well.
Error Analysis: Table TABREF17 shows example cases where incorporation of visual contexts affects prediction of named entities. For example, the token `curry' in the caption “The curry's " is polysemous and may refer to either a type of food or a famous basketball player `Stephen Curry', and the surrounding textual contexts do not provide enough information to disambiguate it. On the other hand, visual contexts (visual tags: `parade', `urban area', ...) provide similarities to the token's distributional semantics from other training examples (snaps from “NBA Championship Parade Story"), and thus the model successfully predicts the token as a named entity. Similarly, while the text-only model erroneously predicts `Apple' in the caption “Grandma w dat lit Apple Crisp" as an organization (Apple Inc.), the visual contexts (describing objects related to food) help disambiguate the token, making the model predict it correctly as a non-named entity (a fruit). Trending entities (musicians or DJs such as `CID', `Duke Dumont', `Marshmello', etc.) are also recognized correctly with strengthened contexts from visual information (describing concert scenes) despite lack of surrounding textual contexts. A few cases where visual contexts harmed the performance mostly include visual tags that are unrelated to a token or its surrounding textual contexts.
Visualization of Modality Attention: Figure FIGREF19 visualizes the modality attention module at each decoding step (each column), where amplified modality is represented with darker color, and attenuated modality is represented with lighter color.
For the image-aided model (W+C+V; upper row in Figure FIGREF19 ), we confirm that the modality attention successfully attenuates irrelevant signals (selfies, etc.) and amplifies relevant modality-based contexts in prediction of a given token. In the example of “disney word essential = coffee" with visual tags selfie, phone, person, the modality attention successfully attenuates distracting visual signals and focuses on textual modalities, consequently making correct predictions. The named entities in the examples of “Beautiful night atop The Space Needle" and “Splash Mountain" are challenging to predict because they are composed of common nouns (space, needle, splash, mountain), and thus they often need additional contexts to correctly predict. In the training data, visual contexts make stronger indicators for these named entities (space needle, splash mountain), and the modality attention module successfully attends more to stronger signals.
For text-only model (W+C), we observe that performance gains mostly come from the modality attention module better handling tokens unseen during training or unknown tokens from the pre-trained word embeddings matrix. For example, while WaRriOoOrs and Kooler Matic are missing tokens in the word embeddings matrix, it successfully amplifies character-based contexts (capitalized first letters, similarity to known entities `Golden State Warriors') and suppresses word-based contexts (word embeddings for unknown tokens `WaRriOoOrs'), leading to correct predictions. This result is significant because it shows performance of the model, with an almost identical architecture, can still improve without having to scale the word embeddings matrix indefinitely.
Figure FIGREF19 (b) shows the cases where the modality attention led to incorrect predictions. For example, the model predicts missing tokens HUUUGE and Shampooer incorrectly as named entities by amplifying misleading character-based contexts (capitalized first letters) or visual contexts (concert scenes, associated contexts of which often include named entities in the training dataset).
Sensitivity to Word Embeddings Vocabulary Size: In order to isolate the effectiveness of the modality attention module on textual models in handling missing tokens, we report the performance with varying word embeddings vocabulary sizes in Table TABREF20 . By increasing the number of missing tokens artificially by randomly removing words from the word embeddings matrix (original vocab size: 400K), we observe that while the overall performance degrades, the modality attention module is able to suppress the peformance degradation. Note also that the performance gap generally gets bigger as we decrease the vocabulary size of the word embeddings matrix. This result is significant in that the modality attention is able to improve the model more robust to missing tokens without having to train an indefinitely large word embeddings matrix for arbitrarily noisy social media text datasets.
## Conclusions
We proposed a new multimodal NER (MNER: image + text) task on short social media posts. We demonstrated for the first time an effective MNER system, where visual information is combined with textual information to outperform traditional text-based NER baselines. Our work can be applied to myriads of social media posts or other articles across multiple platforms which often include both text and accompanying images. In addition, we proposed the modality attention module, a new neural mechanism which learns optimal integration of different modes of correlated information. In essence, the modality attention learns to attenuate irrelevant or uninformative modal information while amplifying the primary modality to extract better overall representations. We showed that the modality attention based model outperforms other state-of-the-art baselines when text was the only modality available, by better combining word and character level information.
| 10
|
1803.07771
|
$\rho$-hot Lexicon Embedding-based Two-level LSTM for Sentiment Analysis
|
# $\rho$-hot Lexicon Embedding-based Two-level LSTM for Sentiment Analysis
## Abstract
Sentiment analysis is a key component in various text mining applications. Numerous sentiment classification techniques, including conventional and deep learning-based methods, have been proposed in the literature. In most existing methods, a high-quality training set is assumed to be given. Nevertheless, constructing a high-quality training set that consists of highly accurate labels is challenging in real applications. This difficulty stems from the fact that text samples usually contain complex sentiment representations, and their annotation is subjective. We address this challenge in this study by leveraging a new labeling strategy and utilizing a two-level long short-term memory network to construct a sentiment classifier. Lexical cues are useful for sentiment analysis, and they have been utilized in conventional studies. For example, polar and privative words play important roles in sentiment analysis. A new encoding strategy, that is, $\rho$-hot encoding, is proposed to alleviate the drawbacks of one-hot encoding and thus effectively incorporate useful lexical cues. We compile three Chinese data sets on the basis of our label strategy and proposed methodology. Experiments on the three data sets demonstrate that the proposed method outperforms state-of-the-art algorithms.
## Introduction
Text is important in many artificial intelligence applications. Among various text mining techniques, sentiment analysis is a key component in applications such as public opinion monitoring and comparative analysis. Sentiment analysis can be divided into three problems according to input texts, namely, sentence, paragraph, and document levels. This study focuses on sentence and paragraph levels.
Text sentiment analysis is usually considered a text classification problem. Almost all existing text classification techniques are applied to text sentiment analysis BIBREF0 . Typical techniques include bag-of-words (BOW)-based BIBREF1 , deep learning-based BIBREF2 , and lexicon-based (or rule-based) methods BIBREF3 .
Although many achievements have been made and sentiment analysis has been successfully used in various commercial applications, its accuracy can be further improved. The construction of a high-accuracy sentiment classification model usually entails the challenging compilation of training sets with numerous samples and sufficiently accurate labels. The reason behind this difficulty is two-fold. First, sentiment is somewhat subjective, and a sample may receive different labels from different users. Second, some texts contain complex sentiment representations, and a single label is difficult to provide. We conduct a statistical analysis of public Chinese sentiment text sets in GitHub. The results show that the average label error is larger than 10%. This error value reflects the degree of difficulty of sentiment labeling.
Privative and interrogative sentences are difficult to classify when deep learning-based methods are applied. Although lexicon-based methods can deal with particular types of privative sentences, their generalization capability is poor.
We address the above issues with a new methodology. First, we introduce a two-stage labeling strategy for sentiment texts. In the first stage, annotators are invited to label a large number of short texts with relatively pure sentiment orientations. Each sample is labeled by only one annotator. In the second stage, a relatively small number of text samples with mixed sentiment orientations are annotated, and each sample is labeled by multiple annotators. Second, we propose a two-level long short-term memory (LSTM) BIBREF4 network to achieve two-level feature representation and classify the sentiment orientations of a text sample to utilize two labeled data sets. Lastly, in the proposed two-level LSTM network, lexicon embedding is leveraged to incorporate linguistic features used in lexicon-based methods.
Three Chinese sentiment data sets are compiled to investigate the performance of the proposed methodology. The experimental results demonstrate the effectiveness of the proposed methods. Our work is new in the following aspects.
The rest of this paper is organized as follows. Section 2 briefly reviews related work. Section 3 describes our methodology. Section 4 reports the experimental results, and Section 5 concludes the study.
## Text Sentiment Analysis
Sentiment analysis aims to predict the sentiment polarity of an input text sample. Sentiment polarity can be divided into negative, neutral, and positive in many applications.
Existing sentiment classification methods can be roughly divided into two categories, namely, lexicon-based and machine learning-based methods BIBREF5 . Lexicon-based methods BIBREF6 construct polar and privative word dictionaries. A set of rules for polar and privative words is compiled to judge the sentiment orientation of a text document. This method cannot effectively predict implicit orientations. Machine learning-based methods BIBREF7 utilize a standard binary or multi-category classification approach. Different feature extraction algorithms, including BOW BIBREF8 and part of speech (POS) BIBREF7 , are used. Word embedding and deep neural networks have recently been applied to sentiment analysis, and promising results have been obtained BIBREF9 BIBREF10 .
## Lexion-based Sentiment Classification
Lexicon-based methods are actually in implemented in an unsupervised manner. They infer the sentiment categories of input texts on the basis of polar and privative words. The primary advantage of these methods is that they do not require labeled training data. The key of lexicon-based methods is the lexical resource construction, which maps words into a category (positive, negative, neutral, or privative). Senti-WordNet BIBREF11 is a lexical resource for English text sentiment classification. For Chinese texts, Senti-HowNet is usually used.
Fig. 1 characterizes a typical lexicon-based sentiment classification approach. The approach iteratively checks each word in an input sentence from left to right. The weight score of each word is calculated according to the procedure shown in Fig. 1. The final sentiment score is the average score of the words with weight scores. The scores of positive, neutral, and negative sentiments are denoted as “+1",“0", and “-1", respectively. According to the lexicon-based algorithm shown in Fig. 1, the sentiment score of “it is not bad" is 0.25, and the sentiment score of “it is good" is 1. However, the score of “it is not so bad" is -0.75, and this score is definitely wrong. Therefore, machine learning (including feature learning) methodologies have become mainstream in sentiment analysis.
## Deep Learning-based Sentiment Classification
Deep learning (including word embedding BIBREF12 ) has been applied to almost all text-related applications, such as translation BIBREF13 , quality assurance BIBREF14 , recommendation BIBREF15 , and categorization BIBREF16 . Popular deep neural networks are divided into convolutional neural networks (CNNs) BIBREF17 and recurrent neural network (RNNs) BIBREF18 BIBREF19 . Both are utilized in sentiment classification BIBREF20 . Kim investigated the use of CNN in sentence sentiment classification and achieved promising results BIBREF2 . LSTM BIBREF21 , a classical type of RNN, is the most popular network used for sentiment classification. A binary-directional LSTM BIBREF22 with an attention mechanism is demonstrated to be effective in sentiment analysis.
Deep learning-based methods rarely utilize the useful resources adopted in lexicon-based methods. Qiao et al. BIBREF23 incorporated lexicon-based cues into the training of an LSTM-based model. Their proposed method relies on a new loss function that considers the relationships between polar or certain types of words (e.g., privative) and those words next to them in input texts. Our study also combines lexical cues into LSTM. Nevertheless, unlike Qiao et al.'s study that implicitly used lexical cues, the present work explicitly uses lexical cues in the LSTM network. Shin et al. BIBREF24 combined the lexicon embeddings of polar words with word embeddings for sentiment classification. The difference between our approach an the method proposed by Shin et al. the is discussed in Section 3.3.5.
Numerous studies on aspect-level sentiment analysis exist BIBREF25 . This problem is different from the sentiment classification investigated in this study.
## METHODOLOGY
This section first introduces our two-stage labeling procedure. A two-level LSTM is then proposed. Lexicon embedding is finally leveraged to incorporate lexical cues.
## Two-stage Labeling
As stated earlier, sentiment is subjective, and texts usually contain mixed sentiment orientations. Therefore, texts¡¯ sentiment orientations are difficult to label. In our study, three sentiment labels, namely, positive, neutral, and negative, are used. The following sentences are taken as examples.
The service is poor. The taste is good, but the rest is not so bad.
The quality of the phone is good, but the appearance is just so-so.
In user annotation, the labels of these two sentences depend on users. If a user is concerned about service, then the label of S1 may be “negative". By contrast, for another user who does not care about service, the label may be “positive". Similarly, a user may label S2 as “positive" if he cares about quality. Another user may label it as “negative" if the conjunction “but" attracts the user¡¯s attention more. Another user may label it as “neutral" if they are concerned about quality and appearance.
The underlying reason is that sentiment is more subjective than semantics. In related research on subjective categorization, such as visual aesthetics, each sample is usually repeatedly annotated by multiple annotators, and the average label is taken as the final label of the sample. This labeling strategy can also be applied to text sentiment annotation. However, we argue that this strategy is unsuitable for a (relatively) large number of samples. The reason lies in the following two aspects.
Multiple annotators for a large number of data sets require a large budget.
In our practice, annotators claim that their judgment criteria on sentiment become fused on texts with mixed sentiment orientations (e.g., S1 and S2) over time during labeling, and they become bored accordingly.
A two-stage labeling strategy is adopted in this study. In the first stage, each sentence/paragraph is divided into several clauses according to punctuation. The sentiment of each partitioned clause is relatively easy to annotate; therefore, each clause is labeled by only one user. In the second stage, a relatively small-sized sentence/paragraph set is labeled, and each sentence is labeled by all involved annotators. We still take the two sentences, S1 and S2, as examples. S1 and S2 are split into clauses, as shown below.
S1:
S1.1: The service is poor
S1.2: The taste is good
S1.3: but the rest is not so bad.
S2:
S2.1: The quality of the phone is good
S2.2: but the appearance is just so-so.
Each of the above clauses is labeled by only one annotator. The annotation in the first stage is easy to perform; thus, the number of clauses can be larger than the number of sentences used in the second labeling stage.
## Two-level LSTM
Given two training data sets (denoted by T1 and T2), a new learning model should be utilized. LSTM is a widely used deep neural network in deep learning-based text classification.
LSTM is a typical RNN model for short-term memory, which can last for a long period of time. An LSTM is applicable to classify, process, and predict time series information with given time lags of unknown size. A common LSTM block is composed of a cell, an input gate, an output gate, and a forget gate. The forward computation of an LSTM block at time INLINEFORM0 or position INLINEFORM1 is as follows BIBREF21 : DISPLAYFORM0
where INLINEFORM0 is the input vector at time INLINEFORM1 (or position INLINEFORM2 ); INLINEFORM3 and INLINEFORM4 are the input vectors of the input unit and input gate, respectively; INLINEFORM5 and INLINEFORM6 are the output and hidden vectors at time INLINEFORM7 , respectively; INLINEFORM8 is the output of the forget gate at time INLINEFORM9 ; INLINEFORM10 is the internal state of the memory cell in an LSTM block at time INLINEFORM11 ; and INLINEFORM12 is the sigmoid active function.
When LSTM is used to classify an input sentence, the hidden vectors of each input vector are summed to form a dense vector that can be considered the feature representation of the input sentence, i.e., DISPLAYFORM0
In many applications, a bi-directional LSTM (bi-LSTM) structure is usually used, as shown in Fig. 2(a). In bi-LSTM, forward and backward information are considered for information at time INLINEFORM0 ; hence, the context is modeled. Bi-LSTM is thus significantly reasonable for text processing tasks. In our two-level LSTM, bi-LSTM is used in each level.
The output hidden state at time INLINEFORM0 of a bi-LSTM block can be described as follows: DISPLAYFORM0
where INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are the corresponding vectors at time INLINEFORM3 in the forward LSTM block; and INLINEFORM4 , INLINEFORM5 , and INLINEFORM6 are the corresponding vectors at time INLINEFORM7 in the backward LSTM block. INLINEFORM8 . When attention is used, the dense feature vector INLINEFORM9 of an input sentence is calculated as follows: DISPLAYFORM0
where INLINEFORM0 is the vector that consists of attention weights. The bi-LSTM with attention is shown in Fig. 2(b).
Our proposed network consists of two levels of LSTM network. In the first level, a bi-LSTM is used and learned on the basis of the first training set T1. This level is a conventional sentiment classification process. The input of this level is a clause, and the input INLINEFORM0 is the embedding of the basic unit of the input texts. The network is shown in Fig. 3(a).
In the second level, a bi-LSTM is also used and learned on the basis of the second training set T2. The input of this level is a sentence or a paragraph. The input INLINEFORM0 consists of two parts. The first part is the feature vector of the INLINEFORM1 -th clause. The feature vector is generated by the first-level network. In other words, the dense feature shown in Fig. 3(a) ( INLINEFORM2 ) is used. The second part is the sentiment score (not predicted label) output by the first-level network. The sentence S1 (The service is poor. The taste is good, but the rest is not so bad.) used in Subsection 3.1 is taken as an illustrative example. S1 contains three clauses. Therefore, the input vector of S1 can be represented by INLINEFORM3
where DISPLAYFORM0
where INLINEFORM0 is the output score of the INLINEFORM1 th clause by the first-level LSTM and INLINEFORM2 is the feature representation of the INLINEFORM3 th clause by the first LSTM. The network of the whole two-level network is shown in Fig. 3(b).
## Lexical Embedding
The proposed lexicon embedding is based on INLINEFORM0 -hot encoding. Therefore, INLINEFORM1 -hot encoding is first described.
For categorical data, one-hot encoding is the most widely used encoding strategy when different categories are independent. For example, if one-hot encoding is used to represent three categories, namely, positive, neutral, and negative, the encoding vectors for the three categories are INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 , respectively.
In this work, many lexical cues are categorical data, and different categories are independent. These lexical cues can directly be represented by one-hot encoding. The encoded vectors for lexical cues are then concatenated with other vectors, such as character/word embedding. However, one-hot encoding presents two main limitations when the encoded vector is concatenated with other vectors.
The value difference between the elements of one-hot encoded vectors and those of other encoded vectors (e.g., word embedding vectors) may be large. Fig. 4 shows the histogram of the values of the elements of the word embedding vectors. The magnitude of most elements are smaller than 1.
The lengths of one-hot encoded vectors are usually shorter than those of other encoded vectors. Consequently, the proportion of one-hot encoded part is small in the concatenated vectors.
The above two limitations affect the final sentiment analysis performance. To this end, we propose a new encoding strategy. DISPLAYFORM0
where INLINEFORM0 is the INLINEFORM1 -hot encoded vector, INLINEFORM2 is the proportion parameter, INLINEFORM3 is the one-hot encoded vector, and INLINEFORM4 is an INLINEFORM5 -dimensional vector. If INLINEFORM6 and INLINEFORM7 are equal to 1, then INLINEFORM8 -hot encoding is reduced to one-hot encoding. The parameter INLINEFORM9 is applied to increase the length of the final encoded vector.
Most lexicon-based sentiment methods rely on four types of words, namely, positive, negative, neutral, and privative. These words are useful cues for predicting the sentiment labels of input texts. The incorporation of these words should also be useful. A previous study has shown that a typical document comprises approximately 8% of such sentences BIBREF26 . Sentiments expressed in a conditional sentence can be difficult to determine due to the semantic condition. The sentiment polarities of interrogative sentences are also difficult to classify according to our empirical study.
Five types of words, namely, positive (Pos), negative (Neg), privative (Pri), suppositive (Sup), and interrogative (Int), are represented by the proposed encoding method. The rest words, which do not belong to any of the above five types, are named “others (Oth)" instead of “neutral" because some words, such as “the", are unrelated to “sentiment". The value of INLINEFORM0 in Eq. (6) is set as 10. The encoded vectors are as follows. INLINEFORM1
In the proposed INLINEFORM0 -hot embedding, the parameter INLINEFORM1 can be learned during training. The representation of the third clause (“but the rest is not so bad") of S1 in Subsection 3.1 is taken as an illustrative example. The new embedding of each word in this clause is as follows. DISPLAYFORM0
Certain types (e.g., positive, negative, and privative) of words should play more important roles than other words do in texts; therefore, their embeddings are also used in the attention layer. A new LSTM based on our lexicon embedding is proposed, as shown in Fig. 5. The attention layer and final dense vector of the network in Fig. 3(a) are calculated as follows. DISPLAYFORM0
where INLINEFORM0 is the attention weight for the INLINEFORM1 -th input, lt is the lexicon embedding for key lexical words for the INLINEFORM2 -th input, and INLINEFORM3 is the final dense vector. Eq. (2) is used in the first-level LSTM.
POS is usually used as a key cue in sentiment analysis BIBREF27 . To this end, we use additional lexicon embedding. The new lexicon embedding includes several major types of POS, namely, interrogative, exclamatory, and others. This new lexicon embedding is also applied to the attention layer. The motivation lies in that certain types of POS should play important roles in sentiment.
The proposed INLINEFORM0 -hot embedding is still applied to POS types in this study. According to our initial case studies, eight POS types are considered. They are noun, adjective, verb, pronoun, adverb, preposition, accessory, and others. The eight POS types are represented by the proposed INLINEFORM1 -hot encoding. We let INLINEFORM2 in Eq. (6) be 10. The first three POS types are as follows. INLINEFORM3
When POS embedding is used, the attention layer and final outputs of the network in Eq. (3) become DISPLAYFORM0
where INLINEFORM0 is the lexicon embedding for key lexical words for the INLINEFORM1 -th input.
Conjunction words play important roles in sentiment analysis BIBREF28 . For example, conjunctions such as “but" and “moreover" usually indicate the focus of texts and attract readers¡¯ attention. Therefore, conjunctions are considered in the input of the second-level LSTM.
Once a set of conjunction words is compiled, INLINEFORM0 -hot embedding is used. In our experiments, the number of conjunction words is 169. Therefore, the parameter INLINEFORM1 in Eq. (2) is set as 1.
When conjunction embedding is used for the second-level layer, the attention layer and final outputs of the network in Fig. 3(b) are calculated as follows. DISPLAYFORM0
where INLINEFORM0 is the attention weight for the INLINEFORM1 -th input clause; INLINEFORM2 is the hidden vector of the second-level LSTM; INLINEFORM3 and INLINEFORM4 are the conjunction embeddings for the first and last words in the INLINEFORM5 -th input clause, respectively; and INLINEFORM6 is the final dense vector used for the final classification.
Shin et al. BIBREF24 also embedded lexical information into sentiment analysis. Three major differences exist between our method and the method proposed by Shin et al. BIBREF24 .
The lexicon embedding proposed by Shin et al. us-es one-hot encoding, whereas the proposed method uses a new encoding strategy that can be considered a soft one-hot encoding.
The lexicon embedding proposed by Shin et al. ex-tends the length of raw encoded vectors. However, the extension aims to keep the lengths of lexical and word embeddings equal. Their extension method also only relies on zero padding and is thus different from the proposed method.
Only sentimental words are considered in the lexicon embedding proposed by Shin et al. On the contrary, sentimental words, POS, and conjunctions are considered in our work.
## The Learning Procedure
The algorithmic steps of the entire learning procedure for the proposed INLINEFORM0 -hot lexicon embedding-based two-level LSTM (called INLINEFORM1 Tl-LSTM) are shown in Algorithm 1. In Algorithm 1, T1 refers to the training data that consist of clauses and the labels obtained in the first-stage labeling procedure. T2 refers to the training data that consist of sentences and the labels obtained in the second-stage labeling procedure. The structure of INLINEFORM2 Tl-LSTM is presented in Fig. 6.
INLINEFORM0 Tl-LSTM Input: Training sets T1 and T2; dictionary of key lexical words; POS for each word; dictionary of conjunction words; character/word embeddings for each character/word.
Output: A trained two-level LSTM for sentiment classification.
Steps:
Construct the embedding vector for each character (including punctuation) in the clauses in T1. The embeddings include the character/word and lexicon embeddings of each character/word;
Train the first-level LSTM on the basis of the input embedding vectors and labels of the T1 text clauses;
Run the learned first-level LSTM on each clause of the text samples in T2. Record the predicted score INLINEFORM0 and the final dense vector INLINEFORM1 for each clause;
Construct the embedding vector for each clause in the text samples in T2. Each embedding vector consists of INLINEFORM0 , INLINEFORM1 , and the lexicon embedding of conjunctions of each clause;
Train the second-level LSTM on the basis of the input embedding vectors and labels of the T2 text samples.
The first-level and second-level LSTM networks consist of the final two-level LSTM.
The proposed two-level LSTM can be applied to texts with arbitrary languages. Word information is required in lexical construction regardless of whether character or word embedding is used. The reason is that the three types of lexicon embeddings are performed at the word level. Therefore, when character embedding is used, the lexicon embedding of each character is the lexicon embedding of the word containing it.
This section shows the evaluation of the proposed methodology in terms of the two-level LSTM network and each part of the lexicon embedding.
We compile three Chinese text corpora from online data for three domains, namely, “hotel", “mobile phone (mobile)", and “travel". All texts are about user reviews. Each text sample collected is first partitioned into clauses according to Chinese tokens. Three clause sets are subsequently obtained from the three text corpora.
The labels “+1", “0.5", and “0" correspond to the three sentiment classes “positive", “neutral", and “negative", respectively. The text data are labeled according to our two-stage labeling strategy.
In the first stage, only one user is invited to label each clause sample as the sentiment orientations for clauses (or sub-sentences) are easy to label.
In the second stage, five users are invited to label each text sample in the three raw data sets. The average score of the five users on each sample is calculated. Samples with average scores located in [0.6, 1] are labeled as “positive". Samples with average scores located in [0, 0.4] are labeled as “negative". Others are labeled as “neutral". The details of the labeling results are shown in Table 1.
All the training and test data and the labels are available online.
In our experiments, the five types of key lexical words introduced in Subsection 3.3.2 are manually constructed. The details of the five types of words are listed in Table 2. The conjunction words are also manually constructed. The number of conjunction words used in the experiments is 169.
In each experimental run, the training set is compiled on the basis of the training data listed in Table 1. The compiling rule is specified before each experimental run. The test data are fixed to facilitate experimental duplication and comparison by other researchers.
In our experiments, three competing algorithms, namely, BOW, CNN, and (conventional) LSTM, are used.
For BOW, term frequency-inverse document frequency is utilized to construct features. Ridge regression BIBREF29 is used as a classifier. For CNN, a three-channel CNN is used. For LSTM, one-layer and two-layer bi-LSTM with attention are adopted, and the results of the network with superior performance are presented. CNN and LSTM are performed on TensorFlow, and default parameter settings are followed.
The key parameters are searched as follows. The embedding dimensions of characters and words are searched in [100, 150, 200, 250, 300]. The parameter INLINEFORM0 in INLINEFORM1 -hot encoding is searched in INLINEFORM2 .
In this subsubsection, each of the three raw data sets (associated with their labels) shown in Table 1 is used. The clause data are not used. In other words, the training data used in this subsubsection are the same as those used in previous studies. For each data corpus, 1000 raw data samples are used as the test data, and the rest are used as the training data. The involved algorithms are detailed as follows.
CNN-C denotes the CNN with (Chinese) character embedding.
CNN-W denotes the CNN with (Chinese) word embedding.
CNN-Lex-C denotes the algorithm which also integrates polar words in CNN which is proposed by Shin et al. BIBREF24 . The (Chinese) character embedding is used.
CNN-Lex-W denotes the algorithm which also integrates polar words in CNN which is proposed by Shin et al. BIBREF24 . The (Chinese) word embedding is used.
Bi-LSTM-C denotes the BI-LSTM with (Chinese) character embedding.
Bi-LSTM-W denotes the Bi-LSTM with (Chinese) word embedding.
Lex-rule denotes the rule-based approach shows in Fig. 1. This approach is unsupervised.
BOW denotes the conventional algorithm which is based of bag-of-words features.
The accuracies of the above algorithms are listed in Table 3. Overall, Bi-LSTM outperforms CNN and BOW. This conclusion is in accordance with the conclusion that RNN performs efficiently against CNN in a broad range of natural language processing (NLP) tasks on the basis of extensive comparative studies BIBREF30 . In addition, CNN-lex outperforms CNN under both character and word embeddings, which suggests that lexicon cues are useful in sentiment analysis. Lex-rule achieves the lowest accuracies on all the three data sets. Considering that the performances of (traditional) CNN, Lex-rule, and BOW are relatively poor, they are not applied in the remaining parts.
In this experimental comparison, the proposed two-level LSTM is evaluated, whereas lexicon embedding is not used in the entire network. The primary goal is to test whether the introduced two-stage labeling and the two-level network structure are useful for sentiment analysis.
The raw and clause data listed in Table 1 are used to perform the two-level LSTM. Tl-LSTM denotes the two-level LSTM. “R+C" refer to the mixed data of raw and clause data. The test data are still the 1000 samples used in section 4.3.1 for each corpus. Table 4 shows the classification accuracies. To ensure that the results differ from those in Table 3, we explicitly add “R+C" after each algorithm in Table 4. In the last line of Table 4, the base results for each corpus in Table 3 are also listed.
On all the three data corpora, the proposed two-level network (without lexicon embedding) with character embedding, Tl-LSTM-C, outperforms all the other involved algorithms. On the travel and the mobile corpora, TI-LSTM-W outperforms Bi-LSTM-W. The results in Table 4 indicate that the performances of Tl-LSTM on the mixed training and test data (R+C) are better than those of Bi-LSTM. This comparison indicates that the proposed two-level LSTM is effective.
In addition, for the involved algorithms, most results achieved on “R+C" are better than the best results only achieved on `R'listed in Table 3. This comparison suggests that the introduced two-stage labeling is useful.
The results also show that in the two-level LSTM, character embedding is more effective than word embedding.
In this experimental run, lexicon embedding is used in the proposed two-level LSTM or INLINEFORM0 Tl-LSTM. Table 5 shows the results. The optimal parameter INLINEFORM1 is about 11.
The performances of TI-LSTM with lexicon embedding (i.e., INLINEFORM0 Tl-LSTM) are consistently better than those of TI-LSTM without lexicon embedding (i.e., Tl-LSTM) listed in Table 5. The improved accuracies of INLINEFORM1 TI-LSTM over Tl-LSTM on the three data corpora are explicitly listed in Table 6.
The experimental evaluation discussed in Subsection 4.3 verifies the effectiveness of the proposed method, INLINEFORM0 Tl-LSTM. Unlike the conventional RNN, INLINEFORM1 Tl-LSTM contains lexicon embedding that consists of new technique and components, including INLINEFORM2 -hot encoding, embedding for polar words, embedding for POS, and embedding for conjunctions. Therefore, this subsection evaluates the performances of the involved technique and embeddings separately.
Our INLINEFORM0 -hot encoding differs from one-hot encoding in two aspects. The first aspect is that the nonzero values in one-hot encoding are only equal to 1, whereas the nonzero values in INLINEFORM1 -hot encoding are INLINEFORM2 . The second aspect is that only one element in one-hot encoding is nonzero, whereas n elements in INLINEFORM3 -hot encoding are nonzero.
In this experiment, we test whether INLINEFORM0 -hot encoding is useful in two experimental runs. In the first run, the value of INLINEFORM1 is manually set to 0.5 and 1 in the experimental run without optimization. The parameter INLINEFORM2 in Eq. (6) is set as 15. The classification accuracies vary according to different INLINEFORM3 values on all the three data corpora. When INLINEFORM4 equals 1, the accuracies are the lowest in most cases shown in Fig. 7.
The results shown in Fig. 7 indicate that the value of INLINEFORM0 does affect the performance of the entire network. Consequently, the classical one-hot encoding, which fixes the value of nonzero elements as 1, is ineffective. In our experiments, the learned value of INLINEFORM1 is approximate 0.4.
In the second run, the performances under different INLINEFORM0 (i.e., 1, 5, 10, 15) are tested. Table 7 shows the comparison results. The value of INLINEFORM1 does affect the performance of the entire network, thereby indicating that the introduction of the INLINEFORM2 -duplicated strategy in encoding is effective. In the experiments, when INLINEFORM3 is increasing, the accuracies first increase and then decrease. The main reason may lie in the fact that when INLINEFORM4 becomes large, the proportion of lexicon embedding becomes large accordingly. An over-length input feature vector may incur “curse of dimensionality" and thus weaken the performance of the proposed two-level network.
In this experimental run, we test whether the labeled polar (negative and positive) words do affect the performance of the entire method when they are used in lexicon embedding. To this end, we order the polar words according to their frequencies in the training data. Top 0%, 50%, 100% polar words are used. The corresponding classification accuracies are depicted in Fig. 8.
In most cases, the accuracies are the lowest when no polar words are used in the lexicon embedding. When all polar words are used, the proposed network achieves the highest accuracies.
In the experiment, only one user is invited to manually compile the dictionary for a data corpus. One and a half hour is needed for each data corpus. In our viewpoint, it is worth manually compiling the polar words for sentiment analysis by considering the performance improvement and time-consumption.
In this experimental run, we test whether POS cues do play positive roles in the entire model. To this end, we remove POS in the lexicon embedding of the proposed method. The results are shown in Fig. 9.
In most cases, the accuracies with POS embedding are greater than those without POS embedding, thereby indicating that the application of POS to lexicon embedding is useful.
In this experimental run, we test whether conjunction cues do play positive roles in the entire model. To this end, the lexicon embedding for conjunction words is also removed from the proposed method. The results are shown in Fig. 10.
The algorithm with conjunction embedding outperforms that without conjunction embedding consistently, thereby indicating that the application of conjunction to lexicon embedding is useful.
High-quality labels are crucial for learning systems. Nevertheless, texts with mixed sentiments are difficult for humans to label in text sentiment classification. In this study, a new labeling strategy is introduced to partition texts into those with pure and mixed sentiment orientations. These two categories of texts are labeled using different processes. A two-level network is accordingly proposed to utilize the two labeled data in our two-stage labeling strategy. Lexical cues (e.g., polar words, POS, conjunction words) are particularly useful in sentiment analysis. These lexical cues are used in our two-level network, and a new encoding strategy, that is, INLINEFORM0 -hot encoding, is introduced. INLINEFORM1 -hot encoding is motivated by one-hot encoding. However, the former alleviates the drawbacks of the latter. Three Chinese sentiment text data corpora are compiled to verify the effectiveness of the proposed methodology. Our proposed method achieves the highest accuracies on these three data corpora.
The proposed two-level network and lexicon embedding can also be applied to other types of deep neural networks. In our future work, we will extend our main idea into several networks and text mining applications.
The authors wish to thank Zefeng Han, Qing Yin, Lei Yang, Xiaonan Wang, Nan Chen, Rujing Yao, Lihong Guo, Pinglong Zhao for the labeling of the experimental data.
| 9
|
1804.00079
|
Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning
|
# Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning
## Abstract
A lot of the recent success in natural language processing (NLP) has been driven by distributed vector representations of words trained on large amounts of text in an unsupervised manner. These representations are typically used as general purpose features for words across a range of NLP problems. However, extending this success to learning representations of sequences of words, such as sentences, remains an open problem. Recent work has explored unsupervised as well as supervised learning techniques with different training objectives to learn general purpose fixed-length sentence representations. In this work, we present a simple, effective multi-task learning framework for sentence representations that combines the inductive biases of diverse training objectives in a single model. We train this model on several data sources with multiple training objectives on over 100 million sentences. Extensive experiments demonstrate that sharing a single recurrent sentence encoder across weakly related tasks leads to consistent improvements over previous methods. We present substantial improvements in the context of transfer learning and low-resource settings using our learned general-purpose representations.
## Introduction
Transfer learning has driven a number of recent successes in computer vision and NLP. Computer vision tasks like image captioning BIBREF0 and visual question answering typically use CNNs pretrained on ImageNet BIBREF1 , BIBREF2 to extract representations of the image, while several natural language tasks such as reading comprehension and sequence labeling BIBREF3 have benefited from pretrained word embeddings BIBREF4 , BIBREF5 that are either fine-tuned for a specific task or held fixed.
Many neural NLP systems are initialized with pretrained word embeddings but learn their representations of words in context from scratch, in a task-specific manner from supervised learning signals. However, learning these representations reliably from scratch is not always feasible, especially in low-resource settings, where we believe that using general purpose sentence representations will be beneficial.
Some recent work has addressed this by learning general-purpose sentence representations BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 . However, there exists no clear consensus yet on what training objective or methodology is best suited to this goal.
Understanding the inductive biases of distinct neural models is important for guiding progress in representation learning. BIBREF14 and BIBREF15 demonstrate that neural machine translation (NMT) systems appear to capture morphology and some syntactic properties. BIBREF14 also present evidence that sequence-to-sequence parsers BIBREF16 more strongly encode source language syntax. Similarly, BIBREF17 probe representations extracted by sequence autoencoders, word embedding averages, and skip-thought vectors with a multi-layer perceptron (MLP) classifier to study whether sentence characteristics such as length, word content and word order are encoded.
To generalize across a diverse set of tasks, it is important to build representations that encode several aspects of a sentence. Neural approaches to tasks such as skip-thoughts, machine translation, natural language inference, and constituency parsing likely have different inductive biases. Our work exploits this in the context of a simple one-to-many multi-task learning (MTL) framework, wherein a single recurrent sentence encoder is shared across multiple tasks. We hypothesize that sentence representations learned by training on a reasonably large number of weakly related tasks will generalize better to novel tasks unseen during training, since this process encodes the inductive biases of multiple models. This hypothesis is based on the theoretical work of BIBREF18 . While our work aims at learning fixed-length distributed sentence representations, it is not always practical to assume that the entire “meaning” of a sentence can be encoded into a fixed-length vector. We merely hope to capture some of its characteristics that could be of use in a variety of tasks.
The primary contribution of our work is to combine the benefits of diverse sentence-representation learning objectives into a single multi-task framework. To the best of our knowledge, this is the first large-scale reusable sentence representation model obtained by combining a set of training objectives with the level of diversity explored here, i.e. multi-lingual NMT, natural language inference, constituency parsing and skip-thought vectors. We demonstrate through extensive experimentation that representations learned in this way lead to improved performance across a diverse set of novel tasks not used in the learning of our representations. Such representations facilitate low-resource learning as exhibited by significant improvements to model performance for new tasks in the low labelled data regime - achieving comparable performance to a few models trained from scratch using only 6% of the available training set on the Quora duplicate question dataset.
## Related Work
The problem of learning distributed representations of phrases and sentences dates back over a decade. For example, BIBREF19 present an additive and multiplicative linear composition function of the distributed representations of individual words. BIBREF20 combine symbolic and distributed representations of words using tensor products. Advances in learning better distributed representations of words BIBREF4 , BIBREF5 combined with deep learning have made it possible to learn complex non-linear composition functions of an arbitrary number of word embeddings using convolutional or recurrent neural networks (RNNs). A network's representation of the last element in a sequence, which is a non-linear composition of all inputs, is typically assumed to contain a squashed “summary” of the sentence. Most work in supervised learning for NLP builds task-specific representations of sentences rather than general-purpose ones.
Notably, skip-thought vectors BIBREF6 , an extension of the skip-gram model for word embeddings BIBREF4 to sentences, learn re-usable sentence representations from weakly labeled data. Unfortunately, these models take weeks or often months to train. BIBREF8 address this by considering faster alternatives such as sequential denoising autoencoders and shallow log-linear models. BIBREF21 , however, demonstrate that simple word embedding averages are comparable to more complicated models like skip-thoughts. More recently, BIBREF9 show that a completely supervised approach to learning sentence representations from natural language inference data outperforms all previous approaches on transfer learning benchmarks. Here we use the terms “transfer learning performance" on “transfer tasks” to mean the performance of sentence representations evaluated on tasks unseen during training. BIBREF10 demonstrated that representations learned by state-of-the-art large-scale NMT systems also generalize well to other tasks. However, their use of an attention mechanism prevents the learning of a single fixed-length vector representation of a sentence. As a result, they present a bi-attentive classification network that composes information present in all of the model's hidden states to achieve improvements over a corresponding model trained from scratch. BIBREF11 and BIBREF12 demonstrate that discourse-based objectives can also be leveraged to learn good sentence representations.
Our work is most similar to that of BIBREF22 , who train a many-to-many sequence-to-sequence model on a diverse set of weakly related tasks that includes machine translation, constituency parsing, image captioning, sequence autoencoding, and intra-sentence skip-thoughts. There are two key differences between that work and our own. First, like BIBREF10 , their use of an attention mechanism prevents learning a fixed-length vector representation for a sentence. Second, their work aims for improvements on the same tasks on which the model is trained, as opposed to learning re-usable sentence representations that transfer elsewhere.
We further present a fine-grained analysis of how different tasks contribute to the encoding of different information signals in our representations following work by BIBREF14 and BIBREF17 .
BIBREF23 similarly present a multi-task framework for textual entailment with task supervision at different levels of learning. “Universal" multi-task models have also been successfully explored in the context of computer vision problems BIBREF24 , BIBREF25 .
## Sequence-to-Sequence Learning
Five out of the six tasks that we consider for multi-task learning are formulated as sequence-to-sequence problems BIBREF26 , BIBREF27 . Briefly, sequence-to-sequence models are a specific case of encoder-decoder models where the inputs and outputs are sequential. They directly model the conditional distribution of outputs given inputs INLINEFORM0 . The input INLINEFORM1 and output INLINEFORM2 are sequences INLINEFORM3 and INLINEFORM4 . The encoder produces a fixed length vector representation INLINEFORM5 of the input, which the decoder then conditions on to generate an output. The decoder is auto-regressive and breaks down the joint probability of outputs into a product of conditional probabilities via the chain rule: INLINEFORM6
BIBREF26 and BIBREF27 use encoders and decoders parameterized as RNN variants such as Long Short-term Memory (LSTMs) BIBREF28 or Gated Recurrent Units (GRUs) BIBREF29 . The hidden representation INLINEFORM0 is typically the last hidden state of the encoder RNN.
BIBREF30 alleviate the gradient bottleneck between the encoder and the decoder by introducing an attention mechanism that allows the decoder to condition on every hidden state of the encoder RNN instead of only the last one. In this work, as in BIBREF6 , BIBREF8 , we do not employ an attention mechanism. This enables us to obtain a single, fixed-length, distributed sentence representation. To diminish the effects of vanishing gradient, we condition every decoding step on the encoder hidden representation INLINEFORM0 . We use a GRU for the encoder and decoder in the interest of computational speed. The encoder is a bidirectional GRU while the decoder is a unidirectional conditional GRU whose parameterization is as follows: DISPLAYFORM0
The encoder representation INLINEFORM0 is provided as conditioning information to the reset gate, update gate and hidden state computation in the GRU via the parameters INLINEFORM1 , INLINEFORM2 and INLINEFORM3 to avoid attenuation of information from the encoder.
## Multi-task Sequence-to-sequence Learning
BIBREF31 present a simple one-to-many multi-task sequence-to-sequence learning model for NMT that uses a shared encoder for English and task-specific decoders for multiple target languages. BIBREF22 extend this by also considering many-to-one (many encoders, one decoder) and many-to-many architectures. In this work, we consider a one-to-many model since it lends itself naturally to the idea of combining inductive biases from different training objectives. The same bidirectional GRU encodes the input sentences from different tasks into a compressed summary INLINEFORM0 which is then used to condition a task-specific GRU to produce the output sentence.
## Training Objectives & Evaluation
Our motivation for multi-task training stems from theoretical insights presented in BIBREF18 . We refer readers to that work for a detailed discussion of results, but the conclusions most relevant to this discussion are (i) that learning multiple related tasks jointly results in good generalization as measured by the number of training examples required per task; and (ii) that inductive biases learned on sufficiently many training tasks are likely to be good for learning novel tasks drawn from the same environment.
We select the following training objectives to learn general-purpose sentence embeddings. Our desiderata for the task collection were: sufficient diversity, existence of fairly large datasets for training, and success as standalone training objectives for sentence representations.
## Multi-task training setup
Multi-task training with different data sources for each task stills poses open questions. For example: When does one switch to training on a different task? Should the switching be periodic? Do we weight each task equally? If not, what training ratios do we use?
BIBREF31 use periodic task alternations with equal training ratios for every task. In contrast, BIBREF22 alter the training ratios for each task based on the size of their respective training sets. Specifically, the training ratio for a particular task, INLINEFORM0 , is the fraction of the number of training examples in that task to the total number of training samples across all tasks. The authors then perform INLINEFORM1 parameter updates on task INLINEFORM2 before selecting a new task at random proportional to the training ratios, where N is a predetermined constant.
We take a simpler approach and pick a new sequence-to-sequence task to train on after every parameter update sampled uniformly. An NLI minibatch is interspersed after every ten parameter updates on sequence-to-sequence tasks (this was chosen so as to complete roughly 6 epochs of the dataset after 7 days of training). Our approach is described formally in the Algorithm below.
Model details can be found in section SECREF7 in the Appendix.
0ptcenterline
A set of INLINEFORM0 tasks with a common source language, a shared encoder INLINEFORM1 across all tasks and a set of INLINEFORM2 task specific decoders INLINEFORM3 . Let INLINEFORM4 denote each model's parameters, INLINEFORM5 a probability vector ( INLINEFORM6 ) denoting the probability of sampling a task such that INLINEFORM7 , datasets for each task INLINEFORM8 and a loss function INLINEFORM9 .
INLINEFORM0 has not converged [1] Sample task INLINEFORM1 . Sample input, output pairs INLINEFORM2 . Input representation INLINEFORM3 . Prediction INLINEFORM4 INLINEFORM5 Adam INLINEFORM6 .
## Evaluation Strategies, Experimental Results & Discussion
In this section, we describe our approach to evaluate the quality of our learned representations, present the results of our evaluation and discuss our findings.
## Evaluation Strategy
We follow a similar evaluation protocol to those presented in BIBREF6 , BIBREF8 , BIBREF9 which is to use our learned representations as features for a low complexity classifier (typically linear) on a novel supervised task/domain unseen during training without updating the parameters of our sentence representation model. We also consider such a transfer learning evaluation in an artificially constructed low-resource setting. In addition, we also evaluate the quality of our learned individual word representations using standard benchmarks BIBREF36 , BIBREF37 .
The choice of transfer tasks and evaluation framework are borrowed largely from BIBREF9 . We provide a condensed summary of the tasks in section SECREF10 in the Appendix but refer readers to their paper for a more detailed description.
https://github.com/kudkudak/word-embeddings-benchmarks/wiki
## Experimental Results & Discussion
Table 2 presents the results of training logistic regression on 10 different supervised transfer tasks using different fixed-length sentence representation. Supervised approaches trained from scratch on some of these tasks are also presented for comparison. We present performance ablations when adding more tasks and increasing the number of hidden units in our GRU (+L). Ablation specifics are presented in section SECREF9 of the Appendix.
It is evident from Table 2 that adding more tasks improves the transfer performance of our model. Increasing the capacity our sentence encoder with more hidden units (+L) as well as an additional layer (+2L) also lead to improved transfer performance. We observe gains of 1.1-2.0% on the sentiment classification tasks (MR, CR, SUBJ & MPQA) over Infersent. We demonstrate substantial gains on TREC (6% over Infersent and roughly 2% over the CNN-LSTM), outperforming even a competitive supervised baseline. We see similar gains (2.3%) on paraphrase identification (MPRC), closing the gap on supervised approaches trained from scratch. The addition of constituency parsing improves performance on sentence relatedness (SICK-R) and entailment (SICK-E) consistent with observations made by BIBREF48 .
In Table TABREF19 , we show that simply training an MLP on top of our fixed sentence representations outperforms several strong & complex supervised approaches that use attention mechanisms, even on this fairly large dataset. For example, we observe a 0.2-0.5% improvement over the decomposable attention model BIBREF49 . When using only a small fraction of the training data, indicated by the columns 1k-25k, we are able to outperform the Siamese and Multi-Perspective CNN using roughly 6% of the available training set. We also outperform the Deconv LVM model proposed by BIBREF47 in this low-resource setting.
Unlike BIBREF9 , who use pretrained GloVe word embeddings, we learn our word embeddings from scratch. Somewhat surprisingly, in Table TABREF18 we observe that the learned word embeddings are competitive with popular methods such as GloVe, word2vec, and fasttext BIBREF50 on the benchmarks presented by BIBREF36 and BIBREF37 .
In Table TABREF20 , we probe our sentence representations to determine if certain sentence characteristics and syntactic properties can be inferred following work by BIBREF17 and BIBREF14 . We observe that syntactic properties are better encoded with the addition of multi-lingual NMT and parsing. Representations learned solely from NLI do appear to encode syntax but incorporation into our multi-task framework does not amplify this signal. Similarly, we observe that sentence characteristics such as length and word order are better encoded with the addition of parsing.
In Appendix Table TABREF30 , we note that our sentence representations outperform skip-thoughts and are on par with Infersent for image-caption retrieval. We also observe that comparing sentences using cosine similarities correlates reasonably well with their relatedness on semantic textual similarity benchmarks (Appendix Table TABREF31 ).
We also present qualitative analysis of our learned representations by visualizations using dimensionality reduction techniques (Figure FIGREF11 ) and nearest neighbor exploration (Appendix Table TABREF32 ). Figure FIGREF11 shows t-sne plots of our sentence representations on three different datasets - SUBJ, TREC and DBpedia. DBpedia is a large corpus of sentences from Wikipedia labeled by category and used by BIBREF51 . Sentences appear to cluster reasonably well according to their labels. The clustering also appears better than that demonstrated in Figure 2 of BIBREF6 on TREC and SUBJ. Appendix Table TABREF32 contains sentences from the BookCorpus and their nearest neighbors. Sentences with some lexical overlap and similar discourse structure appear to be clustered together.
## Conclusion & Future Work
We present a multi-task framework for learning general-purpose fixed-length sentence representations. Our primary motivation is to encapsulate the inductive biases of several diverse training signals used to learn sentence representations into a single model. Our multi-task framework includes a combination of sequence-to-sequence tasks such as multi-lingual NMT, constituency parsing and skip-thought vectors as well as a classification task - natural language inference. We demonstrate that the learned representations yield competitive or superior results to previous general-purpose sentence representation methods. We also observe that this approach produces good word embeddings.
In future work, we would like understand and interpret the inductive biases that our model learns and observe how it changes with the addition of different tasks beyond just our simple analysis of sentence characteristics and syntax. Having a rich, continuous sentence representation space could allow the application of state-of-the-art generative models of images such as that of BIBREF52 to language. One could also consider controllable text generation by directly manipulating the sentence representations and realizing it by decoding with a conditional language model.
## Acknowledgements
The authors would like to thank Chinnadhurai Sankar, Sebastian Ruder, Eric Yuan, Tong Wang, Alessandro Sordoni, Guillaume Lample and Varsha Embar for useful discussions. We are also grateful to the PyTorch development team BIBREF53 . We thank NVIDIA for donating a DGX-1 computer used in this work and Fonds de recherche du Québec - Nature et technologies for funding.
## Model Training
We present some architectural specifics and training details of our multi-task framework. Our shared encoder uses a common word embedding lookup table and GRU. We experiment with unidirectional, bidirectional and 2 layer bidirectional GRUs (details in Appendix section SECREF9 ). For each task, every decoder has its separate word embedding lookups, conditional GRUs and fully connected layers that project the GRU hidden states to the target vocabularies. The last hidden state of the encoder is used as the initial hidden state of the decoder and is also presented as input to all the gates of the GRU at every time step. For natural language inference, the same encoder is used to encode both the premise and hypothesis and a concatenation of their representations along with the absolute difference and hadamard product (as described in BIBREF9 ) are given to a single layer MLP with a dropout BIBREF55 rate of 0.3. All models use word embeddings of 512 dimensions and GRUs with either 1500 or 2048 hidden units. We used minibatches of 48 examples and the Adam BIBREF54 optimizer with a learning rate of 0.002. Models were trained for 7 days on an Nvidia Tesla P100-SXM2-16GB GPU. While BIBREF6 report close to a month of training, we only train for 7 days, made possible by advancements in GPU hardware and software (cuDNN RNNs).
We did not tune any of the architectural details and hyperparameters owing to the fact that we were unable to identify any clear criterion on which to tune them. Gains in performance on a specific task do not often translate to better transfer performance.
## Vocabulary Expansion & Representation Pooling
In addition to performing 10-fold cross-validation to determine the L2 regularization penalty on the logistic regression models, we also tune the way in which our sentence representations are generated from the hidden states corresponding to words in a sentence. For example, BIBREF6 use the last hidden state while BIBREF9 perform max-pooling across all of the hidden states. We consider both of these approaches and pick the one with better performance on the validation set. We note that max-pooling works best on sentiment tasks such as MR, CR, SUBJ and MPQA, while the last hidden state works better on all other tasks.
We also employ vocabulary expansion on all tasks as in BIBREF6 by training a linear regression to map from the space of pre-trained word embeddings (GloVe) to our model's word embeddings.
## Multi-task model details
This section describes the specifics our multi-task ablations in the experiments section. These definitions hold for all tables except for TABREF18 and TABREF20 . We refer to skip-thought next as STN, French and German NMT as Fr and De, natural language inference as NLI, skip-thought previous as STP and parsing as Par.
+STN +Fr +De : The sentence representation INLINEFORM0 is the concatenation of the final hidden vectors from a forward GRU with 1500-dimensional hidden vectors and a bidirectional GRU, also with 1500-dimensional hidden vectors.
+STN +Fr +De +NLI : The sentence representation INLINEFORM0 is the concatenation of the final hidden vectors from a bidirectional GRU with 1500-dimensional hidden vectors and another bidirectional GRU with 1500-dimensional hidden vectors trained without NLI.
+STN +Fr +De +NLI +L : The sentence representation INLINEFORM0 is the concatenation of the final hidden vectors from a bidirectional GRU with 2048-dimensional hidden vectors and another bidirectional GRU with 2048-dimensional hidden vectors trained without NLI.
+STN +Fr +De +NLI +L +STP : The sentence representation INLINEFORM0 is the concatenation of the final hidden vectors from a bidirectional GRU with 2048-dimensional hidden vectors and another bidirectional GRU with 2048-dimensional hidden vectors trained without STP.
+STN +Fr +De +NLI +2L +STP : The sentence representation INLINEFORM0 is the concatenation of the final hidden vectors from a 2-layer bidirectional GRU with 2048-dimensional hidden vectors and a 1-layer bidirectional GRU with 2048-dimensional hidden vectors trained without STP.
+STN +Fr +De +NLI +L +STP +Par : The sentence representation INLINEFORM0 is the concatenation of the final hidden vectors from a bidirectional GRU with 2048-dimensional hidden vectors and another bidirectional GRU with 2048-dimensional hidden vectors trained without Par.
In tables TABREF18 and TABREF20 we do not concatenate the representations of multiple models.
## Description of evaluation tasks
BIBREF6 and BIBREF9 provide a detailed description of tasks that are typically used to evaluate sentence representations. We provide a condensed summary and refer readers to their work for a more thorough description.
## Text Classification
We evaluate on text classification benchmarks - sentiment classification on movie reviews (MR), product reviews (CR) and Stanford sentiment (SST), question type classification (TREC), subjectivity/objectivity classification (SUBJ) and opinion polarity (MPQA). Representations are used to train a logistic regression classifier with 10-fold cross validation to tune the L2 weight penalty. The evaluation metric for all these tasks is classification accuracy.
## Paraphrase Identification
We also evaluate on pairwise text classification tasks such as paraphrase identification on the Microsoft Research Paraphrase Corpus (MRPC) corpus. This is a binary classification problem to identify if two sentences are paraphrases of each other. The evaluation metric is classification accuracy and F1.
## Entailment and Semantic Relatedness
To test if similar sentences share similar representations, we evaluate on the SICK relatedness (SICK-R) task where a linear model is trained to output a score from 1 to 5 indicating the relatedness of two sentences. We also evaluate using the entailment labels in the same dataset (SICK-E) which is a binary classification problem. The evaluation metric for SICK-R is Pearson correlation and classification accuracy for SICK-E.
## Semantic Textual Similarity
In this evaluation, we measure the relatedness of two sentences using only the cosine similarity between their representations. We use the similarity textual similarity (STS) benchmark tasks from 2012-2016 (STS12, STS13, STS14, STS15, STS16, STSB). The STS dataset contains sentences from a diverse set of data sources. The evaluation criteria is Pearson correlation.
## Image-caption retrieval
Image-caption retrieval is typically formulated as a ranking task wherein images are retrieved and ranked based on a textual description and vice-versa. We use 113k training images from MSCOCO with 5k images for validation and 5k for testing. Image features are extracted using a pre-trained 110 layer ResNet. The evaluation criterion is Recall@K and the median K across 5 different splits of the data.
## Quora Duplicate Question Classification
In addition to the above tasks which were considered by BIBREF9 , we also evaluate on the recently published Quora duplicate question dataset since it is an order of magnitude larger than the others (approximately 400,000 question pairs). The task is to correctly identify question pairs that are duplicates of one another, which we formulate as a binary classification problem. We use the same data splits as in BIBREF45 . Given the size of this data, we consider a more expressive classifier on top of the representations of both questions. Specifically, we train a 4 layer MLP with 1024 hidden units, with a dropout rate of 0.5 after every hidden layer. The evaluation criterion is classification accuracy. We also artificially create a low-resource setting by reducing the number of training examples between 1,000 and 25,000 using the same splits as BIBREF47 .
## Sentence Characteristics & Syntax
In an attempt to understand what information is encoded in by sentence representations, we consider six different classification tasks where the objective is to predict sentence characteristics such as length, word content and word order BIBREF17 or syntactic properties such as active/passive, tense and the top syntactic sequence (TSS) from the parse tree of a sentence BIBREF14 .
The sentence characteristic tasks are setup in the same way as described in BIBREF17 . The length task is an 8-way classification problem where sentence lengths are binned into 8 ranges. The content task is formulated as a binary classification problem that takes a concatenation of a sentence representation INLINEFORM0 and a word representation INLINEFORM1 to determine if the word is contained in the sentence. The order task is an extension of the content task where a concatenation of the sentence representation and word representations of two words in sentence is used to determine if the first word occurs before or after the second. We use a random subset of the 1-billion-word dataset for these experiments that were not used to train our multi-task representations.
The syntactic properties tasks are setup in the same way as described in BIBREF14 .The passive and tense tasks are characterized as binary classification problems given a sentence's representation. The former's objective is to determine if a sentence is written in active/passive voice while the latter's objective is to determine if the sentence is in the past tense or not. The top syntactic sequence (TSS) is a 20-way classification problem with 19 most frequent top syntactic sequences and 1 miscellaneous class. We use the same dataset as the authors but different training, validation and test splits.
| 22
|
1804.04225
|
Exploiting Task-Oriented Resources to Learn Word Embeddings for Clinical Abbreviation Expansion
|
# Exploiting Task-Oriented Resources to Learn Word Embeddings for Clinical Abbreviation Expansion
## Abstract
In the medical domain, identifying and expanding abbreviations in clinical texts is a vital task for both better human and machine understanding. It is a challenging task because many abbreviations are ambiguous especially for intensive care medicine texts, in which phrase abbreviations are frequently used. Besides the fact that there is no universal dictionary of clinical abbreviations and no universal rules for abbreviation writing, such texts are difficult to acquire, expensive to annotate and even sometimes, confusing to domain experts. This paper proposes a novel and effective approach - exploiting task-oriented resources to learn word embeddings for expanding abbreviations in clinical notes. We achieved 82.27% accuracy, close to expert human performance.
## Introduction
Abbreviations and acronyms appear frequently in the medical domain. Based on a popular online knowledge base, among the 3,096,346 stored abbreviations, 197,787 records are medical abbreviations, ranked first among all ten domains. An abbreviation can have over 100 possible explanations even within the medical domain. Medical record documentation, the authors of which are mainly physicians, other health professionals, and domain experts, is usually written under the pressure of time and high workload, requiring notation to be frequently compressed with shorthand jargon and acronyms. This is even more evident within intensive care medicine, where it is crucial that information is expressed in the most efficient manner possible to provide time-sensitive care to critically ill patients, but can result in code-like messages with poor readability. For example, given a sentence written by a physician with specialty training in critical care medicine, “STAT TTE c/w RVS. AKI - no CTA. .. etc”, it is difficult for non-experts to understand all abbreviations without specific context and/or knowledge. But when a doctor reads this, he/she would know that although “STAT” is widely used as the abbreviation of “statistic”, “statistics” and “statistical” in most domains, in hospital emergency rooms, it is often used to represent “immediately”. Within the arena of medical research, abbreviation expansion using a natural language processing system to automatically analyze clinical notes may enable knowledge discovery (e.g., relations between diseases) and has potential to improve communication and quality of care.
In this paper, we study the task of abbreviation expansion in clinical notes. As shown in Figure 1, our goal is to normalize all the abbreviations in the intensive care unit (ICU) documentation to reduce misinterpretation and to make the texts accessible to a wider range of readers. For accurately capturing the semantics of an abbreviation in its context, we adopt word embedding, which can be seen as a distributional semantic representation and has been proven to be effective BIBREF0 to compute the semantic similarity between words based on the context without any labeled data. The intuition of distributional semantics BIBREF1 is that if two words share similar contexts, they should have highly similar semantics. For example, in Figure 1, “RF” and “respiratory failure” have very similar contexts so that their semantics should be similar. If we know “respiratory failure” is a possible candidate expansion of “RF” and its semantics is similar to the “RF” in the intensive care medicine texts, we can determine that it should be the correct expansion of “RF”. Due to the limited resource of intensive care medicine texts where full expansions rarely appear, we exploit abundant and easily-accessible task-oriented resources to enrich our dataset for training embeddings. To the best of our knowledge, we are the first to apply word embeddings to this task. Experimental results show that the embeddings trained on the task-oriented corpus are much more useful than those trained on other corpora. By combining the embeddings with domain-specific knowledge, we achieve 82.27% accuracy, which outperforms baselines and is close to human's performance.
## Related Work
The task of abbreviation disambiguation in biomedical documents has been studied by various researchers using supervised machine learning algorithms BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 . However, the performance of these supervised methods mainly depends on a large amount of labeled data which is extremely difficult to obtain for our task since intensive care medicine texts are very rare resources in clinical domain due to the high cost of de-identification and annotation. Tengstrand et al. tengstrand2014eacl proposed a distributional semantics-based approach for abbreviation expansion in Swedish but they focused only on expanding single words and cannot handle multi-word phrases. In contrast, we use word embeddings combined with task-oriented resources and knowledge, which can handle multiword expressions.
## Overview
The overview of our approach is shown in Figure FIGREF6 . Within ICU notes (e.g., text example in top-left box in Figure 2), we first identify all abbreviations using regular expressions and then try to find all possible expansions of these abbreviations from domain-specific knowledge base as candidates. We train word embeddings using the clinical notes data with task-oriented resources such as Wikipedia articles of candidates and medical scientific papers and compute the semantic similarity between an abbreviation and its candidate expansions based on their embeddings (vector representations of words).
## Training embeddings with task oriented resources
Given an abbreviation as input, we expect the correct expansion to be the most semantically similar to the abbreviation, which requires the abbreviation and the expansion share similar contexts. For this reason, we exploit rich task-oriented resources such as the Wikipedia articles of all the possible candidates, research papers and books written by the intensive care medicine fellows. Together with our clinical notes data which functions as a corpus, we train word embeddings since the expansions of abbreviations in the clinical notes are likely to appear in these resources and also share the similar contexts to the abbreviation's contexts.
## Handling MultiWord Phrases
In most cases, an abbreviation's expansion is a multi-word phrase. Therefore, we need to obtain the phrase's embedding so that we can compute its semantic similarity to the abbreviation.
It is proven that a phrase's embedding can be effectively obtained by summing the embeddings of words contained in the phrase BIBREF0 , BIBREF7 . For computing a phrase's embedding, we formally define a candidate INLINEFORM0 as a list of the words contained in the candidate, for example: one of MICU's candidate expansions is medical intensive care unit=[medical,intensive,care,unit]. Then, INLINEFORM1 's embedding can be computed as follows: DISPLAYFORM0
where INLINEFORM0 is a token in the candidate INLINEFORM1 and INLINEFORM2 denotes the embedding of a word/phrase, which is a vector of real-value entries.
## Expansion Candidate Ranking
Even though embeddings are very helpful to compute the semantic similarity between an abbreviation and a candidate expansion, in some cases, context-independent information is also useful to identify the correct expansion. For example, CHF in the clinical notes usually refers to “congestive heart failure”. By using embedding-based semantic similarity, we can find two possible candidates – “congestive heart failure” (similarity=0.595) and “chronic heart failure”(similarity=0.621). These two candidates have close semantic similarity score but their popularity scores in the medical domain are quite different – the former has a rating score of 50 while the latter only has a rating score of 7. Therefore, we can see that the rating score, which can be seen as a kind of domain-specific knowledge, can also contribute to the candidate ranking.
We combine semantic similarity with rating information. Formally, given an abbreviation INLINEFORM0 's candidate list INLINEFORM1 , we rank INLINEFORM2 based on the following formula: DISPLAYFORM0
where INLINEFORM0 denotes the rating of this candidate as an expansion of the abbreviation INLINEFORM1 , which reflects this candidate's popularity, INLINEFORM2 denotes the embedding of a word. The parameter INLINEFORM3 serves to adjust the weights of similarity and popularity
## Data and Evaluation Metrics
The clinical notes we used for the experiment are provided by domain experts, consisting of 1,160 physician logs of Medical ICU admission requests at a tertiary care center affiliated to Mount Sanai. Prospectively collected over one year, these semi-structured logs contain free-text descriptions of patients' clinical presentations, medical history, and required critical care-level interventions. We identify 818 abbreviations and find 42,506 candidates using domain-specific knowledge (i.e., www.allacronym.com/_medical). The enriched corpus contains 42,506 Wikipedia articles, each of which corresponds to one candidate, 6 research papers and 2 critical care medicine textbooks, besides our raw ICU data.
We use word2vec BIBREF0 to train the word embeddings. The dimension of embeddings is empirically set to 100.
Since the goal of our task is to find the correct expansion for an abbreviation, we use accuracy as a metric to evaluate the performance of our approach. For ground-truth, we have 100 physician logs which are manually expanded and normalized by one of the authors Dr. Mathews, a well-trained domain expert, and thus we use these 100 physician logs as the test set to evaluate our approach's performance.
## Baseline Models
For our task, it's difficult to re-implement the supervised methods as in previous works mentioned since we do not have sufficient training data. And a direct comparison is also impossible because all previous work used different data sets which are not publicly available. Alternatively, we use the following baselines to compare with our approach.
Rating: This baseline model chooses the highest rating candidate expansion in the domain specific knowledge base.
Raw Input embeddings: We trained word embeddings only from the 1,160 raw ICU texts and we choose the most semantically related candidate as the answer.
General embeddings: Different from the Raw Input embeddings baseline, we use the embedding trained from a large biomedical data collection that includes knowledge bases like PubMed and PMC and a Wikipedia dump of biomedical related articles BIBREF8 for semantic similarity computation.
## Results
Table 1 shows the performance of abbreviation expansion. Our approach significantly outperforms the baseline methods and achieves 82.27% accuracy.
Figure FIGREF21 shows how our approach improves the performance of a rating-based approach. By using embeddings, we can learn that the meaning of “OD” used in our test cases should be “overdose” rather than “out-of-date” and this semantic information largely benefits the abbreviation expansion model.
Compared with our approach, embeddings trained only from the ICU texts do not significantly contribute to the performance over the rating baseline. The reason is that the size of data for training the embeddings is so small that many candidate expansions of abbreviations do not appear in the corpus, which results in poor performance. It is notable that general embeddings trained from large biomedical data are not effective for this task because many abbreviations within critical care medicine appear in the biomedical corpus with different senses.
For example, “OD” in intensive care medicine texts refers to “overdose” while in the PubMed corpus it usually refers to “optical density”, as shown in Figure FIGREF24 . Therefore, the embeddings trained from the PubMed corpus do not benefit the expansion of abbreviations in the ICU texts.
Moreover, we estimated human performance for this task, shown in Table TABREF26 . Note that the performance is estimated by one of the authors Dr. Mathews who is a board-certified pulmonologist and critical care medicine specialist based on her experience and the human's performance estimated in Table TABREF26 is under the condition that the participants can not use any other external resources. We can see that our approach can achieve a performance close to domain experts and thus it is promising to tackle this challenge.
## Error Analysis
The distribution of errors is shown in Table TABREF28 . There are mainly three reasons that cause the incorrect expansion. In some cases, some certain abbreviations do not exist in the knowledge base. In this case we would not be able to populate the corresponding candidate list. Secondly, in many cases although we have the correct expansion in the candidate list, it's not ranked as the top one due to the lower semantic similarity because there are not enough samples in the training data. Among all the incorrect expansions in our test set, such kind of errors accounted for about 54%. One possible solution may be adding more effective data to the embedding training, which means discovering more task-oriented resources. In a few cases, we failed to identify some abbreviations because of their complicated representations. For example, we have the following sentence in the patient's notes: “ No n/v/f/c.” and the correct expansion should be “No nausea/vomiting/fever/chills.” Such abbreviations are by far the most difficult to expand in our task because they do not exist in any knowledge base and usually only occur once in the training data.
## Conclusions and Future Work
This paper proposes a simple but novel approach for automatic expansion of abbreviations. It achieves very good performance without any manually labeled data. Experiments demonstrate that using task-oriented resources to train word embeddings is much more effective than using general or arbitrary corpus.
In the future, we plan to collectively expand semantically related abbreviations co-occurring in a sentence. In addition, we expect to integrate our work into a natural language processing system for processing the clinical notes for discovering knowledge, which will largely benefit the medical research.
## Acknowledgements
This work is supported by RPI's Tetherless World Constellation, IARPA FUSE Numbers D11PC20154 and J71493 and DARPA DEFT No. FA8750-13-2-0041. Dr. Mathews' effort is supported by Award #1K12HL109005-01 from the National Heart, Lung, and Blood Institute (NHLBI). The content is solely the responsibility of the authors and does not necessarily represent the official views of NHLBI, the National Institutes of Health, IARPA, or DARPA.
| 12
|
1804.10686
|
An Unsupervised Word Sense Disambiguation System for Under-Resourced Languages
|
# An Unsupervised Word Sense Disambiguation System for Under-Resourced Languages
## Abstract
In this paper, we present Watasense, an unsupervised system for word sense disambiguation. Given a sentence, the system chooses the most relevant sense of each input word with respect to the semantic similarity between the given sentence and the synset constituting the sense of the target word. Watasense has two modes of operation. The sparse mode uses the traditional vector space model to estimate the most similar word sense corresponding to its context. The dense mode, instead, uses synset embeddings to cope with the sparsity problem. We describe the architecture of the present system and also conduct its evaluation on three different lexical semantic resources for Russian. We found that the dense mode substantially outperforms the sparse one on all datasets according to the adjusted Rand index.
## Introduction
Word sense disambiguation (WSD) is a natural language processing task of identifying the particular word senses of polysemous words used in a sentence. Recently, a lot of attention was paid to the problem of WSD for the Russian language BIBREF0 , BIBREF1 , BIBREF2 . This problem is especially difficult because of both linguistic issues – namely, the rich morphology of Russian and other Slavic languages in general – and technical challenges like the lack of software and language resources required for addressing the problem.
To address these issues, we present Watasense, an unsupervised system for word sense disambiguation. We describe its architecture and conduct an evaluation on three datasets for Russian. The choice of an unsupervised system is motivated by the absence of resources that would enable a supervised system for under-resourced languages. Watasense is not strictly tied to the Russian language and can be applied to any language for which a tokenizer, part-of-speech tagger, lemmatizer, and a sense inventory are available.
The rest of the paper is organized as follows. Section 2 reviews related work. Section 3 presents the Watasense word sense disambiguation system, presents its architecture, and describes the unsupervised word sense disambiguation methods bundled with it. Section 4 evaluates the system on a gold standard for Russian. Section 5 concludes with final remarks.
## Related Work
Although the problem of WSD has been addressed in many SemEval campaigns BIBREF3 , BIBREF4 , BIBREF5 , we focus here on word sense disambiguation systems rather than on the research methodologies.
Among the freely available systems, IMS (“It Makes Sense”) is a supervised WSD system designed initially for the English language BIBREF6 . The system uses a support vector machine classifier to infer the particular sense of a word in the sentence given its contextual sentence-level features. Pywsd is an implementation of several popular WSD algorithms implemented in a library for the Python programming language. It offers both the classical Lesk algorithm for WSD and path-based algorithms that heavily use the WordNet and similar lexical ontologies. DKPro WSD BIBREF7 is a general-purpose framework for WSD that uses a lexical ontology as the sense inventory and offers the variety of WordNet-based algorithms. Babelfy BIBREF8 is a WSD system that uses BabelNet, a large-scale multilingual lexical ontology available for most natural languages. Due to the broad coverage of BabelNet, Babelfy offers entity linking as part of the WSD functionality.
Panchenko:17:emnlp present an unsupervised WSD system that is also knowledge-free: its sense inventory is induced based on the JoBimText framework, and disambiguation is performed by computing the semantic similarity between the context and the candidate senses BIBREF9 . Pelevina:16 proposed a similar approach to WSD, but based on dense vector representations (word embeddings), called SenseGram. Similarly to SenseGram, our WSD system is based on averaging of word embeddings on the basis of an automatically induced sense inventory. A crucial difference, however, is that we induce our sense inventory from synonymy dictionaries and not distributional word vectors. While this requires more manually created resources, a potential advantage of our approach is that the resulting inventory contains less noise.
## Watasense, an Unsupervised System for Word Sense Disambiguation
Watasense is implemented in the Python programming language using the scikit-learn BIBREF10 and Gensim BIBREF11 libraries. Watasense offers a Web interface (Figure FIGREF2 ), a command-line tool, and an application programming interface (API) for deployment within other applications.
## System Architecture
A sentence is represented as a list of spans. A span is a quadruple: INLINEFORM0 , where INLINEFORM1 is the word or the token, INLINEFORM2 is the part of speech tag, INLINEFORM3 is the lemma, INLINEFORM4 is the position of the word in the sentence. These data are provided by tokenizer, part-of-speech tagger, and lemmatizer that are specific for the given language. The WSD results are represented as a map of spans to the corresponding word sense identifiers.
The sense inventory is a list of synsets. A synset is represented by three bag of words: the synonyms, the hypernyms, and the union of two former – the bag. Due to the performance reasons, on initialization, an inverted index is constructed to map a word to the set of synsets it is included into.
Each word sense disambiguation method extends the BaseWSD class. This class provides the end user with a generic interface for WSD and also encapsulates common routines for data pre-processing. The inherited classes like SparseWSD and DenseWSD should implement the disambiguate_word(...) method that disambiguates the given word in the given sentence. Both classes use the bag representation of synsets on the initialization. As the result, for WSD, not just the synonyms are used, but also the hypernyms corresponding to the synsets. The UML class diagram is presented in Figure FIGREF4 .
Watasense supports two sources of word vectors: it can either read the word vector dataset in the binary Word2Vec format or use Word2Vec-Pyro4, a general-purpose word vector server. The use of a remote word vector server is recommended due to the reduction of memory footprint per each Watasense process.
## User Interface
FIGREF2 shows the Web interface of Watasense. It is composed of two primary activities. The first is the text input and the method selection ( FIGREF2 ). The second is the display of the disambiguation results with part of speech highlighting ( FIGREF7 ). Those words with resolved polysemy are underlined; the tooltips with the details are raised on hover.
## Word Sense Disambiguation
We use two different unsupervised approaches for word sense disambiguation. The first, called `sparse model', uses a straightforward sparse vector space model, as widely used in Information Retrieval, to represent contexts and synsets. The second, called `dense model', represents synsets and contexts in a dense, low-dimensional space by averaging word embeddings.
In the vector space model approach, we follow the sparse context-based disambiguated method BIBREF12 , BIBREF13 . For estimating the sense of the word INLINEFORM0 in a sentence, we search for such a synset INLINEFORM1 that maximizes the cosine similarity to the sentence vector: DISPLAYFORM0
where INLINEFORM0 is the set of words forming the synset, INLINEFORM1 is the set of words forming the sentence. On initialization, the synsets represented in the sense inventory are transformed into the INLINEFORM2 -weighted word-synset sparse matrix efficiently represented in the memory using the compressed sparse row format. Given a sentence, a similar transformation is done to obtain the sparse vector representation of the sentence in the same space as the word-synset matrix. Then, for each word to disambiguate, we retrieve the synset containing this word that maximizes the cosine similarity between the sparse sentence vector and the sparse synset vector. Let INLINEFORM3 be the maximal number of synsets containing a word and INLINEFORM4 be the maximal size of a synset. Therefore, disambiguation of the whole sentence INLINEFORM5 requires INLINEFORM6 operations using the efficient sparse matrix representation.
In the synset embeddings model approach, we follow SenseGram BIBREF14 and apply it to the synsets induced from a graph of synonyms. We transform every synset into its dense vector representation by averaging the word embeddings corresponding to each constituent word: DISPLAYFORM0
where INLINEFORM0 denotes the word embedding of INLINEFORM1 . We do the same transformation for the sentence vectors. Then, given a word INLINEFORM2 , a sentence INLINEFORM3 , we find the synset INLINEFORM4 that maximizes the cosine similarity to the sentence: DISPLAYFORM0
On initialization, we pre-compute the dense synset vectors by averaging the corresponding word embeddings. Given a sentence, we similarly compute the dense sentence vector by averaging the vectors of the words belonging to non-auxiliary parts of speech, i.e., nouns, adjectives, adverbs, verbs, etc. Then, given a word to disambiguate, we retrieve the synset that maximizes the cosine similarity between the dense sentence vector and the dense synset vector. Thus, given the number of dimensions INLINEFORM0 , disambiguation of the whole sentence INLINEFORM1 requires INLINEFORM2 operations.
## Evaluation
We conduct our experiments using the evaluation methodology of SemEval 2010 Task 14: Word Sense Induction & Disambiguation BIBREF5 . In the gold standard, each word is provided with a set of instances, i.e., the sentences containing the word. Each instance is manually annotated with the single sense identifier according to a pre-defined sense inventory. Each participating system estimates the sense labels for these ambiguous words, which can be viewed as a clustering of instances, according to sense labels. The system's clustering is compared to the gold-standard clustering for evaluation.
## Quality Measure
The original SemEval 2010 Task 14 used the V-Measure external clustering measure BIBREF5 . However, this measure is maximized by clustering each sentence into his own distinct cluster, i.e., a `dummy' singleton baseline. This is achieved by the system deciding that every ambiguous word in every sentence corresponds to a different word sense. To cope with this issue, we follow a similar study BIBREF1 and use instead of the adjusted Rand index (ARI) proposed by Hubert:85 as an evaluation measure.
In order to provide the overall value of ARI, we follow the addition approach used in BIBREF1 . Since the quality measure is computed for each lemma individually, the total value is a weighted sum, namely DISPLAYFORM0
where INLINEFORM0 is the lemma, INLINEFORM1 is the set of the instances for the lemma INLINEFORM2 , INLINEFORM3 is the adjusted Rand index computed for the lemma INLINEFORM4 . Thus, the contribution of each lemma to the total score is proportional to the number of instances of this lemma.
## Dataset
We evaluate the word sense disambiguation methods in Watasense against three baselines: an unsupervised approach for learning multi-prototype word embeddings called AdaGram BIBREF15 , same sense for all the instances per lemma (One), and one sense per instance (Singletons). The AdaGram model is trained on the combination of RuWac, Lib.Ru, and the Russian Wikipedia with the overall vocabulary size of 2 billion tokens BIBREF1 .
As the gold-standard dataset, we use the WSD training dataset for Russian created during RUSSE'2018: A Shared Task on Word Sense Induction and Disambiguation for the Russian Language BIBREF16 . The dataset has 31 words covered by INLINEFORM0 instances in the bts-rnc subset and 5 words covered by 439 instances in the wiki-wiki subset.
The following different sense inventories have been used during the evaluation:
[leftmargin=4mm]
Watlink, a word sense network constructed automatically. It uses the synsets induced in an unsupervised way by the Watset[CWnolog, MCL] method BIBREF2 and the semantic relations from such dictionaries as Wiktionary referred as Joint INLINEFORM0 Exp INLINEFORM1 SWN in Ustalov:17:dialogue. This is the only automatically built inventory we use in the evaluation.
RuThes, a large-scale lexical ontology for Russian created by a group of expert lexicographers BIBREF17 .
RuWordNet, a semi-automatic conversion of the RuThes lexical ontology into a WordNet-like structure BIBREF18 .
Since the Dense model requires word embeddings, we used the 500-dimensional word vectors from the Russian Distributional Thesaurus BIBREF19 . These vectors are obtained using the Skip-gram approach trained on the lib.rus.ec text corpus.
## Results
We compare the evaluation results obtained for the Sparse and Dense approaches with three baselines: the AdaGram model (AdaGram), the same sense for all the instances per lemma (One) and one sense per instance (Singletons). The evaluation results are presented in Table TABREF25 . The columns bts-rnc and wiki-wiki represent the overall value of ARI according to Equation ( EQREF15 ). The column Avg. consists of the weighted average of the datasets w.r.t. the number of instances.
We observe that the SenseGram-based approach for word sense disambiguation yields substantially better results in every case (Table TABREF25 ). The primary reason for that is the implicit handling of similar words due to the averaging of dense word vectors for semantically related words. Thus, we recommend using the dense approach in further studies. Although the AdaGram approach trained on a large text corpus showed better results according to the weighted average, this result does not transfer to languages with less available corpus size.
## Conclusion
In this paper, we presented Watasense, an open source unsupervised word sense disambiguation system that is parameterized only by a word sense inventory. It supports both sparse and dense sense representations. We were able to show that the dense approach substantially boosts the performance of the sparse approach on three different sense inventories for Russian. We recommend using the dense approach in further studies due to its smoothing capabilities that reduce sparseness. In further studies, we will look at the problem of phrase neighbors that influence the sentence vector representations.
Finally, we would like to emphasize the fact that Watasense has a simple API for integrating different algorithms for WSD. At the same time, it requires only a basic set of language processing tools to be available: tokenizer, a part-of-speech tagger, lemmatizer, and a sense inventory, which means that low-resourced language can benefit of its usage.
## Acknowledgements
We acknowledge the support of the Deutsche Forschungsgemeinschaft (DFG) under the project “Joining Ontologies and Semantics Induced from Text” (JOIN-T), the RFBR under the projects no. 16-37-00203 mol_a and no. 16-37-00354 mol_a, and the RFH under the project no. 16-04-12019. The research was supported by the Ministry of Education and Science of the Russian Federation Agreement no. 02.A03.21.0006. The calculations were carried out using the supercomputer “Uran” at the Krasovskii Institute of Mathematics and Mechanics.
| 12
|
1805.00760
|
Aspect Term Extraction with History Attention and Selective Transformation
|
# Aspect Term Extraction with History Attention and Selective Transformation
## Abstract
Aspect Term Extraction (ATE), a key sub-task in Aspect-Based Sentiment Analysis, aims to extract explicit aspect expressions from online user reviews. We present a new framework for tackling ATE. It can exploit two useful clues, namely opinion summary and aspect detection history. Opinion summary is distilled from the whole input sentence, conditioned on each current token for aspect prediction, and thus the tailor-made summary can help aspect prediction on this token. Another clue is the information of aspect detection history, and it is distilled from the previous aspect predictions so as to leverage the coordinate structure and tagging schema constraints to upgrade the aspect prediction. Experimental results over four benchmark datasets clearly demonstrate that our framework can outperform all state-of-the-art methods.
## Introduction
Aspect-Based Sentiment Analysis (ABSA) involves detecting opinion targets and locating opinion indicators in sentences in product review texts BIBREF0 . The first sub-task, called Aspect Term Extraction (ATE), is to identify the phrases targeted by opinion indicators in review sentences. For example, in the sentence “I love the operating system and preloaded software”, the words “operating system” and “preloaded software” should be extracted as aspect terms, and the sentiment on them is conveyed by the opinion word “love”. According to the task definition, for a term/phrase being regarded as an aspect, it should co-occur with some “opinion words” that indicate a sentiment polarity on it BIBREF1 .
Many researchers formulated ATE as a sequence labeling problem or a token-level classification problem. Traditional sequence models such as Conditional Random Fields (CRFs) BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , Long Short-Term Memory Networks (LSTMs) BIBREF6 and classification models such as Support Vector Machine (SVM) BIBREF7 have been applied to tackle the ATE task, and achieved reasonable performance. One drawback of these existing works is that they do not exploit the fact that, according to the task definition, aspect terms should co-occur with opinion-indicating words. Thus, the above methods tend to output false positives on those frequently used aspect terms in non-opinionated sentences, e.g., the word “restaurant” in “the restaurant was packed at first, so we waited for 20 minutes”, which should not be extracted because the sentence does not convey any opinion on it.
There are a few works that consider opinion terms when tackling the ATE task. BIBREF8 proposed Recursive Neural Conditional Random Fields (RNCRF) to explicitly extract aspects and opinions in a single framework. Aspect-opinion relation is modeled via joint extraction and dependency-based representation learning. One assumption of RNCRF is that dependency parsing will capture the relation between aspect terms and opinion words in the same sentence so that the joint extraction can benefit. Such assumption is usually valid for simple sentences, but rather fragile for some complicated structures, such as clauses and parenthesis. Moreover, RNCRF suffers from errors of dependency parsing because its network construction hinges on the dependency tree of inputs. CMLA BIBREF9 models aspect-opinion relation without using syntactic information. Instead, it enables the two tasks to share information via attention mechanism. For example, it exploits the global opinion information by directly computing the association score between the aspect prototype and individual opinion hidden representations and then performing weighted aggregation. However, such aggregation may introduce noise. To some extent, this drawback is inherited from the attention mechanism, as also observed in machine translation BIBREF10 and image captioning BIBREF11 .
To make better use of opinion information to assist aspect term extraction, we distill the opinion information of the whole input sentence into opinion summary, and such distillation is conditioned on a particular current token for aspect prediction. Then, the opinion summary is employed as part of features for the current aspect prediction. Taking the sentence “the restaurant is cute but not upscale” as an example, when our model performs the prediction for the word “restaurant”, it first generates an opinion summary of the entire sentence conditioned on “restaurant”. Due to the strong correlation between “restaurant' and “upscale” (an opinion word), the opinion summary will convey more information of “upscale” so that it will help predict “restaurant” as an aspect with high probability. Note that the opinion summary is built on the initial opinion features coming from an auxiliary opinion detection task, and such initial features already distinguish opinion words to some extent. Moreover, we propose a novel transformation network that helps strengthen the favorable correlations, e.g. between “restaurant' and “upscale”, so that the produced opinion summary involves less noise.
Besides the opinion summary, another useful clue we explore is the aspect prediction history due to the inspiration of two observations: (1) In sequential labeling, the predictions at the previous time steps are useful clues for reducing the error space of the current prediction. For example, in the B-I-O tagging (refer to Section SECREF4 ), if the previous prediction is “O”, then the current prediction cannot be “I”; (2) It is observed that some sentences contain multiple aspect terms. For example, “Apple is unmatched in product quality, aesthetics, craftmanship, and customer service” has a coordinate structure of aspects. Under this structure, the previously predicted commonly-used aspect terms (e.g., “product quality”) can guide the model to find the infrequent aspect terms (e.g., “craftmanship”). To capture the above clues, our model distills the information of the previous aspect detection for making a better prediction on the current state.
Concretely, we propose a framework for more accurate aspect term extraction by exploiting the opinion summary and the aspect detection history. Firstly, we employ two standard Long-Short Term Memory Networks (LSTMs) for building the initial aspect and opinion representations recording the sequential information. To encode the historical information into the initial aspect representations at each time step, we propose truncated history attention to distill useful features from the most recent aspect predictions and generate the history-aware aspect representations. We also design a selective transformation network to obtain the opinion summary at each time step. Specifically, we apply the aspect information to transform the initial opinion representations and apply attention over the transformed representations to generate the opinion summary. Experimental results show that our framework can outperform state-of-the-art methods.
## The ATE Task
Given a sequence INLINEFORM0 of INLINEFORM1 words, the ATE task can be formulated as a token/word level sequence labeling problem to predict an aspect label sequence INLINEFORM2 , where each INLINEFORM3 comes from a finite label set INLINEFORM4 which describes the possible aspect labels. As shown in the example below:
INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 denote beginning of, inside and outside of the aspect span respectively. Note that in commonly-used datasets such as BIBREF12 , the gold standard opinions are usually not annotated.
## Model Description
As shown in Figure FIGREF3 , our model contains two key components, namely Truncated History-Attention (THA) and Selective Transformation Network (STN), for capturing aspect detection history and opinion summary respectively. THA and STN are built on two LSTMs that generate the initial word representations for the primary ATE task and the auxiliary opinion detection task respectively. THA is designed to integrate the information of aspect detection history into the current aspect feature to generate a new history-aware aspect representation. STN first calculates a new opinion representation conditioned on the current aspect candidate. Then, we employ a bi-linear attention network to calculate the opinion summary as the weighted sum of the new opinion representations, according to their associations with the current aspect representation. Finally, the history-aware aspect representation and the opinion summary are concatenated as features for aspect prediction of the current time step.
As Recurrent Neural Networks can record the sequential information BIBREF13 , we employ two vanilla LSTMs to build the initial token-level contextualized representations for sequence labeling of the ATE task and the auxiliary opinion word detection task respectively. For simplicity, let INLINEFORM0 denote an LSTM unit where INLINEFORM1 is the task indicator. In the following sections, without specification, the symbols with superscript INLINEFORM2 and INLINEFORM3 are the notations used in the ATE task and the opinion detection task respectively. We use Bi-Directional LSTM to generate the initial token-level representations INLINEFORM4 ( INLINEFORM5 is the dimension of hidden states): DISPLAYFORM0
In principle, RNN can memorize the entire history of the predictions BIBREF13 , but there is no mechanism to exploit the relation between previous predictions and the current prediction. As discussed above, such relation could be useful because of two reasons: (1) reducing the model's error space in predicting the current label by considering the definition of B-I-O schema, (2) improving the prediction accuracy for multiple aspects in one coordinate structure.
We propose a Truncated History-Attention (THA) component (the THA block in Figure FIGREF3 ) to explicitly model the aspect-aspect relation. Specifically, THA caches the most recent INLINEFORM0 hidden states. At the current prediction time step INLINEFORM1 , THA calculates the normalized importance score INLINEFORM2 of each cached state INLINEFORM3 ( INLINEFORM4 ) as follows: DISPLAYFORM0
DISPLAYFORM0
INLINEFORM0 denotes the previous history-aware aspect representation (refer to Eq. EQREF12 ). INLINEFORM1 can be learned during training. INLINEFORM2 are parameters associated with previous aspect representations, current aspect representation and previous history-aware aspect representations respectively. Then, the aspect history INLINEFORM3 is obtained as follows: DISPLAYFORM0
To benefit from the previous aspect detection, we consolidate the hidden aspect representation with the distilled aspect history to generate features for the current prediction. Specifically, we adopt a way similar to the residual block BIBREF14 , which is shown to be useful in refining word-level features in Machine Translation BIBREF15 and Part-Of-Speech tagging BIBREF16 , to calculate the history-aware aspect representations INLINEFORM0 at the time step INLINEFORM1 : DISPLAYFORM0
where ReLU is the relu activation function.
Previous works show that modeling aspect-opinion association is helpful to improve the accuracy of ATE, as exemplified in employing attention mechanism for calculating the opinion information BIBREF9 , BIBREF17 . MIN BIBREF17 focuses on a few surrounding opinion representations and computes their importance scores according to the proximity and the opinion salience derived from a given opinion lexicon. However, it is unable to capture the long-range association between aspects and opinions. Besides, the association is not strong because only the distance information is modeled. Although CMLA BIBREF9 can exploit global opinion information for aspect extraction, it may suffer from the noise brought in by attention-based feature aggregation. Taking the aspect term “fish” in “Furthermore, while the fish is unquestionably fresh, rolls tend to be inexplicably bland.” as an example, it might be enough to tell “fish” is an aspect given the appearance of the strongly related opinion “fresh”. However, CMLA employs conventional attention and does not have a mechanism to suppress the noise caused by other terms such as “rolls”. Dependency parsing seems to be a good solution for finding the most related opinion and indeed it was utilized in BIBREF8 , but the parser is prone to generating mistakes when processing the informal online reviews, as discussed in BIBREF17 .
To make use of opinion information and suppress the possible noise, we propose a novel Selective Transformation Network (STN) (the STN block in Figure FIGREF3 ), and insert it before attending to global opinion features so that more important features with respect to a given aspect candidate will be highlighted. Specifically, STN first calculates a new opinion representation INLINEFORM0 given the current aspect feature INLINEFORM1 as follows: DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are parameters for history-aware aspect representations and opinion representations respectively. They map INLINEFORM2 and INLINEFORM3 to the same subspace. Here the aspect feature INLINEFORM4 acts as a “filter” to keep more important opinion features. Equation EQREF14 also introduces a residual block to obtain a better opinion representation INLINEFORM5 , which is conditioned on the current aspect feature INLINEFORM6 .
For distilling the global opinion summary, we introduce a bi-linear term to calculate the association score between INLINEFORM0 and each INLINEFORM1 : DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are parameters of the Bi-Linear Attention layer. The improved opinion summary INLINEFORM2 at the time INLINEFORM3 is obtained via the weighted sum of the opinion representations: DISPLAYFORM0
Finally, we concatenate the opinion summary INLINEFORM0 and the history-aware aspect representation INLINEFORM1 and feed it into the top-most fully-connected (FC) layer for aspect prediction: DISPLAYFORM0 DISPLAYFORM1
Note that our framework actually performs a multi-task learning, i.e. predicting both aspects and opinions. We regard the initial token-level representations INLINEFORM0 as the features for opinion prediction: DISPLAYFORM0
INLINEFORM0 and INLINEFORM1 are parameters of the FC layers.
## Joint Training
All the components in the proposed framework are differentiable. Thus, our framework can be efficiently trained with gradient methods. We use the token-level cross-entropy error between the predicted distribution INLINEFORM0 ( INLINEFORM1 ) and the gold distribution INLINEFORM2 as the loss function: DISPLAYFORM0
Then, the losses from both tasks are combined to form the training objective of the entire model: DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 represent the loss functions for aspect and opinion extractions respectively.
## Datasets
To evaluate the effectiveness of the proposed framework for the ATE task, we conduct experiments over four benchmark datasets from the SemEval ABSA challenge BIBREF1 , BIBREF18 , BIBREF12 . Table TABREF24 shows their statistics. INLINEFORM0 (SemEval 2014) contains reviews of the laptop domain and those of INLINEFORM1 (SemEval 2014), INLINEFORM2 (SemEval 2015) and INLINEFORM3 (SemEval 2016) are for the restaurant domain. In these datasets, aspect terms have been labeled by the task organizer.
Gold standard annotations for opinion words are not provided. Thus, we choose words with strong subjectivity from MPQA to provide the distant supervision BIBREF19 . To compare with the best SemEval systems and the current state-of-the-art methods, we use the standard train-test split in SemEval challenge as shown in Table TABREF24 .
## Comparisons
We compare our framework with the following methods:
CRF-1: Conditional Random Fields with basic feature templates.
CRF-2: Conditional Random Fields with basic feature templates and word embeddings.
Semi-CRF: First-order Semi-Markov Conditional Random Fields BIBREF20 and the feature templates in BIBREF21 are adopted.
LSTM: Vanilla bi-directional LSTM with pre-trained word embeddings.
IHS_RD BIBREF2 , DLIREC BIBREF3 , EliXa BIBREF22 , NLANGP BIBREF4 : The winning systems in the ATE subtask in SemEval ABSA challenge BIBREF1 , BIBREF18 , BIBREF12 .
WDEmb BIBREF5 : Enhanced CRF with word embeddings, dependency path embeddings and linear context embeddings.
MIN BIBREF17 : MIN consists of three LSTMs. Two LSTMs are employed to model the memory interactions between ATE and opinion detection. The last one is a vanilla LSTM used to predict the subjectivity of the sentence as additional guidance.
RNCRF BIBREF8 : CRF with high-level representations learned from Dependency Tree based Recursive Neural Network.
CMLA BIBREF9 : CMLA is a multi-layer architecture where each layer consists of two coupled GRUs to model the relation between aspect terms and opinion words.
To clarify, our framework aims at extracting aspect terms where the opinion information is employed as auxiliary, while RNCRF and CMLA perform joint extraction of aspects and opinions. Nevertheless, the comparison between our framework and RNCRF/CMLA is still fair, because we do not use manually annotated opinions as used by RNCRF and CMLA, instead, we employ an existing opinion lexicon to provide weak opinion supervision.
## Settings
We pre-processed each dataset by lowercasing all words and replace all punctuations with PUNCT. We use pre-trained GloVe 840B vectors BIBREF23 to initialize the word embeddings and the dimension (i.e., INLINEFORM0 ) is 300. For out-of-vocabulary words, we randomly sample their embeddings from the uniform distribution INLINEFORM1 as done in BIBREF24 . All of the weight matrices except those in LSTMs are initialized from the uniform distribution INLINEFORM2 . For the initialization of the matrices in LSTMs, we adopt Glorot Uniform strategy BIBREF25 . Besides, all biases are initialized as 0's.
The model is trained with SGD. We apply dropout over the ultimate aspect/opinion features and the input word embeddings of LSTMs. The dropout rates are empirically set as 0.5. With 5-fold cross-validation on the training data of INLINEFORM0 , other hyper-parameters are set as follows: INLINEFORM1 , INLINEFORM2 ; the number of cached historical aspect representations INLINEFORM3 is 5; the learning rate of SGD is 0.07.
## Main Results
As shown in Table TABREF39 , the proposed framework consistently obtains the best scores on all of the four datasets. Compared with the winning systems of SemEval ABSA, our framework achieves 5.0%, 1.6%, 1.4%, 1.3% absolute gains on INLINEFORM0 , INLINEFORM1 , INLINEFORM2 and INLINEFORM3 respectively.
Our framework can outperform RNCRF, a state-of-the-art model based on dependency parsing, on all datasets. We also notice that RNCRF does not perform well on INLINEFORM0 and INLINEFORM1 (3.7% and 3.9% inferior than ours). We find that INLINEFORM2 and INLINEFORM3 contain many informal reviews, thus RNCRF's performance degradation is probably due to the errors from the dependency parser when processing such informal texts.
CMLA and MIN do not rely on dependency parsing, instead, they employ attention mechanism to distill opinion information to help aspect extraction. Our framework consistently performs better than them. The gains presumably come from two perspectives: (1) In our model, the opinion summary is exploited after performing the selective transformation conditioned on the current aspect features, thus the summary can to some extent avoid the noise due to directly applying conventional attention. (2) Our model can discover some uncommon aspects under the guidance of some commonly-used aspects in coordinate structures by the history attention.
CRF with basic feature template is not strong, therefore, we add CRF-2 as another baseline. As shown in Table TABREF39 , CRF-2 with word embeddings achieves much better results than CRF-1 on all datasets. WDEmb, which is also an enhanced CRF-based method using additional dependency context embeddings, obtains superior performances than CRF-2. Therefore, the above comparison shows that word embeddings are useful and the embeddings incorporating structure information can further improve the performance.
## Ablation Study
To further investigate the efficacy of the key components in our framework, namely, THA and STN, we perform ablation study as shown in the second block of Table TABREF39 . The results show that each of THA and STN is helpful for improving the performance, and the contribution of STN is slightly larger than THA. “OURS w/o THA & STN” only keeps the basic bi-linear attention. Although it performs not bad, it is still less competitive compared with the strongest baseline (i.e., CMLA), suggesting that only using attention mechanism to distill opinion summary is not enough. After inserting the STN component before the bi-linear attention, i.e. “OURS w/o THA”, we get about 1% absolute gains on each dataset, and then the performance is comparable to CMLA. By adding THA, i.e. “OURS”, the performance is further improved, and all state-of-the-art methods are surpassed.
## Attention Visualization and Case Study
In Figure FIGREF41 , we visualize the opinion attention scores of the words in two example sentences with the candidate aspects “maitre-D” and “bathroom”. The scores in Figures FIGREF41 and FIGREF41 show that our full model captures the related opinion words very accurately with significantly larger scores, i.e. “incredibly”, “unwelcoming” and “arrogant” for “maitre-D”, and “unfriendly” and “filthy” for “bathroom”. “OURS w/o STN” directly applies attention over the opinion hidden states INLINEFORM0 's, similar to what CMLA does. As shown in Figure FIGREF41 , it captures some unrelated opinion words (e.g. “fine”) and even some non-opinionated words. As a result, it brings in some noise into the global opinion summary, and consequently the final prediction accuracy will be affected. This example demonstrates that the proposed STN works pretty well to help attend to more related opinion words given a particular aspect.
Some predictions of our model and those of LSTM and OURS w/o THA & STN are given in Table TABREF43 . The models incorporating attention-based opinion summary (i.e., OURS and OURS w/o THA & STN) can better determine if the commonly-used nouns are aspect terms or not (e.g. “device” in the first input), since they make decisions based on the global opinion information. Besides, they are able to extract some infrequent or even misspelled aspect terms (e.g. “survice” in the second input) based on the indicative clues provided by opinion words. For the last three cases, having aspects in coordinate structures (i.e. the third and the fourth) or long aspects (i.e. the fifth), our model can give precise predictions owing to the previous detection clues captured by THA. Without using these clues, the baseline models fail.
## Related Work
Some initial works BIBREF26 developed a bootstrapping framework for tackling Aspect Term Extraction (ATE) based on the observation that opinion words are usually located around the aspects. BIBREF27 and BIBREF28 performed co-extraction of aspect terms and opinion words based on sophisticated syntactic patterns. However, relying on syntactic patterns suffers from parsing errors when processing informal online reviews. To avoid this drawback, BIBREF29 , BIBREF30 employed word-based translation models. Specifically, these models formulated the ATE task as a monolingual word alignment process and aspect-opinion relation is captured by alignment links rather than word dependencies. The ATE task can also be formulated as a token-level sequence labeling problem. The winning systems BIBREF2 , BIBREF22 , BIBREF4 of SemEval ABSA challenges employed traditional sequence models, such as Conditional Random Fields (CRFs) and Maximum Entropy (ME), to detect aspects. Besides heavy feature engineering, they also ignored the consideration of opinions.
Recently, neural network based models, such as LSTM-based BIBREF6 and CNN-based BIBREF31 methods, become the mainstream approach. Later on, some neural models jointly extracting aspect and opinion were proposed. BIBREF8 performs the two task in a single Tree-Based Recursive Neural Network. Their network structure depends on dependency parsing, which is prone to error on informal reviews. CMLA BIBREF9 consists of multiple attention layers on top of standard GRUs to extract the aspects and opinion words. Similarly, MIN BIBREF17 employs multiple LSTMs to interactively perform aspect term extraction and opinion word extraction in a multi-task learning framework. Our framework is different from them in two perspectives: (1) It filters the opinion summary by incorporating the aspect features at each time step into the original opinion representations; (2) It exploits history information of aspect detection to capture the coordinate structures and previous aspect features.
## Concluding Discussions
For more accurate aspect term extraction, we explored two important types of information, namely aspect detection history, and opinion summary. We design two components, i.e. truncated history attention, and selective transformation network. Experimental results show that our model dominates those joint extraction works such as RNCRF and CMLA on the performance of ATE. It suggests that the joint extraction sacrifices the accuracy of aspect prediction, although the ground-truth opinion words were annotated by these authors. Moreover, one should notice that those joint extraction methods do not care about the correspondence between the extracted aspect terms and opinion words. Therefore, the necessity of such joint extraction should be obelized, given the experimental findings in this paper.
| 12
|
1805.02400
|
Stay On-Topic: Generating Context-specific Fake Restaurant Reviews
|
# Stay On-Topic: Generating Context-specific Fake Restaurant Reviews
## Abstract
Automatically generated fake restaurant reviews are a threat to online review systems. Recent research has shown that users have difficulties in detecting machine-generated fake reviews hiding among real restaurant reviews. The method used in this work (char-LSTM ) has one drawback: it has difficulties staying in context, i.e. when it generates a review for specific target entity, the resulting review may contain phrases that are unrelated to the target, thus increasing its detectability. In this work, we present and evaluate a more sophisticated technique based on neural machine translation (NMT) with which we can generate reviews that stay on-topic. We test multiple variants of our technique using native English speakers on Amazon Mechanical Turk. We demonstrate that reviews generated by the best variant have almost optimal undetectability (class-averaged F-score 47%). We conduct a user study with skeptical users and show that our method evades detection more frequently compared to the state-of-the-art (average evasion 3.2/4 vs 1.5/4) with statistical significance, at level {\alpha} = 1% (Section 4.3). We develop very effective detection tools and reach average F-score of 97% in classifying these. Although fake reviews are very effective in fooling people, effective automatic detection is still feasible.
## Introduction
Automatically generated fake reviews have only recently become natural enough to fool human readers. Yao et al. BIBREF0 use a deep neural network (a so-called 2-layer LSTM BIBREF1 ) to generate fake reviews, and concluded that these fake reviews look sufficiently genuine to fool native English speakers. They train their model using real restaurant reviews from yelp.com BIBREF2 . Once trained, the model is used to generate reviews character-by-character. Due to the generation methodology, it cannot be easily targeted for a specific context (meaningful side information). Consequently, the review generation process may stray off-topic. For instance, when generating a review for a Japanese restaurant in Las Vegas, the review generation process may include references to an Italian restaurant in Baltimore. The authors of BIBREF0 apply a post-processing step (customization), which replaces food-related words with more suitable ones (sampled from the targeted restaurant). The word replacement strategy has drawbacks: it can miss certain words and replace others independent of their surrounding words, which may alert savvy readers. As an example: when we applied the customization technique described in BIBREF0 to a review for a Japanese restaurant it changed the snippet garlic knots for breakfast with garlic knots for sushi).
We propose a methodology based on neural machine translation (NMT) that improves the generation process by defining a context for the each generated fake review. Our context is a clear-text sequence of: the review rating, restaurant name, city, state and food tags (e.g. Japanese, Italian). We show that our technique generates review that stay on topic. We can instantiate our basic technique into several variants. We vet them on Amazon Mechanical Turk and find that native English speakers are very poor at recognizing our fake generated reviews. For one variant, the participants' performance is close to random: the class-averaged F-score of detection is INLINEFORM0 (whereas random would be INLINEFORM1 given the 1:6 imbalance in the test). Via a user study with experienced, highly educated participants, we compare this variant (which we will henceforth refer to as NMT-Fake* reviews) with fake reviews generated using the char-LSTM-based technique from BIBREF0 .
We demonstrate that NMT-Fake* reviews constitute a new category of fake reviews that cannot be detected by classifiers trained only using previously known categories of fake reviews BIBREF0 , BIBREF3 , BIBREF4 . Therefore, NMT-Fake* reviews may go undetected in existing online review sites. To meet this challenge, we develop an effective classifier that detects NMT-Fake* reviews effectively (97% F-score). Our main contributions are:
## Background
Fake reviews User-generated content BIBREF5 is an integral part of the contemporary user experience on the web. Sites like tripadvisor.com, yelp.com and Google Play use user-written reviews to provide rich information that helps other users choose where to spend money and time. User reviews are used for rating services or products, and for providing qualitative opinions. User reviews and ratings may be used to rank services in recommendations. Ratings have an affect on the outwards appearance. Already 8 years ago, researchers estimated that a one-star rating increase affects the business revenue by 5 – 9% on yelp.com BIBREF6 .
Due to monetary impact of user-generated content, some businesses have relied on so-called crowd-turfing agents BIBREF7 that promise to deliver positive ratings written by workers to a customer in exchange for a monetary compensation. Crowd-turfing ethics are complicated. For example, Amazon community guidelines prohibit buying content relating to promotions, but the act of writing fabricated content is not considered illegal, nor is matching workers to customers BIBREF8 . Year 2015, approximately 20% of online reviews on yelp.com were suspected of being fake BIBREF9 .
Nowadays, user-generated review sites like yelp.com use filters and fraudulent review detection techniques. These factors have resulted in an increase in the requirements of crowd-turfed reviews provided to review sites, which in turn has led to an increase in the cost of high-quality review. Due to the cost increase, researchers hypothesize the existence of neural network-generated fake reviews. These neural-network-based fake reviews are statistically different from human-written fake reviews, and are not caught by classifiers trained on these BIBREF0 .
Detecting fake reviews can either be done on an individual level or as a system-wide detection tool (i.e. regulation). Detecting fake online content on a personal level requires knowledge and skills in critical reading. In 2017, the National Literacy Trust assessed that young people in the UK do not have the skillset to differentiate fake news from real news BIBREF10 . For example, 20% of children that use online news sites in age group 12-15 believe that all information on news sites are true.
Neural Networks Neural networks are function compositions that map input data through INLINEFORM0 subsequent layers: DISPLAYFORM0
where the functions INLINEFORM0 are typically non-linear and chosen by experts partly for known good performance on datasets and partly for simplicity of computational evaluation. Language models (LMs) BIBREF11 are generative probability distributions that assign probabilities to sequences of tokens ( INLINEFORM1 ): DISPLAYFORM0
such that the language model can be used to predict how likely a specific token at time step INLINEFORM0 is, based on the INLINEFORM1 previous tokens. Tokens are typically either words or characters.
For decades, deep neural networks were thought to be computationally too difficult to train. However, advances in optimization, hardware and the availability of frameworks have shown otherwise BIBREF1 , BIBREF12 . Neural language models (NLMs) have been one of the promising application areas. NLMs are typically various forms of recurrent neural networks (RNNs), which pass through the data sequentially and maintain a memory representation of the past tokens with a hidden context vector. There are many RNN architectures that focus on different ways of updating and maintaining context vectors: Long Short-Term Memory units (LSTM) and Gated Recurrent Units (GRUs) are perhaps most popular. Neural LMs have been used for free-form text generation. In certain application areas, the quality has been high enough to sometimes fool human readers BIBREF0 . Encoder-decoder (seq2seq) models BIBREF13 are architectures of stacked RNNs, which have the ability to generate output sequences based on input sequences. The encoder network reads in a sequence of tokens, and passes it to a decoder network (a LM). In contrast to simpler NLMs, encoder-decoder networks have the ability to use additional context for generating text, which enables more accurate generation of text. Encoder-decoder models are integral in Neural Machine Translation (NMT) BIBREF14 , where the task is to translate a source text from one language to another language. NMT models additionally use beam search strategies to heuristically search the set of possible translations. Training datasets are parallel corpora; large sets of paired sentences in the source and target languages. The application of NMT techniques for online machine translation has significantly improved the quality of translations, bringing it closer to human performance BIBREF15 .
Neural machine translation models are efficient at mapping one expression to another (one-to-one mapping). Researchers have evaluated these models for conversation generation BIBREF16 , with mixed results. Some researchers attribute poor performance to the use of the negative log likelihood cost function during training, which emphasizes generation of high-confidence phrases rather than diverse phrases BIBREF17 . The results are often generic text, which lacks variation. Li et al. have suggested various augmentations to this, among others suppressing typical responses in the decoder language model to promote response diversity BIBREF17 .
## System Model
We discuss the attack model, our generative machine learning method and controlling the generative process in this section.
## Attack Model
Wang et al. BIBREF7 described a model of crowd-turfing attacks consisting of three entities: customers who desire to have fake reviews for a particular target (e.g. their restaurant) on a particular platform (e.g. Yelp), agents who offer fake review services to customers, and workers who are orchestrated by the agent to compose and post fake reviews.
Automated crowd-turfing attacks (ACA) replace workers by a generative model. This has several benefits including better economy and scalability (human workers are more expensive and slower) and reduced detectability (agent can better control the rate at which fake reviews are generated and posted).
We assume that the agent has access to public reviews on the review platform, by which it can train its generative model. We also assume that it is easy for the agent to create a large number of accounts on the review platform so that account-based detection or rate-limiting techniques are ineffective against fake reviews.
The quality of the generative model plays a crucial role in the attack. Yao et al. BIBREF0 propose the use of a character-based LSTM as base for generative model. LSTMs are not conditioned to generate reviews for a specific target BIBREF1 , and may mix-up concepts from different contexts during free-form generation. Mixing contextually separate words is one of the key criteria that humans use to identify fake reviews. These may result in violations of known indicators for fake content BIBREF18 . For example, the review content may not match prior expectations nor the information need that the reader has. We improve the attack model by considering a more capable generative model that produces more appropriate reviews: a neural machine translation (NMT) model.
## Generative Model
We propose the use of NMT models for fake review generation. The method has several benefits: 1) the ability to learn how to associate context (keywords) to reviews, 2) fast training time, and 3) a high-degree of customization during production time, e.g. introduction of specific waiter or food items names into reviews.
NMT models are constructions of stacked recurrent neural networks (RNNs). They include an encoder network and a decoder network, which are jointly optimized to produce a translation of one sequence to another. The encoder rolls over the input data in sequence and produces one INLINEFORM0 -dimensional context vector representation for the sentence. The decoder then generates output sequences based on the embedding vector and an attention module, which is taught to associate output words with certain input words. The generation typically continues until a specific EOS (end of sentence) token is encountered. The review length can be controlled in many ways, e.g. by setting the probability of generating the EOS token to zero until the required length is reached.
NMT models often also include a beam search BIBREF14 , which generates several hypotheses and chooses the best ones amongst them. In our work, we use the greedy beam search technique. We forgo the use of additional beam searches as we found that the quality of the output was already adequate and the translation phase time consumption increases linearly for each beam used.
We use the Yelp Challenge dataset BIBREF2 for our fake review generation. The dataset (Aug 2017) contains 2.9 million 1 –5 star restaurant reviews. We treat all reviews as genuine human-written reviews for the purpose of this work, since wide-scale deployment of machine-generated review attacks are not yet reported (Sep 2017) BIBREF19 . As preprocessing, we remove non-printable (non-ASCII) characters and excessive white-space. We separate punctuation from words. We reserve 15,000 reviews for validation and 3,000 for testing, and the rest we use for training. NMT models require a parallel corpus of source and target sentences, i.e. a large set of (source, target)-pairs. We set up a parallel corpus by constructing (context, review)-pairs from the dataset. Next, we describe how we created our input context.
The Yelp Challenge dataset includes metadata about restaurants, including their names, food tags, cities and states these restaurants are located in. For each restaurant review, we fetch this metadata and use it as our input context in the NMT model. The corresponding restaurant review is similarly set as the target sentence. This method produced 2.9 million pairs of sentences in our parallel corpus. We show one example of the parallel training corpus in Example 1 below:
5 Public House Las Vegas NV Gastropubs Restaurants > Excellent
food and service . Pricey , but well worth it . I would recommend
the bone marrow and sampler platter for appetizers . \end{verbatim}
\noindent The order {\textbf{[rating name city state tags]}} is kept constant.
Training the model conditions it to associate certain sequences of words in the input sentence with others in the output.
\subsubsection{Training Settings}
We train our NMT model on a commodity PC with a i7-4790k CPU (4.00GHz), with 32GB RAM and one NVidia GeForce GTX 980 GPU. Our system can process approximately 1,300 \textendash 1,500 source tokens/s and approximately 5,730 \textendash 5,830 output tokens/s. Training one epoch takes in average 72 minutes. The model is trained for 8 epochs, i.e. over night. We call fake review generated by this model \emph{NMT-Fake reviews}. We only need to train one model to produce reviews of different ratings.
We use the training settings: adam optimizer \cite{kingma2014adam} with the suggested learning rate 0.001 \cite{klein2017opennmt}. For most parts, parameters are at their default values. Notably, the maximum sentence length of input and output is 50 tokens by default.
We leverage the framework openNMT-py \cite{klein2017opennmt} to teach the our NMT model.
We list used openNMT-py commands in Appendix Table~\ref{table:openNMT-py_commands}.
\begin{figure}[t]
\begin{center}
\begin{tabular}{ | l | }
\hline
Example 2. Greedy NMT \\
Great food, \underline{great} service, \underline{great} \textit{\textit{beer selection}}. I had the \textit{Gastropubs burger} and it
\\
was delicious. The \underline{\textit{beer selection}} was also \underline{great}. \\
\\
Example 3. NMT-Fake* \\
I love this restaurant. Great food, great service. It's \textit{a little pricy} but worth\\
it for the \textit{quality} of the \textit{beer} and atmosphere you can see in \textit{Vegas}
\\
\hline
\end{tabular}
\label{table:output_comparison}
\end{center}
\caption{Na\"{i}ve text generation with NMT vs. generation using our NTM model. Repetitive patterns are \underline{underlined}. Contextual words are \emph{italicized}. Both examples here are generated based on the context given in Example~1.}
\label{fig:comparison}
\end{figure}
\subsection{Controlling generation of fake reviews}
\label{sec:generating}
Greedy NMT beam searches are practical in many NMT cases. However, the results are simply repetitive, when naively applied to fake review generation (See Example~2 in Figure~\ref{fig:comparison}).
The NMT model produces many \emph{high-confidence} word predictions, which are repetitive and obviously fake. We calculated that in fact, 43\% of the generated sentences started with the phrase ``Great food''. The lack of diversity in greedy use of NMTs for text generation is clear.
\begin{algorithm}[!b]
\KwData{Desired review context $C_\mathrm{input}$ (given as cleartext), NMT model}
\KwResult{Generated review $out$ for input context $C_\mathrm{input}$}
set $b=0.3$, $\lambda=-5$, $\alpha=\frac{2}{3}$, $p_\mathrm{typo}$, $p_\mathrm{spell}$ \\
$\log p \leftarrow \text{NMT.decode(NMT.encode(}C_\mathrm{input}\text{))}$ \\
out $\leftarrow$ [~] \\
$i \leftarrow 0$ \\
$\log p \leftarrow \text{Augment}(\log p$, $b$, $\lambda$, $1$, $[~]$, 0)~~~~~~~~~~~~~~~ |~random penalty~\\
\While{$i=0$ or $o_i$ not EOS}{
$\log \Tilde{p} \leftarrow \text{Augment}(\log p$, $b$, $\lambda$, $\alpha$, $o_i$, $i$)~~~~~~~~~~~ |~start \& memory penalty~\\
$o_i \leftarrow$ \text{NMT.beam}($\log \Tilde{p}$, out) \\
out.append($o_i$) \\
$i \leftarrow i+1$
}\text{return}~$\text{Obfuscate}$(out,~$p_\mathrm{typo}$,~$p_\mathrm{spell}$)
\caption{Generation of NMT-Fake* reviews.}
\label{alg:base}
\end{algorithm}
In this work, we describe how we succeeded in creating more diverse and less repetitive generated reviews, such as Example 3 in Figure~\ref{fig:comparison}.
We outline pseudocode for our methodology of generating fake reviews in Algorithm~\ref{alg:base}. There are several parameters in our algorithm.
The details of the algorithm will be shown later.
We modify the openNMT-py translation phase by changing log-probabilities before passing them to the beam search.
We notice that reviews generated with openNMT-py contain almost no language errors. As an optional post-processing step, we obfuscate reviews by introducing natural typos/misspellings randomly. In the next sections, we describe how we succeeded in generating more natural sentences from our NMT model, i.e. generating reviews like Example~3 instead of reviews like Example~2.
\subsubsection{Variation in word content}
Example 2 in Figure~\ref{fig:comparison} repeats commonly occurring words given for a specific context (e.g. \textit{great, food, service, beer, selection, burger} for Example~1). Generic review generation can be avoided by decreasing probabilities (log-likelihoods \cite{murphy2012machine}) of the generators LM, the decoder.
We constrain the generation of sentences by randomly \emph{imposing penalties to words}.
We tried several forms of added randomness, and found that adding constant penalties to a \emph{random subset} of the target words resulted in the most natural sentence flow. We call these penalties \emph{Bernoulli penalties}, since the random variables are chosen as either 1 or 0 (on or off).
\paragraph{Bernoulli penalties to language model}
To avoid generic sentences components, we augment the default language model $p(\cdot)$ of the decoder by
\begin{equation}
\log \Tilde{p}(t_k) = \log p(t_k | t_i, \dots, t_1) + \lambda q,
\end{equation}
where $q \in R^{V}$ is a vector of Bernoulli-distributed random values that obtain values $1$ with probability $b$ and value $0$ with probability $1-b_i$, and $\lambda < 0$. Parameter $b$ controls how much of the vocabulary is forgotten and $\lambda$ is a soft penalty of including ``forgotten'' words in a review.
$\lambda q_k$ emphasizes sentence forming with non-penalized words. The randomness is reset at the start of generating a new review.
Using Bernoulli penalties in the language model, we can ``forget'' a certain proportion of words and essentially ``force'' the creation of less typical sentences. We will test the effect of these two parameters, the Bernoulli probability $b$ and log-likelihood penalty of including ``forgotten'' words $\lambda$, with a user study in Section~\ref{sec:varying}.
\paragraph{Start penalty}
We introduce start penalties to avoid generic sentence starts (e.g. ``Great food, great service''). Inspired by \cite{li2016diversity}, we add a random start penalty $\lambda s^\mathrm{i}$, to our language model, which decreases monotonically for each generated token. We set $\alpha \leftarrow 0.66$ as it's effect decreases by 90\% every 5 words generated.
\paragraph{Penalty for reusing words}
Bernoulli penalties do not prevent excessive use of certain words in a sentence (such as \textit{great} in Example~2).
To avoid excessive reuse of words, we included a memory penalty for previously used words in each translation.
Concretely, we add the penalty $\lambda$ to each word that has been generated by the greedy search.
\subsubsection{Improving sentence coherence}
\label{sec:grammar}
We visually analyzed reviews after applying these penalties to our NMT model. While the models were clearly diverse, they were \emph{incoherent}: the introduction of random penalties had degraded the grammaticality of the sentences. Amongst others, the use of punctuation was erratic, and pronouns were used semantically wrongly (e.g. \emph{he}, \emph{she} might be replaced, as could ``and''/``but''). To improve the authenticity of our reviews, we added several \emph{grammar-based rules}.
English language has several classes of words which are important for the natural flow of sentences.
We built a list of common pronouns (e.g. I, them, our), conjunctions (e.g. and, thus, if), punctuation (e.g. ,/.,..), and apply only half memory penalties for these words. We found that this change made the reviews more coherent. The pseudocode for this and the previous step is shown in Algorithm~\ref{alg:aug}.
The combined effect of grammar-based rules and LM augmentation is visible in Example~3, Figure~\ref{fig:comparison}.
\begin{algorithm}[!t]
\KwData{Initial log LM $\log p$, Bernoulli probability $b$, soft-penalty $\lambda$, monotonic factor $\alpha$, last generated token $o_i$, grammar rules set $G$}
\KwResult{Augmented log LM $\log \Tilde{p}$}
\begin{algorithmic}[1]
\Procedure {Augment}{$\log p$, $b$, $\lambda$, $\alpha$, $o_i$, $i$}{ \\
generate $P_{\mathrm{1:N}} \leftarrow Bernoulli(b)$~~~~~~~~~~~~~~~|~$\text{One value} \in \{0,1\}~\text{per token}$~ \\
$I \leftarrow P>0$ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~|~Select positive indices~\\
$\log \Tilde{p} \leftarrow$ $\text{Discount}$($\log p$, $I$, $\lambda \cdot \alpha^i$,$G$) ~~~~~~ |~start penalty~\\
$\log \Tilde{p} \leftarrow$ $\text{Discount}$($\log \Tilde{p}$, $[o_i]$, $\lambda$,$G$) ~~~~~~~~~ |~memory penalty~\\
\textbf{return}~$\log \Tilde{p}$
}
\EndProcedure
\\
\Procedure {Discount}{$\log p$, $I$, $\lambda$, $G$}{
\State{\For{$i \in I$}{
\eIf{$o_i \in G$}{
$\log p_{i} \leftarrow \log p_{i} + \lambda/2$
}{
$\log p_{i} \leftarrow \log p_{i} + \lambda$}
}\textbf{return}~$\log p$
\EndProcedure
}}
\end{algorithmic}
\caption{Pseudocode for augmenting language model. }
\label{alg:aug}
\end{algorithm}
\subsubsection{Human-like errors}
\label{sec:obfuscation}
We notice that our NMT model produces reviews without grammar mistakes.
This is unlike real human writers, whose sentences contain two types of language mistakes 1) \emph{typos} that are caused by mistakes in the human motoric input, and 2) \emph{common spelling mistakes}.
We scraped a list of common English language spelling mistakes from Oxford dictionary\footnote{\url{https://en.oxforddictionaries.com/spelling/common-misspellings}} and created 80 rules for randomly \emph{re-introducing spelling mistakes}.
Similarly, typos are randomly reintroduced based on the weighted edit distance\footnote{\url{https://pypi.python.org/pypi/weighted-levenshtein/0.1}}, such that typos resulting in real English words with small perturbations are emphasized.
We use autocorrection tools\footnote{\url{https://pypi.python.org/pypi/autocorrect/0.1.0}} for finding these words.
We call these augmentations \emph{obfuscations}, since they aim to confound the reader to think a human has written them. We omit the pseudocode description for brevity.
\subsection{Experiment: Varying generation parameters in our NMT model}
\label{sec:varying}
Parameters $b$ and $\lambda$ control different aspects in fake reviews.
We show six different examples of generated fake reviews in Table~\ref{table:categories}.
Here, the largest differences occur with increasing values of $b$: visibly, the restaurant reviews become more extreme.
This occurs because a large portion of vocabulary is ``forgotten''. Reviews with $b \geq 0.7$ contain more rare word combinations, e.g. ``!!!!!'' as punctuation, and they occasionally break grammaticality (''experience was awesome'').
Reviews with lower $b$ are more generic: they contain safe word combinations like ``Great place, good service'' that occur in many reviews. Parameter $\lambda$'s is more subtle: it affects how random review starts are and to a degree, the discontinuation between statements within the review.
We conducted an Amazon Mechanical Turk (MTurk) survey in order to determine what kind of NMT-Fake reviews are convincing to native English speakers. We describe the survey and results in the next section.
\begin{table}[!b]
\caption{Six different parametrizations of our NMT reviews and one example for each. The context is ``5 P~.~F~.~Chang ' s Scottsdale AZ'' in all examples.}
\begin{center}
\begin{tabular}{ | l | l | }
\hline
$(b, \lambda)$ & Example review for context \\ \hline
\hline
$(0.3, -3)$ & I love this location! Great service, great food and the best drinks in Scottsdale. \\
& The staff is very friendly and always remembers u when we come in\\\hline
$(0.3, -5)$ & Love love the food here! I always go for lunch. They have a great menu and \\
& they make it fresh to order. Great place, good service and nice staff\\\hline
$(0.5, -4)$ & I love their chicken lettuce wraps and fried rice!! The service is good, they are\\
& always so polite. They have great happy hour specials and they have a lot\\
& of options.\\\hline
$(0.7, -3)$ & Great place to go with friends! They always make sure your dining \\
& experience was awesome.\\ \hline
$(0.7, -5)$ & Still haven't ordered an entree before but today we tried them once..\\
& both of us love this restaurant....\\\hline
$(0.9, -4)$ & AMAZING!!!!! Food was awesome with excellent service. Loved the lettuce \\
& wraps. Great drinks and wine! Can't wait to go back so soon!!\\ \hline
\end{tabular}
\label{table:categories}
\end{center}
\end{table}
\subsubsection{MTurk study}
\label{sec:amt}
We created 20 jobs, each with 100 questions, and requested master workers in MTurk to complete the jobs.
We randomly generated each survey for the participants. Each review had a 50\% chance to be real or fake. The fake ones further were chosen among six (6) categories of fake reviews (Table~\ref{table:categories}).
The restaurant and the city was given as contextual information to the participants. Our aim was to use this survey to understand how well English-speakers react to different parametrizations of NMT-Fake reviews.
Table~\ref{table:amt_pop} in Appendix summarizes the statistics for respondents in the survey. All participants were native English speakers from America. The base rate (50\%) was revealed to the participants prior to the study.
We first investigated overall detection of any NMT-Fake reviews (1,006 fake reviews and 994 real reviews). We found that the participants had big difficulties in detecting our fake reviews. In average, the reviews were detected with class-averaged \emph{F-score of only 56\%}, with 53\% F-score for fake review detection and 59\% F-score for real review detection. The results are very close to \emph{random detection}, where precision, recall and F-score would each be 50\%. Results are recorded in Table~\ref{table:MTurk_super}. Overall, the fake review generation is very successful, since human detection rate across categories is close to random.
\begin{table}[t]
\caption{Effectiveness of Mechanical Turkers in distinguishing human-written reviews from fake reviews generated by our NMT model (all variants).}
\begin{center}
\begin{tabular}{ | c | c |c |c | c | }
\hline
\multicolumn{5}{|c|}{Classification report}
\\ \hline
Review Type & Precision & Recall & F-score & Support \\ \hline
\hline
Human & 55\% & 63\% & 59\% & 994\\
NMT-Fake & 57\% & 50\% & 53\% & 1006 \\
\hline
\end{tabular}
\label{table:MTurk_super}
\end{center}
\end{table}
We noticed some variation in the detection of different fake review categories. The respondents in our MTurk survey had most difficulties recognizing reviews of category $(b=0.3, \lambda=-5)$, where true positive rate was $40.4\%$, while the true negative rate of the real class was $62.7\%$. The precision were $16\%$ and $86\%$, respectively. The class-averaged F-score is $47.6\%$, which is close to random. Detailed classification reports are shown in Table~\ref{table:MTurk_sub} in Appendix. Our MTurk-study shows that \emph{our NMT-Fake reviews pose a significant threat to review systems}, since \emph{ordinary native English-speakers have very big difficulties in separating real reviews from fake reviews}. We use the review category $(b=0.3, \lambda=-5)$ for future user tests in this paper, since MTurk participants had most difficulties detecting these reviews. We refer to this category as NMT-Fake* in this paper.
\section{Evaluation}
\graphicspath{ {figures/}}
We evaluate our fake reviews by first comparing them statistically to previously proposed types of fake reviews, and proceed with a user study with experienced participants. We demonstrate the statistical difference to existing fake review types \cite{yao2017automated,mukherjee2013yelp,rayana2015collective} by training classifiers to detect previous types and investigate classification performance.
\subsection{Replication of state-of-the-art model: LSTM}
\label{sec:repl}
Yao et al. \cite{yao2017automated} presented the current state-of-the-art generative model for fake reviews. The model is trained over the Yelp Challenge dataset using a two-layer character-based LSTM model.
We requested the authors of \cite{yao2017automated} for access to their LSTM model or a fake review dataset generated by their model. Unfortunately they were not able to share either of these with us. We therefore replicated their model as closely as we could, based on their paper and e-mail correspondence\footnote{We are committed to sharing our code with bonafide researchers for the sake of reproducibility.}.
We used the same graphics card (GeForce GTX) and trained using the same framework (torch-RNN in lua). We downloaded the reviews from Yelp Challenge and preprocessed the data to only contain printable ASCII characters, and filtered out non-restaurant reviews. We trained the model for approximately 72 hours. We post-processed the reviews using the customization methodology described in \cite{yao2017automated} and email correspondence. We call fake reviews generated by this model LSTM-Fake reviews.
\subsection{Similarity to existing fake reviews}
\label{sec:automated}
We now want to understand how NMT-Fake* reviews compare to a) LSTM fake reviews and b) human-generated fake reviews. We do this by comparing the statistical similarity between these classes.
For `a' (Figure~\ref{fig:lstm}), we use the Yelp Challenge dataset. We trained a classifier using 5,000 random reviews from the Yelp Challenge dataset (``human'') and 5,000 fake reviews generated by LSTM-Fake. Yao et al. \cite{yao2017automated} found that character features are essential in identifying LSTM-Fake reviews. Consequently, we use character features (n-grams up to 3).
For `b' (Figure~\ref{fig:shill}),we the ``Yelp Shills'' dataset (combination of YelpZip \cite{mukherjee2013yelp}, YelpNYC \cite{mukherjee2013yelp}, YelpChi \cite{rayana2015collective}). This dataset labels entries that are identified as fraudulent by Yelp's filtering mechanism (''shill reviews'')\footnote{Note that shill reviews are probably generated by human shills \cite{zhao2017news}.}. The rest are treated as genuine reviews from human users (''genuine''). We use 100,000 reviews from each category to train a classifier. We use features from the commercial psychometric tool LIWC2015 \cite{pennebaker2015development} to generated features.
In both cases, we use AdaBoost (with 200 shallow decision trees) for training. For testing each classifier, we use a held out test set of 1,000 reviews from both classes in each case. In addition, we test 1,000 NMT-Fake* reviews. Figures~\ref{fig:lstm} and~\ref{fig:shill} show the results. The classification threshold of 50\% is marked with a dashed line.
\begin{figure}
\begin{subfigure}[b]{0.5\columnwidth}
\includegraphics[width=\columnwidth]{figures/lstm.png}
\caption{Human--LSTM reviews.}
\label{fig:lstm}
\end{subfigure}
\begin{subfigure}[b]{0.5\columnwidth}
\includegraphics[width=\columnwidth]{figures/distribution_shill.png}
\caption{Genuine--Shill reviews.}
\label{fig:shill}
\end{subfigure}
\caption{
Histogram comparison of NMT-Fake* reviews with LSTM-Fake reviews and human-generated (\emph{genuine} and \emph{shill}) reviews. Figure~\ref{fig:lstm} shows that a classifier trained to distinguish ``human'' vs. LSTM-Fake cannot distinguish ``human'' vs NMT-Fake* reviews. Figure~\ref{fig:shill} shows NMT-Fake* reviews are more similar to \emph{genuine} reviews than \emph{shill} reviews.
}
\label{fig:statistical_similarity}
\end{figure}
We can see that our new generated reviews do not share strong attributes with previous known categories of fake reviews. If anything, our fake reviews are more similar to genuine reviews than previous fake reviews. We thus conjecture that our NMT-Fake* fake reviews present a category of fake reviews that may go undetected on online review sites.
\subsection{Comparative user study}
\label{sec:comparison}
We wanted to evaluate the effectiveness of fake reviews againsttech-savvy users who understand and know to expect machine-generated fake reviews. We conducted a user study with 20 participants, all with computer science education and at least one university degree. Participant demographics are shown in Table~\ref{table:amt_pop} in the Appendix. Each participant first attended a training session where they were asked to label reviews (fake and genuine) and could later compare them to the correct answers -- we call these participants \emph{experienced participants}.
No personal data was collected during the user study.
Each person was given two randomly selected sets of 30 of reviews (a total of 60 reviews per person) with reviews containing 10 \textendash 50 words each.
Each set contained 26 (87\%) real reviews from Yelp and 4 (13\%) machine-generated reviews,
numbers chosen based on suspicious review prevalence on Yelp~\cite{mukherjee2013yelp,rayana2015collective}.
One set contained machine-generated reviews from one of the two models (NMT ($b=0.3, \lambda=-5$) or LSTM),
and the other set reviews from the other in randomized order. The number of fake reviews was revealed to each participant in the study description. Each participant was requested to mark four (4) reviews as fake.
Each review targeted a real restaurant. A screenshot of that restaurant's Yelp page was shown to each participant prior to the study. Each participant evaluated reviews for one specific, randomly selected, restaurant. An example of the first page of the user study is shown in Figure~\ref{fig:screenshot} in Appendix.
\begin{figure}[!ht]
\centering
\includegraphics[width=.7\columnwidth]{detection2.png}
\caption{Violin plots of detection rate in comparative study. Mean and standard deviations for number of detected fakes are $0.8\pm0.7$ for NMT-Fake* and $2.5\pm1.0$ for LSTM-Fake. $n=20$. A sample of random detection is shown as comparison.}
\label{fig:aalto}
\end{figure}
Figure~\ref{fig:aalto} shows the distribution of detected reviews of both types. A hypothetical random detector is shown for comparison.
NMT-Fake* reviews are significantly more difficult to detect for our experienced participants. In average, detection rate (recall) is $20\%$ for NMT-Fake* reviews, compared to $61\%$ for LSTM-based reviews.
The precision (and F-score) is the same as the recall in our study, since participants labeled 4 fakes in each set of 30 reviews \cite{murphy2012machine}.
The distribution of the detection across participants is shown in Figure~\ref{fig:aalto}. \emph{The difference is statistically significant with confidence level $99\%$} (Welch's t-test).
We compared the detection rate of NMT-Fake* reviews to a random detector, and find that \emph{our participants detection rate of NMT-Fake* reviews is not statistically different from random predictions with 95\% confidence level} (Welch's t-test).
\section{Defenses}
\label{sec:detection}
We developed an AdaBoost-based classifier to detect our new fake reviews, consisting of 200 shallow decision trees (depth 2). The features we used are recorded in Table~\ref{table:features_adaboost} (Appendix).
We used word-level features based on spaCy-tokenization \cite{honnibal-johnson:2015:EMNLP} and constructed n-gram representation of POS-tags and dependency tree tags. We added readability features from NLTK~\cite{bird2004nltk}.
\begin{figure}[ht]
\centering
\includegraphics[width=.7\columnwidth]{obf_score_fair_2.png}
\caption{
Adaboost-based classification of NMT-Fake and human-written reviews.
Effect of varying $b$ and $\lambda$ in fake review generation.
The variant native speakers had most difficulties detecting is well detectable by AdaBoost (97\%).}
\label{fig:adaboost_matrix_b_lambda}
\end{figure}
Figure~\ref{fig:adaboost_matrix_b_lambda} shows our AdaBoost classifier's class-averaged F-score at detecting different kind of fake reviews. The classifier is very effective in detecting reviews that humans have difficulties detecting. For example, the fake reviews MTurk users had most difficulty detecting ($b=0.3, \lambda=-5$) are detected with an excellent 97\% F-score.
The most important features for the classification were counts for frequently occurring words in fake reviews (such as punctuation, pronouns, articles) as well as the readability feature ``Automated Readability Index''. We thus conclude that while NMT-Fake reviews are difficult to detect for humans, they can be well detected with the right tools.
\section{Related Work}
Kumar and Shah~\cite{kumar2018false} survey and categorize false information research. Automatically generated fake reviews are a form of \emph{opinion-based false information}, where the creator of the review may influence reader's opinions or decisions.
Yao et al. \cite{yao2017automated} presented their study on machine-generated fake reviews. Contrary to us, they investigated character-level language models, without specifying a specific context before generation. We leverage existing NMT tools to encode a specific context to the restaurant before generating reviews.
Supporting our study, Everett et al~\cite{Everett2016Automated} found that security researchers were less likely to be fooled by Markov chain-generated Reddit comments compared to ordinary Internet users.
Diversification of NMT model outputs has been studied in \cite{li2016diversity}. The authors proposed the use of a penalty to commonly occurring sentences (\emph{n-grams}) in order to emphasize maximum mutual information-based generation.
The authors investigated the use of NMT models in chatbot systems.
We found that unigram penalties to random tokens (Algorithm~\ref{alg:aug}) was easy to implement and produced sufficiently diverse responses.
\section {Discussion and Future Work}
\paragraph{What makes NMT-Fake* reviews difficult to detect?} First, NMT models allow the encoding of a relevant context for each review, which narrows down the possible choices of words that the model has to choose from. Our NMT model had a perplexity of approximately $25$, while the model of \cite{yao2017automated} had a perplexity of approximately $90$ \footnote{Personal communication with the authors}. Second, the beam search in NMT models narrows down choices to natural-looking sentences. Third, we observed that the NMT model produced \emph{better structure} in the generated sentences (i.e. a more coherent story).
\paragraph{Cost of generating reviews} With our setup, generating one review took less than one second. The cost of generation stems mainly from the overnight training. Assuming an electricity cost of 16 cents / kWh (California) and 8 hours of training, training the NMT model requires approximately 1.30 USD. This is a 90\% reduction in time compared to the state-of-the-art \cite{yao2017automated}. Furthermore, it is possible to generate both positive and negative reviews with the same model.
\paragraph{Ease of customization} We experimented with inserting specific words into the text by increasing their log likelihoods in the beam search. We noticed that the success depended on the prevalence of the word in the training set. For example, adding a +5 to \emph{Mike} in the log-likelihood resulted in approximately 10\% prevalence of this word in the reviews. An attacker can therefore easily insert specific keywords to reviews, which can increase evasion probability.
\paragraph{Ease of testing} Our diversification scheme is applicable during \emph{generation phase}, and does not affect the training setup of the network in any way. Once the NMT model is obtained, it is easy to obtain several different variants of NMT-Fake reviews by varying parameters $b$ and $\lambda$.
\paragraph{Languages} The generation methodology is not per-se language-dependent. The requirement for successful generation is that sufficiently much data exists in the targeted language. However, our language model modifications require some knowledge of that target language's grammar to produce high-quality reviews.
\paragraph{Generalizability of detection techniques} Currently, fake reviews are not universally detectable. Our results highlight that it is difficult to claim detection performance on unseen types of fake reviews (Section~\ref{sec:automated}). We see this an open problem that deserves more attention in fake reviews research.
\paragraph{Generalizability to other types of datasets} Our technique can be applied to any dataset, as long as there is sufficient training data for the NMT model. We used approximately 2.9 million reviews for this work.
\section{Conclusion}
In this paper, we showed that neural machine translation models can be used to generate fake reviews that are very effective in deceiving even experienced, tech-savvy users.
This supports anecdotal evidence \cite{national2017commission}.
Our technique is more effective than state-of-the-art \cite{yao2017automated}.
We conclude that machine-aided fake review detection is necessary since human users are ineffective in identifying fake reviews.
We also showed that detectors trained using one type of fake reviews are not effective in identifying other types of fake reviews.
Robust detection of fake reviews is thus still an open problem.
\section*{Acknowledgments}
We thank Tommi Gr\"{o}ndahl for assistance in planning user studies and the
participants of the user study for their time and feedback. We also thank
Luiza Sayfullina for comments that improved the manuscript.
We thank the authors of \cite{yao2017automated} for answering questions about
their work.
\bibliographystyle{splncs}
\begin{thebibliography}{10}
\bibitem{yao2017automated}
Yao, Y., Viswanath, B., Cryan, J., Zheng, H., Zhao, B.Y.:
\newblock Automated crowdturfing attacks and defenses in online review systems.
\newblock In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and
Communications Security, ACM (2017)
\bibitem{murphy2012machine}
Murphy, K.:
\newblock Machine learning: a probabilistic approach.
\newblock Massachusetts Institute of Technology (2012)
\bibitem{challenge2013yelp}
Yelp:
\newblock {Yelp Challenge Dataset} (2013)
\bibitem{mukherjee2013yelp}
Mukherjee, A., Venkataraman, V., Liu, B., Glance, N.:
\newblock What yelp fake review filter might be doing?
\newblock In: Seventh International AAAI Conference on Weblogs and Social Media
(ICWSM). (2013)
\bibitem{rayana2015collective}
Rayana, S., Akoglu, L.:
\newblock Collective opinion spam detection: Bridging review networks and
metadata.
\newblock In: {}Proceedings of the 21th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining
\bibitem{o2008user}
{O'Connor}, P.:
\newblock {User-generated content and travel: A case study on Tripadvisor.com}.
\newblock Information and communication technologies in tourism 2008 (2008)
\bibitem{luca2010reviews}
Luca, M.:
\newblock {Reviews, Reputation, and Revenue: The Case of Yelp. com}.
\newblock {Harvard Business School} (2010)
\bibitem{wang2012serf}
Wang, G., Wilson, C., Zhao, X., Zhu, Y., Mohanlal, M., Zheng, H., Zhao, B.Y.:
\newblock Serf and turf: crowdturfing for fun and profit.
\newblock In: Proceedings of the 21st international conference on World Wide
Web (WWW), ACM (2012)
\bibitem{rinta2017understanding}
Rinta-Kahila, T., Soliman, W.:
\newblock Understanding crowdturfing: The different ethical logics behind the
clandestine industry of deception.
\newblock In: ECIS 2017: Proceedings of the 25th European Conference on
Information Systems. (2017)
\bibitem{luca2016fake}
Luca, M., Zervas, G.:
\newblock Fake it till you make it: Reputation, competition, and yelp review
fraud.
\newblock Management Science (2016)
\bibitem{national2017commission}
{National Literacy Trust}:
\newblock Commission on fake news and the teaching of critical literacy skills
in schools URL:
\url{https://literacytrust.org.uk/policy-and-campaigns/all-party-parliamentary-group-literacy/fakenews/}.
\bibitem{jurafsky2014speech}
Jurafsky, D., Martin, J.H.:
\newblock Speech and language processing. Volume~3.
\newblock Pearson London: (2014)
\bibitem{kingma2014adam}
Kingma, D.P., Ba, J.:
\newblock Adam: A method for stochastic optimization.
\newblock arXiv preprint arXiv:1412.6980 (2014)
\bibitem{cho2014learning}
Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F.,
Schwenk, H., Bengio, Y.:
\newblock Learning phrase representations using rnn encoder--decoder for
statistical machine translation.
\newblock In: Proceedings of the 2014 Conference on Empirical Methods in
Natural Language Processing (EMNLP). (2014)
\bibitem{klein2017opennmt}
Klein, G., Kim, Y., Deng, Y., Senellart, J., Rush, A.:
\newblock Opennmt: Open-source toolkit for neural machine translation.
\newblock Proceedings of ACL, System Demonstrations (2017)
\bibitem{wu2016google}
Wu, Y., Schuster, M., Chen, Z., Le, Q.V., Norouzi, M., Macherey, W., Krikun,
M., Cao, Y., Gao, Q., Macherey, K., et~al.:
\newblock Google's neural machine translation system: Bridging the gap between
human and machine translation.
\newblock arXiv preprint arXiv:1609.08144 (2016)
\bibitem{mei2017coherent}
Mei, H., Bansal, M., Walter, M.R.:
\newblock Coherent dialogue with attention-based language models.
\newblock In: AAAI. (2017) 3252--3258
\bibitem{li2016diversity}
Li, J., Galley, M., Brockett, C., Gao, J., Dolan, B.:
\newblock A diversity-promoting objective function for neural conversation
models.
\newblock In: Proceedings of NAACL-HLT. (2016)
\bibitem{rubin2006assessing}
Rubin, V.L., Liddy, E.D.:
\newblock Assessing credibility of weblogs.
\newblock In: AAAI Spring Symposium: Computational Approaches to Analyzing
Weblogs. (2006)
\bibitem{zhao2017news}
news.com.au:
\newblock {The potential of AI generated 'crowdturfing' could undermine online
reviews and dramatically erode public trust} URL:
\url{http://www.news.com.au/technology/online/security/the-potential-of-ai-generated-crowdturfing-could-undermine-online-reviews-and-dramatically-erode-public-trust/news-story/e1c84ad909b586f8a08238d5f80b6982}.
\bibitem{pennebaker2015development}
Pennebaker, J.W., Boyd, R.L., Jordan, K., Blackburn, K.:
\newblock {The development and psychometric properties of LIWC2015}.
\newblock Technical report (2015)
\bibitem{honnibal-johnson:2015:EMNLP}
Honnibal, M., Johnson, M.:
\newblock An improved non-monotonic transition system for dependency parsing.
\newblock In: Proceedings of the 2015 Conference on Empirical Methods in
Natural Language Processing (EMNLP), ACM (2015)
\bibitem{bird2004nltk}
Bird, S., Loper, E.:
\newblock {NLTK: the natural language toolkit}.
\newblock In: Proceedings of the ACL 2004 on Interactive poster and
demonstration sessions, Association for Computational Linguistics (2004)
\bibitem{kumar2018false}
Kumar, S., Shah, N.:
\newblock False information on web and social media: A survey.
\newblock arXiv preprint arXiv:1804.08559 (2018)
\bibitem{Everett2016Automated}
Everett, R.M., Nurse, J.R.C., Erola, A.:
\newblock The anatomy of online deception: What makes automated text
convincing?
\newblock In: Proceedings of the 31st Annual ACM Symposium on Applied
Computing. SAC '16, ACM (2016)
\end{thebibliography}
\section*{Appendix}
We present basic demographics of our MTurk study and the comparative study with experienced users in Table~\ref{table:amt_pop}.
\begin{table}
\caption{User study statistics.}
\begin{center}
\begin{tabular}{ | l | c | c | }
\hline
Quality & Mechanical Turk users & Experienced users\\
\hline
Native English Speaker & Yes (20) & Yes (1) No (19) \\
Fluent in English & Yes (20) & Yes (20) \\
Age & 21-40 (17) 41-60 (3) & 21-25 (8) 26-30 (7) 31-35 (4) 41-45 (1)\\
Gender & Male (14) Female (6) & Male (17) Female (3)\\
Highest Education & High School (10) Bachelor (10) & Bachelor (9) Master (6) Ph.D. (5) \\
\hline
\end{tabular}
\label{table:amt_pop}
\end{center}
\end{table}
Table~\ref{table:openNMT-py_commands} shows a listing of the openNMT-py commands we used to create our NMT model and to generate fake reviews.
\begin{table}[t]
\caption{Listing of used openNMT-py commands.}
\begin{center}
\begin{tabular}{ | l | l | }
\hline
Phase & Bash command \\
\hline
Preprocessing & \begin{lstlisting}[language=bash]
python preprocess.py -train_src context-train.txt
-train_tgt reviews-train.txt -valid_src context-val.txt
-valid_tgt reviews-val.txt -save_data model
-lower -tgt_words_min_frequency 10
\end{lstlisting}
\\ & \\
Training & \begin{lstlisting}[language=bash]
python train.py -data model -save_model model -epochs 8
-gpuid 0 -learning_rate_decay 0.5 -optim adam
-learning_rate 0.001 -start_decay_at 3\end{lstlisting}
\\ & \\
Generation & \begin{lstlisting}[language=bash]
python translate.py -model model_acc_35.54_ppl_25.68_e8.pt
-src context-tst.txt -output pred-e8.txt -replace_unk
-verbose -max_length 50 -gpu 0
\end{lstlisting} \\
\hline
\end{tabular}
\label{table:openNMT-py_commands}
\end{center}
\end{table}
Table~\ref{table:MTurk_sub} shows the classification performance of Amazon Mechanical Turkers, separated across different categories of NMT-Fake reviews. The category with best performance ($b=0.3, \lambda=-5$) is denoted as NMT-Fake*.
\begin{table}[b]
\caption{MTurk study subclass classification reports. Classes are imbalanced in ratio 1:6. Random predictions are $p_\mathrm{human} = 86\%$ and $p_\mathrm{machine} = 14\%$, with $r_\mathrm{human} = r_\mathrm{machine} = 50\%$. Class-averaged F-scores for random predictions are $42\%$.}
\begin{center}
\begin{tabular}{ | c || c |c |c | c | }
\hline
$(b=0.3, \lambda = -3)$ & Precision & Recall & F-score & Support \\ \hline
Human & 89\% & 63\% & 73\% & 994\\
NMT-Fake & 15\% & 45\% & 22\% & 146 \\
\hline
\hline
$(b=0.3, \lambda = -5)$ & Precision & Recall & F-score & Support \\ \hline
Human & 86\% & 63\% & 73\% & 994\\
NMT-Fake* & 16\% & 40\% & 23\% & 171 \\
\hline
\hline
$(b=0.5, \lambda = -4)$ & Precision & Recall & F-score & Support \\ \hline
Human & 88\% & 63\% & 73\% & 994\\
NMT-Fake & 21\% & 55\% & 30\% & 181 \\
\hline
\hline
$(b=0.7, \lambda = -3)$ & Precision & Recall & F-score & Support \\ \hline
Human & 88\% & 63\% & 73\% & 994\\
NMT-Fake & 19\% & 50\% & 27\% & 170 \\
\hline
\hline
$(b=0.7, \lambda = -5)$ & Precision & Recall & F-score & Support \\ \hline
Human & 89\% & 63\% & 74\% & 994\\
NMT-Fake & 21\% & 57\% & 31\% & 174 \\
\hline
\hline
$(b=0.9, \lambda = -4)$ & Precision & Recall & F-score & Support \\ \hline
Human & 88\% & 63\% & 73\% & 994\\
NMT-Fake & 18\% & 50\% & 27\% & 164 \\
\hline
\end{tabular}
\label{table:MTurk_sub}
\end{center}
\end{table}
Figure~\ref{fig:screenshot} shows screenshots of the first two pages of our user study with experienced participants.
\begin{figure}[ht]
\centering
\includegraphics[width=1.\columnwidth]{figures/screenshot_7-3.png}
\caption{
Screenshots of the first two pages in the user study. Example 1 is a NMT-Fake* review, the rest are human-written.
}
\label{fig:screenshot}
\end{figure}
Table~\ref{table:features_adaboost} shows the features used to detect NMT-Fake reviews using the AdaBoost classifier.
\begin{table}
\caption{Features used in NMT-Fake review detector.}
\begin{center}
\begin{tabular}{ | l | c | }
\hline
Feature type & Number of features \\ \hline
\hline
Readability features & 13 \\ \hline
Unique POS tags & $~20$ \\ \hline
Word unigrams & 22,831 \\ \hline
1/2/3/4-grams of simple part-of-speech tags & 54,240 \\ \hline
1/2/3-grams of detailed part-of-speech tags & 112,944 \\ \hline
1/2/3-grams of syntactic dependency tags & 93,195 \\ \hline
\end{tabular}
\label{table:features_adaboost}
\end{center}
\end{table}
\end{document}
| 5
|
1805.10824
|
UG18 at SemEval-2018 Task 1: Generating Additional Training Data for Predicting Emotion Intensity in Spanish
|
# UG18 at SemEval-2018 Task 1: Generating Additional Training Data for Predicting Emotion Intensity in Spanish
## Abstract
The present study describes our submission to SemEval 2018 Task 1: Affect in Tweets. Our Spanish-only approach aimed to demonstrate that it is beneficial to automatically generate additional training data by (i) translating training data from other languages and (ii) applying a semi-supervised learning method. We find strong support for both approaches, with those models outperforming our regular models in all subtasks. However, creating a stepwise ensemble of different models as opposed to simply averaging did not result in an increase in performance. We placed second (EI-Reg), second (EI-Oc), fourth (V-Reg) and fifth (V-Oc) in the four Spanish subtasks we participated in.
## Introduction
Understanding the emotions expressed in a text or message is of high relevance nowadays. Companies are interested in this to get an understanding of the sentiment of their current customers regarding their products and the sentiment of their potential customers to attract new ones. Moreover, changes in a product or a company may also affect the sentiment of a customer. However, the intensity of an emotion is crucial in determining the urgency and importance of that sentiment. If someone is only slightly happy about a product, is a customer willing to buy it again? Conversely, if someone is very angry about customer service, his or her complaint might be given priority over somewhat milder complaints.
BIBREF0 present four tasks in which systems have to automatically determine the intensity of emotions (EI) or the intensity of the sentiment (Valence) of tweets in the languages English, Arabic, and Spanish. The goal is to either predict a continuous regression (reg) value or to do ordinal classification (oc) based on a number of predefined categories. The EI tasks have separate training sets for four different emotions: anger, fear, joy and sadness. Due to the large number of subtasks and the fact that this language does not have many resources readily available, we only focus on the Spanish subtasks. Our work makes the following contributions:
Our submissions ranked second (EI-Reg), second (EI-Oc), fourth (V-Reg) and fifth (V-Oc), demonstrating that the proposed method is accurate in automatically determining the intensity of emotions and sentiment of Spanish tweets. This paper will first focus on the datasets, the data generation procedure, and the techniques and tools used. Then we present the results in detail, after which we perform a small error analysis on the largest mistakes our model made. We conclude with some possible ideas for future work.
## Data
For each task, the training data that was made available by the organizers is used, which is a selection of tweets with for each tweet a label describing the intensity of the emotion or sentiment BIBREF1 . Links and usernames were replaced by the general tokens URL and @username, after which the tweets were tokenized by using TweetTokenizer. All text was lowercased. In a post-processing step, it was ensured that each emoji is tokenized as a single token.
## Word Embeddings
To be able to train word embeddings, Spanish tweets were scraped between November 8, 2017 and January 12, 2018. We chose to create our own embeddings instead of using pre-trained embeddings, because this way the embeddings would resemble the provided data set: both are based on Twitter data. Added to this set was the Affect in Tweets Distant Supervision Corpus (DISC) made available by the organizers BIBREF0 and a set of 4.1 million tweets from 2015, obtained from BIBREF2 . After removing duplicate tweets and tweets with fewer than ten tokens, this resulted in a set of 58.7 million tweets, containing 1.1 billion tokens. The tweets were preprocessed using the method described in Section SECREF6 . The word embeddings were created using word2vec in the gensim library BIBREF3 , using CBOW, a window size of 40 and a minimum count of 5. The feature vectors for each tweet were then created by using the AffectiveTweets WEKA package BIBREF4 .
## Translating Lexicons
Most lexical resources for sentiment analysis are in English. To still be able to benefit from these sources, the lexicons in the AffectiveTweets package were translated to Spanish, using the machine translation platform Apertium BIBREF5 .
All lexicons from the AffectiveTweets package were translated, except for SentiStrength. Instead of translating this lexicon, the English version was replaced by the Spanish variant made available by BIBREF6 .
For each subtask, the optimal combination of lexicons was determined. This was done by first calculating the benefits of adding each lexicon individually, after which only beneficial lexicons were added until the score did not increase anymore (e.g. after adding the best four lexicons the fifth one did not help anymore, so only four were added). The tests were performed using a default SVM model, with the set of word embeddings described in the previous section. Each subtask thus uses a different set of lexicons (see Table TABREF1 for an overview of the lexicons used in our final ensemble). For each subtask, this resulted in a (modest) increase on the development set, between 0.01 and 0.05.
## Translating Data
The training set provided by BIBREF0 is not very large, so it was interesting to find a way to augment the training set. A possible method is to simply translate the datasets into other languages, leaving the labels intact. Since the present study focuses on Spanish tweets, all tweets from the English datasets were translated into Spanish. This new set of “Spanish” data was then added to our original training set. Again, the machine translation platform Apertium BIBREF5 was used for the translation of the datasets.
## Algorithms Used
Three types of models were used in our system, a feed-forward neural network, an LSTM network and an SVM regressor. The neural nets were inspired by the work of Prayas BIBREF7 in the previous shared task. Different regression algorithms (e.g. AdaBoost, XGBoost) were also tried due to the success of SeerNet BIBREF8 , but our study was not able to reproduce their results for Spanish.
For both the LSTM network and the feed-forward network, a parameter search was done for the number of layers, the number of nodes and dropout used. This was done for each subtask, i.e. different tasks can have a different number of layers. All models were implemented using Keras BIBREF9 . After the best parameter settings were found, the results of 10 system runs to produce our predictions were averaged (note that this is different from averaging our different type of models in Section SECREF16 ). For the SVM (implemented in scikit-learn BIBREF10 ), the RBF kernel was used and a parameter search was conducted for epsilon. Detailed parameter settings for each subtask are shown in Table TABREF12 . Each parameter search was performed using 10-fold cross validation, as to not overfit on the development set.
## Semi-supervised Learning
One of the aims of this study was to see if using semi-supervised learning is beneficial for emotion intensity tasks. For this purpose, the DISC BIBREF0 corpus was used. This corpus was created by querying certain emotion-related words, which makes it very suitable as a semi-supervised corpus. However, the specific emotion the tweet belonged to was not made public. Therefore, a method was applied to automatically assign the tweets to an emotion by comparing our scraped tweets to this new data set.
First, in an attempt to obtain the query-terms, we selected the 100 words which occurred most frequently in the DISC corpus, in comparison with their frequencies in our own scraped tweets corpus. Words that were clearly not indicators of emotion were removed. The rest was annotated per emotion or removed if it was unclear to which emotion the word belonged. This allowed us to create silver datasets per emotion, assigning tweets to an emotion if an annotated emotion-word occurred in the tweet.
Our semi-supervised approach is quite straightforward: first a model is trained on the training set and then this model is used to predict the labels of the silver data. This silver data is then simply added to our training set, after which the model is retrained. However, an extra step is applied to ensure that the silver data is of reasonable quality. Instead of training a single model initially, ten different models were trained which predict the labels of the silver instances. If the highest and lowest prediction do not differ more than a certain threshold the silver instance is maintained, otherwise it is discarded.
This results in two parameters that could be optimized: the threshold and the number of silver instances that would be added. This method can be applied to both the LSTM and feed-forward networks that were used. An overview of the characteristics of our data set with the final parameter settings is shown in Table TABREF14 . Usually, only a small subset of data was added to our training set, meaning that most of the silver data is not used in the experiments. Note that since only the emotions were annotated, this method is only applicable to the EI tasks.
## Ensembling
To boost performance, the SVM, LSTM, and feed-forward models were combined into an ensemble. For both the LSTM and feed-forward approach, three different models were trained. The first model was trained on the training data (regular), the second model was trained on both the training and translated training data (translated) and the third one was trained on both the training data and the semi-supervised data (silver). Due to the nature of the SVM algorithm, semi-supervised learning does not help, so only the regular and translated model were trained in this case. This results in 8 different models per subtask. Note that for the valence tasks no silver training data was obtained, meaning that for those tasks the semi-supervised models could not be used.
Per task, the LSTM and feed-forward model's predictions were averaged over 10 prediction runs. Subsequently, the predictions of all individual models were combined into an average. Finally, models were removed from the ensemble in a stepwise manner if the removal increased the average score. This was done based on their original scores, i.e. starting out by trying to remove the worst individual model and working our way up to the best model. We only consider it an increase in score if the difference is larger than 0.002 (i.e. the difference between 0.716 and 0.718). If at some point the score does not increase and we are therefore unable to remove a model, the process is stopped and our best ensemble of models has been found. This process uses the scores on the development set of different combinations of models. Note that this means that the ensembles for different subtasks can contain different sets of models. The final model selections can be found in Table TABREF17 .
## Results and Discussion
Table TABREF18 shows the results on the development set of all individuals models, distinguishing the three types of training: regular (r), translated (t) and semi-supervised (s). In Tables TABREF17 and TABREF18 , the letter behind each model (e.g. SVM-r, LSTM-r) corresponds to the type of training used. Comparing the regular and translated columns for the three algorithms, it shows that in 22 out of 30 cases, using translated instances as extra training data resulted in an improvement. For the semi-supervised learning approach, an improvement is found in 15 out of 16 cases. Moreover, our best individual model for each subtask (bolded scores in Table TABREF18 ) is always either a translated or semi-supervised model. Table TABREF18 also shows that, in general, our feed-forward network obtained the best results, having the highest F-score for 8 out of 10 subtasks.
However, Table TABREF19 shows that these scores can still be improved by averaging or ensembling the individual models. On the dev set, averaging our 8 individual models results in a better score for 8 out of 10 subtasks, while creating an ensemble beats all of the individual models as well as the average for each subtask. On the test set, however, only a small increase in score (if any) is found for stepwise ensembling, compared to averaging. Even though the results do not get worse, we cannot conclude that stepwise ensembling is a better method than simply averaging.
Our official scores (column Ens Test in Table TABREF19 ) have placed us second (EI-Reg, EI-Oc), fourth (V-Reg) and fifth (V-Oc) on the SemEval AIT-2018 leaderboard. However, it is evident that the results obtained on the test set are not always in line with those achieved on the development set. Especially on the anger subtask for both EI-Reg and EI-Oc, the scores are considerably lower on the test set in comparison with the results on the development set. Therefore, a small error analysis was performed on the instances where our final model made the largest errors.
## Error Analysis
Due to some large differences between our results on the dev and test set of this task, we performed a small error analysis in order to see what caused these differences. For EI-Reg-anger, the gold labels were compared to our own predictions, and we manually checked 50 instances for which our system made the largest errors.
Some examples that were indicative of the shortcomings of our system are shown in Table TABREF20 . First of all, our system did not take into account capitalization. The implications of this are shown in the first sentence, where capitalization intensifies the emotion used in the sentence. In the second sentence, the name Imperator Furiosa is not understood. Since our texts were lowercased, our system was unable to capture the named entity and thought the sentence was about an angry emperor instead. In the third sentence, our system fails to capture that when you are so angry that it makes you laugh, it results in a reduced intensity of the angriness. Finally, in the fourth sentence, it is the figurative language me infla la vena (it inflates my vein) that the system is not able to understand.
The first two error-categories might be solved by including smart features regarding capitalization and named entity recognition. However, the last two categories are problems of natural language understanding and will be very difficult to fix.
## Conclusion
To conclude, the present study described our submission for the Semeval 2018 Shared Task on Affect in Tweets. We participated in four Spanish subtasks and our submissions ranked second, second, fourth and fifth place. Our study aimed to investigate whether the automatic generation of additional training data through translation and semi-supervised learning, as well as the creation of stepwise ensembles, increase the performance of our Spanish-language models. Strong support was found for the translation and semi-supervised learning approaches; our best models for all subtasks use either one of these approaches. These results suggest that both of these additional data resources are beneficial when determining emotion intensity (for Spanish). However, the creation of a stepwise ensemble from the best models did not result in better performance compared to simply averaging the models. In addition, some signs of overfitting on the dev set were found. In future work, we would like to apply the methods (translation and semi-supervised learning) used on Spanish on other low-resource languages and potentially also on other tasks.
| 11
|
1806.00722
|
Dense Information Flow for Neural Machine Translation
|
# Dense Information Flow for Neural Machine Translation
## Abstract
Recently, neural machine translation has achieved remarkable progress by introducing well-designed deep neural networks into its encoder-decoder framework. From the optimization perspective, residual connections are adopted to improve learning performance for both encoder and decoder in most of these deep architectures, and advanced attention connections are applied as well. Inspired by the success of the DenseNet model in computer vision problems, in this paper, we propose a densely connected NMT architecture (DenseNMT) that is able to train more efficiently for NMT. The proposed DenseNMT not only allows dense connection in creating new features for both encoder and decoder, but also uses the dense attention structure to improve attention quality. Our experiments on multiple datasets show that DenseNMT structure is more competitive and efficient.
## Introduction
Neural machine translation (NMT) is a challenging task that attracts lots of attention in recent years. Starting from the encoder-decoder framework BIBREF0 , NMT starts to show promising results in many language pairs. The evolving structures of NMT models in recent years have made them achieve higher scores and become more favorable. The attention mechanism BIBREF1 added on top of encoder-decoder framework is shown to be very useful to automatically find alignment structure, and single-layer RNN-based structure has evolved into deeper models with more efficient transformation functions BIBREF2 , BIBREF3 , BIBREF4 .
One major challenge of NMT is that its models are hard to train in general due to the complexity of both the deep models and languages. From the optimization perspective, deeper models are hard to efficiently back-propagate the gradients, and this phenomenon as well as its solution is better explored in the computer vision society. Residual networks (ResNet) BIBREF5 achieve great performance in a wide range of tasks, including image classification and image segmentation. Residual connections allow features from previous layers to be accumulated to the next layer easily, and make the optimization of the model efficiently focus on refining upper layer features.
NMT is considered as a challenging problem due to its sequence-to-sequence generation framework, and the goal of comprehension and reorganizing from one language to the other. Apart from the encoder block that works as a feature generator, the decoder network combining with the attention mechanism bring new challenges to the optimization of the models. While nowadays best-performing NMT systems use residual connections, we question whether this is the most efficient way to propagate information through deep models. In this paper, inspired by the idea of using dense connections for training computer vision tasks BIBREF6 , we propose a densely connected NMT framework (DenseNMT) that efficiently propagates information from the encoder to the decoder through the attention component. Taking the CNN-based deep architecture as an example, we verify the efficiency of DenseNMT. Our contributions in this work include: (i) by comparing the loss curve, we show that DenseNMT allows the model to pass information more efficiently, and speeds up training; (ii) we show through ablation study that dense connections in all three blocks altogether help improve the performance, while not increasing the number of parameters; (iii) DenseNMT allows the models to achieve similar performance with much smaller embedding size; (iv) DenseNMT on IWSLT14 German-English and Turkish-English translation tasks achieves new benchmark BLEU scores, and the result on WMT14 English-German task is more competitive than the residual connections based baseline model.
## DenseNMT
In this section, we introduce our DenseNMT architecture. In general, compared with residual connected NMT models, DenseNMT allows each layer to provide its information to all subsequent layers directly. Figure FIGREF9 - FIGREF15 show the design of our model structure by parts.
We start with the formulation of a regular NMT model. Given a set of sentence pairs INLINEFORM0 , an NMT model learns parameter INLINEFORM1 by maximizing the log-likelihood function: DISPLAYFORM0
For every sentence pair INLINEFORM0 , INLINEFORM1 is calculated based on the decomposition: DISPLAYFORM0
where INLINEFORM0 is the length of sentence INLINEFORM1 . Typically, NMT models use the encoder-attention-decoder framework BIBREF1 , and potentially use multi-layer structure for both encoder and decoder. Given a source sentence INLINEFORM2 with length INLINEFORM3 , the encoder calculates hidden representations by layer. We denote the representation in the INLINEFORM4 -th layer as INLINEFORM5 , with dimension INLINEFORM6 , where INLINEFORM7 is the dimension of features in layer INLINEFORM8 . The hidden representation at each position INLINEFORM9 is either calculated by: DISPLAYFORM0
for recurrent transformation INLINEFORM0 such as LSTM and GRU, or by: DISPLAYFORM0
for parallel transformation INLINEFORM0 . On the other hand, the decoder layers INLINEFORM1 follow similar structure, while getting extra representations from the encoder side. These extra representations are also called attention, and are especially useful for capturing alignment information.
In our experiments, we use convolution based transformation for INLINEFORM0 due to both its efficiency and high performance, more formally, DISPLAYFORM0
INLINEFORM0 is the gated linear unit proposed in BIBREF11 and the kernel size is INLINEFORM1 . DenseNMT is agnostic to the transformation function, and we expect it to also work well combining with other transformations, such as LSTM, self-attention and depthwise separable convolution.
## Dense encoder and decoder
Different from residual connections, later layers in the dense encoder are able to use features from all previous layers by concatenating them: DISPLAYFORM0
Here, INLINEFORM0 is defined in Eq. ( EQREF10 ), INLINEFORM1 represents concatenation operation. Although this brings extra connections to the network, with smaller number of features per layer, the architecture encourages feature reuse, and can be more compact and expressive. As shown in Figure FIGREF9 , when designing the model, the hidden size in each layer is much smaller than the hidden size of the corresponding layer in the residual-connected model.
While each encoder layer perceives information from its previous layers, each decoder layer INLINEFORM0 has two information sources: previous layers INLINEFORM1 , and attention values INLINEFORM2 . Therefore, in order to allow dense information flow, we redefine the generation of INLINEFORM3 -th layer as a nonlinear function over all its previous decoder layers and previous attentions. This can be written as: DISPLAYFORM0
where INLINEFORM0 is the attention value using INLINEFORM1 -th decoder layer and information from encoder side, which will be specified later. Figure FIGREF13 shows the comparison of a dense decoder with a regular residual decoder. The dimensions of both attention values and hidden layers are chosen with smaller values, yet the perceived information for each layer consists of a higher dimension vector with more representation power. The output of the decoder is a linear transformation of the concatenation of all layers by default. To compromise to the increment of dimensions, we use summary layers, which will be introduced in Section 3.3. With summary layers, the output of the decoder is only a linear transformation of the concatenation of the upper few layers.
## Dense attention
Prior works show a trend of designing more expressive attention mechanisms (as discussed in Section 2). However, most of them only use the last encoder layer. In order to pass more abundant information from the encoder side to the decoder side, the attention block needs to be more expressive. Following the recent development of designing attention architectures, we propose DenseAtt as the dense attention block, which serves for the dense connection between the encoder and the decoder side. More specifically, two options are proposed accordingly. For each decoding step in the corresponding decoder layer, the two options both calculate attention using multiple encoder layers. The first option is more compressed, while the second option is more expressive and flexible. We name them as DenseAtt-1 and DenseAtt-2 respectively. Figure FIGREF15 shows the architecture of (a) multi-step attention BIBREF2 , (b) DenseAtt-1, and (c) DenseAtt-2 in order. In general, a popular multiplicative attention module can be written as: DISPLAYFORM0
where INLINEFORM0 represent query, key, value respectively. We will use this function INLINEFORM1 in the following descriptions.
In the decoding phase, we use a layer-wise attention mechanism, such that each decoder layer absorbs different attention information to adjust its output. Instead of treating the last hidden layer as the encoder's output, we treat the concatenation of all hidden layers from encoder side as the output. The decoder layer multiplies with the encoder output to obtain the attention weights, which is then multiplied by a linear combination of the encoder output and the sentence embedding. The attention output of each layer INLINEFORM0 can be formally written as: DISPLAYFORM0
where INLINEFORM0 is the multiplicative attention function, INLINEFORM1 is a concatenation operation that combines all features, and INLINEFORM2 is a linear transformation function that maps each variable to a fixed dimension in order to calculate the attention value. Notice that we explicitly write the INLINEFORM3 term in ( EQREF19 ) to keep consistent with the multi-step attention mechanism, as pictorially shown in Figure FIGREF15 (a).
Notice that the transformation INLINEFORM0 in DenseAtt-1 forces the encoder layers to be mixed before doing attention. Since we use multiple hidden layers from the encoder side to get an attention value, we can alternatively calculate multiple attention values before concatenating them. In another word, the decoder layer can get different attention values from different encoder layers. This can be formally expressed as: DISPLAYFORM0
where the only difference from Eq. ( EQREF19 ) is that the concatenation operation is substituted by a summation operation, and is put after the attention function INLINEFORM0 . This method further increases the representation power in the attention block, while maintaining the same number of parameters in the model.
## Summary layers
Since the number of features fed into nonlinear operation is accumulated along the path, the parameter size increases accordingly. For example, for the INLINEFORM0 -th encoder layer, the input dimension of features is INLINEFORM1 , where INLINEFORM2 is the feature dimension in previous layers, INLINEFORM3 is the embedding size. In order to avoid the calculation bottleneck for later layers due to large INLINEFORM4 , we introduce the summary layer for deeper models. It summarizes the features for all previous layers and projects back to the embedding size, so that later layers of both the encoder and the decoder side do not need to look back further. The summary layers can be considered as contextualized word vectors in a given sentence BIBREF12 . We add one summary layer after every INLINEFORM5 layers, where INLINEFORM6 is the hyperparameter we introduce. Accordingly, the input dimension of features is at most INLINEFORM7 for the last layer of the encoder. Moreover, combined with the summary layer setting, our DenseAtt mechanism allows each decoder layer to calculate the attention value focusing on the last few encoder layers, which consists of the last contextual embedding layer and several dense connected layers with low dimension. In practice, we set INLINEFORM8 as 5 or 6.
## Analysis of information flow
Figure FIGREF9 and Figure FIGREF13 show the difference of information flow compared with a residual-based encoder/decoder. For residual-based models, each layer can absorb a single high-dimensional vector from its previous layer as the only information, while for DenseNMT, each layer can utilize several low-dimensional vectors from its previous layers and a high-dimensional vector from the first layer (embedding layer) as its information. In DenseNMT, each layer directly provides information to its later layers. Therefore, the structure allows feature reuse, and encourages upper layers to focus on creating new features. Furthermore, the attention block allows the embedding vectors (as well as other hidden layers) to guide the decoder's generation more directly; therefore, during back-propagation, the gradient information can be passed directly to all encoder layers simultaneously.
## Datasets
We use three datasets for our experiments: IWSLT14 German-English, Turkish-English, and WMT14 English-German.
We preprocess the IWSLT14 German-English dataset following byte-pair-encoding (BPE) method BIBREF13 . We learn 25k BPE codes using the joint corpus of source and target languages. We randomly select 7k from IWSLT14 German-English as the development set , and the test set is a concatenation of dev2010, tst2010, tst2011 and tst2012, which is widely used in prior works BIBREF14 , BIBREF15 , BIBREF16 .
For the Turkish-English translation task, we use the data provided by IWSLT14 BIBREF17 and the SETimes corpus BIBREF17 following BIBREF18 . After removing sentence pairs with length ratio over 9, we obtain 360k sentence pairs. Since there is little commonality between the two languages, we learn 30k size BPE codes separately for Turkish and English. In addition to this, we give another preprocessing for Turkish sentences and use word-level English corpus. For Turkish sentences, following BIBREF19 , BIBREF18 , we use the morphology tool Zemberek with disambiguation by the morphological analysis BIBREF20 and removal of non-surface tokens. Following BIBREF18 , we concatenate tst2011, tst2012, tst2013, tst2014 as our test set. We concatenate dev2010 and tst2010 as the development set.
We preprocess the WMT14 English-German dataset using a BPE code size of 40k. We use the concatenation of newstest2013 and newstest2012 as the development set.
## Model and architect design
As the baseline model (BASE-4L) for IWSLT14 German-English and Turkish-English, we use a 4-layer encoder, 4-layer decoder, residual-connected model, with embedding and hidden size set as 256 by default. As a comparison, we design a densely connected model with same number of layers, but the hidden size is set as 128 in order to keep the model size consistent. The models adopting DenseAtt-1, DenseAtt-2 are named as DenseNMT-4L-1 and DenseNMT-4L-2 respectively. In order to check the effect of dense connections on deeper models, we also construct a series of 8-layer models. We set the hidden number to be 192, such that both 4-layer models and 8-layer models have similar number of parameters. For dense structured models, we set the dimension of hidden states to be 96.
Since NMT model usually allocates a large proportion of its parameters to the source/target sentence embedding and softmax matrix, we explore in our experiments to what extent decreasing the dimensions of the three parts would harm the BLEU score. We change the dimensions of the source embedding, the target embedding as well as the softmax matrix simultaneously to smaller values, and then project each word back to the original embedding dimension through a linear transformation. This significantly reduces the number of total parameters, while not influencing the upper layer structure of the model.
We also introduce three additional models we use for ablation study, all using 4-layer structure. Based on the residual connected BASE-4L model, (1) DenseENC-4L only makes encoder side dense, (2) DenseDEC-4L only makes decoder side dense, and (3) DenseAtt-4L only makes the attention dense using DenseAtt-2. There is no summary layer in the models, and both DenseENC-4L and DenseDEC-4L use hidden size 128. Again, by reducing the hidden size, we ensure that different 4-layer models have similar model sizes.
Our design for the WMT14 English-German model follows the best performance model provided in BIBREF2 . The construction of our model is straightforward: our 15-layer model DenseNMT-En-De-15 uses dense connection with DenseAtt-2, INLINEFORM0 . The hidden number in each layer is INLINEFORM1 that of the original model, while the kernel size maintains the same.
## Training setting
We use Nesterov Accelerated Gradient (NAG) BIBREF21 as our optimizer, and the initial learning rate is set to INLINEFORM0 . For German-English and Turkish-English experiments, the learning rate will shrink by 10 every time the validation loss increases. For the English-German dataset, in consistent with BIBREF2 , the learning rate will shrink by 10 every epoch since the first increment of validation loss. The system stops training until the learning rate is less than INLINEFORM1 . All models are trained end-to-end without any warmstart techniques. We set our batch size for the WMT14 English-German dataset to be 48, and additionally tune the length penalty parameter, in consistent with BIBREF2 . For other datasets, we set batch size to be 32. During inference, we use a beam size of 5.
## Training curve
We first show that DenseNMT helps information flow more efficiently by presenting the training loss curve. All hyperparameters are fixed in each plot, only the models are different. In Figure FIGREF30 , the loss curves for both training and dev sets (before entering the finetuning period) are provided for De-En, Tr-En and Tr-En-morph. For clarity, we compare DenseNMT-4L-2 with BASE-4L. We observe that DenseNMT models are consistently better than residual-connected models, since their loss curves are always below those of the baseline models. The effect is more obvious on the WMT14 English-German dataset. We rerun the best model provided by BIBREF2 and compare with our model. In Figure FIGREF33 , where train/test loss curve are provided, DenseNMT-En-De-15 reaches the same level of loss and starts finetuning (validation loss starts to increase) at epoch 13, which is 35% faster than the baseline.
Adding dense connections changes the architecture, and would slightly influence training speed. For the WMT14 En-De experiments, the computing time for both DenseNMT and the baseline (with similar number of parameters and same batch size) tested on single M40 GPU card are 1571 and 1710 word/s, respectively. While adding dense connections influences the per-iteration training slightly (8.1% reduction of speed), it uses many fewer epochs, and achieves a better BLEU score. In terms of training time, DenseNMT uses 29.3%(before finetuning)/22.9%(total) less time than the baseline.
## DenseNMT improves accuracy with similar architectures and model sizes
Table TABREF32 shows the results for De-En, Tr-En, Tr-En-morph datasets, where the best accuracy for models with the same depth and of similar sizes are marked in boldface. In almost all genres, DenseNMT models are significantly better than the baselines. With embedding size 256, where all models achieve their best scores, DenseNMT outperforms baselines by 0.7-1.0 BLEU on De-En, 0.5-1.3 BLEU on Tr-En, 0.8-1.5 BLEU on Tr-En-morph. We observe significant gain using other embedding sizes as well.
Furthermore, in Table TABREF36 , we investigate DenseNMT models through ablation study. In order to make the comparison fair, six models listed have roughly the same number of parameters. On De-En, Tr-En and Tr-En-morph, we see improvement by making the encoder dense, making the decoder dense, and making the attention dense. Fully dense-connected model DenseNMT-4L-1 further improves the translation accuracy. By allowing more flexibility in dense attention, DenseNMT-4L-2 provides the highest BLEU scores for all three experiments.
From the experiments, we have seen that enlarging the information flow in the attention block benefits the models. The dense attention block provides multi-layer information transmission from the encoder to the decoder, and to the output as well. Meanwhile, as shown by the ablation study, the dense-connected encoder and decoder both give more powerful representations than the residual-connected counterparts. As a result, the integration of the three parts improve the accuracy significantly.
## DenseNMT with smaller embedding size
From Table TABREF32 , we also observe that DenseNMT performs better with small embedding sizes compared to residual-connected models with regular embedding size. For example, on Tr-En model, the 8-layer DenseNMT-8L-2 model with embedding size 64 matches the BLEU score of the 8-layer BASE model with embedding size 256, while the number of parameter of the former one is only INLINEFORM0 of the later one. In all genres, DenseNMT model with embedding size 128 is comparable or even better than the baseline model with embedding size 256.
While overlarge embedding sizes hurt accuracy because of overfitting issues, smaller sizes are not preferable because of insufficient representation power. However, our dense models show that with better model design, the embedding information can be well concentrated on fewer dimensions, e.g., 64. This is extremely helpful when building models on mobile and small devices where the model size is critical. While there are other works that stress the efficiency issue by using techniques such as separable convolution BIBREF3 , and shared embedding BIBREF4 , our DenseNMT framework is orthogonal to those approaches. We believe that other techniques would produce more efficient models through combining with our DenseNMT framework.
## DenseNMT compares with state-of-the-art results
For the IWSLT14 German-English dataset, we compare with the best results reported from literatures. To be consistent with prior works, we also provide results using our model directly on the dataset without BPE preprocessing. As shown in Table TABREF39 , DenseNMT outperforms the phrase-structure based network NPMT BIBREF16 (with beam size 10) by 1.2 BLEU, using a smaller beam size, and outperforms the actor-critic method based algorithm BIBREF15 by 2.8 BLEU. For reference, our model trained on the BPE preprocessed dataset achieves 32.26 BLEU, which is 1.93 BLEU higher than our word-based model. For Turkish-English task, we compare with BIBREF19 which uses the same morphology preprocessing as our Tr-En-morph. As shown in Table TABREF37 , our baseline is higher than the previous result, and we further achieve new benchmark result with 24.36 BLEU average score. For WMT14 English-German, from Table TABREF41 , we can see that DenseNMT outperforms ConvS2S model by 0.36 BLEU score using 35% fewer training iterations and 20% fewer parameters. We also compare with another convolution based NMT model: SliceNet BIBREF3 , which explores depthwise separable convolution architectures. SliceNet-Full matches our result, and SliceNet-Super outperforms by 0.58 BLEU score. However, both models have 2.2x more parameters than our model. We expect DenseNMT structure could help improve their performance as well.
## Conclusion
In this work, we have proposed DenseNMT as a dense-connection framework for translation tasks, which uses the information from embeddings more efficiently, and passes abundant information from the encoder side to the decoder side. Our experiments have shown that DenseNMT is able to speed up the information flow and improve translation accuracy. For the future work, we will combine dense connections with other deep architectures, such as RNNs BIBREF7 and self-attention networks BIBREF4 .
| 14
|
1806.04511
|
Multilingual Sentiment Analysis: An RNN-Based Framework for Limited Data
|
# Multilingual Sentiment Analysis: An RNN-Based Framework for Limited Data
## Abstract
Sentiment analysis is a widely studied NLP task where the goal is to determine opinions, emotions, and evaluations of users towards a product, an entity or a service that they are reviewing. One of the biggest challenges for sentiment analysis is that it is highly language dependent. Word embeddings, sentiment lexicons, and even annotated data are language specific. Further, optimizing models for each language is very time consuming and labor intensive especially for recurrent neural network models. From a resource perspective, it is very challenging to collect data for different languages. In this paper, we look for an answer to the following research question: can a sentiment analysis model trained on a language be reused for sentiment analysis in other languages, Russian, Spanish, Turkish, and Dutch, where the data is more limited? Our goal is to build a single model in the language with the largest dataset available for the task, and reuse it for languages that have limited resources. For this purpose, we train a sentiment analysis model using recurrent neural networks with reviews in English. We then translate reviews in other languages and reuse this model to evaluate the sentiments. Experimental results show that our robust approach of single model trained on English reviews statistically significantly outperforms the baselines in several different languages.
## Introduction
With the steady growth in the commercial websites and social media venues, the access to users' reviews have become easier. As the amount of data that can be mined for opinion increased, commercial companies' interests for sentiment analysis increased as well. Sentiment analysis is an important part of understanding user behavior and opinions on products, places, or services.
Sentiment analysis has long been studied by the research community, leading to several sentiment-related resources such as sentiment dictionaries that can be used as features for machine learning models BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . These resources help increase sentiment analysis accuracies; however, they are highly dependent on language and require researchers to build such resources for every language to process.
Feature engineering is a large part of the model building phase for most sentiment analysis and emotion detection models BIBREF4 . Determining the correct set of features is a task that requires thorough investigation. Furthermore, these features are mostly language and dataset dependent making it even further challenging to build models for different languages. For example, the sentiment and emotion lexicons, as well as pre-trained word embeddings are not completely transferable to other languages which replicates the efforts for every language that users would like to build sentiment classification models on. For languages and tasks where the data is limited, extracting these features, building language models, training word embeddings, and creating lexicons are big challenges. In addition to the feature engineering effort, the machine learning models' parameters also need to be tuned separately for each language to get the optimal results.
In this paper, we take a different approach. We build a reusable sentiment analysis model that does not utilize any lexicons. Our goal is to evaluate how well a generic model can be used to mine opinion in different languages where data is more limited than the language where the generic model is trained on. To that end, we build a training set that contains reviews from different domains in English (e.g., movie reviews, product reviews) and train a recurrent neural network (RNN) model to predict polarity of those reviews. Then focusing on a domain, we make the model specialized in that domain by using the trained weights from the larger data and further training with data on a specific domain. To evaluate the reusability of the sentiment analysis model, we test with non-English datasets. We first translate the test set to English and use the pre-trained model to score polarity in the translated text. In this way, our proposed approach eliminates the need to train language-dependent models, use of sentiment lexicons and word embeddings for each language. Our experiments show that a generalizable sentiment analysis model can be utilized successfully to perform opinion mining for languages that do not have enough resources to train specific models.
The contributions of this study are; 1) a robust approach that utilizes machine translation to reuse a model trained on one language in other languages, 2) an RNN-based approach to eliminate feature extraction as well as resource requirements for sentiment analysis, and 3) a technique that statistically significantly outperforms baselines for multilingual sentiment analysis task when data is limited. To the best of our knowledge, this study is the first to apply a deep learning model to the multilingual sentiment analysis task.
## Related Work
There is a rich body of work in sentiment analysis including social media platforms such as Twitter BIBREF5 and Facebook BIBREF4 . One common factor in most of the sentiment analysis work is that features that are specific to sentiment analysis are extracted (e.g., sentiment lexicons) and used in different machine learning models. Lexical resources BIBREF0 , BIBREF1 , BIBREF4 for sentiment analysis such as SentiWordNet BIBREF6 , BIBREF7 , linguistic features and expressions BIBREF8 , polarity dictionaries BIBREF2 , BIBREF3 , other features such as topic-oriented features and syntax BIBREF9 , emotion tokens BIBREF10 , word vectors BIBREF11 , and emographics BIBREF12 are some of the information that are found useful for improving sentiment analysis accuracies. Although these features are beneficial, extracting them requires language-dependent data (e.g., a sentiment dictionary for Spanish is trained on Spanish data instead of using all data from different languages).
Our goal in this work is to streamline the feature engineering phase by not relying on any dictionary other than English word embeddings that are trained on any data (i.e. not necessarily sentiment analysis corpus). To that end, we utilize off-the-shelf machine translation tools to first translate corpora to the language where more training data is available and use the translated corpora to do inference on.
Machine translation for multilingual sentiment analysis has also seen attention from researchers. Hiroshi et al. BIBREF13 translated only sentiment units with a pattern-based approach. Balahur and Turchi BIBREF14 used uni-grams, bi-grams and tf-idf features for building support vector machines on translated text. Boyd-Graber and Resnik BIBREF15 built Latent Dirichlet Allocation models to investigate how multilingual concepts are clustered into topics. Mohammed et al. BIBREF16 translate Twitter posts to English as well as the English sentiment lexicons. Tellez et al. BIBREF17 propose a framework where language-dependent and independent features are used with an SVM classifier. These machine learning approaches also require a feature extraction phase where we eliminate by incorporating a deep learning approach that does the feature learning intrinsically. Further, Wan BIBREF18 uses an ensemble approach where the resources (e.g., lexicons) in both the original language and the translated language are used – requiring resources to be present in both languages. Brooke et al. BIBREF19 also use multiple dictionaries.
In this paper, we address the resource bottleneck of these translation-based approaches and propose a deep learning approach that does not require any dictionaries.
## Methodology
In order to eliminate the need to find data and build separate models for each language, we propose a multilingual approach where a single model is built in the language where the largest resources are available. In this paper we focus on English as there are several sentiment analysis datasets in English. To make the English sentiment analysis model as generalizable as possible, we first start by training with a large dataset that has product reviews for different categories. Then, using the trained weights from the larger generic dataset, we make the model more specialized for a specific domain. We further train the model with domain-specific English reviews and use this trained model to score reviews that share the same domain from different languages. To be able to employ the trained model, test sets are first translated to English via machine translation and then inference takes place. Figure FIGREF1 shows our multilingual sentiment analysis approach. It is important to note that this approach does not utilize any resource in any of the languages of the test sets (e.g., word embeddings, lexicons, training set).
Deep learning approaches have been successful in many applications ranging from computer vision to natural language processing BIBREF20 . Recurrent neural network (RNN) including Long Short Term Memory (LSTM) and Gated Recurrent Units (GRU) are subsets of deep learning algorithms where the dependencies between tokens can be used by the model. These models can also be used with variable length input vectors which makes them suitable for text input. LSTM and GRU models allow operations of sequences of vectors over time and have the capability to `remember' previous information BIBREF20 . RNN have been found useful for several natural language processing tasks including language modeling, text classification, machine translation. RNN can also utilize pre-trained word embeddings (numeric vector representations of words trained on unlabeled data) without requiring hand-crafted features. Therefore in this paper, we employ an RNN architecture that takes text and pre-trained word embeddings as inputs and generates a classification result. Word embeddings represent words as numeric vectors and capture semantic information. They are trained in an unsupervised fashion making it useful for our task.
The sentiment analysis model that is trained on English reviews has two bidirectional layers, each with 40 neurons and a dropout BIBREF21 of 0.2 is used. The training phase takes pre-trained word embeddings and reviews in textual format, then predicts the polarity of the reviews. For this study, an embedding length of 100 is used (i.e., each word is represented by a vector of length 100). We utilized pre-trained global vectors BIBREF22 . The training phase is depicted in Figure FIGREF2 .
## Experiments
To evaluate the proposed approach for multilingual sentiment analysis task, we conducted experiments. This section first presents the corpora used in this study followed by experimental results.
Throughout our experiments, we use SAS Deep Learning Toolkit. For machine translation, Google translation API is used.
## Corpora
Two sets of corpora are used in this study, both are publicly available. The first set consists of English reviews and the second set contains restaurant reviews from four different languages (Spanish, Turkish, Dutch, Russian). We focus on polarity detection in reviews, therefore all datasets in this study have two class values (positive, negative).
With the goal of building a generalizable sentiment analysis model, we used three different training sets as provided in Table TABREF5 . One of these three datasets (Amazon reviews BIBREF23 , BIBREF24 ) is larger and has product reviews from several different categories including book reviews, electronics products reviews, and application reviews. The other two datasets are to make the model more specialized in the domain. In this paper we focus on restaurant reviews as our domain and use Yelp restaurant reviews dataset extracted from Yelp Dataset Challenge BIBREF25 and restaurant reviews dataset as part of a Kaggle competition BIBREF26 .
For evaluation of the multilingual approach, we use four languages. These datasets are part of SemEval-2016 Challenge Task 5 BIBREF27 , BIBREF28 . Table TABREF7 shows the number of observations in each test corpus.
## Experimental Results
For experimental results, we report majority baseline for each language where the majority baseline corresponds to a model's accuracy if it always predicts the majority class in the dataset. For example, if the dataset has 60% of all reviews positive and 40% negative, majority baseline would be 60% because a model that always predicts “positive” will be 60% accurate and will make mistakes 40% of the time.
In addition to the majority baseline, we also compare our results with a lexicon-based approach. We use SentiWordNet BIBREF29 to obtain a positive and a negative sentiment score for each token in a review. Then sum of positive sentiment scores and negative sentiment scores for each review is obtained by summing up the scores for each token. If the positive sum score for a given review is greater than the negative sum score, we accept that review as a positive review. If negative sum is larger than or equal to the positive sum, the review is labeled as a negative review.
RNN outperforms both baselines in all four datasets (see Table TABREF9 ). Also for Spanish restaurant review, the lexicon-based baseline is below the majority baseline which shows that solely translating data and using lexicons is not sufficient to achieve good results in multilingual sentiment analysis.
Among the wrong classifications for each test set, we calculated the percentage of false positives and false negatives. Table TABREF10 shows the distribution of false positives and false negatives for each class. In all four classes, the number of false negatives are more than the number of false positives. This can be explained by the unbalanced training dataset where the number of positive reviews are more than the number of negative reviews (59,577 vs 17,132).
To be able to see the difference between baseline and RNN, we took each method's results as a group (4 values: one for each language) and compared the means. Post hoc comparisons using the Tukey HSD test indicated that the mean accuracies for baselines (majority and lexicon-based) are significantly different than RNN accuracies as can be seen in Table TABREF12 (family-wise error rate=0.06). When RNN is compared with lexicon-based baseline and majority baseline, the null hypothesis can be rejected meaning that each test is significant. In addition to these comparisons, we also calculated the effect sizes (using Cohen's d) between the baselines and our method. The results are aligning with Tukey HSD results such that while our method versus baselines have very large effect sizes, lexicon-based baseline and majority baseline have negligible effect size.
Figure FIGREF11 shows the differences in minimum and maximum values of all three approaches. As the figure shows, RNN significantly outperforms both baselines for the sentiment classification task.
## Discussion
One of the crucial elements while using machine translation is to have highly accurate translations. It is likely that non-English words would not have word embeddings, which will dramatically affect the effectiveness of the system. We analyzed the effect of incorrect translations into our approach. To that end, we extracted all wrong predictions from the test set and computed the ratio of misclassifications that have non-English words in them. We first extracted all misclassifications for a given language and for each observation in the misclassification set, we iterated through each token to check if the token is in English. In this way, we counted the number of observations that contained at least one non-English word and divided that with the size of the misclassifications set. We used this ratio to investigate the effect of machine translation errors.
We found that 25.84% of Dutch, 21.76% of Turkish, 24.46% Spanish, and 10.71% of Russian reviews that were misclassified had non-English words in them. These non-English words might be causing the misclassifications. However, a large portion of the missclassifications is not caused due to not-translated words. At the end, the machine translation errors has some but not noticeable effects on our model. Therefore, we can claim that machine translation preserves most of the information necessary for sentiment analysis.
We also evaluated our model with an English corpus BIBREF27 to see its performance without any interference from machine translation errors. Using the English data for testing, the model achieved 87.06% accuracy where a majority baseline was 68.37% and the lexicon-based baseline was 60.10%.
Considering the improvements over the majority baseline achieved by the RNN model for both non-English (on the average 22.76% relative improvement; 15.82% relative improvement on Spanish, 72.71% vs. 84.21%, 30.53% relative improvement on Turkish, 56.97% vs. 74.36%, 37.13% relative improvement on Dutch, 59.63% vs. 81.77%, and 7.55% relative improvement on Russian, 79.60% vs. 85.62%) and English test sets (27.34% relative improvement), we can draw the conclusion that our model is robust to handle multiple languages. Building separate models for each language requires both labeled and unlabeled data. Even though having lots of labeled data in every language is the perfect case, it is unrealistic. Therefore, eliminating the resource requirement in this resource-constrained task is crucial. The fact that machine translation can be used in reusing models from different languages is promising for reducing the data requirements.
## Conclusion
Building effective machine learning models for text requires data and different resources such as pre-trained word embeddings and reusable lexicons. Unfortunately, most of these resources are not entirely transferable to different domains, tasks or languages. Sentiment analysis is one such task that requires additional effort to transfer knowledge between languages.
In this paper, we studied the research question: Can we build reusable sentiment analysis models that can be utilized for making inferences in different languages without requiring separate models and resources for each language? To that end, we built a recurrent neural network model in the language that had largest data available. We took a general-to-specific model building strategy where the larger corpus that had reviews from different domains was first used to train the RNN model and a smaller single-domain corpus of sentiment reviews was used to specialize the model on the given domain. During scoring time, we used corpora for the given domain in different languages and translated them to English to be able to classify sentiments with the trained model. Experimental results showed that the proposed multilingual approach outperforms both the majority baseline and the lexicon-based baseline.
In this paper we made the sentiment analysis model specific to a single domain. For future work, we would like to investigate the effectiveness of our model on different review domains including hotel reviews and on different problems such as detecting stance.
| 8
|
1807.03367
|
Talk the Walk: Navigating New York City through Grounded Dialogue
|
# Talk the Walk: Navigating New York City through Grounded Dialogue
## Abstract
We introduce"Talk The Walk", the first large-scale dialogue dataset grounded in action and perception. The task involves two agents (a"guide"and a"tourist") that communicate via natural language in order to achieve a common goal: having the tourist navigate to a given target location. The task and dataset, which are described in detail, are challenging and their full solution is an open problem that we pose to the community. We (i) focus on the task of tourist localization and develop the novel Masked Attention for Spatial Convolutions (MASC) mechanism that allows for grounding tourist utterances into the guide's map, (ii) show it yields significant improvements for both emergent and natural language communication, and (iii) using this method, we establish non-trivial baselines on the full task.
0pt0.03.03 *
0pt0.030.03 *
0pt0.030.03
We introduce “Talk The Walk”, the first large-scale dialogue dataset grounded in action and perception. The task involves two agents (a “guide” and a “tourist”) that communicate via natural language in order to achieve a common goal: having the tourist navigate to a given target location. The task and dataset, which are described in detail, are challenging and their full solution is an open problem that we pose to the community. We (i) focus on the task of tourist localization and develop the novel Masked Attention for Spatial Convolutions (MASC) mechanism that allows for grounding tourist utterances into the guide's map, (ii) show it yields significant improvements for both emergent and natural language communication, and (iii) using this method, we establish non-trivial baselines on the full task.
## Introduction
As artificial intelligence plays an ever more prominent role in everyday human lives, it becomes increasingly important to enable machines to communicate via natural language—not only with humans, but also with each other. Learning algorithms for natural language understanding, such as in machine translation and reading comprehension, have progressed at an unprecedented rate in recent years, but still rely on static, large-scale, text-only datasets that lack crucial aspects of how humans understand and produce natural language. Namely, humans develop language capabilities by being embodied in an environment which they can perceive, manipulate and move around in; and by interacting with other humans. Hence, we argue that we should incorporate all three fundamental aspects of human language acquisition—perception, action and interactive communication—and develop a task and dataset to that effect.
We introduce the Talk the Walk dataset, where the aim is for two agents, a “guide” and a “tourist”, to interact with each other via natural language in order to achieve a common goal: having the tourist navigate towards the correct location. The guide has access to a map and knows the target location, but does not know where the tourist is; the tourist has a 360-degree view of the world, but knows neither the target location on the map nor the way to it. The agents need to work together through communication in order to successfully solve the task. An example of the task is given in Figure FIGREF3 .
Grounded language learning has (re-)gained traction in the AI community, and much attention is currently devoted to virtual embodiment—the development of multi-agent communication tasks in virtual environments—which has been argued to be a viable strategy for acquiring natural language semantics BIBREF0 . Various related tasks have recently been introduced, but in each case with some limitations. Although visually grounded dialogue tasks BIBREF1 , BIBREF2 comprise perceptual grounding and multi-agent interaction, their agents are passive observers and do not act in the environment. By contrast, instruction-following tasks, such as VNL BIBREF3 , involve action and perception but lack natural language interaction with other agents. Furthermore, some of these works use simulated environments BIBREF4 and/or templated language BIBREF5 , which arguably oversimplifies real perception or natural language, respectively. See Table TABREF15 for a comparison.
Talk The Walk is the first task to bring all three aspects together: perception for the tourist observing the world, action for the tourist to navigate through the environment, and interactive dialogue for the tourist and guide to work towards their common goal. To collect grounded dialogues, we constructed a virtual 2D grid environment by manually capturing 360-views of several neighborhoods in New York City (NYC). As the main focus of our task is on interactive dialogue, we limit the difficulty of the control problem by having the tourist navigating a 2D grid via discrete actions (turning left, turning right and moving forward). Our street view environment was integrated into ParlAI BIBREF6 and used to collect a large-scale dataset on Mechanical Turk involving human perception, action and communication.
We argue that for artificial agents to solve this challenging problem, some fundamental architecture designs are missing, and our hope is that this task motivates their innovation. To that end, we focus on the task of localization and develop the novel Masked Attention for Spatial Convolutions (MASC) mechanism. To model the interaction between language and action, this architecture repeatedly conditions the spatial dimensions of a convolution on the communicated message sequence.
This work makes the following contributions: 1) We present the first large scale dialogue dataset grounded in action and perception; 2) We introduce the MASC architecture for localization and show it yields improvements for both emergent and natural language; 4) Using localization models, we establish initial baselines on the full task; 5) We show that our best model exceeds human performance under the assumption of “perfect perception” and with a learned emergent communication protocol, and sets a non-trivial baseline with natural language.
## Talk The Walk
We create a perceptual environment by manually capturing several neighborhoods of New York City (NYC) with a 360 camera. Most parts of the city are grid-like and uniform, which makes it well-suited for obtaining a 2D grid. For Talk The Walk, we capture parts of Hell's Kitchen, East Village, the Financial District, Williamsburg and the Upper East Side—see Figure FIGREF66 in Appendix SECREF14 for their respective locations within NYC. For each neighborhood, we choose an approximately 5x5 grid and capture a 360 view on all four corners of each intersection, leading to a grid-size of roughly 10x10 per neighborhood.
The tourist's location is given as a tuple INLINEFORM0 , where INLINEFORM1 are the coordinates and INLINEFORM2 signifies the orientation (north, east, south or west). The tourist can take three actions: turn left, turn right and go forward. For moving forward, we add INLINEFORM3 , INLINEFORM4 , INLINEFORM5 , INLINEFORM6 to the INLINEFORM7 coordinates for the respective orientations. Upon a turning action, the orientation is updated by INLINEFORM8 where INLINEFORM9 for left and INLINEFORM10 for right. If the tourist moves outside the grid, we issue a warning that they cannot go in that direction and do not update the location. Moreover, tourists are shown different types of transitions: a short transition for actions that bring the tourist to a different corner of the same intersection; and a longer transition for actions that bring them to a new intersection.
The guide observes a map that corresponds to the tourist's environment. We exploit the fact that urban areas like NYC are full of local businesses, and overlay the map with these landmarks as localization points for our task. Specifically, we manually annotate each corner of the intersection with a set of landmarks INLINEFORM0 , each coming from one of the following categories:
Bar Playfield Bank Hotel Shop Subway Coffee Shop Restaurant Theater
The right-side of Figure FIGREF3 illustrates how the map is presented. Note that within-intersection transitions have a smaller grid distance than transitions to new intersections. To ensure that the localization task is not too easy, we do not include street names in the overhead map and keep the landmark categories coarse. That is, the dialogue is driven by uncertainty in the tourist's current location and the properties of the target location: if the exact location and orientation of the tourist were known, it would suffice to communicate a sequence of actions.
## Task
For the Talk The Walk task, we randomly choose one of the five neighborhoods, and subsample a 4x4 grid (one block with four complete intersections) from the entire grid. We specify the boundaries of the grid by the top-left and bottom-right corners INLINEFORM0 . Next, we construct the overhead map of the environment, i.e. INLINEFORM1 with INLINEFORM2 and INLINEFORM3 . We subsequently sample a start location and orientation INLINEFORM4 and a target location INLINEFORM5 at random.
The shared goal of the two agents is to navigate the tourist to the target location INLINEFORM0 , which is only known to the guide. The tourist perceives a “street view” planar projection INLINEFORM1 of the 360 image at location INLINEFORM2 and can simultaneously chat with the guide and navigate through the environment. The guide's role consists of reading the tourist description of the environment, building a “mental map” of their current position and providing instructions for navigating towards the target location. Whenever the guide believes that the tourist has reached the target location, they instruct the system to evaluate the tourist's location. The task ends when the evaluation is successful—i.e., when INLINEFORM3 —or otherwise continues until a total of three failed attempts. The additional attempts are meant to ease the task for humans, as we found that they otherwise often fail at the task but still end up close to the target location, e.g., at the wrong corner of the correct intersection.
## Data Collection
We crowd-sourced the collection of the dataset on Amazon Mechanical Turk (MTurk). We use the MTurk interface of ParlAI BIBREF6 to render 360 images via WebGL and dynamically display neighborhood maps with an HTML5 canvas. Detailed task instructions, which were also given to our workers before they started their task, are shown in Appendix SECREF15 . We paired Turkers at random and let them alternate between the tourist and guide role across different HITs.
## Dataset Statistics
The Talk The Walk dataset consists of over 10k successful dialogues—see Table FIGREF66 in the appendix for the dataset statistics split by neighborhood. Turkers successfully completed INLINEFORM0 of all finished tasks (we use this statistic as the human success rate). More than six hundred participants successfully completed at least one Talk The Walk HIT. Although the Visual Dialog BIBREF2 and GuessWhat BIBREF1 datasets are larger, the collected Talk The Walk dialogs are significantly longer. On average, Turkers needed more than 62 acts (i.e utterances and actions) before they successfully completed the task, whereas Visual Dialog requires 20 acts. The majority of acts comprise the tourist's actions, with on average more than 44 actions per dialogue. The guide produces roughly 9 utterances per dialogue, slightly more than the tourist's 8 utterances. Turkers use diverse discourse, with a vocabulary size of more than 10K (calculated over all successful dialogues). An example from the dataset is shown in Appendix SECREF14 . The dataset is available at https://github.com/facebookresearch/talkthewalk.
## Experiments
We investigate the difficulty of the proposed task by establishing initial baselines. The final Talk The Walk task is challenging and encompasses several important sub-tasks, ranging from landmark recognition to tourist localization and natural language instruction-giving. Arguably the most important sub-task is localization: without such capabilities the guide can not tell whether the tourist reached the target location. In this work, we establish a minimal baseline for Talk The Walk by utilizing agents trained for localization. Specifically, we let trained tourist models undertake random walks, using the following protocol: at each step, the tourist communicates its observations and actions to the guide, who predicts the tourist's location. If the guide predicts that the tourist is at target, we evaluate its location. If successful, the task ends, otherwise we continue until there have been three wrong evaluations. The protocol is given as pseudo-code in Appendix SECREF12 .
## Tourist Localization
The designed navigation protocol relies on a trained localization model that predicts the tourist's location from a communicated message. Before we formalize this localization sub-task in Section UID21 , we further introduce two simplifying assumptions—perfect perception and orientation-agnosticism—so as to overcome some of the difficulties we encountered in preliminary experiments.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Perfect Perception Early experiments revealed that perceptual grounding of landmarks is difficult: we set up a landmark classification problem, on which models with extracted CNN BIBREF7 or text recognition features BIBREF8 barely outperform a random baseline—see Appendix SECREF13 for full details. This finding implies that localization models from image input are limited by their ability to recognize landmarks, and, as a result, would not generalize to unseen environments. To ensure that perception is not the limiting factor when investigating the landmark-grounding and action-grounding capabilities of localization models, we assume “perfect perception”: in lieu of the 360 image view, the tourist is given the landmarks at its current location. More formally, each state observation INLINEFORM0 now equals the set of landmarks at the INLINEFORM1 -location, i.e. INLINEFORM2 . If the INLINEFORM3 -location does not have any visible landmarks, we return a single “empty corner” symbol. We stress that our findings—including a novel architecture for grounding actions into an overhead map, see Section UID28 —should carry over to settings without the perfect perception assumption.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Orientation-agnosticism We opt to ignore the tourist's orientation, which simplifies the set of actions to [Left, Right, Up, Down], corresponding to adding [(-1, 0), (1, 0), (0, 1), (0, -1)] to the current INLINEFORM0 coordinates, respectively. Note that actions are now coupled to an orientation on the map—e.g. up is equal to going north—and this implicitly assumes that the tourist has access to a compass. This also affects perception, since the tourist now has access to views from all orientations: in conjunction with “perfect perception”, implying that only landmarks at the current corner are given, whereas landmarks from different corners (e.g. across the street) are not visible.
Even with these simplifications, the localization-based baseline comes with its own set of challenges. As we show in Section SECREF34 , the task requires communication about a short (random) path—i.e., not only a sequence of observations but also actions—in order to achieve high localization accuracy. This means that the guide needs to decode observations from multiple time steps, as well as understand their 2D spatial arrangement as communicated via the sequence of actions. Thus, in order to get to a good understanding of the task, we thoroughly examine whether the agents can learn a communication protocol that simultaneously grounds observations and actions into the guide's map. In doing so, we thoroughly study the role of the communication channel in the localization task, by investigating increasingly constrained forms of communication: from differentiable continuous vectors to emergent discrete symbols to the full complexity of natural language.
The full navigation baseline hinges on a localization model from random trajectories. While we can sample random actions in the emergent communication setup, this is not possible for the natural language setup because the messages are coupled to the trajectories of the human annotators. This leads to slightly different problem setups, as described below.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Emergent language A tourist, starting from a random location, takes INLINEFORM0 random actions INLINEFORM1 to reach target location INLINEFORM2 . Every location in the environment has a corresponding set of landmarks INLINEFORM3 for each of the INLINEFORM4 coordinates. As the tourist navigates, the agent perceives INLINEFORM5 state-observations INLINEFORM6 where each observation INLINEFORM7 consists of a set of INLINEFORM8 landmark symbols INLINEFORM9 . Given the observations INLINEFORM10 and actions INLINEFORM11 , the tourist generates a message INLINEFORM12 which is communicated to the other agent. The objective of the guide is to predict the location INLINEFORM13 from the tourist's message INLINEFORM14 .
paragraph4 0.1ex plus0.1ex minus.1ex-1em Natural language In contrast to our emergent communication experiments, we do not take random actions but instead extract actions, observations, and messages from the dataset. Specifically, we consider each tourist utterance (i.e. at any point in the dialogue), obtain the current tourist location as target location INLINEFORM0 , the utterance itself as message INLINEFORM1 , and the sequence of observations and actions that took place between the current and previous tourist utterance as INLINEFORM2 and INLINEFORM3 , respectively. Similar to the emergent language setting, the guide's objective is to predict the target location INLINEFORM4 models from the tourist message INLINEFORM5 . We conduct experiments with INLINEFORM6 taken from the dataset and with INLINEFORM7 generated from the extracted observations INLINEFORM8 and actions INLINEFORM9 .
## Model
This section outlines the tourist and guide architectures. We first describe how the tourist produces messages for the various communication channels across which the messages are sent. We subsequently describe how these messages are processed by the guide, and introduce the novel Masked Attention for Spatial Convolutions (MASC) mechanism that allows for grounding into the 2D overhead map in order to predict the tourist's location.
## The Tourist
For each of the communication channels, we outline the procedure for generating a message INLINEFORM0 . Given a set of state observations INLINEFORM1 , we represent each observation by summing the INLINEFORM2 -dimensional embeddings of the observed landmarks, i.e. for INLINEFORM3 , INLINEFORM4 , where INLINEFORM5 is the landmark embedding lookup table. In addition, we embed action INLINEFORM6 into a INLINEFORM7 -dimensional embedding INLINEFORM8 via a look-up table INLINEFORM9 . We experiment with three types of communication channel.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Continuous vectors The tourist has access to observations of several time steps, whose order is important for accurate localization. Because summing embeddings is order-invariant, we introduce a sum over positionally-gated embeddings, which, conditioned on time step INLINEFORM0 , pushes embedding information into the appropriate dimensions. More specifically, we generate an observation message INLINEFORM1 , where INLINEFORM2 is a learned gating vector for time step INLINEFORM3 . In a similar fashion, we produce action message INLINEFORM4 and send the concatenated vectors INLINEFORM5 as message to the guide. We can interpret continuous vector communication as a single, monolithic model because its architecture is end-to-end differentiable, enabling gradient-based optimization for training.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Discrete symbols Like the continuous vector communication model, with discrete communication the tourist also uses separate channels for observations and actions, as well as a sum over positionally gated embeddings to generate observation embedding INLINEFORM0 . We pass this embedding through a sigmoid and generate a message INLINEFORM1 by sampling from the resulting Bernoulli distributions:
INLINEFORM0
The action message INLINEFORM0 is produced in the same way, and we obtain the final tourist message INLINEFORM1 through concatenating the messages.
The communication channel's sampling operation yields the model non-differentiable, so we use policy gradients BIBREF9 , BIBREF10 to train the parameters INLINEFORM0 of the tourist model. That is, we estimate the gradient by INLINEFORM1
where the reward function INLINEFORM0 is the negative guide's loss (see Section SECREF25 ) and INLINEFORM1 a state-value baseline to reduce variance. We use a linear transformation over the concatenated embeddings as baseline prediction, i.e. INLINEFORM2 , and train it with a mean squared error loss.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Natural Language Because observations and actions are of variable-length, we use an LSTM encoder over the sequence of observations embeddings INLINEFORM0 , and extract its last hidden state INLINEFORM1 . We use a separate LSTM encoder for action embeddings INLINEFORM2 , and concatenate both INLINEFORM3 and INLINEFORM4 to the input of the LSTM decoder at each time step: DISPLAYFORM0
where INLINEFORM0 a look-up table, taking input tokens INLINEFORM1 . We train with teacher-forcing, i.e. we optimize the cross-entropy loss: INLINEFORM2 . At test time, we explore the following decoding strategies: greedy, sampling and a beam-search. We also fine-tune a trained tourist model (starting from a pre-trained model) with policy gradients in order to minimize the guide's prediction loss.
## The Guide
Given a tourist message INLINEFORM0 describing their observations and actions, the objective of the guide is to predict the tourist's location on the map. First, we outline the procedure for extracting observation embedding INLINEFORM1 and action embeddings INLINEFORM2 from the message INLINEFORM3 for each of the types of communication. Next, we discuss the MASC mechanism that takes the observations and actions in order to ground them on the guide's map in order to predict the tourist's location.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Continuous For the continuous communication model, we assign the observation message to the observation embedding, i.e. INLINEFORM0 . To extract the action embedding for time step INLINEFORM1 , we apply a linear layer to the action message, i.e. INLINEFORM2 .
paragraph4 0.1ex plus0.1ex minus.1ex-1em Discrete For discrete communication, we obtain observation INLINEFORM0 by applying a linear layer to the observation message, i.e. INLINEFORM1 . Similar to the continuous communication model, we use a linear layer over action message INLINEFORM2 to obtain action embedding INLINEFORM3 for time step INLINEFORM4 .
paragraph4 0.1ex plus0.1ex minus.1ex-1em Natural Language The message INLINEFORM0 contains information about observations and actions, so we use a recurrent neural network with attention mechanism to extract the relevant observation and action embeddings. Specifically, we encode the message INLINEFORM1 , consisting of INLINEFORM2 tokens INLINEFORM3 taken from vocabulary INLINEFORM4 , with a bidirectional LSTM: DISPLAYFORM0
where INLINEFORM0 is the word embedding look-up table. We obtain observation embedding INLINEFORM1 through an attention mechanism over the hidden states INLINEFORM2 : DISPLAYFORM0
where INLINEFORM0 is a learned control embedding who is updated through a linear transformation of the previous control and observation embedding: INLINEFORM1 . We use the same mechanism to extract the action embedding INLINEFORM2 from the hidden states. For the observation embedding, we obtain the final representation by summing positionally gated embeddings, i.e., INLINEFORM3 .
We represent the guide's map as INLINEFORM0 , where in this case INLINEFORM1 , where each INLINEFORM2 -dimensional INLINEFORM3 location embedding INLINEFORM4 is computed as the sum of the guide's landmark embeddings for that location.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Motivation While the guide's map representation contains only local landmark information, the tourist communicates a trajectory of the map (i.e. actions and observations from multiple locations), implying that directly comparing the tourist's message with the individual landmark embeddings is probably suboptimal. Instead, we want to aggregate landmark information from surrounding locations by imputing trajectories over the map to predict locations. We propose a mechanism for translating landmark embeddings according to state transitions (left, right, up, down), which can be expressed as a 2D convolution over the map embeddings. For simplicity, let us assume that the map embedding INLINEFORM0 is 1-dimensional, then a left action can be realized through application of the following INLINEFORM1 kernel: INLINEFORM2 which effectively shifts all values of INLINEFORM3 one position to the left. We propose to learn such state-transitions from the tourist message through a differentiable attention-mask over the spatial dimensions of a 3x3 convolution.
paragraph4 0.1ex plus0.1ex minus.1ex-1em MASC We linearly project each predicted action embedding INLINEFORM0 to a 9-dimensional vector INLINEFORM1 , normalize it by a softmax and subsequently reshape the vector into a 3x3 mask INLINEFORM2 : DISPLAYFORM0
We learn a 3x3 convolutional kernel INLINEFORM0 , with INLINEFORM1 features, and apply the mask INLINEFORM2 to the spatial dimensions of the convolution by first broadcasting its values along the feature dimensions, i.e. INLINEFORM3 , and subsequently taking the Hadamard product: INLINEFORM4 . For each action step INLINEFORM5 , we then apply a 2D convolution with masked weight INLINEFORM6 to obtain a new map embedding INLINEFORM7 , where we zero-pad the input to maintain identical spatial dimensions.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Prediction model We repeat the MASC operation INLINEFORM0 times (i.e. once for each action), and then aggregate the map embeddings by a sum over positionally-gated embeddings: INLINEFORM1 . We score locations by taking the dot-product of the observation embedding INLINEFORM2 , which contains information about the sequence of observed landmarks by the tourist, and the map. We compute a distribution over the locations of the map INLINEFORM3 by taking a softmax over the computed scores: DISPLAYFORM0
paragraph4 0.1ex plus0.1ex minus.1ex-1em Predicting T While emergent communication models use a fixed length trasjectory INLINEFORM0 , natural language messages may differ in the number of communicated observations and actions. Hence, we predict INLINEFORM1 from the communicated message. Specifically, we use a softmax regression layer over the last hidden state INLINEFORM2 of the RNN, and subsequently sample INLINEFORM3 from the resulting multinomial distribution: DISPLAYFORM0
We jointly train the INLINEFORM0 -prediction model via REINFORCE, with the guide's loss as reward function and a mean-reward baseline.
## Comparisons
To better analyze the performance of the models incorporating MASC, we compare against a no-MASC baseline in our experiments, as well as a prediction upper bound.
paragraph4 0.1ex plus0.1ex minus.1ex-1em No MASC We compare the proposed MASC model with a model that does not include this mechanism. Whereas MASC predicts a convolution mask from the tourist message, the “No MASC” model uses INLINEFORM0 , the ordinary convolutional kernel to convolve the map embedding INLINEFORM1 to obtain INLINEFORM2 . We also share the weights of this convolution at each time step.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Prediction upper-bound Because we have access to the class-conditional likelihood INLINEFORM0 , we are able to compute the Bayes error rate (or irreducible error). No model (no matter how expressive) with any amount of data can ever obtain better localization accuracy as there are multiple locations consistent with the observations and actions.
## Results and Discussion
In this section, we describe the findings of various experiments. First, we analyze how much information needs to be communicated for accurate localization in the Talk The Walk environment, and find that a short random path (including actions) is necessary. Next, for emergent language, we show that the MASC architecture can achieve very high localization accuracy, significantly outperforming the baseline that does not include this mechanism. We then turn our attention to the natural language experiments, and find that localization from human utterances is much harder, reaching an accuracy level that is below communicating a single landmark observation. We show that generated utterances from a conditional language model leads to significantly better localization performance, by successfully grounding the utterance on a single landmark observation (but not yet on multiple observations and actions). Finally, we show performance of the localization baseline on the full task, which can be used for future comparisons to this work.
## Analysis of Localization Task
paragraph4 0.1ex plus0.1ex minus.1ex-1em Task is not too easy The upper-bound on localization performance in Table TABREF32 suggest that communicating a single landmark observation is not sufficient for accurate localization of the tourist ( INLINEFORM0 35% accuracy). This is an important result for the full navigation task because the need for two-way communication disappears if localization is too easy; if the guide knows the exact location of the tourist it suffices to communicate a list of instructions, which is then executed by the tourist. The uncertainty in the tourist's location is what drives the dialogue between the two agents.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Importance of actions We observe that the upperbound for only communicating observations plateaus around 57% (even for INLINEFORM0 actions), whereas it exceeds 90% when we also take actions into account. This implies that, at least for random walks, it is essential to communicate a trajectory, including observations and actions, in order to achieve high localization accuracy.
## Emergent Language Localization
We first report the results for tourist localization with emergent language in Table TABREF32 .
paragraph4 0.1ex plus0.1ex minus.1ex-1em MASC improves performance The MASC architecture significantly improves performance compared to models that do not include this mechanism. For instance, for INLINEFORM0 action, MASC already achieves 56.09 % on the test set and this further increases to 69.85% for INLINEFORM1 . On the other hand, no-MASC models hit a plateau at 43%. In Appendix SECREF11 , we analyze learned MASC values, and show that communicated actions are often mapped to corresponding state-transitions.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Continuous vs discrete We observe similar performance for continuous and discrete emergent communication models, implying that a discrete communication channel is not a limiting factor for localization performance.
## Natural Language Localization
We report the results of tourist localization with natural language in Table TABREF36 . We compare accuracy of the guide model (with MASC) trained on utterances from (i) humans, (ii) a supervised model with various decoding strategies, and (iii) a policy gradient model optimized with respect to the loss of a frozen, pre-trained guide model on human utterances.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Human utterances Compared to emergent language, localization from human utterances is much harder, achieving only INLINEFORM0 on the test set. Here, we report localization from a single utterance, but in Appendix SECREF45 we show that including up to five dialogue utterances only improves performance to INLINEFORM1 . We also show that MASC outperform no-MASC models for natural language communication.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Generated utterances We also investigate generated tourist utterances from conditional language models. Interestingly, we observe that the supervised model (with greedy and beam-search decoding) as well as the policy gradient model leads to an improvement of more than 10 accuracy points over the human utterances. However, their level of accuracy is slightly below the baseline of communicating a single observation, indicating that these models only learn to ground utterances in a single landmark observation.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Better grounding of generated utterances We analyze natural language samples in Table TABREF38 , and confirm that, unlike human utterances, the generated utterances are talking about the observed landmarks. This observation explains why the generated utterances obtain higher localization accuracy. The current language models are most successful when conditioned on a single landmark observation; We show in Appendix UID43 that performance quickly deteriorates when the model is conditioned on more observations, suggesting that it can not produce natural language utterances about multiple time steps.
## Localization-based Baseline
Table TABREF36 shows results for the best localization models on the full task, evaluated via the random walk protocol defined in Algorithm SECREF12 .
paragraph4 0.1ex plus0.1ex minus.1ex-1em Comparison with human annotators Interestingly, our best localization model (continuous communication, with MASC, and INLINEFORM0 ) achieves 88.33% on the test set and thus exceed human performance of 76.74% on the full task. While emergent models appear to be stronger localizers, humans might cope with their localization uncertainty through other mechanisms (e.g. better guidance, bias towards taking particular paths, etc). The simplifying assumption of perfect perception also helps.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Number of actions Unsurprisingly, humans take fewer steps (roughly 15) than our best random walk model (roughly 34). Our human annotators likely used some form of guidance to navigate faster to the target.
## Conclusion
We introduced the Talk The Walk task and dataset, which consists of crowd-sourced dialogues in which two human annotators collaborate to navigate to target locations in the virtual streets of NYC. For the important localization sub-task, we proposed MASC—a novel grounding mechanism to learn state-transition from the tourist's message—and showed that it improves localization performance for emergent and natural language. We use the localization model to provide baseline numbers on the Talk The Walk task, in order to facilitate future research.
## Related Work
The Talk the Walk task and dataset facilitate future research on various important subfields of artificial intelligence, including grounded language learning, goal-oriented dialogue research and situated navigation. Here, we describe related previous work in these areas.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Related tasks There has been a long line of work involving related tasks. Early work on task-oriented dialogue dates back to the early 90s with the introduction of the Map Task BIBREF11 and Maze Game BIBREF25 corpora. Recent efforts have led to larger-scale goal-oriented dialogue datasets, for instance to aid research on visually-grounded dialogue BIBREF2 , BIBREF1 , knowledge-base-grounded discourse BIBREF29 or negotiation tasks BIBREF36 . At the same time, there has been a big push to develop environments for embodied AI, many of which involve agents following natural language instructions with respect to an environment BIBREF13 , BIBREF50 , BIBREF5 , BIBREF39 , BIBREF19 , BIBREF18 , following-up on early work in this area BIBREF38 , BIBREF20 . An early example of navigation using neural networks is BIBREF28 , who propose an online learning approach for robot navigation. Recently, there has been increased interest in using end-to-end trainable neural networks for learning to navigate indoor scenes BIBREF27 , BIBREF26 or large cities BIBREF17 , BIBREF40 , but, unlike our work, without multi-agent communication. Also the task of localization (without multi-agent communication) has recently been studied BIBREF18 , BIBREF48 .
paragraph4 0.1ex plus0.1ex minus.1ex-1em Grounded language learning Grounded language learning is motivated by the observation that humans learn language embodied (grounded) in sensorimotor experience of the physical world BIBREF15 , BIBREF45 . On the one hand, work in multi-modal semantics has shown that grounding can lead to practical improvements on various natural language understanding tasks BIBREF14 , BIBREF31 . In robotics, researchers dissatisfied with purely symbolic accounts of meaning attempted to build robotic systems with the aim of grounding meaning in physical experience of the world BIBREF44 , BIBREF46 . Recently, grounding has also been applied to the learning of sentence representations BIBREF32 , image captioning BIBREF37 , BIBREF49 , visual question answering BIBREF12 , BIBREF22 , visual reasoning BIBREF30 , BIBREF42 , and grounded machine translation BIBREF43 , BIBREF23 . Grounding also plays a crucial role in the emergent research of multi-agent communication, where, agents communicate (in natural language or otherwise) in order to solve a task, with respect to their shared environment BIBREF35 , BIBREF21 , BIBREF41 , BIBREF24 , BIBREF36 , BIBREF47 , BIBREF34 .
## Implementation Details
For the emergent communication models, we use an embedding size INLINEFORM0 . The natural language experiments use 128-dimensional word embeddings and a bidirectional RNN with 256 units. In all experiments, we train the guide with a cross entropy loss using the ADAM optimizer with default hyper-parameters BIBREF33 . We perform early stopping on the validation accuracy, and report the corresponding train, valid and test accuracy. We optimize the localization models with continuous, discrete and natural language communication channels for 200, 200, and 25 epochs, respectively. To facilitate further research on Talk The Walk, we make our code base for reproducing experiments publicly available at https://github.com/facebookresearch/talkthewalk.
## Additional Natural Language Experiments
First, we investigate the sensitivity of tourist generation models to the trajectory length, finding that the model conditioned on a single observation (i.e. INLINEFORM0 ) achieves best performance. In the next subsection, we further analyze localization models from human utterances by investigating MASC and no-MASC models with increasing dialogue context.
## Tourist Generation Models
After training the supervised tourist model (conditioned on observations and action from human expert trajectories), there are two ways to train an accompanying guide model. We can optimize a location prediction model on either (i) extracted human trajectories (as in the localization setup from human utterances) or (ii) on all random paths of length INLINEFORM0 (as in the full task evaluation). Here, we investigate the impact of (1) using either human or random trajectories for training the guide model, and (2) the effect of varying the path length INLINEFORM1 during the full-task evaluation. For random trajectories, guide training uses the same path length INLINEFORM2 as is used during evaluation. We use a pre-trained tourist model with greedy decoding for generating the tourist utterances. Table TABREF40 summarizes the results.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Human vs random trajectories We only observe small improvements for training on random trajectories. Human trajectories are thus diverse enough to generalize to random trajectories.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Effect of path length There is a strong negative correlation between task success and the conditioned trajectory length. We observe that the full task performance quickly deteriorates for both human and random trajectories. This suggests that the tourist generation model can not produce natural language utterances that describe multiple observations and actions. Although it is possible that the guide model can not process such utterances, this is not very likely because the MASC architectures handles such messages successfully for emergent communication.
We report localization performance of tourist utterances generated by beam search decoding of varying beam size in Table TABREF40 . We find that performance decreases from 29.05% to 20.87% accuracy on the test set when we increase the beam-size from one to eight.
## Localization from Human Utterances
We conduct an ablation study for MASC on natural language with varying dialogue context. Specifically, we compare localization accuracy of MASC and no-MASC models trained on the last [1, 3, 5] utterances of the dialogue (including guide utterances). We report these results in Table TABREF41 . In all cases, MASC outperforms the no-MASC models by several accuracy points. We also observe that mean predicted INLINEFORM0 (over the test set) increases from 1 to 2 when more dialogue context is included.
## Visualizing MASC predictions
Figure FIGREF46 shows the MASC values for a learned model with emergent discrete communications and INLINEFORM0 actions. Specifically, we look at the predicted MASC values for different action sequences taken by the tourist. We observe that the first action is always mapped to the correct state-transition, but that the second and third MASC values do not always correspond to right state-transitions.
## Evaluation on Full Setup
We provide pseudo-code for evaluation of localization models on the full task in Algorithm SECREF12 , as well as results for all emergent communication models in Table TABREF55 .
INLINEFORM0 INLINEFORM1
INLINEFORM0 take new action INLINEFORM1 INLINEFORM2
Performance evaluation of location prediction model on full Talk The Walk setup
## Landmark Classification
While the guide has access to the landmark labels, the tourist needs to recognize these landmarks from raw perceptual information. In this section, we study landmark classification as a supervised learning problem to investigate the difficulty of perceptual grounding in Talk The Walk.
The Talk The Walk dataset contains a total of 307 different landmarks divided among nine classes, see Figure FIGREF62 for how they are distributed. The class distribution is fairly imbalanced, with shops and restaurants as the most frequent landmarks and relatively few play fields and theaters. We treat landmark recognition as a multi-label classification problem as there can be multiple landmarks on a corner.
For the task of landmark classification, we extract the relevant views of the 360 image from which a landmark is visible. Because landmarks are labeled to be on a specific corner of an intersection, we assume that they are visible from one of the orientations facing away from the intersection. For example, for a landmark on the northwest corner of an intersection, we extract views from both the north and west direction. The orientation-specific views are obtained by a planar projection of the full 360-image with a small field of view (60 degrees) to limit distortions. To cover the full field of view, we extract two images per orientation, with their horizontal focus point 30 degrees apart. Hence, we obtain eight images per 360 image with corresponding orientation INLINEFORM0 .
We run the following pre-trained feature extractors over the extracted images:
For the text recognition model, we use a learned look-up table INLINEFORM0 to embed the extracted text features INLINEFORM1 , and fuse all embeddings of four images through a bag of embeddings, i.e., INLINEFORM2 . We use a linear layer followed by a sigmoid to predict the probability for each class, i.e. INLINEFORM3 . We also experiment with replacing the look-up embeddings with pre-trained FastText embeddings BIBREF16 . For the ResNet model, we use a bag of embeddings over the four ResNet features, i.e. INLINEFORM4 , before we pass it through a linear layer to predict the class probabilities: INLINEFORM5 . We also conduct experiments where we first apply PCA to the extracted ResNet and FastText features before we feed them to the model.
To account for class imbalance, we train all described models with a binary cross entropy loss weighted by the inverted class frequency. We create a 80-20 class-conditional split of the dataset into a training and validation set. We train for 100 epochs and perform early stopping on the validation loss.
The F1 scores for the described methods in Table TABREF65 . We compare to an “all positive” baseline that always predicts that the landmark class is visible and observe that all presented models struggle to outperform this baseline. Although 256-dimensional ResNet features achieve slightly better precision on the validation set, it results in much worse recall and a lower F1 score. Our results indicate that perceptual grounding is a difficult task, which easily merits a paper of its own right, and so we leave further improvements (e.g. better text recognizers) for future work.
## Dataset Details
paragraph4 0.1ex plus0.1ex minus.1ex-1em Dataset split We split the full dataset by assigning entire 4x4 grids (independent of the target location) to the train, valid or test set. Specifically, we design the split such that the valid set contains at least one intersection (out of four) is not part of the train set. For the test set, all four intersections are novel. See our source code, available at URL ANONYMIZED, for more details on how this split is realized.
paragraph4 0.1ex plus0.1ex minus.1ex-1em Example
Tourist: ACTION:TURNRIGHT ACTION:TURNRIGHT
Guide: Hello, what are you near?
Tourist: ACTION:TURNLEFT ACTION:TURNLEFT ACTION:TURNLEFT
Tourist: Hello, in front of me is a Brooks Brothers
Tourist: ACTION:TURNLEFT ACTION:FORWARD ACTION:TURNLEFT ACTION:TURNLEFT
Guide: Is that a shop or restaurant?
Tourist: ACTION:TURNLEFT
Tourist: It is a clothing shop.
Tourist: ACTION:TURNLEFT
Guide: You need to go to the intersection in the northwest corner of the map
Tourist: ACTION:TURNLEFT
Tourist: There appears to be a bank behind me.
Tourist: ACTION:TURNLEFT ACTION:TURNLEFT ACTION:TURNRIGHT ACTION:TURNRIGHT
Guide: Ok, turn left then go straight up that road
Tourist: ACTION:TURNLEFT ACTION:TURNLEFT ACTION:TURNLEFT ACTION:FORWARD ACTION:TURNRIGHT
ACTION:FORWARD ACTION:FORWARD ACTION:TURNLEFT ACTION:TURNLEFT ACTION:TURNLEFT
Guide: There should be shops on two of the corners but you
need to go to the corner without a shop.
Tourist: ACTION:FORWARD ACTION:FORWARD ACTION:FORWARD ACTION:TURNLEFT ACTION:TURNLEFT
Guide: let me know when you get there.
Tourist: on my left is Radio city Music hall
Tourist: ACTION:TURNLEFT ACTION:FORWARD ACTION:TURNLEFT ACTION:TURNRIGHT ACTION:TURNRIGHT
Tourist: I can't go straight any further.
Guide: ok. turn so that the theater is on your right.
Guide: then go straight
Tourist: That would be going back the way I came
Guide: yeah. I was looking at the wrong bank
Tourist: I'll notify when I am back at the brooks brothers, and the bank.
Tourist: ACTION:TURNRIGHT
Guide: make a right when the bank is on your left
Tourist: ACTION:FORWARD ACTION:FORWARD ACTION:TURNRIGHT
Tourist: Making the right at the bank.
Tourist: ACTION:FORWARD ACTION:FORWARD
Tourist: I can't go that way.
Tourist: ACTION:TURNLEFT
Tourist: Bank is ahead of me on the right
Tourist: ACTION:FORWARD ACTION:FORWARD ACTION:TURNLEFT
Guide: turn around on that intersection
Tourist: I can only go to the left or back the way I just came.
Tourist: ACTION:TURNLEFT
Guide: you're in the right place. do you see shops on the corners?
Guide: If you're on the corner with the bank, cross the street
Tourist: I'm back where I started by the shop and the bank.
Tourist: ACTION:TURNRIGHT
Guide: on the same side of the street?
Tourist: crossing the street now
Tourist: ACTION:FORWARD ACTION:FORWARD ACTION:TURNLEFT
Tourist: there is an I love new york shop across the street on the left from me now
Tourist: ACTION:TURNRIGHT ACTION:FORWARD
Guide: ok. I'll see if it's right.
Guide: EVALUATE_LOCATION
Guide: It's not right.
Tourist: What should I be on the look for?
Tourist: ACTION:TURNRIGHT ACTION:TURNRIGHT ACTION:TURNRIGHT
Guide: There should be shops on two corners but you need to be on one of the corners
without the shop.
Guide: Try the other corner.
Tourist: this intersection has 2 shop corners and a bank corner
Guide: yes. that's what I see on the map.
Tourist: should I go to the bank corner? or one of the shop corners?
or the blank corner (perhaps a hotel)
Tourist: ACTION:TURNLEFT ACTION:TURNLEFT ACTION:TURNRIGHT ACTION:TURNRIGHT
Guide: Go to the one near the hotel. The map says the hotel is a little
further down but it might be a little off.
Tourist: It's a big hotel it's possible.
Tourist: ACTION:FORWARD ACTION:TURNLEFT ACTION:FORWARD ACTION:TURNRIGHT
Tourist: I'm on the hotel corner
Guide: EVALUATE_LOCATION
| 27
|
1807.07961
|
Twitter Sentiment Analysis via Bi-sense Emoji Embedding and Attention-based LSTM
|
# Twitter Sentiment Analysis via Bi-sense Emoji Embedding and Attention-based LSTM
## Abstract
Sentiment analysis on large-scale social media data is important to bridge the gaps between social media contents and real world activities including political election prediction, individual and public emotional status monitoring and analysis, and so on. Although textual sentiment analysis has been well studied based on platforms such as Twitter and Instagram, analysis of the role of extensive emoji uses in sentiment analysis remains light. In this paper, we propose a novel scheme for Twitter sentiment analysis with extra attention on emojis. We first learn bi-sense emoji embeddings under positive and negative sentimental tweets individually, and then train a sentiment classifier by attending on these bi-sense emoji embeddings with an attention-based long short-term memory network (LSTM). Our experiments show that the bi-sense embedding is effective for extracting sentiment-aware embeddings of emojis and outperforms the state-of-the-art models. We also visualize the attentions to show that the bi-sense emoji embedding provides better guidance on the attention mechanism to obtain a more robust understanding of the semantics and sentiments.
## Introduction
The rapid growth of social media platforms such as Twitter provides rich multimedia data in large scales for various research opportunities, such as sentiment analysis which focuses on automatically sentiment (positive and negative) prediction on given contents. Sentiment analysis has been widely used in real world applications by analyzing the online user-generated data, such as election prediction, opinion mining and business-related activity analysis. Emojis, which consist of various symbols ranging from cartoon facial expressions to figures such as flags and sports, are widely used in daily communications to express people's feelings . Since their first release in 2010, emojis have taken the place of emoticons (such as “:- INLINEFORM0 ” and “:-P”) BIBREF0 to create a new form of language for social media users BIBREF1 . According to recent science reports, there are 2,823 emojis in unicode standard in Emoji 11.0 , with over 50% of the Instagram posts containing one or more emojis BIBREF2 and 92% of the online population using emojis BIBREF3 .
The extensive use of emojis has drawn a growing attention from researchers BIBREF4 , BIBREF5 because the emojis convey fruitful semantical and sentimental information to visually complement the textual information which is significantly useful in understanding the embedded emotional signals in texts BIBREF6 . For example, emoji embeddings have been proposed to understand the semantics behind the emojis BIBREF7 , BIBREF8 , and the embedding vectors can be used to visualize and predict emoji usages given their corresponding contexts. Previous work also shows that, it is useful to pre-train a deep neural network on an emoji prediction task with pre-trained emoji embeddings to learn the emotional signals of emojis for other tasks including sentiment, emotion and sarcasm prediction BIBREF9 . However, the previous literatures lack in considerations of the linguistic complexities and diversity of emoji. Therefore, previous emoji embedding methods fail to handle the situation when the semantics or sentiments of the learned emoji embeddings contradict the information from the corresponding contexts BIBREF5 , or when the emojis convey multiple senses of semantics and sentiments such as ( and ). In practice, emojis can either summarize and emphasis the original tune of their contexts, or express more complex semantics such as irony and sarcasm by being combined with contexts of contradictory semantics or sentiments. For the examples shown in Table TABREF3 , the emoji () is of consistent sentiment with text to emphasis the sentiment, but is of the opposite sentiment (positive) to the text sentiment (negative) example 3 and 4 to deliver a sense of sarcasm. Conventional emoji analysis can only extract single embedding of each emoji, and such embeddings will confuse the following sentiment analysis model by inconsistent sentiment signals from the input texts and emojis. Moreover, we consider the emoji effect modeling different from the conventional multimodal sentiment analysis which usually includes images and texts in that, image sentiment and text sentiment are usually assumed to be consistent BIBREF10 while it carries no such assumption for texts and emojis.
To tackle such limitations, we propose a novel scheme that consists of an attention-based recurrent neural network (RNN) with robust bi-sense emoji embeddings. Inspired by the word sense embedding task in natural language processing (NLP) BIBREF11 , BIBREF12 , BIBREF13 where each sense of an ambiguous word responds to one unique embedding vector, the proposed bi-sense embedding is a more robust and fine-grained representation of the complicated semantics for emojis where each emoji is embedded into two distinct vectors, namely positive-sense and negative-sense vector, respectively. For our specific task which is Twitter sentiment analysis BIBREF14 , BIBREF15 , we initialize the bi-sense embedding vectors together with word embedding vectors using word embedding algorithm fasttext BIBREF16 by extracting two distinct embeddings for each emoji according to the sentiment of its corresponding textual contexts, namely bi-sense embedding. A long short-term memory (LSTM) based recurrent neural network is then used for predicting sentiments which is integrated with the pre-trained emoji embedding features by a context-guide and self-selected attention mechanism. Because most of the previous Twitter sentiment datasets exclude emojis and there exists little resource that contains sufficient emoji-tweets with sentiment labels, we construct our own emoji-tweets dataset by automatically generating weak labels using a rule-based sentiment analysis algorithm Vader BIBREF17 for pre-traning the networks, and manually labeling a subset of tweets for fine tuning and testing purposes. The experimental results demonstrate that the bi-sense emoji embedding is capable of extracting more distinguished information from emojis and outperforms the state-of-the-art sentiment analysis models with the proposed attention-based LSTM networks. We further visualize the bi-sense emoji embedding to obtain the sentiments and semantics learned by the proposed approach.
The main contributions of this paper are summarized as follows.
## Sentiment Analysis
Sentiment analysis is to extract and quantify subjective information including the status of attitudes, emotions and opinions from a variety of contents such as texts, images and audios BIBREF18 . Sentiment analysis has been drawing great attentions because of its wide applications in business and government intelligence, political science, sociology and psychology BIBREF19 , BIBREF20 , BIBREF21 , BIBREF22 . From a technical perspective, textual sentiment analysis is first explored by researchers as an NLP task. Methods range from lexical-based approaches using features including keywords BIBREF23 , BIBREF24 where each word corresponds to a sentiment vector with entries representing the possibility of the word and each sentiment and phase-level features (n-grams and unigrams) BIBREF25 , BIBREF26 , to deep neural network based embedding approaches including skip-grams, continuous bag-of-words (CBoW) and skip-thoughts BIBREF27 , BIBREF28 , BIBREF16 , BIBREF29 . It was until recent years when researchers start focusing on image and multimodal sentiments BIBREF30 , BIBREF31 and analyzing how to take advantage of the cross-modality resources BIBREF10 , BIBREF32 . For multimodal sentiment analysis, an underlying assumption is that both modalities express similar sentiment and such similarity is enforced in order to train a robust sentiment inference model BIBREF10 . However, the same assumption does not stand in modeling textual tweets and emojis because the complexities of natural language exist extensively, such as the use of irony, jokes, sarcasm, etc. BIBREF9 .
Models
We set up the baselines and proposed models as follows:
LSTM with text embedding: CNNs and LSTMs are widely used to encode textual contents for sentiment analysis in BIBREF45 , BIBREF46 and many online tutorials. Here we select the standard LSTM with pre-trained word embedding as input, and add one fully-connected layer with sigmoid activation top of the LSTM encoder (same as all other models), denoted as T-LSTM.
LSTM with emoji embedding: We consider the emoji as one special word and input both pre-trained text and emoji embeddings into the same LSTM network, namely E-LSTM. Similarly, we concatenate the pre-trained bi-sense emoji embedding as one special word to feed into the LSTM network. This model is called BiE-LSTM.
Attention-based LSTM with emojis:We also use the word-emoji embedding to calculate the emoji-word attention following Equation EQREF20 and EQREF21 , and the only difference is that we replace the attention-derived senti-emoji embedding with the pre-trained word-emoji embedding by fasttext, denoted as ATT-E-LSTM.
LSTM with bi-sense emoji embedding (proposed): As we have introduced in Section SECREF13 , we propose two attention-based LSTM networks based on bi-sense emoji embedding, denoted as MATT-BiE-LSTM and WATT-BiE-LSTM.
Evaluation
We evaluate the baseline and proposed models on sentiment analysis by F1 scores and accuracies based on the auto-annotated testing set (AA-Sentiment) and human-annotated testing set (HA-Sentiment), as shown in Table TABREF25 . We only test the models after fine-tuning with a subset of the samples with human annotations because training exclusively on the samples with auto-generated weak labels results in relatively poor performances when tested with human annotated data indicating the models after fine-tuning are more robust. The F1 scores and accuracies are overall higher with the AA-Sentiment than the results with HA-sentiment, indicating that the HA-Sentiment is a more challenging task and the sentiments involved are more difficult to identify supported by their relatively lower sentiment scores returned from Vader. We still, however, observe competitive results from HA-Sentiment showing that the models are well-trained and robust to noisy labels with the help of fine-tuning with human annotated data. The T-LSTM baseline achieves decent performance in both experiments with accuracies of 86.6% and 70.7% showing that LSTM is an effective encoder for sentiment analysis as suggested by the references. The models with proposed bi-sense emoji embedding obtain accuracies over 82.4% and we observe improvements on the performance with the attention-based LSTM from our proposed model MATT-BiE-LSTM and WATT-BiE-LSTM, which is consistent with that ATT-E-LSTM (F1@84.6%, accuracy@82.0% on HA-Sentiment) outperforms significantly T-LSTM and E-LSTM.
Emoji information is useful in sentiment analysis. Most models outperforms the baseline T-LSTM in both dataset suggesting that the emoji information is useful for sentiment analysis as a complement to the textual contents, even with the naive use of emoji embeddings (E-LSTM) when tested with HA-Sentiment. We observe that E-LSTM obtains similar performance to T-LSTM with AA-Sentiment but a significant gain over the T-LSTM when tested with HA-Sentiment indicating that sentiment information is helpful and necessary when the hidden sentiment is relatively subtle and the task is more challenging.
Bi-sense emoji embedding helps. All the models using bi-sense emoji embedding perform significantly better than the baseline models without emoji feature or with word-emoji embedding. BiE-LSTM outperforms T-LSTM and E-LSTM significantly with the same utilization of emoji embedding indicates that the proposed bi-sense emoji embedding is capable of extracting more informative and distinguishable vectors over the use of conventional word embedding algorithms, which is consistent based on the comparisons between the proposed models (MATT-BiE-LSTM and WATT-BiE-LSTM) with bi-sense emoji embedding and the baseline model ATT-E-LSTM with word-emoji embedding and attention.
Attention mechanism aligns and performs well with bi-sense embedding. MATT-BiE-LSTM and WATT-BiE-LSTM obtain similar performances when tested on both Vader and human annotated samples, though their ways of computing the attention (weights and vectors) are different that WATT computes attention weights and the senti-emoji embeddings guided by each word, and MATT obtains the senti-emoji embedding based on the LSTM encoder on the whole contexts and computes the attention weights of the senti-emoji embedding across all words. Both models outperforms the state-of-the-art baseline models including ATT-E-LSTM. The proposed attention-based LSTM can be further extended to handle tasks involving multi-sense embedding as inputs, such as the word-sense embedding in NLP, by using context-guide attention to self-select how much to attend on each sense of the embeddings each of which correspond to a distinct sense of semantics or sentiments. In this way we are able to take advantage of the more robust and fine-grained embeddings.
## Emojis and Sentiment Analysis
With the overwhelming development of Internet of Things (IOT), the growing accessibility and popularity of subjective contents have provided new opportunities and challenges for sentiment analysis BIBREF33 . For example, social medias such as Twitter and Instagram have been explored because the massive user-generated contents with rich user sentiments BIBREF25 , BIBREF34 , BIBREF35 where emojis (and emoticons) are extensively used. Non-verbal cues of sentiment, such as emoticon which is considered as the previous generation of emoji, has been studied for their sentiment effect before emojis take over BIBREF36 , BIBREF37 , BIBREF38 . For instance, BIBREF36 , BIBREF38 pre-define sentiment labels to emoticons and construct a emoticon-sentiment dictionary. BIBREF37 applies emoticons for smoothing noisy sentiment labels. Similar work from BIBREF39 first considers emoji as a component in extracting the lexical feature for further sentiment analysis. BIBREF40 constructs an emoji sentiment ranking based on the occurrences of emojis and the human-annotated sentiments of the corresponding tweets where each emoji is assigned with a sentiment score from negative to positive , similar to the SentiWordNet BIBREF41 . However, the relatively intuitive use of emojis by lexical- and dictionary-based approaches lacks insightful understanding of the complexed semantics of emojis. Therefore, inspired by the success of word semantic embedding algorithms such as BIBREF28 , BIBREF16 , BIBREF7 obtains semantic embeddings of each emoji by averaging the words from its descriptions and shows it is effective to take advantage of the emoji embedding for the task of Twitter sentiment analysis. BIBREF8 proposes a convoluntional neural network to predict the emoji occurrence and jointly learns the emoji embedding via a matching layer based on cosine similarities. Despite the growing popularity of Twitter sentiment analysis, there is a limited number of emoji datasets with sentiment labels available because previous studies usually filter out urls, emojis and sometimes emoticons. However, BIBREF9 shows that it is effective to extract sentiment information from emojis for emotion classification and sarcasm detection tasks in the absence of learning vector-based emoji representations by pre-training a deep neural network to predict the emoji occurrence.
## Methodology
We propose two mechanisms, namely Word-guide Attention-based LSTM and Multi-level Attention-based LSTM, to take advantage of bi-sense emoji embedding for the sentiment analysis task. The frameworks of these two methods are shown in Figure FIGREF10 and Figure FIGREF19 , respectively. Our workflow includes the following steps: initialization of bi-sense emoji embedding, generating senti-emoji embedding based on self-selected attention, and sentiment classification via the proposed attention-based LSTM networks.
## Bi-sense Embedding
Recent research shows great success in word embedding task such as word2vec and fasttext BIBREF27 , BIBREF16 . We use fasttext to initialize emoji embeddings by considering each emoji as a special word, together with word embeddings. The catch is, different from conventional approaches where each emoji responds to one embedding vector (as we call word-emoji embedding), we embed each emoji into two distinct vectors (bi-sense emoji embedding): we first assign two distinct tokens to each emoji, of which one is for the particular emoji used in positive sentimental contexts and the other one is for this emoji used in negative sentimental contexts (text sentiment initialized by Vader BIBREF17 , details will be discussed in Section SECREF23 ), respectively; the same fasttext training process is used to embed each token into a distinct vector, and we thus obtain the positive-sense and negative-sense embeddings for each emoji.
The word2vec is based on the skip-gram model whose objective is to maximize the log likelihood calculated by summing the probabilities of current word occurrences given a set of the surrounding words. The fasttext model is different by formatting the problem as a binary classification task to predict the occurrence of each context word, with negative samples being randomly selected from the absent context words. Given an input word sequence INLINEFORM0 , and the context word set INLINEFORM1 and the set of negative word samples INLINEFORM2 of the current word INLINEFORM3 , the objective function is obtained based on binary logistic loss as in Equation EQREF12 : DISPLAYFORM0
where INLINEFORM0 denotes the logistic loss of the score function INLINEFORM1 which is computed by summing up the scalar products between the n-gram embeddings of the current word and the context word embedding which is different from word2vec where the score is the scalar product between the current word and the context word embedding. We select fasttext over word2vec mainly because its computational efficiency. In general, the two models yield competitive performances and the comparison between word embeddings is beyond our discussion. Therefore we only show the results derived by the fasttext initialization within the scope of this work.
## Word-guide Attention-based LSTM
Long short-term memory (LSTM) units have been extensively used to encode textual contents. The basic encoder model consists of a text embedding layer, LSTMs layer, and fully-connected layers for further tasks such as text classifications based on the encoded feature. The operations in an LSTM unit for time step INLINEFORM0 is formulated in Equation EQREF14 : DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 represent the current and previous hidden states, INLINEFORM2 denotes the current LSTM input and here we use the embedding INLINEFORM3 of the current word INLINEFORM4 , and INLINEFORM5 and INLINEFORM6 denote the weight matrices BIBREF42 . In order to take advantage of the bi-sense emoji embedding, we modify the input layer into the LSTM units. We first obtain the senti-emoji embedding as an weighted average of the bi-sense emoji embedding based on the self-selected attention mechanism. Let INLINEFORM7 represent the INLINEFORM8 -th sense embedding of emoji INLINEFORM9 ( INLINEFORM10 in our bi-sense embedding), and INLINEFORM11 denote the attention function conditioned on the current word embedding, the attention weight INLINEFORM12 and senti-emoji embedding vector INLINEFORM13 is formulated as follows: DISPLAYFORM0
We choose a fully-connected layer with ReLU activation as the attention function, and the attention vector INLINEFORM0 is concatenated with the word embedding as the new input of the LSTM. Thus the input vector INLINEFORM1 in Equation EQREF14 becomes INLINEFORM2 . The output of the final LSTM unit is then fed into a fully-connected layer with INLINEFORM3 activation to output the tweet sentiment and binary cross-entropy loss is used as the objection function (Equation EQREF16 ) where INLINEFORM4 is the total number of samples. The motivation behind this model is that each context word guides the attention weights in order to enforce the model to self-select which embedding sense it should attend on. Therefore we denote this model as the Word-guide Attention-based LSTM with Bi-sense emoji embedding (WATT-BiE-LSTM). DISPLAYFORM0
## Multi-level Attention-based LSTM
There is another way of formulating the attention mechanism where the attention weights indicate how the image information (which is emoji in our case) is distributed through the context words as proposed in BIBREF43 , BIBREF44 . The modified senti-emoji embedding vector INLINEFORM0 is thus at the tweet-level instead of the word-level in Equation EQREF15 by replacing the INLINEFORM1 with the final state vector INLINEFORM2 outputted from the last LSTM unit, as shown in Equation EQREF18 : DISPLAYFORM0
The derived senti-emoji embedding INLINEFORM0 is then used to calculate an additional layer of attention following BIBREF43 , BIBREF44 . Given the input tweet sequence INLINEFORM1 , the attention weight INLINEFORM2 conditioned on the senti-emoji embedding is formulated as follows: DISPLAYFORM0
Therefore, we construct the new input INLINEFORM0 to each LSTM unit by concatenating the original word embedding and the attention vector in Equation EQREF21 to distribute the senti-emoji information to each step. This model is called Multi-level Attention-based LSTM with Bi-sense Emoji Embedding (MATT-BiE-LSTM). We choose the same binary cross-entropy as the loss function with the same network configuration with WATT-BiE-LSTM. DISPLAYFORM0
## Data Collection and Annotation
Data Collection
We construct our own Twitter sentiment dataset by crawling tweets through the REST API which consists of 350,000 users and is magnitude larger comparing to previous work. We collect up to 3,200 tweets from each user and follow the standard tweet preprocessing procedures to remove the tweets without emojis and tweets containing less than ten words, and contents including the urls, mentions, and emails.
Data Annotation
For acquiring the sentiment annotations, we first use Vader which is a rule-based sentiment analysis algorithm BIBREF17 for text tweets only to generate weak sentiment labels. The algorithm outputs sentiment scores ranging from -1 (negative) to 1 (positive) with neutral in the middle. We consider the sentiment analysis as a binary classification problem (positive sentiment and negative sentiment), we filter out samples with weak prediction scores within INLINEFORM0 and keep the tweets with strong sentiment signals. Emoji occurrences are calculated separately for positive tweets and negative tweets, and threshold is set to 2,000 to further filter out emojis which are less frequently used in at least one type of sentimental text. In the end, we have constructed a dataset with 1,492,065 tweets and 55 frequently used emojis in total. For the tweets with an absolute sentiment score over 0.70, we keep the auto-generated sentiment label as ground truth because the automatic annotation is reliable with high sentiment scores. On the other hand, we select a subset of the tweets with absolute sentiment scores between INLINEFORM1 for manual labeling by randomly sampling, following the distribution of emoji occurrences where each tweet is labeled by two graduate students. Tweets are discarded if the two annotations disagree with each other or they are labeled as neutral. In the end, we have obtained 4,183 manually labeled tweets among which 60% are used for fine-tuning and 40% are used for testing purposes. The remainder of the tweets with automatic annotations are divided into three sets: 60% are used for pre-training the bi-sense and conventional emoji embedding, 10% for validation and 30% are for testing. We do not include a “neutral” class because it is difficult to obtain valid neutral samples. For auto-generated labels, the neutrals are the samples with low absolute confidence scores and their sentiments are more likely to be model failures other than “true neutrals”. Moreover, based on the human annotations, most of the tweets with emojis convey non-neutral sentiment and only few neutral samples are observed during the manual labeling which are excluded from the manually labeled subset.
In order to valid our motivation that emojis are also extensively used in tweets that contain contradictory information to the emoji sentiments, we calculate the emoji usage in Table TABREF22 according to the sentiment labels where Pos-Ratio means the percentage of each emoji occurs in the positive tweets over its total number of occurrences, AA and HA indicate automatic-annotation and human-annotation, respectively. We present the top-10 most frequently used emojis in our dataset and observe a slight difference in the Pos-Ratios between AA and HA dataset because of the randomness involved in the sampling process. Results from both of the datasets show a fair amount of emoji use in both positive and negative tweets. For example, it is interesting to notice that emoji () occurs more in the positive tweets in with the automatic annotations, while emojis with strong positive sentiment have also been used in negative tweets with about 5% occurrences, such as (, , and ). Given the averaged positive ratio among all emojis in the whole dataset is about 74% and that most emojis have been extensively used in tweets containing both positive and negative sentiments, it suggests that distinguishing the emoji occurrences in both sentiments via bi-sense embedding is worth investigating. Additionally, we observe the Pos-Ratios of the AA-sentiment and HA-sentiment have little differences which are due to two main reasons: 1) Some tweets we sampled to construct the HA-sentiment are discarded because the annotators have disagreements and we only keep the samples that we are confident about; 2) Tweets with absolute sentiment scores between (0.60,0.70) are selected for manual labeling as discussed in Section SECREF23 , which are lower than the tweets used to construct the AA-sentiment (0.7 and above). The lower sentiment scores indicate that Vader is less reliable on the samples of HA-sentiment dataset and the sentiments of these tweets are more likely to be affected by emojis.
## Qualitative Analysis
In order to obtain insights about why the more fine-grained bi-sense emoji embedding helps in understanding the complexed sentiments behind tweets, we visualize the attention weights for ATT-E-LSTM and MATT-BiE-LSTM for comparison. The example tweets with corresponding attention weights calculated by word-emoji embedding and senti-emoji embedding are shown in Figure FIGREF27 , where the contexts are presented in the captions. The emojis used are , , and , respectively.
In Figure FIGREF27 (a), the ATT-E-LSTM model (baseline) assigns relatively more weights on the word “no” and “pressure”, while MATT-BiE-LSTM attends mostly on the word “happy” and “lovely”. The different attention distributions suggest that the proposed senti-emoji embedding is capable of recognizing words with strong sentiments that are closely related to the true sentiment even with the presence of words with conflicting sentiments, such as “pressure” and “happy”. while ATT-E-LSTM tends to pick up all sentimental words which could raise confusions. The senti-emoji embedding is capable of extracting representations of complexed semantics and sentiments which help guide the attentions even in cases when the word sentiment and emoji sentiment are somewhat contradictory to each other. From Figure FIGREF27 (b) and (c) we can observe that the ATT-E-LSTM assigns more weights on the sentiment-irrelevant words than the MATT-BiE-LSTM such as “hoodies”, “wait” and “after”, indicating that the proposed model is more robust to irrelevant words and concentrates better on important words. Because of the senti-emoji embedding obtained through bi-sense emoji embedding and the sentence-level LSTM encoding on the text input (described in Section SECREF13 ), we are able to construct a more robust embedding based on the semantic and sentiment information from the whole context compared to the word-emoji embedding used in ATT-E-LSTM which takes only word-level information into account.
## Bi-sense Emoji Embedding Visualization
To gain further insights on the bi-sense emoji embedding, we use t-SNE BIBREF47 to project high-dimensional bi-sense embedding vectors into a two-dimensional space and preserving the relative distances between the embedding vectors at the same time. In Figure FIGREF28 we visualize the bi-sense emoji embedding, positive-sense embedding, negative-sense embedding and the subtraction between positive and negative sense embeddings of each emoji, respectively. The subtraction of an emoji between its two sense embeddings indicates the semantic differences between emoji usages in positive and negative sentimental contexts, similarly to the objective of word embeddings BIBREF28 . The positive-sense of emoji ( and ), and the negative-sense of emoji (, and ) are embedded far from the two main clusters as observed in Figure FIGREF28 (a), suggesting that the semantics of these emojis are different from the other popular emojis. The positive-sense embedding and negative-sense embeddings are clustered well with no intersection with each other. Such observation supports our objective of applying bi-sense emoji embedding because there exist such significant differences in the semantics of each emoji when appears in positive and negative sentimental contexts, and it is well-motivated to consider the emoji usages individually according to the sentiment of the contexts to extract the more fine-grained bi-sense embedding. Additionally, we observe consistent patterns in the Figure FIGREF28 (b), (c) and (d) where the sentiments conveyed in the emojis become an important factor. For example, emojis with positive sentiments such as (, and ), and emojis with negative sentiment such as (, and ) are embedded into one clusters in both positive-sense and negative-sense space. The embedding subtractions of emojis in Figure FIGREF28 (d) shows the different usages of emojis across sentiments are similar between emojis and preserve the cluster patterns observed in Figure FIGREF28 (b) and (c).
## Conclusions
In this paper, we present a novel approach to the task of sentiment analysis and achieve the state-of-the-art performance. Different from the previous work, our method combines a more robust and fine-grained bi-sense emoji embedding that effectively represents complex semantic and sentiment information, with attention-based LSTM networks that selectively attend on the correlated sense of the emoji embeddings, and seamlessly fuse the obtained senti-emoji embeddings with the word embeddings for a better understanding of the rich semantics and sentiments involved. In the future, we plan to further extend our attention-based LSTM with bi-embedding work frame to tackle tasks involving multi-sense embedding such as the learning and applications of word-sense embedding.
## Acknowledgement
We would like to thank the support of New York State through the Goergen Institute for Data Science, and NSF Award #1704309.
| 12
|
1808.03986
|
Multimodal Differential Network for Visual Question Generation
|
# Multimodal Differential Network for Visual Question Generation
## Abstract
Generating natural questions from an image is a semantic task that requires using visual and language modality to learn multimodal representations. Images can have multiple visual and language contexts that are relevant for generating questions namely places, captions, and tags. In this paper, we propose the use of exemplars for obtaining the relevant context. We obtain this by using a Multimodal Differential Network to produce natural and engaging questions. The generated questions show a remarkable similarity to the natural questions as validated by a human study. Further, we observe that the proposed approach substantially improves over state-of-the-art benchmarks on the quantitative metrics (BLEU, METEOR, ROUGE, and CIDEr).
## Introduction
To understand the progress towards multimedia vision and language understanding, a visual Turing test was proposed by BIBREF0 that was aimed at visual question answering BIBREF1 . Visual Dialog BIBREF2 is a natural extension for VQA. Current dialog systems as evaluated in BIBREF3 show that when trained between bots, AI-AI dialog systems show improvement, but that does not translate to actual improvement for Human-AI dialog. This is because, the questions generated by bots are not natural (human-like) and therefore does not translate to improved human dialog. Therefore it is imperative that improvement in the quality of questions will enable dialog agents to perform well in human interactions. Further, BIBREF4 show that unanswered questions can be used for improving VQA, Image captioning and Object Classification.
An interesting line of work in this respect is the work of BIBREF5 . Here the authors have proposed the challenging task of generating natural questions for an image. One aspect that is central to a question is the context that is relevant to generate it. However, this context changes for every image. As can be seen in Figure FIGREF1 , an image with a person on a skateboard would result in questions related to the event. Whereas for a little girl, the questions could be related to age rather than the action. How can one have widely varying context provided for generating questions? To solve this problem, we use the context obtained by considering exemplars, specifically we use the difference between relevant and irrelevant exemplars. We consider different contexts in the form of Location, Caption, and Part of Speech tags.
The human annotated questions are (b) for the first image and (a) for the second image.
Our method implicitly uses a differential context obtained through supporting and contrasting exemplars to obtain a differentiable embedding. This embedding is used by a question decoder to decode the appropriate question. As discussed further, we observe this implicit differential context to perform better than an explicit keyword based context. The difference between the two approaches is illustrated in Figure FIGREF2 . This also allows for better optimization as we can backpropagate through the whole network. We provide detailed empirical evidence to support our hypothesis. As seen in Figure FIGREF1 our method generates natural questions and improves over the state-of-the-art techniques for this problem.
To summarize, we propose a multimodal differential network to solve the task of visual question generation. Our contributions are: (1) A method to incorporate exemplars to learn differential embeddings that captures the subtle differences between supporting and contrasting examples and aid in generating natural questions. (2) We provide Multimodal differential embeddings, as image or text alone does not capture the whole context and we show that these embeddings outperform the ablations which incorporate cues such as only image, or tags or place information. (3) We provide a thorough comparison of the proposed network against state-of-the-art benchmarks along with a user study and statistical significance test.
## Related Work
Generating a natural and engaging question is an interesting and challenging task for a smart robot (like chat-bot). It is a step towards having a natural visual dialog instead of the widely prevalent visual question answering bots. Further, having the ability to ask natural questions based on different contexts is also useful for artificial agents that can interact with visually impaired people. While the task of generating question automatically is well studied in NLP community, it has been relatively less studied for image-related natural questions. This is still a difficult task BIBREF5 that has gained recent interest in the community.
Recently there have been many deep learning based approaches as well for solving the text-based question generation task such as BIBREF6 . Further, BIBREF7 have proposed a method to generate a factoid based question based on triplet set {subject, relation and object} to capture the structural representation of text and the corresponding generated question.
These methods, however, were limited to text-based question generation. There has been extensive work done in the Vision and Language domain for solving image captioning, paragraph generation, Visual Question Answering (VQA) and Visual Dialog. BIBREF8 , BIBREF9 , BIBREF10 proposed conventional machine learning methods for image description. BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 , BIBREF17 , BIBREF18 have generated descriptive sentences from images with the help of Deep Networks. There have been many works for solving Visual Dialog BIBREF19 , BIBREF20 , BIBREF2 , BIBREF21 , BIBREF22 . A variety of methods have been proposed by BIBREF23 , BIBREF24 , BIBREF1 , BIBREF25 , BIBREF26 , BIBREF27 for solving VQA task including attention-based methods BIBREF28 , BIBREF29 , BIBREF30 , BIBREF31 , BIBREF32 , BIBREF33 , BIBREF34 . However, Visual Question Generation (VQG) is a separate task which is of interest in its own right and has not been so well explored BIBREF5 . This is a vision based novel task aimed at generating natural and engaging question for an image. BIBREF35 proposed a method for continuously generating questions from an image and subsequently answering those questions. The works closely related to ours are that of BIBREF5 and BIBREF36 . In the former work, the authors used an encoder-decoder based framework whereas in the latter work, the authors extend it by using a variational autoencoder based sequential routine to obtain natural questions by performing sampling of the latent variable.
## Approach
In this section, we clarify the basis for our approach of using exemplars for question generation. To use exemplars for our method, we need to ensure that our exemplars can provide context and that our method generates valid exemplars.
We first analyze whether the exemplars are valid or not. We illustrate this in figure FIGREF3 . We used a pre-trained RESNET-101 BIBREF37 object classification network on the target, supporting and contrasting images. We observed that the supporting image and target image have quite similar probability scores. The contrasting exemplar image, on the other hand, has completely different probability scores.
Exemplars aim to provide appropriate context. To better understand the context, we experimented by analysing the questions generated through an exemplar. We observed that indeed a supporting exemplar could identify relevant tags (cows in Figure FIGREF3 ) for generating questions.
We improve use of exemplars by using a triplet network. This network ensures that the joint image-caption embedding for the supporting exemplar are closer to that of the target image-caption and vice-versa. We empirically evaluated whether an explicit approach that uses the differential set of tags as a one-hot encoding improves the question generation, or the implicit embedding obtained based on the triplet network. We observed that the implicit multimodal differential network empirically provided better context for generating questions. Our understanding of this phenomenon is that both target and supporting exemplars generate similar questions whereas contrasting exemplars generate very different questions from the target question. The triplet network that enhances the joint embedding thus aids to improve the generation of target question. These are observed to be better than the explicitly obtained context tags as can be seen in Figure FIGREF2 . We now explain our method in detail.
## Method
The task in visual question generation (VQG) is to generate a natural language question INLINEFORM0 , for an image INLINEFORM1 . We consider a set of pre-generated context INLINEFORM2 from image INLINEFORM3 . We maximize the conditional probability of generated question given image and context as follows: DISPLAYFORM0
where INLINEFORM0 is a vector for all possible parameters of our model. INLINEFORM1 is the ground truth question. The log probability for the question is calculated by using joint probability over INLINEFORM2 with the help of chain rule. For a particular question, the above term is obtained as: INLINEFORM3
where INLINEFORM0 is length of the sequence, and INLINEFORM1 is the INLINEFORM2 word of the question. We have removed INLINEFORM3 for simplicity.
Our method is based on a sequence to sequence network BIBREF38 , BIBREF12 , BIBREF39 . The sequence to sequence network has a text sequence as input and output. In our method, we take an image as input and generate a natural question as output. The architecture for our model is shown in Figure FIGREF4 . Our model contains three main modules, (a) Representation Module that extracts multimodal features (b) Mixture Module that fuses the multimodal representation and (c) Decoder that generates question using an LSTM-based language model.
During inference, we sample a question word INLINEFORM0 from the softmax distribution and continue sampling until the end token or maximum length for the question is reached. We experimented with both sampling and argmax and found out that argmax works better. This result is provided in the supplementary material.
## Multimodal Differential Network
The proposed Multimodal Differential Network (MDN) consists of a representation module and a joint mixture module.
We used an efficient KNN-based approach (k-d tree) with Euclidean metric to obtain the exemplars. This is obtained through a coarse quantization of nearest neighbors of the training examples into 50 clusters, and selecting the nearest as supporting and farthest as the contrasting exemplars. We experimented with ITML based metric learning BIBREF40 for image features. Surprisingly, the KNN-based approach outperforms the latter one. We also tried random exemplars and different number of exemplars and found that INLINEFORM0 works best. We provide these results in the supplementary material.
We use a triplet network BIBREF41 , BIBREF42 in our representation module. We refereed a similar kind of work done in BIBREF34 for building our triplet network. The triplet network consists of three sub-parts: target, supporting, and contrasting networks. All three networks share the same parameters. Given an image INLINEFORM0 we obtain an embedding INLINEFORM1 using a CNN parameterized by a function INLINEFORM2 where INLINEFORM3 are the weights for the CNN. The caption INLINEFORM4 results in a caption embedding INLINEFORM5 through an LSTM parameterized by a function INLINEFORM6 where INLINEFORM7 are the weights for the LSTM. This is shown in part 1 of Figure FIGREF4 . Similarly we obtain image embeddings INLINEFORM8 & INLINEFORM9 and caption embeddings INLINEFORM10 & INLINEFORM11 . DISPLAYFORM0
The Mixture module brings the image and caption embeddings to a joint feature embedding space. The input to the module is the embeddings obtained from the representation module. We have evaluated four different approaches for fusion viz., joint, element-wise addition, hadamard and attention method. Each of these variants receives image features INLINEFORM0 & the caption embedding INLINEFORM1 , and outputs a fixed dimensional feature vector INLINEFORM2 . The Joint method concatenates INLINEFORM3 & INLINEFORM4 and maps them to a fixed length feature vector INLINEFORM5 as follows: DISPLAYFORM0
where INLINEFORM0 is the 4096-dimensional convolutional feature from the FC7 layer of pretrained VGG-19 Net BIBREF43 . INLINEFORM1 are the weights and INLINEFORM2 is the bias for different layers. INLINEFORM3 is the concatenation operator.
Similarly, We obtain context vectors INLINEFORM0 & INLINEFORM1 for the supporting and contrasting exemplars. Details for other fusion methods are present in supplementary.The aim of the triplet network BIBREF44 is to obtain context vectors that bring the supporting exemplar embeddings closer to the target embedding and vice-versa. This is obtained as follows: DISPLAYFORM0
where INLINEFORM0 is the euclidean distance between two embeddings INLINEFORM1 and INLINEFORM2 . M is the training dataset that contains all set of possible triplets. INLINEFORM3 is the triplet loss function. This is decomposed into two terms, one that brings the supporting sample closer and one that pushes the contrasting sample further. This is given by DISPLAYFORM0
Here INLINEFORM0 represent the euclidean distance between the target and supporting sample, and target and opposing sample respectively. The parameter INLINEFORM1 controls the separation margin between these and is obtained through validation data.
## Decoder: Question Generator
The role of decoder is to predict the probability for a question, given INLINEFORM0 . RNN provides a nice way to perform conditioning on previous state value using a fixed length hidden vector. The conditional probability of a question token at particular time step INLINEFORM1 is modeled using an LSTM as used in machine translation BIBREF38 . At time step INLINEFORM2 , the conditional probability is denoted by INLINEFORM3 , where INLINEFORM4 is the hidden state of the LSTM cell at time step INLINEFORM5 , which is conditioned on all the previously generated words INLINEFORM6 . The word with maximum probability in the probability distribution of the LSTM cell at step INLINEFORM7 is fed as an input to the LSTM cell at step INLINEFORM8 as shown in part 3 of Figure FIGREF4 . At INLINEFORM9 , we are feeding the output of the mixture module to LSTM. INLINEFORM10 are the predicted question tokens for the input image INLINEFORM11 . Here, we are using INLINEFORM12 and INLINEFORM13 as the special token START and STOP respectively. The softmax probability for the predicted question token at different time steps is given by the following equations where LSTM refers to the standard LSTM cell equations: INLINEFORM14
Where INLINEFORM0 is the probability distribution over all question tokens. INLINEFORM1 is cross entropy loss.
## Cost function
Our objective is to minimize the total loss, that is the sum of cross entropy loss and triplet loss over all training examples. The total loss is: DISPLAYFORM0
where INLINEFORM0 is the total number of samples, INLINEFORM1 is a constant, which controls both the loss. INLINEFORM2 is the triplet loss function EQREF13 . INLINEFORM3 is the cross entropy loss between the predicted and ground truth questions and is given by: INLINEFORM4
where, INLINEFORM0 is the total number of question tokens, INLINEFORM1 is the ground truth label. The code for MDN-VQG model is provided .
## Variations of Proposed Method
While, we advocate the use of multimodal differential network for generating embeddings that can be used by the decoder for generating questions, we also evaluate several variants of this architecture. These are as follows:
Tag Net: In this variant, we consider extracting the part-of-speech (POS) tags for the words present in the caption and obtaining a Tag embedding by considering different methods of combining the one-hot vectors. Further details and experimental results are present in the supplementary. This Tag embedding is then combined with the image embedding and provided to the decoder network.
Place Net: In this variant we explore obtaining embeddings based on the visual scene understanding. This is obtained using a pre-trained PlaceCNN BIBREF45 that is trained to classify 365 different types of scene categories. We then combine the activation map for the input image and the VGG-19 based place embedding to obtain the joint embedding used by the decoder.
Differential Image Network: Instead of using multimodal differential network for generating embeddings, we also evaluate differential image network for the same. In this case, the embedding does not include the caption but is based only on the image feature. We also experimented with using multiple exemplars and random exemplars.
Further details, pseudocode and results regarding these are present in the supplementary material.
## Dataset
We conduct our experiments on Visual Question Generation (VQG) dataset BIBREF5 , which contains human annotated questions based on images of MS-COCO dataset. This dataset was developed for generating natural and engaging questions based on common sense reasoning. We use VQG-COCO dataset for our experiments which contains a total of 2500 training images, 1250 validation images, and 1250 testing images. Each image in the dataset contains five natural questions and five ground truth captions. It is worth noting that the work of BIBREF36 also used the questions from VQA dataset BIBREF1 for training purpose, whereas the work by BIBREF5 uses only the VQG-COCO dataset. VQA-1.0 dataset is also built on images from MS-COCO dataset. It contains a total of 82783 images for training, 40504 for validation and 81434 for testing. Each image is associated with 3 questions. We used pretrained caption generation model BIBREF13 to extract captions for VQA dataset as the human annotated captions are not there in the dataset. We also get good results on the VQA dataset (as shown in Table TABREF26 ) which shows that our method doesn't necessitate the presence of ground truth captions. We train our model separately for VQG-COCO and VQA dataset.
## Inference
We made use of the 1250 validation images to tune the hyperparameters and are providing the results on test set of VQG-COCO dataset. During inference, We use the Representation module to find the embeddings for the image and ground truth caption without using the supporting and contrasting exemplars. The mixture module provides the joint representation of the target image and ground truth caption. Finally, the decoder takes in the joint features and generates the question. We also experimented with the captions generated by an Image-Captioning network BIBREF13 for VQG-COCO dataset and the result for that and training details are present in the supplementary material.
## Experiments
We evaluate our proposed MDN method in the following ways: First, we evaluate it against other variants described in section SECREF19 and SECREF10 . Second, we further compare our network with state-of-the-art methods for VQA 1.0 and VQG-COCO dataset. We perform a user study to gauge human opinion on naturalness of the generated question and analyze the word statistics in Figure FIGREF22 . This is an important test as humans are the best deciders of naturalness. We further consider the statistical significance for the various ablations as well as the state-of-the-art models. The quantitative evaluation is conducted using standard metrics like BLEU BIBREF46 , METEOR BIBREF47 , ROUGE BIBREF48 , CIDEr BIBREF49 . Although these metrics have not been shown to correlate with `naturalness' of the question these still provide a reasonable quantitative measure for comparison. Here we only provide the BLEU1 scores, but the remaining BLEU-n metric scores are present in the supplementary. We observe that the proposed MDN provides improved embeddings to the decoder. We believe that these embeddings capture instance specific differential information that helps in guiding the question generation. Details regarding the metrics are given in the supplementary material.
## Ablation Analysis
We considered different variations of our method mentioned in section SECREF19 and the various ways to obtain the joint multimodal embedding as described in section SECREF10 . The results for the VQG-COCO test set are given in table TABREF24 . In this table, every block provides the results for one of the variations of obtaining the embeddings and different ways of combining them. We observe that the Joint Method (JM) of combining the embeddings works the best in all cases except the Tag Embeddings. Among the ablations, the proposed MDN method works way better than the other variants in terms of BLEU, METEOR and ROUGE metrics by achieving an improvement of 6%, 12% and 18% in the scores respectively over the best other variant.
## Baseline and State-of-the-Art
The comparison of our method with various baselines and state-of-the-art methods is provided in table TABREF26 for VQA 1.0 and table TABREF27 for VQG-COCO dataset. The comparable baselines for our method are the image based and caption based models in which we use either only the image or the caption embedding and generate the question. In both the tables, the first block consists of the current state-of-the-art methods on that dataset and the second contains the baselines. We observe that for the VQA dataset we achieve an improvement of 8% in BLEU and 7% in METEOR metric scores over the baselines, whereas for VQG-COCO dataset this is 15% for both the metrics. We improve over the previous state-of-the-art BIBREF35 for VQA dataset by around 6% in BLEU score and 10% in METEOR score. In the VQG-COCO dataset, we improve over BIBREF5 by 3.7% and BIBREF36 by 3.5% in terms of METEOR scores.
## Statistical Significance Analysis
We have analysed Statistical Significance BIBREF50 of our MDN model for VQG for different variations of the mixture module mentioned in section SECREF10 and also against the state-of-the-art methods. The Critical Difference (CD) for Nemenyi BIBREF51 test depends upon the given INLINEFORM0 (confidence level, which is 0.05 in our case) for average ranks and N (number of tested datasets). If the difference in the rank of the two methods lies within CD, then they are not significantly different and vice-versa. Figure FIGREF29 visualizes the post-hoc analysis using the CD diagram. From the figure, it is clear that MDN-Joint works best and is statistically significantly different from the state-of-the-art methods.
## Perceptual Realism
A human is the best judge of naturalness of any question, We evaluated our proposed MDN method using a `Naturalness' Turing test BIBREF52 on 175 people. People were shown an image with 2 questions just as in figure FIGREF1 and were asked to rate the naturalness of both the questions on a scale of 1 to 5 where 1 means `Least Natural' and 5 is the `Most Natural'. We provided 175 people with 100 such images from the VQG-COCO validation dataset which has 1250 images. Figure FIGREF30 indicates the number of people who were fooled (rated the generated question more or equal to the ground truth question). For the 100 images, on an average 59.7% people were fooled in this experiment and this shows that our model is able to generate natural questions.
## Conclusion
In this paper we have proposed a novel method for generating natural questions for an image. The approach relies on obtaining multimodal differential embeddings from image and its caption. We also provide ablation analysis and a detailed comparison with state-of-the-art methods, perform a user study to evaluate the naturalness of our generated questions and also ensure that the results are statistically significant. In future, we would like to analyse means of obtaining composite embeddings. We also aim to consider the generalisation of this approach to other vision and language tasks. Supplementary Material Section SECREF8 will provide details about training configuration for MDN, Section SECREF9 will explain the various Proposed Methods and we also provide a discussion in section regarding some important questions related to our method. We report BLEU1, BLEU2, BLEU3, BLEU4, METEOR, ROUGE and CIDER metric scores for VQG-COCO dataset. We present different experiments with Tag Net in which we explore the performance of various tags (Noun, Verb, and Question tags) and different ways of combining them to get the context vectors. Multimodal Differential Network [1] MDN INLINEFORM0 Finding Exemplars: INLINEFORM1 INLINEFORM2 Compute Triplet Embedding: INLINEFORM3 INLINEFORM4 Compute Triplet Fusion Embedding : INLINEFORM5 INLINEFORM6 INLINEFORM7 Compute Triplet Loss: INLINEFORM8 Compute Decode Question Sentence: INLINEFORM9 INLINEFORM10 —————————————————– Triplet Fusion INLINEFORM11 , INLINEFORM12 INLINEFORM13 :Image feature,14x14x512 INLINEFORM14 : Caption feature,1x512 Match Dimension: INLINEFORM15 ,196x512 INLINEFORM16 196x512 If flag==Joint Fusion: INLINEFORM17 INLINEFORM18 , [ INLINEFORM19 (MDN-Mul), INLINEFORM20 (MDN-Add)] If flag==Attention Fusion : INLINEFORM21 Semb INLINEFORM22 Dataset and Training Details Dataset We conduct our experiments on two types of dataset: VQA dataset BIBREF1 , which contains human annotated questions based on images on MS-COCO dataset. Second one is VQG-COCO dataset based on natural question BIBREF55 . VQA dataset VQA dataset BIBREF1 is built on complex images from MS-COCO dataset. It contains a total of 204721 images, out of which 82783 are for training, 40504 for validation and 81434 for testing. Each image in the MS-COCO dataset is associated with 3 questions and each question has 10 possible answers. So there are 248349 QA pair for training, 121512 QA pairs for validating and 244302 QA pairs for testing. We used pre-trained caption generation model BIBREF53 to extract captions for VQA dataset. VQG dataset The VQG-COCO dataset BIBREF55 , is developed for generating natural and engaging questions that are based on common sense reasoning. This dataset contains a total of 2500 training images, 1250 validation images and 1250 testing images. Each image in the dataset contains 5 natural questions. Training Configuration We have used RMSPROP optimizer to update the model parameter and configured hyper-parameter values to be as follows: INLINEFORM23 to train the classification network . In order to train a triplet model, we have used RMSPROP to optimize the triplet model model parameter and configure hyper-parameter values to be: INLINEFORM24 . We also used learning rate decay to decrease the learning rate on every epoch by a factor given by: INLINEFORM25 where values of a=1500 and b=1250 are set empirically. Ablation Analysis of Model While, we advocate the use of multimodal differential network (MDN) for generating embeddings that can be used by the decoder for generating questions, we also evaluate several variants of this architecture namely (a) Differential Image Network, (b) Tag net and (c) Place net. These are described in detail as follows: Differential Image Network For obtaining the exemplar image based context embedding, we propose a triplet network consist of three network, one is target net, supporting net and opposing net. All these three networks designed with convolution neural network and shared the same parameters. The weights of this network are learnt through end-to-end learning using a triplet loss. The aim is to obtain latent weight vectors that bring the supporting exemplar close to the target image and enhances the difference between opposing examples. More formally, given an image INLINEFORM26 we obtain an embedding INLINEFORM27 using a CNN that we parameterize through a function INLINEFORM28 where INLINEFORM29 are the weights of the CNN. This is illustrated in figure FIGREF43 . Tag net The tag net consists of two parts Context Extractor & Tag Embedding Net. This is illustrated in figure FIGREF45 . Extract Context: The first step is to extract the caption of the image using NeuralTalk2 BIBREF53 model. We find the part-of-speech(POS) tag present in the caption. POS taggers have been developed for two well known corpuses, the Brown Corpus and the Penn Treebanks. For our work, we are using the Brown Corpus tags. The tags are clustered into three category namely Noun tag, Verb tag and Question tags (What, Where, ...). Noun tag consists of all the noun & pronouns present in the caption sentence and similarly, verb tag consists of verb & adverbs present in the caption sentence. The question tags consists of the 7-well know question words i.e., why, how, what, when, where, who and which. Each tag token is represented as a one-hot vector of the dimension of vocabulary size. For generalization, we have considered 5 tokens from each category of the Tags. Tag Embedding Net: The embedding network consists of word embedding followed by temporal convolutions neural network followed by max-pooling network. In the first step, sparse high dimensional one-hot vector is transformed to dense low dimension vector using word embedding. After this, we apply temporal convolution on the word embedding vector. The uni-gram, bi-gram and tri-gram feature are computed by applying convolution filter of size 1, 2 and 3 respectability. Finally, we applied max-pooling on this to get a vector representation of the tags as shown figure FIGREF45 . We concatenated all the tag words followed by fully connected layer to get feature dimension of 512. We also explored joint networks based on concatenation of all the tags, on element-wise addition and element-wise multiplication of the tag vectors. However, we observed that convolution over max pooling and joint concatenation gives better performance based on CIDer score. INLINEFORM30 Where, T_CNN is Temporally Convolution Neural Network applied on word embedding vector with kernel size three. Place net Visual object and scene recognition plays a crucial role in the image. Here, places in the image are labeled with scene semantic categories BIBREF45 , comprise of large and diverse type of environment in the world, such as (amusement park, tower, swimming pool, shoe shop, cafeteria, rain-forest, conference center, fish pond, etc.). So we have used different type of scene semantic categories present in the image as a place based context to generate natural question. A place365 is a convolution neural network is modeled to classify 365 types of scene categories, which is trained on the place2 dataset consist of 1.8 million of scene images. We have used a pre-trained VGG16-places365 network to obtain place based context embedding feature for various type scene categories present in the image. The context feature INLINEFORM31 is obtained by: INLINEFORM32 Where, INLINEFORM33 is Place365_CNN. We have extracted INLINEFORM34 features of dimension 14x14x512 for attention model and FC8 features of dimension 365 for joint, addition and hadamard model of places365. Finally, we use a linear transformation to obtain a 512 dimensional vector. We explored using the CONV5 having feature dimension 14x14 512, FC7 having 4096 and FC8 having feature dimension of 365 of places365. Ablation Analysis Sampling Exemplar: KNN vs ITML Our method is aimed at using efficient exemplar-based retrieval techniques. We have experimented with various exemplar methods, such as ITML BIBREF40 based metric learning for image features and KNN based approaches. We observed KNN based approach (K-D tree) with Euclidean metric is a efficient method for finding exemplars. Also we observed that ITML is computationally expensive and also depends on the training procedure. The table provides the experimental result for Differential Image Network variant with k (number of exemplars) = 2 and Hadamard method: Question Generation approaches: Sampling vs Argmax We obtained the decoding using standard practice followed in the literature BIBREF38 . This method selects the argmax sentence. Also, we evaluated our method by sampling from the probability distributions and provide the results for our proposed MDN-Joint method for VQG dataset as follows: How are exemplars improving Embedding In Multimodel differential network, we use exemplars and train them using a triplet loss. It is known that using a triplet network, we can learn a representation that accentuates how the image is closer to a supporting exemplar as against the opposing exemplar BIBREF42 , BIBREF41 . The Joint embedding is obtained between the image and language representations. Therefore the improved representation helps in obtaining an improved context vector. Further we show that this also results in improving VQG. Are exemplars required? We had similar concerns and validated this point by using random exemplars for the nearest neighbor for MDN. (k=R in table TABREF35 ) In this case the method is similar to the baseline. This suggests that with random exemplar, the model learns to ignore the cue. Are captions necessary for our method? This is not actually necessary. In our method, we have used an existing image captioning method BIBREF13 to generate captions for images that did not have them. For VQG dataset, captions were available and we have used that, but, for VQA dataset captions were not available and we have generated captions while training. We provide detailed evidence with respect to caption-question pairs to ensure that we are generating novel questions. While the caption generates scene description, our proposed method generates semantically meaningful and novel questions. Examples for Figure 1 of main paper: First Image:- Caption- A young man skateboarding around little cones. Our Question- Is this a skateboard competition? Second Image:- Caption- A small child is standing on a pair of skis. Our Question:- How old is that little girl? Intuition behind Triplet Network: The intuition behind use of triplet networks is clear through this paper BIBREF41 that first advocated its use. The main idea is that when we learn distance functions that are “close” for similar and “far” from dissimilar representations, it is not clear that close and far are with respect to what measure. By incorporating a triplet we learn distance functions that learn that “A is more similar to B as compared to C”. Learning such measures allows us to bring target image-caption joint embeddings that are closer to supporting exemplars as compared to contrasting exemplars. Analysis of Network Analysis of Tag Context Tag is language based context. These tags are extracted from caption, except question-tags which is fixed as the 7 'Wh words' (What, Why, Where, Who, When, Which and How). We have experimented with Noun tag, Verb tag and 'Wh-word' tag as shown in tables. Also, we have experimented in each tag by varying the number of tags from 1 to 7. We combined different tags using 1D-convolution, concatenation, and addition of all the tags and observed that the concatenation mechanism gives better results. As we can see in the table TABREF33 that taking Nouns, Verbs and Wh-Words as context, we achieve significant improvement in the BLEU, METEOR and CIDEr scores from the basic models which only takes the image and the caption respectively. Taking Nouns generated from the captions and questions of the corresponding training example as context, we achieve an increase of 1.6% in Bleu Score and 2% in METEOR and 34.4% in CIDEr Score from the basic Image model. Similarly taking Verbs as context gives us an increase of 1.3% in Bleu Score and 2.1% in METEOR and 33.5% in CIDEr Score from the basic Image model. And the best result comes when we take 3 Wh-Words as context and apply the Hadamard Model with concatenating the 3 WH-words. Also in Table TABREF34 we have shown the results when we take more than one words as context. Here we show that for 3 words i.e 3 nouns, 3 verbs and 3 Wh-words, the Concatenation model performs the best. In this table the conv model is using 1D convolution to combine the tags and the joint model combine all the tags. Analysis of Context: Exemplars In Multimodel Differential Network and Differential Image Network, we use exemplar images(target, supporting and opposing image) to obtain the differential context. We have performed the experiment based on the single exemplar(K=1), which is one supporting and one opposing image along with target image, based on two exemplar(K=2), i.e. two supporting and two opposing image along with single target image. similarly we have performed experiment for K=3 and K=4 as shown in table- TABREF35 . Mixture Module: Other Variations Hadamard method uses element-wise multiplication whereas Addition method uses element-wise addition in place of the concatenation operator of the Joint method. The Hadamard method finds a correlation between image feature and caption feature vector while the Addition method learns a resultant vector. In the attention method, the output INLINEFORM35 is the weighted average of attention probability vector INLINEFORM36 and convolutional features INLINEFORM37 . The attention probability vector weights the contribution of each convolutional feature based on the caption vector. This attention method is similar to work stack attention method BIBREF54 . The attention mechanism is given by: DISPLAYFORM0 where INLINEFORM38 is the 14x14x512-dimensional convolution feature map from the fifth convolution layer of VGG-19 Net of image INLINEFORM39 and INLINEFORM40 is the caption context vector. The attention probability vector INLINEFORM41 is a 196-dimensional vector. INLINEFORM42 are the weights and INLINEFORM43 is the bias for different layers. We evaluate the different approaches and provide results for the same. Here INLINEFORM44 represents element-wise addition. Evaluation Metrics Our task is similar to encoder -decoder framework of machine translation. we have used same evaluation metric is used in machine translation. BLEU BIBREF46 is the first metric to find the correlation between generated question with ground truth question. BLEU score is used to measure the precision value, i.e That is how much words in the predicted question is appeared in reference question. BLEU-n score measures the n-gram precision for counting co-occurrence on reference sentences. we have evaluated BLEU score from n is 1 to 4. The mechanism of ROUGE-n BIBREF48 score is similar to BLEU-n,where as, it measures recall value instead of precision value in BLEU. That is how much words in the reference question is appeared in predicted question.Another version ROUGE metric is ROUGE-L, which measures longest common sub-sequence present in the generated question. METEOR BIBREF47 score is another useful evaluation metric to calculate the similarity between generated question with reference one by considering synonyms, stemming and paraphrases. the output of the METEOR score measure the word matches between predicted question and reference question. In VQG, it compute the word match score between predicted question with five reference question. CIDer BIBREF49 score is a consensus based evaluation metric. It measure human-likeness, that is the sentence is written by human or not. The consensus is measured, how often n-grams in the predicted question are appeared in the reference question. If the n-grams in the predicted question sentence is appeared more frequently in reference question then question is less informative and have low CIDer score. We provide our results using all these metrics and compare it with existing baselines.
| 16
|
1809.04960
|
Unsupervised Machine Commenting with Neural Variational Topic Model
|
# Unsupervised Machine Commenting with Neural Variational Topic Model
## Abstract
Article comments can provide supplementary opinions and facts for readers, thereby increase the attraction and engagement of articles. Therefore, automatically commenting is helpful in improving the activeness of the community, such as online forums and news websites. Previous work shows that training an automatic commenting system requires large parallel corpora. Although part of articles are naturally paired with the comments on some websites, most articles and comments are unpaired on the Internet. To fully exploit the unpaired data, we completely remove the need for parallel data and propose a novel unsupervised approach to train an automatic article commenting model, relying on nothing but unpaired articles and comments. Our model is based on a retrieval-based commenting framework, which uses news to retrieve comments based on the similarity of their topics. The topic representation is obtained from a neural variational topic model, which is trained in an unsupervised manner. We evaluate our model on a news comment dataset. Experiments show that our proposed topic-based approach significantly outperforms previous lexicon-based models. The model also profits from paired corpora and achieves state-of-the-art performance under semi-supervised scenarios.
## Introduction
Making article comments is a fundamental ability for an intelligent machine to understand the article and interact with humans. It provides more challenges because commenting requires the abilities of comprehending the article, summarizing the main ideas, mining the opinions, and generating the natural language. Therefore, machine commenting is an important problem faced in building an intelligent and interactive agent. Machine commenting is also useful in improving the activeness of communities, including online forums and news websites. Article comments can provide extended information and external opinions for the readers to have a more comprehensive understanding of the article. Therefore, an article with more informative and interesting comments will attract more attention from readers. Moreover, machine commenting can kick off the discussion about an article or a topic, which helps increase user engagement and interaction between the readers and authors.
Because of the advantage and importance described above, more recent studies have focused on building a machine commenting system with neural models BIBREF0 . One bottleneck of neural machine commenting models is the requirement of a large parallel dataset. However, the naturally paired commenting dataset is loosely paired. Qin et al. QinEA2018 were the first to propose the article commenting task and an article-comment dataset. The dataset is crawled from a news website, and they sample 1,610 article-comment pairs to annotate the relevance score between articles and comments. The relevance score ranges from 1 to 5, and we find that only 6.8% of the pairs have an average score greater than 4. It indicates that the naturally paired article-comment dataset contains a lot of loose pairs, which is a potential harm to the supervised models. Besides, most articles and comments are unpaired on the Internet. For example, a lot of articles do not have the corresponding comments on the news websites, and the comments regarding the news are more likely to appear on social media like Twitter. Since comments on social media are more various and recent, it is important to exploit these unpaired data.
Another issue is that there is a semantic gap between articles and comments. In machine translation and text summarization, the target output mainly shares the same points with the source input. However, in article commenting, the comment does not always tell the same thing as the corresponding article. Table TABREF1 shows an example of an article and several corresponding comments. The comments do not directly tell what happened in the news, but talk about the underlying topics (e.g. NBA Christmas Day games, LeBron James). However, existing methods for machine commenting do not model the topics of articles, which is a potential harm to the generated comments.
To this end, we propose an unsupervised neural topic model to address both problems. For the first problem, we completely remove the need of parallel data and propose a novel unsupervised approach to train a machine commenting system, relying on nothing but unpaired articles and comments. For the second issue, we bridge the articles and comments with their topics. Our model is based on a retrieval-based commenting framework, which uses the news as the query to retrieve the comments by the similarity of their topics. The topic is represented with a variational topic, which is trained in an unsupervised manner.
The contributions of this work are as follows:
## Machine Commenting
In this section, we highlight the research challenges of machine commenting, and provide some solutions to deal with these challenges.
## Challenges
Here, we first introduce the challenges of building a well-performed machine commenting system.
The generative model, such as the popular sequence-to-sequence model, is a direct choice for supervised machine commenting. One can use the title or the content of the article as the encoder input, and the comments as the decoder output. However, we find that the mode collapse problem is severed with the sequence-to-sequence model. Despite the input articles being various, the outputs of the model are very similar. The reason mainly comes from the contradiction between the complex pattern of generating comments and the limited parallel data. In other natural language generation tasks, such as machine translation and text summarization, the target output of these tasks is strongly related to the input, and most of the required information is involved in the input text. However, the comments are often weakly related to the input articles, and part of the information in the comments is external. Therefore, it requires much more paired data for the supervised model to alleviate the mode collapse problem.
One article can have multiple correct comments, and these comments can be very semantically different from each other. However, in the training set, there is only a part of the correct comments, so the other correct comments will be falsely regarded as the negative samples by the supervised model. Therefore, many interesting and informative comments will be discouraged or neglected, because they are not paired with the articles in the training set.
There is a semantic gap between articles and comments. In machine translation and text summarization, the target output mainly shares the same points with the source input. However, in article commenting, the comments often have some external information, or even tell an opposite opinion from the articles. Therefore, it is difficult to automatically mine the relationship between articles and comments.
## Solutions
Facing the above challenges, we provide three solutions to the problems.
Given a large set of candidate comments, the retrieval model can select some comments by matching articles with comments. Compared with the generative model, the retrieval model can achieve more promising performance. First, the retrieval model is less likely to suffer from the mode collapse problem. Second, the generated comments are more predictable and controllable (by changing the candidate set). Third, the retrieval model can be combined with the generative model to produce new comments (by adding the outputs of generative models to the candidate set).
The unsupervised learning method is also important for machine commenting to alleviate the problems descried above. Unsupervised learning allows the model to exploit more data, which helps the model to learn more complex patterns of commenting and improves the generalization of the model. Many comments provide some unique opinions, but they do not have paired articles. For example, many interesting comments on social media (e.g. Twitter) are about recent news, but require redundant work to match these comments with the corresponding news articles. With the help of the unsupervised learning method, the model can also learn to generate these interesting comments. Additionally, the unsupervised learning method does not require negative samples in the training stage, so that it can alleviate the negative sampling bias.
Although there is semantic gap between the articles and the comments, we find that most articles and comments share the same topics. Therefore, it is possible to bridge the semantic gap by modeling the topics of both articles and comments. It is also similar to how humans generate comments. Humans do not need to go through the whole article but are capable of making a comment after capturing the general topics.
## Proposed Approach
We now introduce our proposed approach as an implementation of the solutions above. We first give the definition and the denotation of the problem. Then, we introduce the retrieval-based commenting framework. After that, a neural variational topic model is introduced to model the topics of the comments and the articles. Finally, semi-supervised training is used to combine the advantage of both supervised and unsupervised learning.
## Retrieval-based Commenting
Given an article, the retrieval-based method aims to retrieve a comment from a large pool of candidate comments. The article consists of a title INLINEFORM0 and a body INLINEFORM1 . The comment pool is formed from a large scale of candidate comments INLINEFORM2 , where INLINEFORM3 is the number of the unique comments in the pool. In this work, we have 4.5 million human comments in the candidate set, and the comments are various, covering different topics from pets to sports.
The retrieval-based model should score the matching between the upcoming article and each comments, and return the comments which is matched with the articles the most. Therefore, there are two main challenges in retrieval-based commenting. One is how to evaluate the matching of the articles and comments. The other is how to efficiently compute the matching scores because the number of comments in the pool is large.
To address both problems, we select the “dot-product” operation to compute matching scores. More specifically, the model first computes the representations of the article INLINEFORM0 and the comments INLINEFORM1 . Then the score between article INLINEFORM2 and comment INLINEFORM3 is computed with the “dot-product” operation: DISPLAYFORM0
The dot-product scoring method has proven a successful in a matching model BIBREF1 . The problem of finding datapoints with the largest dot-product values is called Maximum Inner Product Search (MIPS), and there are lots of solutions to improve the efficiency of solving this problem. Therefore, even when the number of candidate comments is very large, the model can still find comments with the highest efficiency. However, the study of the MIPS is out of the discussion in this work. We refer the readers to relevant articles for more details about the MIPS BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 . Another advantage of the dot-product scoring method is that it does not require any extra parameters, so it is more suitable as a part of the unsupervised model.
## Neural Variational Topic Model
We obtain the representations of articles INLINEFORM0 and comments INLINEFORM1 with a neural variational topic model. The neural variational topic model is based on the variational autoencoder framework, so it can be trained in an unsupervised manner. The model encodes the source text into a representation, from which it reconstructs the text.
We concatenate the title and the body to represent the article. In our model, the representations of the article and the comment are obtained in the same way. For simplicity, we denote both the article and the comment as “document”. Since the articles are often very long (more than 200 words), we represent the documents into bag-of-words, for saving both the time and memory cost. We denote the bag-of-words representation as INLINEFORM0 , where INLINEFORM1 is the one-hot representation of the word at INLINEFORM2 position, and INLINEFORM3 is the number of words in the vocabulary. The encoder INLINEFORM4 compresses the bag-of-words representations INLINEFORM5 into topic representations INLINEFORM6 : DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , and INLINEFORM3 are the trainable parameters. Then the decoder INLINEFORM4 reconstructs the documents by independently generating each words in the bag-of-words: DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 is the number of words in the bag-of-words, and INLINEFORM1 is a trainable matrix to map the topic representation into the word distribution.
In order to model the topic information, we use a Dirichlet prior rather than the standard Gaussian prior. However, it is difficult to develop an effective reparameterization function for the Dirichlet prior to train VAE. Therefore, following BIBREF6 , we use the Laplace approximation BIBREF7 to Dirichlet prior INLINEFORM0 : DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 denotes the logistic normal distribution, INLINEFORM1 is the number of topics, and INLINEFORM2 is a parameter vector. Then, the variational lower bound is written as: DISPLAYFORM0
where the first term is the KL-divergence loss and the second term is the reconstruction loss. The mean INLINEFORM0 and the variance INLINEFORM1 are computed as follows: DISPLAYFORM0 DISPLAYFORM1
We use the INLINEFORM0 and INLINEFORM1 to generate the samples INLINEFORM2 by sampling INLINEFORM3 , from which we reconstruct the input INLINEFORM4 .
At the training stage, we train the neural variational topic model with the Eq. EQREF22 . At the testing stage, we use INLINEFORM0 to compute the topic representations of the article INLINEFORM1 and the comment INLINEFORM2 .
## Training
In addition to the unsupervised training, we explore a semi-supervised training framework to combine the proposed unsupervised model and the supervised model. In this scenario we have a paired dataset that contains article-comment parallel contents INLINEFORM0 , and an unpaired dataset that contains the documents (articles or comments) INLINEFORM1 . The supervised model is trained on INLINEFORM2 so that we can learn the matching or mapping between articles and comments. By sharing the encoder of the supervised model and the unsupervised model, we can jointly train both the models with a joint objective function: DISPLAYFORM0
where INLINEFORM0 is the loss function of the unsupervised learning (Eq. refloss), INLINEFORM1 is the loss function of the supervised learning (e.g. the cross-entropy loss of Seq2Seq model), and INLINEFORM2 is a hyper-parameter to balance two parts of the loss function. Hence, the model is trained on both unpaired data INLINEFORM3 , and paired data INLINEFORM4 .
## Datasets
We select a large-scale Chinese dataset BIBREF0 with millions of real comments and a human-annotated test set to evaluate our model. The dataset is collected from Tencent News, which is one of the most popular Chinese websites for news and opinion articles. The dataset consists of 198,112 news articles. Each piece of news contains a title, the content of the article, and a list of the users' comments. Following the previous work BIBREF0 , we tokenize all text with the popular python package Jieba, and filter out short articles with less than 30 words in content and those with less than 20 comments. The dataset is split into training/validation/test sets, and they contain 191,502/5,000/1,610 pieces of news, respectively. The whole dataset has a vocabulary size of 1,858,452. The average lengths of the article titles and content are 15 and 554 Chinese words. The average comment length is 17 words.
## Implementation Details
The hidden size of the model is 512, and the batch size is 64. The number of topics INLINEFORM0 is 100. The weight INLINEFORM1 in Eq. EQREF26 is 1.0 under the semi-supervised setting. We prune the vocabulary, and only leave 30,000 most frequent words in the vocabulary. We train the model for 20 epochs with the Adam optimizing algorithms BIBREF8 . In order to alleviate the KL vanishing problem, we set the initial learning to INLINEFORM2 , and use batch normalization BIBREF9 in each layer. We also gradually increase the KL term from 0 to 1 after each epoch.
## Baselines
We compare our model with several unsupervised models and supervised models.
Unsupervised baseline models are as follows:
TF-IDF (Lexical, Non-Neural) is an important unsupervised baseline. We use the concatenation of the title and the body as the query to retrieve the candidate comment set by means of the similarity of the tf-idf value. The model is trained on unpaired articles and comments, which is the same as our proposed model.
LDA (Topic, Non-Neural) is a popular unsupervised topic model, which discovers the abstract "topics" that occur in a collection of documents. We train the LDA with the articles and comments in the training set. The model retrieves the comments by the similarity of the topic representations.
NVDM (Lexical, Neural) is a VAE-based approach for document modeling BIBREF10 . We compare our model with this baseline to demonstrate the effect of modeling topic.
The supervised baseline models are:
S2S (Generative) BIBREF11 is a supervised generative model based on the sequence-to-sequence network with the attention mechanism BIBREF12 . The model uses the titles and the bodies of the articles as the encoder input, and generates the comments with the decoder.
IR (Retrieval) BIBREF0 is a supervised retrieval-based model, which trains a convolutional neural network (CNN) to take the articles and a comment as inputs, and output the relevance score. The positive instances for training are the pairs in the training set, and the negative instances are randomly sampled using the negative sampling technique BIBREF13 .
## Retrieval Evaluation
For text generation, automatically evaluate the quality of the generated text is an open problem. In particular, the comment of a piece of news can be various, so it is intractable to find out all the possible references to be compared with the model outputs. Inspired by the evaluation methods of dialogue models, we formulate the evaluation as a ranking problem. Given a piece of news and a set of candidate comments, the comment model should return the rank of the candidate comments. The candidate comment set consists of the following parts:
Correct: The ground-truth comments of the corresponding news provided by the human.
Plausible: The 50 most similar comments to the news. We use the news as the query to retrieve the comments that appear in the training set based on the cosine similarity of their tf-idf values. We select the top 50 comments that are not the correct comments as the plausible comments.
Popular: The 50 most popular comments from the dataset. We count the frequency of each comments in the training set, and select the 50 most frequent comments to form the popular comment set. The popular comments are the general and meaningless comments, such as “Yes”, “Great”, “That's right', and “Make Sense”. These comments are dull and do not carry any information, so they are regarded as incorrect comments.
Random: After selecting the correct, plausible, and popular comments, we fill the candidate set with randomly selected comments from the training set so that there are 200 unique comments in the candidate set.
Following previous work, we measure the rank in terms of the following metrics:
Recall@k: The proportion of human comments found in the top-k recommendations.
Mean Rank (MR): The mean rank of the human comments.
Mean Reciprocal Rank (MRR): The mean reciprocal rank of the human comments.
The evaluation protocol is compatible with both retrieval models and generative models. The retrieval model can directly rank the comments by assigning a score for each comment, while the generative model can rank the candidates by the model's log-likelihood score.
Table TABREF31 shows the performance of our models and the baselines in retrieval evaluation. We first compare our proposed model with other popular unsupervised methods, including TF-IDF, LDA, and NVDM. TF-IDF retrieves the comments by similarity of words rather than the semantic meaning, so it achieves low scores on all the retrieval metrics. The neural variational document model is based on the neural VAE framework. It can capture the semantic information, so it has better performance than the TF-IDF model. LDA models the topic information, and captures the deeper relationship between the article and comments, so it achieves improvement in all relevance metrics. Finally, our proposed model outperforms all these unsupervised methods, mainly because the proposed model learns both the semantics and the topic information.
We also evaluate two popular supervised models, i.e. seq2seq and IR. Since the articles are very long, we find either RNN-based or CNN-based encoders cannot hold all the words in the articles, so it requires limiting the length of the input articles. Therefore, we use an MLP-based encoder, which is the same as our model, to encode the full length of articles. In our preliminary experiments, the MLP-based encoder with full length articles achieves better scores than the RNN/CNN-based encoder with limited length articles. It shows that the seq2seq model gets low scores on all relevant metrics, mainly because of the mode collapse problem as described in Section Challenges. Unlike seq2seq, IR is based on a retrieval framework, so it achieves much better performance.
## Generative Evaluation
Following previous work BIBREF0 , we evaluate the models under the generative evaluation setting. The retrieval-based models generate the comments by selecting a comment from the candidate set. The candidate set contains the comments in the training set. Unlike the retrieval evaluation, the reference comments may not appear in the candidate set, which is closer to real-world settings. Generative-based models directly generate comments without a candidate set. We compare the generated comments of either the retrieval-based models or the generative models with the five reference comments. We select four popular metrics in text generation to compare the model outputs with the references: BLEU BIBREF14 , METEOR BIBREF15 , ROUGE BIBREF16 , CIDEr BIBREF17 .
Table TABREF32 shows the performance for our models and the baselines in generative evaluation. Similar to the retrieval evaluation, our proposed model outperforms the other unsupervised methods, which are TF-IDF, NVDM, and LDA, in generative evaluation. Still, the supervised IR achieves better scores than the seq2seq model. With the help of our proposed model, both IR and S2S achieve an improvement under the semi-supervised scenarios.
## Analysis and Discussion
We analyze the performance of the proposed method under the semi-supervised setting. We train the supervised IR model with different numbers of paired data. Figure FIGREF39 shows the curve (blue) of the recall1 score. As expected, the performance grows as the paired dataset becomes larger. We further combine the supervised IR with our unsupervised model, which is trained with full unpaired data (4.8M) and different number of paired data (from 50K to 4.8M). It shows that IR+Proposed can outperform the supervised IR model given the same paired dataset. It concludes that the proposed model can exploit the unpaired data to further improve the performance of the supervised model.
Although our proposed model can achieve better performance than previous models, there are still remaining two questions: why our model can outperform them, and how to further improve the performance. To address these queries, we perform error analysis to analyze the error types of our model and the baseline models. We select TF-IDF, S2S, and IR as the representative baseline models. We provide 200 unique comments as the candidate sets, which consists of four types of comments as described in the above retrieval evaluation setting: Correct, Plausible, Popular, and Random. We rank the candidate comment set with four models (TF-IDF, S2S, IR, and Proposed+IR), and record the types of top-1 comments.
Figure FIGREF40 shows the percentage of different types of top-1 comments generated by each model. It shows that TF-IDF prefers to rank the plausible comments as the top-1 comments, mainly because it matches articles with the comments based on the similarity of the lexicon. Therefore, the plausible comments, which are more similar in the lexicon, are more likely to achieve higher scores than the correct comments. It also shows that the S2S model is more likely to rank popular comments as the top-1 comments. The reason is the S2S model suffers from the mode collapse problem and data sparsity, so it prefers short and general comments like “Great” or “That's right”, which appear frequently in the training set. The correct comments often contain new information and different language models from the training set, so they do not obtain a high score from S2S.
IR achieves better performance than TF-IDF and S2S. However, it still suffers from the discrimination between the plausible comments and correct comments. This is mainly because IR does not explicitly model the underlying topics. Therefore, the correct comments which are more relevant in topic with the articles get lower scores than the plausible comments which are more literally relevant with the articles. With the help of our proposed model, proposed+IR achieves the best performance, and achieves a better accuracy to discriminate the plausible comments and the correct comments. Our proposed model incorporates the topic information, so the correct comments which are more similar to the articles in topic obtain higher scores than the other types of comments. According to the analysis of the error types of our model, we still need to focus on avoiding predicting the plausible comments.
## Article Comment
There are few studies regarding machine commenting. Qin et al. QinEA2018 is the first to propose the article commenting task and a dataset, which is used to evaluate our model in this work. More studies about the comments aim to automatically evaluate the quality of the comments. Park et al. ParkSDE16 propose a system called CommentIQ, which assist the comment moderators in identifying high quality comments. Napoles et al. NapolesTPRP17 propose to discriminating engaging, respectful, and informative conversations. They present a Yahoo news comment threads dataset and annotation scheme for the new task of identifying “good” online conversations. More recently, Kolhaatkar and Taboada KolhatkarT17 propose a model to classify the comments into constructive comments and non-constructive comments. In this work, we are also inspired by the recent related work of natural language generation models BIBREF18 , BIBREF19 .
## Topic Model and Variational Auto-Encoder
Topic models BIBREF20 are among the most widely used models for learning unsupervised representations of text. One of the most popular approaches for modeling the topics of the documents is the Latent Dirichlet Allocation BIBREF21 , which assumes a discrete mixture distribution over topics is sampled from a Dirichlet prior shared by all documents. In order to explore the space of different modeling assumptions, some black-box inference methods BIBREF22 , BIBREF23 are proposed and applied to the topic models.
Kingma and Welling vae propose the Variational Auto-Encoder (VAE) where the generative model and the variational posterior are based on neural networks. VAE has recently been applied to modeling the representation and the topic of the documents. Miao et al. NVDM model the representation of the document with a VAE-based approach called the Neural Variational Document Model (NVDM). However, the representation of NVDM is a vector generated from a Gaussian distribution, so it is not very interpretable unlike the multinomial mixture in the standard LDA model. To address this issue, Srivastava and Sutton nvlda propose the NVLDA model that replaces the Gaussian prior with the Logistic Normal distribution to approximate the Dirichlet prior and bring the document vector into the multinomial space. More recently, Nallapati et al. sengen present a variational auto-encoder approach which models the posterior over the topic assignments to sentences using an RNN.
## Conclusion
We explore a novel way to train a machine commenting model in an unsupervised manner. According to the properties of the task, we propose using the topics to bridge the semantic gap between articles and comments. We introduce a variation topic model to represent the topics, and match the articles and comments by the similarity of their topics. Experiments show that our topic-based approach significantly outperforms previous lexicon-based models. The model can also profit from paired corpora and achieves state-of-the-art performance under semi-supervised scenarios.
| 17
|
1809.06537
|
Automatic Judgment Prediction via Legal Reading Comprehension
|
# Automatic Judgment Prediction via Legal Reading Comprehension
## Abstract
Automatic judgment prediction aims to predict the judicial results based on case materials. It has been studied for several decades mainly by lawyers and judges, considered as a novel and prospective application of artificial intelligence techniques in the legal field. Most existing methods follow the text classification framework, which fails to model the complex interactions among complementary case materials. To address this issue, we formalize the task as Legal Reading Comprehension according to the legal scenario. Following the working protocol of human judges, LRC predicts the final judgment results based on three types of information, including fact description, plaintiffs' pleas, and law articles. Moreover, we propose a novel LRC model, AutoJudge, which captures the complex semantic interactions among facts, pleas, and laws. In experiments, we construct a real-world civil case dataset for LRC. Experimental results on this dataset demonstrate that our model achieves significant improvement over state-of-the-art models. We will publish all source codes and datasets of this work on \urlgithub.com for further research.
## Introduction
Automatic judgment prediction is to train a machine judge to determine whether a certain plea in a given civil case would be supported or rejected. In countries with civil law system, e.g. mainland China, such process should be done with reference to related law articles and the fact description, as is performed by a human judge. The intuition comes from the fact that under civil law system, law articles act as principles for juridical judgments. Such techniques would have a wide range of promising applications. On the one hand, legal consulting systems could provide better access to high-quality legal resources in a low-cost way to legal outsiders, who suffer from the complicated terminologies. On the other hand, machine judge assistants for professionals would help improve the efficiency of the judicial system. Besides, automated judgment system can help in improving juridical equality and transparency. From another perspective, there are currently 7 times much more civil cases than criminal cases in mainland China, with annual rates of increase of INLINEFORM0 and INLINEFORM1 respectively, making judgment prediction in civil cases a promising application BIBREF0 .
Previous works BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 formalize judgment prediction as the text classification task, regarding either charge names or binary judgments, i.e., support or reject, as the target classes. These works focus on the situation where only one result is expected, e.g., the US Supreme Court's decisions BIBREF2 , and the charge name prediction for criminal cases BIBREF3 . Despite these recent efforts and their progress, automatic judgment prediction in civil law system is still confronted with two main challenges:
One-to-Many Relation between Case and Plea. Every single civil case may contain multiple pleas and the result of each plea is co-determined by related law articles and specific aspects of the involved case. For example, in divorce proceedings, judgment of alienation of mutual affection is the key factor for granting divorce but custody of children depends on which side can provide better an environment for children's growth as well as parents' financial condition. Here, different pleas are independent.
Heterogeneity of Input Triple. Inputs to a judgment prediction system consist of three heterogeneous yet complementary parts, i.e., fact description, plaintiff's plea, and related law articles. Concatenating them together and treating them simply as a sequence of words as in previous works BIBREF2 , BIBREF1 would cause a great loss of information. This is the same in question-answering where the dual inputs, i.e., query and passage, should be modeled separately.
Despite the introduction of the neural networks that can learn better semantic representations of input text, it remains unsolved to incorporate proper mechanisms to integrate the complementary triple of pleas, fact descriptions, and law articles together.
Inspired by recent advances in question answering (QA) based reading comprehension (RC) BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , we propose the Legal Reading Comprehension (LRC) framework for automatic judgment prediction. LRC incorporates the reading mechanism for better modeling of the complementary inputs above-mentioned, as is done by human judges when referring to legal materials in search of supporting law articles. Reading mechanism, by simulating how human connects and integrates multiple text, has proven an effective module in RC tasks. We argue that applying the reading mechanism in a proper way among the triplets can obtain a better understanding and more informative representation of the original text, and further improve performance . To instantiate the framework, we propose an end-to-end neural network model named AutoJudge.
For experiments, we train and evaluate our models in the civil law system of mainland China. We collect and construct a large-scale real-world data set of INLINEFORM0 case documents that the Supreme People's Court of People's Republic of China has made publicly available. Fact description, pleas, and results can be extracted easily from these case documents with regular expressions, since the original documents have special typographical characteristics indicating the discourse structure. We also take into account law articles and their corresponding juridical interpretations. We also implement and evaluate previous methods on our dataset, which prove to be strong baselines.
Our experiment results show significant improvements over previous methods. Further experiments demonstrate that our model also achieves considerable improvement over other off-the-shelf state-of-the-art models under classification and question answering framework respectively. Ablation tests carried out by taking off some components of our model further prove its robustness and effectiveness.
To sum up, our contributions are as follows:
(1) We introduce reading mechanism and re-formalize judgment prediction as Legal Reading Comprehension to better model the complementary inputs.
(2) We construct a real-world dataset for experiments, and plan to publish it for further research.
(3) Besides baselines from previous works, we also carry out comprehensive experiments comparing different existing deep neural network methods on our dataset. Supported by these experiments, improvements achieved by LRC prove to be robust.
## Judgment Prediction
Automatic judgment prediction has been studied for decades. At the very first stage of judgment prediction studies, researchers focus on mathematical and statistical analysis of existing cases, without any conclusions or methodologies on how to predict them BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 . Recent attempts consider judgment prediction under the text classification framework. Most of these works extract efficient features from text (e.g., N-grams) BIBREF15 , BIBREF4 , BIBREF1 , BIBREF16 , BIBREF17 or case profiles (e.g., dates, terms, locations and types) BIBREF2 . All these methods require a large amount of human effort to design features or annotate cases. Besides, they also suffer from generalization issue when applied to other scenarios.
Motivated by the successful application of deep neural networks, Luo et al. BIBREF3 introduce an attention-based neural model to predict charges of criminal cases, and verify the effectiveness of taking law articles into consideration. Nevertheless, they still fall into the text classification framework and lack the ability to handle multiple inputs with more complicated structures.
## Text Classification
As the basis of previous judgment prediction works, typical text classification task takes a single text content as input and predicts the category it belongs to. Recent works usually employ neural networks to model the internal structure of a single input BIBREF18 , BIBREF19 , BIBREF20 , BIBREF21 .
There also exists another thread of text classification called entailment prediction. Methods proposed in BIBREF22 , BIBREF23 are intended for complementary inputs, but the mechanisms can be considered as a simplified version of reading comprehension.
## Reading Comprehension
Reading comprehension is a relevant task to model heterogeneous and complementary inputs, where an answer is predicted given two channels of inputs, i.e. a textual passage and a query. Considerable progress has been made BIBREF6 , BIBREF24 , BIBREF5 . These models employ various attention mechanism to model the interaction between passage and query. Inspired by the advantage of reading comprehension models on modeling multiple inputs, we apply this idea into the legal area and propose legal reading comprehension for judgment prediction.
## Conventional Reading Comprehension
Conventional reading comprehension BIBREF25 , BIBREF26 , BIBREF7 , BIBREF8 usually considers reading comprehension as predicting the answer given a passage and a query, where the answer could be a single word, a text span of the original passage, chosen from answer candidates, or generated by human annotators.
Generally, an instance in RC is represented as a triple INLINEFORM0 , where INLINEFORM1 , INLINEFORM2 and INLINEFORM3 correspond to INLINEFORM4 , INLINEFORM5 and INLINEFORM6 respectively. Given a triple INLINEFORM7 , RC takes the pair INLINEFORM8 as the input and employs attention-based neural models to construct an efficient representation. Afterwards, the representation is fed into the output layer to select or generate an INLINEFORM9 .
## Legal Reading Comprehension
Existing works usually formalize judgment prediction as a text classification task and focus on extracting well-designed features of specific cases. Such simplification ignores that the judgment of a case is determined by its fact description and multiple pleas. Moreover, the final judgment should act up to the legal provisions, especially in civil law systems. Therefore, how to integrate the information (i.e., fact descriptions, pleas, and law articles) in a reasonable way is critical for judgment prediction.
Inspired by the successful application of RC, we propose a framework of Legal Reading Comprehension(LRC) for judgment prediction in the legal area. As illustrated in Fig. FIGREF1 , for each plea in a given case, the prediction of judgment result is made based the fact description and the potentially relevant law articles.
In a nutshell, LRC can be formalized as the following quadruplet task: DISPLAYFORM0
where INLINEFORM0 is the fact description, INLINEFORM1 is the plea, INLINEFORM2 is the law articles and INLINEFORM3 is the result. Given INLINEFORM4 , LRC aims to predict the judgment result as DISPLAYFORM0
The probability is calculated with respect to the interaction among the triple INLINEFORM0 , which will draw on the experience of the interaction between INLINEFORM1 pairs in RC.
To summarize, LRC is innovative in the following aspects:
(1) While previous works fit the problem into text classification framework, LRC re-formalizes the way to approach such problems. This new framework provides the ability to deal with the heterogeneity of the complementary inputs.
(2) Rather than employing conventional RC models to handle pair-wise text information in the legal area, LRC takes the critical law articles into consideration and models the facts, pleas, and law articles jointly for judgment prediction, which is more suitable to simulate the human mode of dealing with cases.
## Methods
We propose a novel judgment prediction model AutoJudge to instantiate the LRC framework. As shown in Fig. FIGREF6 , AutoJudge consists of three flexible modules, including a text encoder, a pair-wise attentive reader, and an output module.
In the following parts, we give a detailed introduction to these three modules.
## Text Encoder
As illustrated in Fig. FIGREF6 , Text Encoder aims to encode the word sequences of inputs into continuous representation sequences.
Formally, consider a fact description INLINEFORM0 , a plea INLINEFORM1 , and the relevant law articles INLINEFORM2 , where INLINEFORM3 denotes the INLINEFORM4 -th word in the sequence and INLINEFORM5 are the lengths of word sequences INLINEFORM6 respectively. First, we convert the words to their respective word embeddings to obtain INLINEFORM7 , INLINEFORM8 and INLINEFORM9 , where INLINEFORM10 . Afterwards, we employ bi-directional GRU BIBREF27 , BIBREF28 , BIBREF29 to produce the encoded representation INLINEFORM11 of all words as follows: DISPLAYFORM0
Note that, we adopt different bi-directional GRUs to encode fact descriptions, pleas, and law articles respectively(denoted as INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 ). With these text encoders, INLINEFORM3 , INLINEFORM4 , and INLINEFORM5 are converting into INLINEFORM6 , INLINEFORM7 , and INLINEFORM8 .
## Pair-Wise Attentive Reader
How to model the interactions among the input text is the most important problem in reading comprehension. In AutoJudge, we employ a pair-wise attentive reader to process INLINEFORM0 and INLINEFORM1 respectively. More specifically, we propose to use pair-wise mutual attention mechanism to capture the complex semantic interaction between text pairs, as well as increasing the interpretability of AutoJudge.
For each input pair INLINEFORM0 or INLINEFORM1 , we employ pair-wise mutual attention to select relevant information from fact descriptions INLINEFORM2 and produce more informative representation sequences.
As a variant of the original attention mechanism BIBREF28 , we design the pair-wise mutual attention unit as a GRU with internal memories denoted as mGRU.
Taking the representation sequence pair INLINEFORM0 for instance, mGRU stores the fact sequence INLINEFORM1 into its memories. For each timestamp INLINEFORM2 , it selects relevant fact information INLINEFORM3 from the memories as follows, DISPLAYFORM0
Here, the weight INLINEFORM0 is the softmax value as DISPLAYFORM0
Note that, INLINEFORM0 represents the relevance between INLINEFORM1 and INLINEFORM2 . It is calculated as follows, DISPLAYFORM0
Here, INLINEFORM0 is the last hidden state in the GRU, which will be introduced in the following part. INLINEFORM1 is a weight vector, and INLINEFORM2 , INLINEFORM3 , INLINEFORM4 are attention metrics of our proposed pair-wise attention mechanism.
With the relevant fact information INLINEFORM0 and INLINEFORM1 , we get the INLINEFORM2 -th input of mGRU as DISPLAYFORM0
where INLINEFORM0 indicates the concatenation operation.
Then, we feed INLINEFORM0 into GRU to get more informative representation sequence INLINEFORM1 as follows, DISPLAYFORM0
For the input pair INLINEFORM0 , we can get INLINEFORM1 in the same way. Therefore, we omit the implementation details Here.
Similar structures with attention mechanism are also applied in BIBREF5 , BIBREF30 , BIBREF31 , BIBREF28 to obtain mutually aware representations in reading comprehension models, which significantly improve the performance of this task.
## Output Layer
Using text encoder and pair-wise attentive reader, the initial input triple INLINEFORM0 has been converted into two sequences, i.e., INLINEFORM1 and INLINEFORM2 , where INLINEFORM3 is defined similarly to INLINEFORM4 . These sequences reserve complex semantic information about the pleas and law articles, and filter out irrelevant information in fact descriptions.
With these two sequences, we concatenate INLINEFORM0 and INLINEFORM1 along the sequence length dimension to generate the sequence INLINEFORM2 . Since we have employed several GRU layers to encode the sequential inputs, another recurrent layer may be redundant. Therefore, we utilize a 1-layer CNN BIBREF18 to capture the local structure and generate the representation vector for the final prediction.
Assuming INLINEFORM0 is the predicted probability that the plea in the case sample would be supported and INLINEFORM1 is the gold standard, AutoJudge aims to minimize the cross-entropy as follows, DISPLAYFORM0
where INLINEFORM0 is the number of training data. As all the calculation in our model is differentiable, we employ Adam BIBREF32 for optimization.
## Experiments
To evaluate the proposed LRC framework and the AutoJudge model, we carry out a series of experiments on the divorce proceedings, a typical yet complex field of civil cases. Divorce proceedings often come with several kinds of pleas, e.g. seeking divorce, custody of children, compensation, and maintenance, which focuses on different aspects and thus makes it a challenge for judgment prediction.
## Dataset Construction for Evaluation
Since none of the datasets from previous works have been published, we decide to build a new one. We randomly collect INLINEFORM0 cases from China Judgments Online, among which INLINEFORM1 cases are for training, INLINEFORM2 each for validation and testing. Among the original cases, INLINEFORM3 are granted divorce and others not. There are INLINEFORM4 valid pleas in total, with INLINEFORM5 supported and INLINEFORM6 rejected. Note that, if the divorce plea in a case is not granted, the other pleas of this case will not be considered by the judge. Case materials are all natural language sentences, with averagely INLINEFORM7 tokens per fact description and INLINEFORM8 per plea. There are 62 relevant law articles in total, each with INLINEFORM9 tokens averagely. Note that the case documents include special typographical signals, making it easy to extract labeled data with regular expression.
We apply some rules with legal prior to preprocess the dataset according to previous works BIBREF33 , BIBREF34 , BIBREF35 , which have proved effective in our experiments.
Name Replacement: All names in case documents are replaced with marks indicating their roles, instead of simply anonymizing them, e.g. <Plantiff>, <Defendant>, <Daughter_x> and so on. Since “all are equal before the law”, names should make no more difference than what role they take.
Law Article Filtration : Since most accessible divorce proceeding documents do not contain ground-truth fine-grained articles, we use an unsupervised method instead. First, we extract all the articles from the law text with regular expression. Afterwards, we select the most relevant 10 articles according to the fact descriptions as follows. We obtain sentence representation with CBOW BIBREF36 , BIBREF37 weighted by inverse document frequency, and calculate cosine distance between cases and law articles. Word embeddings are pre-trained with Chinese Wikipedia pages. As the final step, we extract top 5 relevant articles for each sample respectively from the main marriage law articles and their interpretations, which are equally important. We manually check the extracted articles for 100 cases to ensure that the extraction quality is fairly good and acceptable.
The filtration process is automatic and fully unsupervised since the original documents have no ground-truth labels for fine-grained law articles, and coarse-grained law-articles only provide limited information. We also experiment with the ground-truth articles, but only a small fraction of them has fine-grained ones, and they are usually not available in real-world scenarios.
## Implementation Details
We employ Jieba for Chinese word segmentation and keep the top INLINEFORM0 frequent words. The word embedding size is set to 128 and the other low-frequency words are replaced with the mark <UNK>. The hidden size of GRU is set to 128 for each direction in Bi-GRU. In the pair-wise attentive reader, the hidden state is set to 256 for mGRu. In the CNN layer, filter windows are set to 1, 3, 4, and 5 with each filter containing 200 feature maps. We add a dropout layer BIBREF38 after the CNN layer with a dropout rate of INLINEFORM1 . We use Adam BIBREF32 for training and set learning rate to INLINEFORM2 , INLINEFORM3 to INLINEFORM4 , INLINEFORM5 to INLINEFORM6 , INLINEFORM7 to INLINEFORM8 , batch size to 64. We employ precision, recall, F1 and accuracy for evaluation metrics. We repeat all the experiments for 10 times, and report the average results.
## Baselines
For comparison, we adopt and re-implement three kinds of baselines as follows:
We implement an SVM with lexical features in accordance with previous works BIBREF16 , BIBREF17 , BIBREF1 , BIBREF15 , BIBREF4 and select the best feature set on the development set.
We implement and fine-tune a series of neural text classifiers, including attention-based method BIBREF3 and other methods we deem important. CNN BIBREF18 and GRU BIBREF27 , BIBREF21 take as input the concatenation of fact description and plea. Similarly, CNN/GRU+law refers to using the concatenation of fact description, plea and law articles as inputs.
We implement and train some off-the-shelf RC models, including r-net BIBREF5 and AoA BIBREF6 , which are the leading models on SQuAD leaderboard. In our initial experiments, these models take fact description as passage and plea as query. Further, Law articles are added to the fact description as a part of the reading materials, which is a simple way to consider them as well.
## Results and Analysis
From Table TABREF37 , we have the following observations:
(1) AutoJudge consistently and significantly outperforms all the baselines, including RC models and other neural text classification models, which shows the effectiveness and robustness of our model.
(2) RC models achieve better performance than most text classification models (excluding GRU+Attention), which indicates that reading mechanism is a better way to integrate information from heterogeneous yet complementary inputs. On the contrary, simply adding law articles as a part of the reading materials makes no difference in performance. Note that, GRU+Attention employ similar attention mechanism as RC does and takes additional law articles into consideration, thus achieves comparable performance with RC models.
(3) Comparing with conventional RC models, AutoJudge achieves significant improvement with the consideration of additional law articles. It reflects the difference between LRC and conventional RC models. We re-formalize LRC in legal area to incorporate law articles via the reading mechanism, which can enhance judgment prediction. Moreover, CNN/GRU+law decrease the performance by simply concatenating original text with law articles, while GRU+Attention/AutoJudge increase the performance by integrating law articles with attention mechanism. It shows the importance and rationality of using attention mechanism to capture the interaction between multiple inputs.
The experiments support our hypothesis as proposed in the Introduction part that in civil cases, it's important to model the interactions among case materials. Reading mechanism can well perform the matching among them.
## Ablation Test
AutoJudge is characterized by the incorporation of pair-wise attentive reader, law articles, and a CNN output layer, as well as some pre-processing with legal prior. We design ablation tests respectively to evaluate the effectiveness of these modules. When taken off the attention mechanism, AutoJudge degrades into a GRU on which a CNN is stacked. When taken off law articles, the CNN output layer only takes INLINEFORM0 as input. Besides, our model is tested respectively without name-replacement or unsupervised selection of law articles (i.e. passing the whole law text). As mentioned above, we system use law articles extracted with unsupervised method, so we also experiment with ground-truth law articles.
Results are shown in Table TABREF38 . We can infer that:
(1) The performance drops significantly after removing the attention layer or excluding the law articles, which is consistent with the comparison between AutoJudge and baselines. The result verifies that both the reading mechanism and incorporation of law articles are important and effective.
(2) After replacing CNN with an LSTM layer, performance drops as much as INLINEFORM0 in accuracy and INLINEFORM1 in F1 score. The reason may be the redundancy of RNNs. AutoJudge has employed several GRU layers to encode text sequences. Another RNN layer may be useless to capture sequential dependencies, while CNN can catch the local structure in convolution windows.
(3) Motivated by existing rule-based works, we conduct data pre-processing on cases, including name replacement and law article filtration. If we remove the pre-processing operations, the performance drops considerably. It demonstrates that applying the prior knowledge in legal filed would benefit the understanding of legal cases.
It's intuitive that the quality of the retrieved law articles would affect the final performance. As is shown in Table TABREF38 , feeding the whole law text without filtration results in worse performance. However, when we train and evaluate our model with ground truth articles, the performance is boosted by nearly INLINEFORM0 in both F1 and Acc. The performance improvement is quite limited compared to that in previous work BIBREF3 for the following reasons: (1) As mentioned above, most case documents only contain coarse-grained articles, and only a small number of them contain fine-grained ones, which has limited information in themselves. (2) Unlike in criminal cases where the application of an article indicates the corresponding crime, law articles in civil cases work as reference, and can be applied in both the cases of supports and rejects. As law articles cut both ways for the judgment result, this is one of the characteristics that distinguishes civil cases from criminal ones. We also need to remember that, the performance of INLINEFORM1 in accuracy or INLINEFORM2 in F1 score is unattainable in real-world setting for automatic prediction where ground-truth articles are not available.
In the area of civil cases, the understanding of the case materials and how they interact is a critical factor. The inclusion of law articles is not enough. As is shown in Table TABREF38 , compared to feeding the model with an un-selected set of law articles, taking away the reading mechanism results in greater performance drop. Therefore, the ability to read, understand and select relevant information from the complex multi-sourced case materials is necessary. It's even more important in real world since we don't have access to ground-truth law articles to make predictions.
## Case Study
We visualize the heat maps of attention results. As shown in Fig. FIGREF47 , deeper background color represents larger attention score. The attention score is calculated with Eq. ( EQREF15 ). We take the average of the resulting INLINEFORM0 attention matrix over the time dimension to obtain attention values for each word.
The visualization demonstrates that the attention mechanism can capture relevant patterns and semantics in accordance with different pleas in different cases.
As for the failed samples, the most common reason comes from the anonymity issue, which is also shown in Fig. FIGREF47 . As mentioned above, we conduct name replacement. However, some critical elements are also anonymized by the government, due to the privacy issue. These elements are sometimes important to judgment prediction. For example, determination of the key factor long-time separation is relevant to the explicit dates, which are anonymized.
## Conclusion
In this paper, we explore the task of predicting judgments of civil cases. Comparing with conventional text classification framework, we propose Legal Reading Comprehension framework to handle multiple and complex textual inputs. Moreover, we present a novel neural model, AutoJudge, to incorporate law articles for judgment prediction. In experiments, we compare our model on divorce proceedings with various state-of-the-art baselines of various frameworks. Experimental results show that our model achieves considerable improvement than all the baselines. Besides, visualization results also demonstrate the effectiveness and interpretability of our proposed model.
In the future, we can explore the following directions: (1) Limited by the datasets, we can only verify our proposed model on divorce proceedings. A more general and larger dataset will benefit the research on judgment prediction. (2) Judicial decisions in some civil cases are not always binary, but more diverse and flexible ones, e.g. compensation amount. Thus, it is critical for judgment prediction to manage various judgment forms.
| 18
|
1809.08652
|
Mind Your Language: Abuse and Offense Detection for Code-Switched Languages
|
# Mind Your Language: Abuse and Offense Detection for Code-Switched Languages
## Abstract
In multilingual societies like the Indian subcontinent, use of code-switched languages is much popular and convenient for the users. In this paper, we study offense and abuse detection in the code-switched pair of Hindi and English (i.e. Hinglish), the pair that is the most spoken. The task is made difficult due to non-fixed grammar, vocabulary, semantics and spellings of Hinglish language. We apply transfer learning and make a LSTM based model for hate speech classification. This model surpasses the performance shown by the current best models to establish itself as the state-of-the-art in the unexplored domain of Hinglish offensive text classification.We also release our model and the embeddings trained for research purposes
## Introduction
With the penetration of internet among masses, the content being posted on social media channels has uptaken. Specifically, in the Indian subcontinent, number of Internet users has crossed 500 mi, and is rising rapidly due to inexpensive data. With this rise, comes the problem of hate speech, offensive and abusive posts on social media. Although there are many previous works which deal with Hindi and English hate speech (the top two languages in India), but very few on the code-switched version (Hinglish) of the two BIBREF0 . This is partially due to the following reasons: (i) Hinglish consists of no-fixed grammar and vocabulary. It derives a part of its semantics from Devnagari and another part from the Roman script. (ii) Hinglish speech and written text consists of a concoction of words spoken in Hindi as well as English, but written in the Roman script. This makes the spellings variable and dependent on the writer of the text. Hence code-switched languages present tough challenges in terms of parsing and getting the meaning out of the text. For instance, the sentence, “Modiji foreign yatra par hai”, is in the Hinglish language. Somewhat correct translation of this would be, “Mr. Modi is on a foriegn tour”. However, even this translation has some flaws due to no direct translation available for the word ji, which is used to show respect. Verbatim translation would lead to “Mr. Modi foreign tour on is”. Moreover, the word yatra here, can have phonetic variations, which would result in multiple spellings of the word as yatra, yaatra, yaatraa, etc. Also, the problem of hate speech has been rising in India, and according to the policies of the government and the various social networks, one is not allowed to misuse his right to speech to abuse some other community or religion. Due to the various difficulties associated with the Hinglish language, it is challenging to automatically detect and ban such kind of speech.
Thus, with this in mind, we build a transfer learning based model for the code-switched language Hinglish, which outperforms the baseline model of BIBREF0 . We also release the embeddings and the model trained.
## Methodology
Our methodology primarily consists of these steps: Pre-processing of the dataset, training of word embeddings, training of the classifier model and then using that on HEOT dataset.
## Pre-Processing
In this work, we use the datasets released by BIBREF1 and HEOT dataset provided by BIBREF0 . The datasets obtained pass through these steps of processing: (i) Removal of punctuatios, stopwords, URLs, numbers, emoticons, etc. This was then followed by transliteration using the Xlit-Crowd conversion dictionary and translation of each word to English using Hindi to English dictionary. To deal with the spelling variations, we manually added some common variations of popular Hinglish words. Final dictionary comprised of 7200 word pairs. Additionally, to deal with profane words, which are not present in Xlit-Crowd, we had to make a profanity dictionary (with 209 profane words) as well. Table TABREF3 gives some examples from the dictionary.
## Training Word Embeddings
We tried Glove BIBREF2 and Twitter word2vec BIBREF3 code for training embeddings for the processed tweets. The embeddings were trained on both the datasets provided by BIBREF1 and HEOT. These embeddings help to learn distributed representations of tweets. After experimentation, we kept the size of embeddings fixed to 100.
## Classifier Model
Both the HEOT and BIBREF1 datasets contain tweets which are annotated in three categories: offensive, abusive and none (or benign). Some examples from the dataset are shown in Table TABREF4 . We use a LSTM based classifier model for training our model to classify these tweets into these three categories. An overview of the model is given in the Figure FIGREF12 . The model consists of one layer of LSTM followed by three dense layers. The LSTM layer uses a dropout value of 0.2. Categorical crossentropy loss was used for the last layer due to the presence of multiple classes. We use Adam optimizer along with L2 regularisation to prevent overfitting. As indicated by the Figure FIGREF12 , the model was initially trained on the dataset provided by BIBREF1 , and then re-trained on the HEOT dataset so as to benefit from the transfer of learned features in the last stage. The model hyperparameters were experimentally selected by trying out a large number of combinations through grid search.
## Results
Table TABREF9 shows the performance of our model (after getting trained on BIBREF1 ) with two types of embeddings in comparison to the models by BIBREF0 and BIBREF1 on the HEOT dataset averaged over three runs. We also compare results on pre-trained embeddings. As shown in the table, our model when given Glove embeddings performs better than all other models. For comparison purposes, in Table TABREF10 we have also evaluated our results on the dataset by BIBREF1 .
## Conclusion
In this paper, we presented a pipeline which given Hinglish text can classify it into three categories: offensive, abusive and benign. This LSTM based model performs better than the other systems present. We also release the code, the dictionary made and the embeddings trained in the process. We believe this model would be useful in hate speech detection tasks for code-switched languages.
| 7
|
1809.09795
|
Deep contextualized word representations for detecting sarcasm and irony
|
# Deep contextualized word representations for detecting sarcasm and irony
## Abstract
Predicting context-dependent and non-literal utterances like sarcastic and ironic expressions still remains a challenging task in NLP, as it goes beyond linguistic patterns, encompassing common sense and shared knowledge as crucial components. To capture complex morpho-syntactic features that can usually serve as indicators for irony or sarcasm across dynamic contexts, we propose a model that uses character-level vector representations of words, based on ELMo. We test our model on 7 different datasets derived from 3 different data sources, providing state-of-the-art performance in 6 of them, and otherwise offering competitive results.
## Introduction
Sarcastic and ironic expressions are prevalent in social media and, due to the tendency to invert polarity, play an important role in the context of opinion mining, emotion recognition and sentiment analysis BIBREF0 . Sarcasm and irony are two closely related linguistic phenomena, with the concept of meaning the opposite of what is literally expressed at its core. There is no consensus in academic research on the formal definition, both terms are non-static, depending on different factors such as context, domain and even region in some cases BIBREF1 .
In light of the general complexity of natural language, this presents a range of challenges, from the initial dataset design and annotation to computational methods and evaluation BIBREF2 . The difficulties lie in capturing linguistic nuances, context-dependencies and latent meaning, due to richness of dynamic variants and figurative use of language BIBREF3 .
The automatic detection of sarcastic expressions often relies on the contrast between positive and negative sentiment BIBREF4 . This incongruence can be found on a lexical level with sentiment-bearing words, as in "I love being ignored". In more complex linguistic settings an action or a situation can be perceived as negative, without revealing any affect-related lexical elements. The intention of the speaker as well as common knowledge or shared experience can be key aspects, as in "I love waking up at 5 am", which can be sarcastic, but not necessarily. Similarly, verbal irony is referred to as saying the opposite of what is meant and based on sentiment contrast BIBREF5 , whereas situational irony is seen as describing circumstances with unexpected consequences BIBREF6 , BIBREF7 .
Empirical studies have shown that there are specific linguistic cues and combinations of such that can serve as indicators for sarcastic and ironic expressions. Lexical and morpho-syntactic cues include exclamations and interjections, typographic markers such as all caps, quotation marks and emoticons, intensifiers and hyperboles BIBREF8 , BIBREF9 . In the case of Twitter, the usage of emojis and hashtags has also proven to help automatic irony detection.
We propose a purely character-based architecture which tackles these challenges by allowing us to use a learned representation that models features derived from morpho-syntactic cues. To do so, we use deep contextualized word representations, which have recently been used to achieve the state of the art on six NLP tasks, including sentiment analysis BIBREF10 . We test our proposed architecture on 7 different irony/sarcasm datasets derived from 3 different data sources, providing state-of-the-art performance in 6 of them and otherwise offering competitive results, showing the effectiveness of our proposal. We make our code available at https://github.com/epochx/elmo4irony.
## Related work
Apart from the relevance for industry applications related to sentiment analysis, sarcasm and irony detection has received great traction within the NLP research community, resulting in a variety of methods, shared tasks and benchmark datasets. Computational approaches for the classification task range from rule-based systems BIBREF4 , BIBREF11 and statistical methods and machine learning algorithms such as Support Vector Machines BIBREF3 , BIBREF12 , Naive Bayes and Decision Trees BIBREF13 leveraging extensive feature sets, to deep learning-based approaches. In this context, BIBREF14 . delivered state-of-the-art results by using an intra-attentional component in addition to a recurrent neural network. Previous work such as the one by BIBREF15 had proposed a convolutional long-short-term memory network (CNN-LSTM-DNN) that also achieved excellent results. A comprehensive survey on automatic sarcasm detection was done by BIBREF16 , while computational irony detection was reviewed by BIBREF17 .
Further improvements both in terms of classic and deep models came as a result of the SemEval 2018 Shared Task on Irony in English Tweets BIBREF18 . The system that achieved the best results was hybrid, namely, a densely-connected BiLSTM with a multi-task learning strategy, which also makes use of features such as POS tags and lexicons BIBREF19 .
## Proposed Approach
The wide spectrum of linguistic cues that can serve as indicators for sarcastic and ironic expressions has been usually exploited for automatic sarcasm or irony detection by modeling them in the form of binary features in traditional machine learning.
On the other hand, deep models for irony and sarcasm detection, which are currently offer state-of-the-art performance, have exploited sequential neural networks such as LSTMs and GRUs BIBREF15 , BIBREF23 on top of distributed word representations. Recently, in addition to using a sequential model, BIBREF14 proposed to use intra-attention to compare elements in a sequence against themselves. This allowed the model to better capture word-to-word level interactions that could also be useful for detecting sarcasm, such as the incongruity phenomenon BIBREF3 . Despite this, all models in the literature rely on word-level representations, which keeps the models from being able to easily capture some of the lexical and morpho-syntactic cues known to denote irony, such as all caps, quotation marks and emoticons, and in Twitter, also emojis and hashtags.
The usage of a purely character-based input would allow us to directly recover and model these features. Consequently, our architecture is based on Embeddings from Language Model or ELMo BIBREF10 . The ELMo layer allows to recover a rich 1,024-dimensional dense vector for each word. Using CNNs, each vector is built upon the characters that compose the underlying words. As ELMo also contains a deep bi-directional LSTM on top of this character-derived vectors, each word-level embedding contains contextual information from their surroundings. Concretely, we use a pre-trained ELMo model, obtained using the 1 Billion Word Benchmark which contains about 800M tokens of news crawl data from WMT 2011 BIBREF24 .
Subsequently, the contextualized embeddings are passed on to a BiLSTM with 2,048 hidden units. We aggregate the LSTM hidden states using max-pooling, which in our preliminary experiments offered us better results, and feed the resulting vector to a 2-layer feed-forward network, where each layer has 512 units. The output of this is then fed to the final layer of the model, which performs the binary classification.
## Experimental Setup
We test our proposed approach for binary classification on either sarcasm or irony, on seven benchmark datasets retrieved from different media sources. Below we describe each dataset, please see Table TABREF1 below for a summary.
Twitter: We use the Twitter dataset provided for the SemEval 2018 Task 3, Irony Detection in English Tweets BIBREF18 . The dataset was manually annotated using binary labels. We also use the dataset by BIBREF4 , which is manually annotated for sarcasm. Finally, we use the dataset by BIBREF20 , who collected a user self-annotated corpus of tweets with the #sarcasm hashtag.
Reddit: BIBREF21 collected SARC, a corpus comprising of 600.000 sarcastic comments on Reddit. We use main subset, SARC 2.0, and the political subset, SARC 2.0 pol.
Online Dialogues: We utilize the Sarcasm Corpus V1 (SC-V1) and the Sarcasm Corpus V2 (SC-V2), which are subsets of the Internet Argument Corpus (IAC). Compared to other datasets in our selection, these differ mainly in text length and structure complexity BIBREF22 .
In Table TABREF1 , we see a notable difference in terms of size among the Twitter datasets. Given this circumstance, and in light of the findings by BIBREF18 , we are interested in studying how the addition of external soft-annotated data impacts on the performance. Thus, in addition to the datasets introduced before, we use two corpora for augmentation purposes. The first dataset was collected using the Twitter API, targeting tweets with the hashtags #sarcasm or #irony, resulting on a total of 180,000 and 45,000 tweets respectively. On the other hand, to obtain non-sarcastic and non-ironic tweets, we relied on the SemEval 2018 Task 1 dataset BIBREF25 . To augment each dataset with our external data, we first filter out tweets that are not in English using language guessing systems. We later extract all the hashtags in each target dataset and proceed to augment only using those external tweets that contain any of these hashtags. This allows us to, for each class, add a total of 36,835 tweets for the Ptáček corpus, 8,095 for the Riloff corpus and 26,168 for the SemEval-2018 corpus.
In terms of pre-processing, as in our case the preservation of morphological structures is crucial, the amount of normalization is minimal. Concretely, we forgo stemming or lemmatizing, punctuation removal and lowercasing. We limit ourselves to replacing user mentions and URLs with one generic token respectively. In the case of the SemEval-2018 dataset, an additional step was to remove the hashtags #sarcasm, #irony and #not, as they are the artifacts used for creating the dataset. For tokenizing, we use a variation of the Twokenizer BIBREF26 to better deal with emojis.
Our models are trained using Adam with a learning rate of 0.001 and a decay rate of 0.5 when there is no improvement on the accuracy on the validation set, which we use to select the best models. We also experimented using a slanted triangular learning rate scheme, which was shown by BIBREF27 to deliver excellent results on several tasks, but in practice we did not obtain significant differences. We experimented with batch sizes of 16, 32 and 64, and dropouts ranging from 0.1 to 0.5. The size of the LSTM hidden layer was fixed to 1,024, based on our preliminary experiments. We do not train the ELMo embeddings, but allow their dropouts to be active during training.
## Results
Table TABREF2 summarizes our results. For each dataset, the top row denotes our baseline and the second row shows our best comparable model. Rows with FULL models denote our best single model trained with all the development available data, without any other preprocessing other than mentioned in the previous section. In the case of the Twitter datasets, rows indicated as AUG refer to our the models trained using the augmented version of the corresponding datasets.
For the case of the SemEval-2018 dataset we use the best performing model from the Shared Task as a baseline, taken from the task description paper BIBREF18 . As the winning system is a voting-based ensemble of 10 models, for comparison, we report results using an equivalent setting. For the Riloff, Ptáček, SC-V1 and SC-V2 datasets, our baseline models are taken directly from BIBREF14 . As their pre-processing includes truncating sentence lengths at 40 and 80 tokens for the Twitter and Dialog datasets respectively, while always removing examples with less than 5 tokens, we replicate those steps and report our results under these settings. Finally, for the Reddit datasets, our baselines are taken from BIBREF21 . Although their models are trained for binary classification, instead of reporting the performance in terms of standard classification evaluation metrics, their proposed evaluation task is predicting which of two given statements that share the same context is sarcastic, with performance measured solely by accuracy. We follow this and report our results.
In summary, we see our introduced models are able to outperform all previously proposed methods for all metrics, except for the SemEval-2018 best system. Although our approach yields higher Precision, it is not able to reach the given Recall and F1-Score. We note that in terms of single-model architectures, our setting offers increased performance compared to BIBREF19 and their obtained F1-score of 0.674. Moreover, our system does so without requiring external features or multi-task learning. For the other tasks we are able to outperform BIBREF14 without requiring any kind of intra-attention. This shows the effectiveness of using pre-trained character-based word representations, that allow us to recover many of the morpho-syntactic cues that tend to denote irony and sarcasm.
Finally, our experiments showed that enlarging existing Twitter datasets by adding external soft-labeled data from the same media source does not yield improvements in the overall performance. This complies with the observations made by BIBREF18 . Since we have designed our augmentation tactics to maximize the overlap in terms of topic, we believe the soft-annotated nature of the additional data we have used is the reason that keeps the model from improving further.
## Conclusions
We have presented a deep learning model based on character-level word representations obtained from ELMo. It is able to obtain the state of the art in sarcasm and irony detection in 6 out of 7 datasets derived from 3 different data sources. Our results also showed that the model does not benefit from using additional soft-labeled data in any of the three tested Twitter datasets, showing that manually-annotated data may be needed in order to improve the performance in this way.
| 6
|
1809.10644
|
Predictive Embeddings for Hate Speech Detection on Twitter
|
# Predictive Embeddings for Hate Speech Detection on Twitter
## Abstract
We present a neural-network based approach to classifying online hate speech in general, as well as racist and sexist speech in particular. Using pre-trained word embeddings and max/mean pooling from simple, fully-connected transformations of these embeddings, we are able to predict the occurrence of hate speech on three commonly used publicly available datasets. Our models match or outperform state of the art F1 performance on all three datasets using significantly fewer parameters and minimal feature preprocessing compared to previous methods.
## Introduction
The increasing popularity of social media platforms like Twitter for both personal and political communication BIBREF0 has seen a well-acknowledged rise in the presence of toxic and abusive speech on these platforms BIBREF1 , BIBREF2 . Although the terms of services on these platforms typically forbid hateful and harassing speech, enforcing these rules has proved challenging, as identifying hate speech speech at scale is still a largely unsolved problem in the NLP community. BIBREF3 , for example, identify many ambiguities in classifying abusive communications, and highlight the difficulty of clearly defining the parameters of such speech. This problem is compounded by the fact that identifying abusive or harassing speech is a challenge for humans as well as automated systems.
Despite the lack of consensus around what constitutes abusive speech, some definition of hate speech must be used to build automated systems to address it. We rely on BIBREF4 's definition of hate speech, specifically: “language that is used to express hatred towards a targeted group or is intended to be derogatory, to humiliate, or to insult the members of the group.”
In this paper, we present a neural classification system that uses minimal preprocessing to take advantage of a modified Simple Word Embeddings-based Model BIBREF5 to predict the occurrence of hate speech. Our classifier features:
In the following sections, we discuss related work on hate speech classification, followed by a description of the datasets, methods and results of our study.
## Related Work
Many efforts have been made to classify hate speech using data scraped from online message forums and popular social media sites such as Twitter and Facebook. BIBREF3 applied a logistic regression model that used one- to four-character n-grams for classification of tweets labeled as racist, sexist or neither. BIBREF4 experimented in classification of hateful as well as offensive but not hateful tweets. They applied a logistic regression classifier with L2 regularization using word level n-grams and various part-of-speech, sentiment, and tweet-level metadata features.
Additional projects have built upon the data sets created by Waseem and/or Davidson. For example, BIBREF6 used a neural network approach with two binary classifiers: one to predict the presence abusive speech more generally, and another to discern the form of abusive speech.
BIBREF7 , meanwhile, used pre-trained word2vec embeddings, which were then fed into a convolutional neural network (CNN) with max pooling to produce input vectors for a Gated Recurrent Unit (GRU) neural network. Other researchers have experimented with using metadata features from tweets. BIBREF8 built a classifier composed of two separate neural networks, one for the text and the other for metadata of the Twitter user, that were trained jointly in interleaved fashion. Both networks used in combination - and especially when trained using transfer learning - achieved higher F1 scores than either neural network classifier alone.
In contrast to the methods described above, our approach relies on a simple word embedding (SWEM)-based architecture BIBREF5 , reducing the number of required parameters and length of training required, while still yielding improved performance and resilience across related classification tasks. Moreover, our network is able to learn flexible vector representations that demonstrate associations among words typically used in hateful communication. Finally, while metadata-based augmentation is intriguing, here we sought to develop an approach that would function well even in cases where such additional data was missing due to the deletion, suspension, or deactivation of accounts.
## Data
In this paper, we use three data sets from the literature to train and evaluate our own classifier. Although all address the category of hateful speech, they used different strategies of labeling the collected data. Table TABREF5 shows the characteristics of the datasets.
Data collected by BIBREF3 , which we term the Sexist/Racist (SR) data set, was collected using an initial Twitter search followed by analysis and filtering by the authors and their team who identified 17 common phrases, hashtags, and users that were indicative of abusive speech. BIBREF4 collected the HATE dataset by searching for tweets using a lexicon provided by Hatebase.org. The final data set we used, which we call HAR, was collected by BIBREF9 ; we removed all retweets reducing the dataset to 20,000 tweets. Tweets were labeled as “Harrassing” or “Non-Harrassing”; hate speech was not explicitly labeled, but treated as an unlabeled subset of the broader “Harrassing” category BIBREF9 .
## Transformed Word Embedding Model (TWEM)
Our training set consists of INLINEFORM0 examples INLINEFORM1 where the input INLINEFORM2 is a sequence of tokens INLINEFORM3 , and the output INLINEFORM4 is the numerical class for the hate speech class. Each input instance represents a Twitter post and thus, is not limited to a single sentence.
We modify the SWEM-concat BIBREF5 architecture to allow better handling of infrequent and unknown words and to capture non-linear word combinations.
## Word Embeddings
Each token in the input is mapped to an embedding. We used the 300 dimensional embeddings for all our experiments, so each word INLINEFORM0 is mapped to INLINEFORM1 . We denote the full embedded sequence as INLINEFORM2 . We then transform each word embedding by applying 300 dimensional 1-layer Multi Layer Perceptron (MLP) INLINEFORM3 with a Rectified Liner Unit (ReLU) activation to form an updated embedding space INLINEFORM4 . We find this better handles unseen or rare tokens in our training data by projecting the pretrained embedding into a space that the encoder can understand.
## Pooling
We make use of two pooling methods on the updated embedding space INLINEFORM0 . We employ a max pooling operation on INLINEFORM1 to capture salient word features from our input; this representation is denoted as INLINEFORM2 . This forces words that are highly indicative of hate speech to higher positive values within the updated embedding space. We also average the embeddings INLINEFORM3 to capture the overall meaning of the sentence, denoted as INLINEFORM4 , which provides a strong conditional factor in conjunction with the max pooling output. This also helps regularize gradient updates from the max pooling operation.
## Output
We concatenate INLINEFORM0 and INLINEFORM1 to form a document representation INLINEFORM2 and feed the representation into a 50 node 2 layer MLP followed by ReLU Activation to allow for increased nonlinear representation learning. This representation forms the preterminal layer and is passed to a fully connected softmax layer whose output is the probability distribution over labels.
## Experimental Setup
We tokenize the data using Spacy BIBREF10 . We use 300 Dimensional Glove Common Crawl Embeddings (840B Token) BIBREF11 and fine tune them for the task. We experimented extensively with pre-processing variants and our results showed better performance without lemmatization and lower-casing (see supplement for details). We pad each input to 50 words. We train using RMSprop with a learning rate of .001 and a batch size of 512. We add dropout with a drop rate of 0.1 in the final layer to reduce overfitting BIBREF12 , batch size, and input length empirically through random hyperparameter search.
All of our results are produced from 10-fold cross validation to allow comparison with previous results. We trained a logistic regression baseline model (line 1 in Table TABREF10 ) using character ngrams and word unigrams using TF*IDF weighting BIBREF13 , to provide a baseline since HAR has no reported results. For the SR and HATE datasets, the authors reported their trained best logistic regression model's results on their respective datasets.
SR: Sexist/Racist BIBREF3 , HATE: Hate BIBREF4 HAR: Harassment BIBREF9
## Results and Discussion
The approach we have developed establishes a new state of the art for classifying hate speech, outperforming previous results by as much as 12 F1 points. Table TABREF10 illustrates the robustness of our method, which often outperform previous results, measured by weighted F1.
Using the Approximate Randomization (AR) Test BIBREF14 , we perform significance testing using a 75/25 train and test split
to compare against BIBREF3 and BIBREF4 , whose models we re-implemented. We found 0.001 significance compared to both methods. We also include in-depth precision and recall results for all three datasets in the supplement.
Our results indicate better performance than several more complex approaches, including BIBREF4 's best model (which used word and part-of-speech ngrams, sentiment, readability, text, and Twitter specific features), BIBREF6 (which used two fold classification and a hybrid of word and character CNNs, using approximately twice the parameters we use excluding the word embeddings) and even recent work by BIBREF8 , (whose best model relies on GRUs, metadata including popularity, network reciprocity, and subscribed lists).
On the SR dataset, we outperform BIBREF8 's text based model by 3 F1 points, while just falling short of the Text + Metadata Interleaved Training model. While we appreciate the potential added value of metadata, we believe a tweet-only classifier has merits because retrieving features from the social graph is not always tractable in production settings. Excluding the embedding weights, our model requires 100k parameters , while BIBREF8 requires 250k parameters.
## Error Analysis
False negatives
Many of the false negatives we see are specific references to characters in the TV show “My Kitchen Rules”, rather than something about women in general. Such examples may be innocuous in isolation but could potentially be sexist or racist in context. While this may be a limitation of considering only the content of the tweet, it could also be a mislabel.
Debra are now my most hated team on #mkr after least night's ep. Snakes in the grass those two.
Along these lines, we also see correct predictions of innocuous speech, but find data mislabeled as hate speech:
@LoveAndLonging ...how is that example "sexism"?
@amberhasalamb ...in what way?
Another case our classifier misses is problematic speech within a hashtag:
:D @nkrause11 Dudes who go to culinary school: #why #findawife #notsexist :)
This limitation could be potentially improved through the use of character convolutions or subword tokenization.
False Positives
In certain cases, our model seems to be learning user names instead of semantic content:
RT @GrantLeeStone: @MT8_9 I don't even know what that is, or where it's from. Was that supposed to be funny? It wasn't.
Since the bulk of our model's weights are in the embedding and embedding-transformation matrices, we cluster the SR vocabulary using these transformed embeddings to clarify our intuitions about the model ( TABREF14 ). We elaborate on our clustering approach in the supplement. We see that the model learned general semantic groupings of words associated with hate speech as well as specific idiosyncrasies related to the dataset itself (e.g. katieandnikki)
## Conclusion
Despite minimal tuning of hyper-parameters, fewer weight parameters, minimal text preprocessing, and no additional metadata, the model performs remarkably well on standard hate speech datasets. Our clustering analysis adds interpretability enabling inspection of results.
Our results indicate that the majority of recent deep learning models in hate speech may rely on word embeddings for the bulk of predictive power and the addition of sequence-based parameters provide minimal utility. Sequence based approaches are typically important when phenomena such as negation, co-reference, and context-dependent phrases are salient in the text and thus, we suspect these cases are in the minority for publicly available datasets. We think it would be valuable to study the occurrence of such linguistic phenomena in existing datasets and construct new datasets that have a better representation of subtle forms of hate speech. In the future, we plan to investigate character based representations, using character CNNs and highway layers BIBREF15 along with word embeddings to allow robust representations for sparse words such as hashtags.
## Supplemental Material
We experimented with several different preprocessing variants and were surprised to find that reducing preprocessing improved the performance on the task for all of our tasks. We go through each preprocessing variant with an example and then describe our analysis to compare and evaluate each of them.
## Preprocessing
Original text
RT @AGuyNamed_Nick Now, I'm not sexist in any way shape or form but I think women are better at gift wrapping. It's the XX chromosome thing
Tokenize (Basic Tokenize: Keeps case and words intact with limited sanitizing)
RT @AGuyNamed_Nick Now , I 'm not sexist in any way shape or form but I think women are better at gift wrapping . It 's the XX chromosome thing
Tokenize Lowercase: Lowercase the basic tokenize scheme
rt @aguynamed_nick now , i 'm not sexist in any way shape or form but i think women are better at gift wrapping . it 's the xx chromosome thing
Token Replace: Replaces entities and user names with placeholder)
ENT USER now , I 'm not sexist in any way shape or form but I think women are better at gift wrapping . It 's the xx chromosome thing
Token Replace Lowercase: Lowercase the Token Replace Scheme
ENT USER now , i 'm not sexist in any way shape or form but i think women are better at gift wrapping . it 's the xx chromosome thing
We did analysis on a validation set across multiple datasets to find that the "Tokenize" scheme was by far the best. We believe that keeping the case in tact provides useful information about the user. For example, saying something in all CAPS is a useful signal that the model can take advantage of.
## Embedding Analysis
Since our method was a simple word embedding based model, we explored the learned embedding space to analyze results. For this analysis, we only use the max pooling part of our architecture to help analyze the learned embedding space because it encourages salient words to increase their values to be selected. We projected the original pre-trained embeddings to the learned space using the time distributed MLP. We summed the embedding dimensions for each word and sorted by the sum in descending order to find the 1000 most salient word embeddings from our vocabulary. We then ran PCA BIBREF16 to reduce the dimensionality of the projected embeddings from 300 dimensions to 75 dimensions. This captured about 60% of the variance. Finally, we ran K means clustering for INLINEFORM0 clusters to organize the most salient embeddings in the projected space.
The learned clusters from the SR vocabulary were very illuminating (see Table TABREF14 ); they gave insights to how hate speech surfaced in the datasets. One clear grouping we found is the misogynistic and pornographic group, which contained words like breasts, blonds, and skank. Two other clusters had references to geopolitical and religious issues in the Middle East and disparaging and resentful epithets that could be seen as having an intellectual tone. This hints towards the subtle pedagogic forms of hate speech that surface. We ran silhouette analysis BIBREF17 on the learned clusters to find that the clusters from the learned representations had a 35% higher silhouette coefficient using the projected embeddings compared to the clusters created from the original pre-trained embeddings. This reinforces the claim that our training process pushed hate-speech related words together, and words from other clusters further away, thus, structuring the embedding space effectively for detecting hate speech.
| 14
|
1810.00663
|
Translating Navigation Instructions in Natural Language to a High-Level Plan for Behavioral Robot Navigation
|
# Translating Navigation Instructions in Natural Language to a High-Level Plan for Behavioral Robot Navigation
## Abstract
We propose an end-to-end deep learning model for translating free-form natural language instructions to a high-level plan for behavioral robot navigation. We use attention models to connect information from both the user instructions and a topological representation of the environment. We evaluate our model's performance on a new dataset containing 10,050 pairs of navigation instructions. Our model significantly outperforms baseline approaches. Furthermore, our results suggest that it is possible to leverage the environment map as a relevant knowledge base to facilitate the translation of free-form navigational instruction.
## Introduction
Enabling robots to follow navigation instructions in natural language can facilitate human-robot interaction across a variety of applications. For instance, within the service robotics domain, robots can follow navigation instructions to help with mobile manipulation BIBREF0 and delivery tasks BIBREF1 .
Interpreting navigation instructions in natural language is difficult due to the high variability in the way people describe routes BIBREF2 . For example, there are a variety of ways to describe the route in Fig. FIGREF4 (a):
Each fragment of a sentence within these instructions can be mapped to one or more than one navigation behaviors. For instance, assume that a robot counts with a number of primitive, navigation behaviors, such as “enter the room on the left (or on right)” , “follow the corridor”, “cross the intersection”, etc. Then, the fragment “advance forward” in a navigation instruction could be interpreted as a “follow the corridor” behavior, or as a sequence of “follow the corridor” interspersed with “cross the intersection” behaviors depending on the topology of the environment. Resolving such ambiguities often requires reasoning about “common-sense” concepts, as well as interpreting spatial information and landmarks, e.g., in sentences such as “the room on the left right before the end of the corridor” and “the room which is in the middle of two vases”.
In this work, we pose the problem of interpreting navigation instructions as finding a mapping (or grounding) of the commands into an executable navigation plan. While the plan is typically modeled as a formal specification of low-level motions BIBREF2 or a grammar BIBREF3 , BIBREF4 , we focus specifically on translating instructions to a high-level navigation plan based on a topological representation of the environment. This representation is a behavioral navigation graph, as recently proposed by BIBREF5 , designed to take advantage of the semantic structure typical of human environments. The nodes of the graph correspond to semantically meaningful locations for the navigation task, such as kitchens or entrances to rooms in corridors. The edges are parameterized, visuo-motor behaviors that allow a robot to navigate between neighboring nodes, as illustrated in Fig. FIGREF4 (b). Under this framework, complex navigation routes can be achieved by sequencing behaviors without an explicit metric representation of the world.
We formulate the problem of following instructions under the framework of BIBREF5 as finding a path in the behavioral navigation graph that follows the desired route, given a known starting location. The edges (behaviors) along this path serve to reach the – sometimes implicit – destination requested by the user. As in BIBREF6 , our focus is on the problem of interpreting navigation directions. We assume that a robot can realize valid navigation plans according to the graph.
We contribute a new end-to-end model for following directions in natural language under the behavioral navigation framework. Inspired by the information retrieval and question answering literature BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , we propose to leverage the behavioral graph as a knowledge base to facilitate the interpretation of navigation commands. More specifically, the proposed model takes as input user directions in text form, the behavioral graph of the environment encoded as INLINEFORM0 node; edge; node INLINEFORM1 triplets, and the initial location of the robot in the graph. The model then predicts a set of behaviors to reach the desired destination according to the instructions and the map (Fig. FIGREF4 (c)). Our main insight is that using attention mechanisms to correlate navigation instructions with the topological map of the environment can facilitate predicting correct navigation plans.
This work also contributes a new dataset of INLINEFORM0 pairs of free-form natural language instructions and high-level navigation plans. This dataset was collected through Mechanical Turk using 100 simulated environments with a corresponding topological map and, to the best of our knowledge, it is the first of its kind for behavioral navigation. The dataset opens up opportunities to explore data-driven methods for grounding navigation commands into high-level motion plans.
We conduct extensive experiments to study the generalization capabilities of the proposed model for following natural language instructions. We investigate both generalization to new instructions in known and in new environments. We conclude this paper by discussing the benefits of the proposed approach as well as opportunities for future research based on our findings.
## Related work
This section reviews relevant prior work on following navigation instructions. Readers interested in an in-depth review of methods to interpret spatial natural language for robotics are encouraged to refer to BIBREF11 .
Typical approaches to follow navigation commands deal with the complexity of natural language by manually parsing commands, constraining language descriptions, or using statistical machine translation methods. While manually parsing commands is often impractical, the first type of approaches are foundational: they showed that it is possible to leverage the compositionality of semantic units to interpret spatial language BIBREF12 , BIBREF13 .
Constraining language descriptions can reduce the size of the input space to facilitate the interpretation of user commands. For example, BIBREF14 explored using structured, symbolic language phrases for navigation. As in this earlier work, we are also interested in navigation with a topological map of the environment. However, we do not process symbolic phrases. Our aim is to translate free-form natural language instructions to a navigation plan using information from a high-level representation of the environment. This translation problem requires dealing with missing actions in navigation instructions and actions with preconditions, such as “at the end of the corridor, turn right” BIBREF15 .
Statistical machine translation BIBREF16 is at the core of recent approaches to enable robots to follow navigation instructions. These methods aim to automatically discover translation rules from a corpus of data, and often leverage the fact that navigation directions are composed of sequential commands. For instance, BIBREF17 , BIBREF4 , BIBREF2 used statistical machine translation to map instructions to a formal language defined by a grammar. Likewise, BIBREF18 , BIBREF0 mapped commands to spatial description clauses based on the hierarchical structure of language in the navigation problem. Our approach to machine translation builds on insights from these prior efforts. In particular, we focus on end-to-end learning for statistical machine translation due to the recent success of Neural Networks in Natural Language Processing BIBREF19 .
Our work is inspired by methods that reduce the task of interpreting user commands to a sequential prediction problem BIBREF20 , BIBREF21 , BIBREF22 . Similar to BIBREF21 and BIBREF22 , we use a sequence-to-sequence model to enable a mobile agent to follow routes. But instead leveraging visual information to output low-level navigation commands, we focus on using a topological map of the environment to output a high-level navigation plan. This plan is a sequence of behaviors that can be executed by a robot to reach a desired destination BIBREF5 , BIBREF6 .
We explore machine translation from the perspective of automatic question answering. Following BIBREF8 , BIBREF9 , our approach uses attention mechanisms to learn alignments between different input modalities. In our case, the inputs to our model are navigation instructions, a topological environment map, and the start location of the robot (Fig. FIGREF4 (c)). Our results show that the map can serve as an effective source of contextual information for the translation task. Additionally, it is possible to leverage this kind of information in an end-to-end fashion.
## Problem Formulation
Our goal is to translate navigation instructions in text form into a sequence of behaviors that a robot can execute to reach a desired destination from a known start location. We frame this problem under a behavioral approach to indoor autonomous navigation BIBREF5 and assume that prior knowledge about the environment is available for the translation task. This prior knowledge is a topological map, in the form of a behavioral navigation graph (Fig. FIGREF4 (b)). The nodes of the graph correspond to semantically-meaningful locations for the navigation task, and its directed edges are visuo-motor behaviors that a robot can use to move between nodes. This formulation takes advantage of the rich semantic structure behind man-made environments, resulting in a compact route representation for robot navigation.
Fig. FIGREF4 (c) provides a schematic view of the problem setting. The inputs are: (1) a navigation graph INLINEFORM0 , (2) the starting node INLINEFORM1 of the robot in INLINEFORM2 , and (3) a set of free-form navigation instructions INLINEFORM3 in natural language. The instructions describe a path in the graph to reach from INLINEFORM4 to a – potentially implicit – destination node INLINEFORM5 . Using this information, the objective is to predict a suitable sequence of robot behaviors INLINEFORM6 to navigate from INLINEFORM7 to INLINEFORM8 according to INLINEFORM9 . From a supervised learning perspective, the goal is then to estimate: DISPLAYFORM0
based on a dataset of input-target pairs INLINEFORM0 , where INLINEFORM1 and INLINEFORM2 , respectively. The sequential execution of the behaviors INLINEFORM3 should replicate the route intended by the instructions INLINEFORM4 . We assume no prior linguistic knowledge. Thus, translation approaches have to cope with the semantics and syntax of the language by discovering corresponding patterns in the data.
## The Behavioral Graph: A Knowledge Base For Navigation
We view the behavioral graph INLINEFORM0 as a knowledge base that encodes a set of navigational rules as triplets INLINEFORM1 , where INLINEFORM2 and INLINEFORM3 are adjacent nodes in the graph, and the edge INLINEFORM4 is an executable behavior to navigate from INLINEFORM5 to INLINEFORM6 . In general, each behaviors includes a list of relevant navigational attributes INLINEFORM7 that the robot might encounter when moving between nodes.
We consider 7 types of semantic locations, 11 types of behaviors, and 20 different types of landmarks. A location in the navigation graph can be a room, a lab, an office, a kitchen, a hall, a corridor, or a bathroom. These places are labeled with unique tags, such as "room-1" or "lab-2", except for bathrooms and kitchens which people do not typically refer to by unique names when describing navigation routes.
Table TABREF7 lists the navigation behaviors that we consider in this work. These behaviors can be described in reference to visual landmarks or objects, such as paintings, book shelfs, tables, etc. As in Fig. FIGREF4 , maps might contain multiple landmarks of the same type. Please see the supplementary material (Appendix A) for more details.
## Approach
We leverage recent advances in deep learning to translate natural language instructions to a sequence of navigation behaviors in an end-to-end fashion. Our proposed model builds on the sequence-to-sequence translation model of BIBREF23 , which computes a soft-alignment between a source sequence (natural language instructions in our case) and the corresponding target sequence (navigation behaviors).
As one of our main contributions, we augment the neural machine translation approach of BIBREF23 to take as input not only natural language instructions, but also the corresponding behavioral navigation graph INLINEFORM0 of the environment where navigation should take place. Specifically, at each step, the graph INLINEFORM1 operates as a knowledge base that the model can access to obtain information about path connectivity, facilitating the grounding of navigation commands.
Figure FIGREF8 shows the structure of the proposed model for interpreting navigation instructions. The model consists of six layers:
Embed layer: The model first encodes each word and symbol in the input sequences INLINEFORM0 and INLINEFORM1 into fixed-length representations. The instructions INLINEFORM2 are embedded into a 100-dimensional pre-trained GloVe vector BIBREF24 . Each of the triplet components, INLINEFORM3 , INLINEFORM4 , and INLINEFORM5 of the graph INLINEFORM6 , are one-hot encoded into vectors of dimensionality INLINEFORM7 , where INLINEFORM8 and INLINEFORM9 are the number of nodes and edges in INLINEFORM10 , respectively.
Encoder layer: The model then uses two bidirectional Gated Recurrent Units (GRUs) BIBREF25 to independently process the information from INLINEFORM0 and INLINEFORM1 , and incorporate contextual cues from the surrounding embeddings in each sequence. The outputs of the encoder layer are the matrix INLINEFORM2 for the navigational commands and the matrix INLINEFORM3 for the behavioral graph, where INLINEFORM4 is the hidden size of each GRU, INLINEFORM5 is the number of words in the instruction INLINEFORM6 , and INLINEFORM7 is the number of triplets in the graph INLINEFORM8 .
Attention layer: Matrices INLINEFORM0 and INLINEFORM1 generated by the encoder layer are combined using an attention mechanism. We use one-way attention because the graph contains information about the whole environment, while the instruction has (potentially incomplete) local information about the route of interest. The use of attention provides our model with a two-step strategy to interpret commands. This resembles the way people find paths on a map: first, relevant parts on the map are selected according to their affinity to each of the words in the input instruction (attention layer); second, the selected parts are connected to assemble a valid path (decoder layer). More formally, let INLINEFORM2 ( INLINEFORM3 ) be the INLINEFORM4 -th row of INLINEFORM5 , and INLINEFORM6 ( INLINEFORM7 ) the INLINEFORM8 -th row of INLINEFORM9 . We use each encoded triplet INLINEFORM10 in INLINEFORM11 to calculate its associated attention distribution INLINEFORM12 over all the atomic instructions INLINEFORM13 : DISPLAYFORM0
where the matrix INLINEFORM0 serves to combine the different sources of information INLINEFORM1 and INLINEFORM2 . Each component INLINEFORM3 of the attention distributions INLINEFORM4 quantifies the affinity between the INLINEFORM5 -th triplet in INLINEFORM6 and the INLINEFORM7 -th word in the corresponding input INLINEFORM8 .
The model then uses each attention distribution INLINEFORM0 to obtain a weighted sum of the encodings of the words in INLINEFORM1 , according to their relevance to the corresponding triplet INLINEFORM2 . This results in L attention vectors INLINEFORM3 , INLINEFORM4 .
The final step in the attention layer concatenates each INLINEFORM0 with INLINEFORM1 to generate the outputs INLINEFORM2 , INLINEFORM3 . Following BIBREF8 , we include the encoded triplet INLINEFORM4 in the output tensor INLINEFORM5 of this layer to prevent early summaries of relevant map information.
FC layer: The model reduces the dimensionality of each individual vector INLINEFORM0 from INLINEFORM1 to INLINEFORM2 with a fully-connected (FC) layer. The resulting L vectors are output to the next layer as columns of a context matrix INLINEFORM3 .
Decoder layer: After the FC layer, the model predicts likelihoods over the sequence of behaviors that correspond to the input instructions with a GRU network. Without loss of generality, consider the INLINEFORM0 -th recurrent cell in the GRU network. This cell takes two inputs: a hidden state vector INLINEFORM1 from the prior cell, and a one-hot embedding of the previous behavior INLINEFORM2 that was predicted by the model. Based on these inputs, the GRU cell outputs a new hidden state INLINEFORM3 to compute likelihoods for the next behavior. These likelihoods are estimated by combining the output state INLINEFORM4 with relevant information from the context INLINEFORM5 : DISPLAYFORM0
where INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are trainable parameters. The attention vector INLINEFORM3 in Eq. () quantifies the affinity of INLINEFORM4 with respect to each of the columns INLINEFORM5 of INLINEFORM6 , where INLINEFORM7 . The attention vector also helps to estimate a dynamic contextual vector INLINEFORM8 that the INLINEFORM9 -th GRU cell uses to compute logits for the next behavior: DISPLAYFORM0
with INLINEFORM0 trainable parameters. Note that INLINEFORM1 includes a value for each of the pre-defined behaviors in the graph INLINEFORM2 , as well as for a special “stop” symbol to identify the end of the output sequence.
Output layer: The final layer of the model searches for a valid sequence of robot behaviors based on the robot's initial node, the connectivity of the graph INLINEFORM0 , and the output logits from the previous decoder layer. Again, without loss of generality, consider the INLINEFORM1 -th behavior INLINEFORM2 that is finally predicted by the model. The search for this behavior is implemented as: DISPLAYFORM0
with INLINEFORM0 a masking function that takes as input the graph INLINEFORM1 and the node INLINEFORM2 that the robot reaches after following the sequence of behaviors INLINEFORM3 previously predicted by the model. The INLINEFORM4 function returns a vector of the same dimensionality as the logits INLINEFORM5 , but with zeros for the valid behaviors after the last location INLINEFORM6 and for the special stop symbol, and INLINEFORM7 for any invalid predictions according to the connectivity of the behavioral navigation graph.
## Dataset
We created a new dataset for the problem of following navigation instructions under the behavioral navigation framework of BIBREF5 . This dataset was created using Amazon Mechanical Turk and 100 maps of simulated indoor environments, each with 6 to 65 rooms. To the best of our knowledge, this is the first benchmark for comparing translation models in the context of behavioral robot navigation.
As shown in Table TABREF16 , the dataset consists of 8066 pairs of free-form natural language instructions and navigation plans for training. This training data was collected from 88 unique simulated environments, totaling 6064 distinct navigation plans (2002 plans have two different navigation instructions each; the rest has one). The dataset contains two test set variants:
While the dataset was collected with simulated environments, no structure was imposed on the navigation instructions while crowd-sourcing data. Thus, many instructions in our dataset are ambiguous. Moreover, the order of the behaviors in the instructions is not always the same. For instance, a person said “turn right and advance” to describe part of a route, while another person said “go straight after turning right” in a similar situation. The high variability present in the natural language descriptions of our dataset makes the problem of decoding instructions into behaviors not trivial. See Appendix A of the supplementary material for additional details on our data collection effort.
## Experiments
This section describes our evaluation of the proposed approach for interpreting navigation commands in natural language. We provide both quantitative and qualitative results.
## Evaluation Metrics
While computing evaluation metrics, we only consider the behaviors present in the route because they are sufficient to recover the high-level navigation plan from the graph. Our metrics treat each behavior as a single token. For example, the sample plan “R-1 oor C-1 cf C-1 lt C-0 cf C-0 iol O-3" is considered to have 5 tokens, each corresponding to one of its behaviors (“oor", “cf", “lt", “cf", “iol"). In this plan, “R-1",“C-1", “C-0", and “O-3" are symbols for locations (nodes) in the graph.
We compare the performance of translation approaches based on four metrics:
[align=left,leftmargin=0em,labelsep=0.4em,font=]
As in BIBREF20 , EM is 1 if a predicted plan matches exactly the ground truth; otherwise it is 0.
The harmonic average of the precision and recall over all the test set BIBREF26 .
The minimum number of insertions, deletions or swap operations required to transform a predicted sequence of behaviors into the ground truth sequence BIBREF27 .
GM is 1 if a predicted plan reaches the ground truth destination (even if the full sequence of behaviors does not match exactly the ground truth). Otherwise, GM is 0.
## Models Used in the Evaluation
We compare the proposed approach for translating natural language instructions into a navigation plan against alternative deep-learning models:
[align=left,leftmargin=0em,labelsep=0.4em,font=]
The baseline approach is based on BIBREF20 . It divides the task of interpreting commands for behavioral navigation into two steps: path generation, and path verification. For path generation, this baseline uses a standard sequence-to-sequence model augmented with an attention mechanism, similar to BIBREF23 , BIBREF6 . For path verification, the baseline uses depth-first search to find a route in the graph that matches the sequence of predicted behaviors. If no route matches perfectly, the baseline changes up to three behaviors in the predicted sequence to try to turn it into a valid path.
To test the impact of using the behavioral graphs as an extra input to our translation model, we implemented a version of our approach that only takes natural language instructions as input. In this ablation model, the output of the bidirectional GRU that encodes the input instruction INLINEFORM0 is directly fed to the decoder layer. This model does not have the attention and FC layers described in Sec. SECREF4 , nor uses the masking function in the output layer.
This model is the same as the previous Ablation model, but with the masking function in the output layer.
## Implementation Details
We pre-processed the inputs to the various models that are considered in our experiment. In particular, we lowercased, tokenized, spell-checked and lemmatized the input instructions in text-form using WordNet BIBREF28 . We also truncated the graphs to a maximum of 300 triplets, and the navigational instructions to a maximum of 150 words. Only 6.4% (5.4%) of the unique graphs in the training (validation) set had more than 300 triplets, and less than 0.15% of the natural language instructions in these sets had more than 150 tokens.
The dimensionality of the hidden state of the GRU networks was set to 128 in all the experiments. In general, we used 12.5% of the training set as validation for choosing models' hyper-parameters. In particular, we used dropout after the encoder and the fully-connected layers of the proposed model to reduce overfitting. Best performance was achieved with a dropout rate of 0.5 and batch size equal to 256. We also used scheduled sampling BIBREF29 at training time for all models except the baseline.
We input the triplets from the graph to our proposed model in alphabetical order, and consider a modification where the triplets that surround the start location of the robot are provided first in the input graph sequence. We hypothesized that such rearrangement would help identify the starting location (node) of the robot in the graph. In turn, this could facilitate the prediction of correct output sequences. In the remaining of the paper, we refer to models that were provided a rearranged graph, beginning with the starting location of the robot, as models with “Ordered Triplets”.
## Quantitative Evaluation
Table TABREF28 shows the performance of the models considered in our evaluation on both test sets. The next two sections discuss the results in detail.
First, we can observe that the final model “Ours with Mask and Ordered Triplets” outperforms the Baseline and Ablation models on all metrics in previously seen environments. The difference in performance is particularly evident for the Exact Match and Goal Match metrics, with our model increasing accuracy by 35% and 25% in comparison to the Baseline and Ablation models, respectively. These results suggest that providing the behavioral navigation graph to the model and allowing it to process this information as a knowledge base in an end-to-end fashion is beneficial.
We can also observe from Table TABREF28 that the masking function of Eq. ( EQREF12 ) tends to increase performance in the Test-Repeated Set by constraining the output sequence to a valid set of navigation behaviors. For the Ablation model, using the masking function leads to about INLINEFORM0 increase in EM and GM accuracy. For the proposed model (with or without reordering the graph triplets), the increase in accuracy is around INLINEFORM1 . Note that the impact of the masking function is less evident in terms of the F1 score because this metric considers if a predicted behavior exists in the ground truth navigation plan, irrespective of its specific position in the output sequence.
The results in the last four rows of Table TABREF28 suggest that ordering the graph triplets can facilitate predicting correct navigation plans in previously seen environments. Providing the triplets that surround the starting location of the robot first to the model leads to a boost of INLINEFORM0 in EM and GM performance. The rearrangement of the graph triplets also helps to reduce ED and increase F1.
Lastly, it is worth noting that our proposed model (last row of Table TABREF28 ) outperforms all other models in previously seen environments. In particular, we obtain over INLINEFORM0 increase in EM and GM between our model and the next best two models.
The previous section evaluated model performance on new instructions (and corresponding navigation plans) for environments that were previously seen at training time. Here, we examine whether the trained models succeed on environments that are completely new.
The evaluation on the Test-New Set helps understand the generalization capabilities of the models under consideration. This experiment is more challenging than the one in the previous section, as can be seen in performance drops in Table TABREF28 for the new environments. Nonetheless, the insights from the previous section still hold: masking in the output layer and reordering the graph triplets tend to increase performance.
Even though the results in Table TABREF28 suggest that there is room for future work on decoding natural language instructions, our model still outperforms the baselines by a clear margin in new environments. For instance, the difference between our model and the second best model in the Test-New set is about INLINEFORM0 EM and GM. Note that the average number of actions in the ground truth output sequences is 7.07 for the Test-New set. Our model's predictions are just INLINEFORM1 edits off on average from the correct navigation plans.
## Qualitative Evaluation
This section discusses qualitative results to better understand how the proposed model uses the navigation graph.
We analyze the evolution of the attention weights INLINEFORM0 in Eq. () to assess if the decoder layer of the proposed model is attending to the correct parts of the behavioral graph when making predictions. Fig FIGREF33 (b) shows an example of the resulting attention map for the case of a correct prediction. In the Figure, the attention map is depicted as a scaled and normalized 2D array of color codes. Each column in the array shows the attention distribution INLINEFORM1 used to generate the predicted output at step INLINEFORM2 . Consequently, each row in the array represents a triplet in the corresponding behavioral graph. This graph consists of 72 triplets for Fig FIGREF33 (b).
We observe a locality effect associated to the attention coefficients corresponding to high values (bright areas) in each column of Fig FIGREF33 (b). This suggests that the decoder is paying attention to graph triplets associated to particular neighborhoods of the environment in each prediction step. We include additional attention visualizations in the supplementary Appendix, including cases where the dynamics of the attention distribution are harder to interpret.
All the routes in our dataset are the shortest paths from a start location to a given destination. Thus, we collected a few additional natural language instructions to check if our model was able to follow navigation instructions describing sub-optimal paths. One such example is shown in Fig. FIGREF37 , where the blue route (shortest path) and the red route (alternative path) are described by:
[leftmargin=*, labelsep=0.2em, itemsep=0em]
“Go out the office and make a left. Turn right at the corner and go down the hall. Make a right at the next corner and enter the kitchen in front of table.”
“Exit the room 0 and turn right, go to the end of the corridor and turn left, go straight to the end of the corridor and turn left again. After passing bookshelf on your left and table on your right, Enter the kitchen on your right.”
For both routes, the proposed model was able to predict the correct sequence of navigation behaviors. This result suggests that the model is indeed using the input instructions and is not just approximating shortest paths in the behavioral graph. Other examples on the prediction of sub-obtimal paths are described in the Appendix.
## Conclusion
This work introduced behavioral navigation through free-form natural language instructions as a challenging and a novel task that falls at the intersection of natural language processing and robotics. This problem has a range of interesting cross-domain applications, including information retrieval.
We proposed an end-to-end system to translate user instructions to a high-level navigation plan. Our model utilized an attention mechanism to merge relevant information from the navigation instructions with a behavioral graph of the environment. The model then used a decoder to predict a sequence of navigation behaviors that matched the input commands.
As part of this effort, we contributed a new dataset of 11,051 pairs of user instructions and navigation plans from 100 different environments. Our model achieved the best performance in this dataset in comparison to a two-step baseline approach for interpreting navigation instructions, and a sequence-to-sequence model that does not consider the behavioral graph. Our quantitative and qualitative results suggest that attention mechanisms can help leverage the behavioral graph as a relevant knowledge base to facilitate the translation of free-form navigation instructions. Overall, our approach demonstrated practical form of learning for a complex and useful task. In future work, we are interested in investigating mechanisms to improve generalization to new environments. For example, pointer and graph networks BIBREF30 , BIBREF31 are a promising direction to help supervise translation models and predict motion behaviors.
## Acknowledgments
The Toyota Research Institute (TRI) provided funds to assist with this research, but this paper solely reflects the opinions and conclusions of its authors and not TRI or any other Toyota entity. This work is also partially funded by Fondecyt grant 1181739, Conicyt, Chile. The authors would also like to thank Gabriel Sepúlveda for his assistance with parts of this project.
| 14
|
1810.04428
|
Improving Neural Text Simplification Model with Simplified Corpora
|
# Improving Neural Text Simplification Model with Simplified Corpora
## Abstract
Text simplification (TS) can be viewed as monolingual translation task, translating between text variations within a single language. Recent neural TS models draw on insights from neural machine translation to learn lexical simplification and content reduction using encoder-decoder model. But different from neural machine translation, we cannot obtain enough ordinary and simplified sentence pairs for TS, which are expensive and time-consuming to build. Target-side simplified sentences plays an important role in boosting fluency for statistical TS, and we investigate the use of simplified sentences to train, with no changes to the network architecture. We propose to pair simple training sentence with a synthetic ordinary sentence via back-translation, and treating this synthetic data as additional training data. We train encoder-decoder model using synthetic sentence pairs and original sentence pairs, which can obtain substantial improvements on the available WikiLarge data and WikiSmall data compared with the state-of-the-art methods.
## Introduction
Text simplification aims to reduce the lexical and structural complexity of a text, while still retaining the semantic meaning, which can help children, non-native speakers, and people with cognitive disabilities, to understand text better. One of the methods of automatic text simplification can be generally divided into three categories: lexical simplification (LS) BIBREF0 , BIBREF1 , rule-based BIBREF2 , and machine translation (MT) BIBREF3 , BIBREF4 . LS is mainly used to simplify text by substituting infrequent and difficult words with frequent and easier words. However, there are several challenges for the LS approach: a great number of transformation rules are required for reasonable coverage and should be applied based on the specific context; third, the syntax and semantic meaning of the sentence is hard to retain. Rule-based approaches use hand-crafted rules for lexical and syntactic simplification, for example, substituting difficult words in a predefined vocabulary. However, such approaches need a lot of human-involvement to manually define these rules, and it is impossible to give all possible simplification rules. MT-based approach has attracted great attention in the last several years, which addresses text simplification as a monolingual machine translation problem translating from 'ordinary' and 'simplified' sentences.
In recent years, neural Machine Translation (NMT) is a newly-proposed deep learning approach and achieves very impressive results BIBREF5 , BIBREF6 , BIBREF7 . Unlike the traditional phrased-based machine translation system which operates on small components separately, NMT system is being trained end-to-end, without the need to have external decoders, language models or phrase tables. Therefore, the existing architectures in NMT are used for text simplification BIBREF8 , BIBREF4 . However, most recent work using NMT is limited to the training data that are scarce and expensive to build. Language models trained on simplified corpora have played a central role in statistical text simplification BIBREF9 , BIBREF10 . One main reason is the amount of available simplified corpora typically far exceeds the amount of parallel data. The performance of models can be typically improved when trained on more data. Therefore, we expect simplified corpora to be especially helpful for NMT models.
In contrast to previous work, which uses the existing NMT models, we explore strategy to include simplified training corpora in the training process without changing the neural network architecture. We first propose to pair simplified training sentences with synthetic ordinary sentences during training, and treat this synthetic data as additional training data. We obtain synthetic ordinary sentences through back-translation, i.e. an automatic translation of the simplified sentence into the ordinary sentence BIBREF11 . Then, we mix the synthetic data into the original (simplified-ordinary) data to train NMT model. Experimental results on two publicly available datasets show that we can improve the text simplification quality of NMT models by mixing simplified sentences into the training set over NMT model only using the original training data.
## Related Work
Automatic TS is a complicated natural language processing (NLP) task, which consists of lexical and syntactic simplification levels BIBREF12 . It has attracted much attention recently as it could make texts more accessible to wider audiences, and used as a pre-processing step, improve performances of various NLP tasks and systems BIBREF13 , BIBREF14 , BIBREF15 . Usually, hand-crafted, supervised, and unsupervised methods based on resources like English Wikipedia and Simple English Wikipedia (EW-SEW) BIBREF10 are utilized for extracting simplification rules. It is very easy to mix up the automatic TS task and the automatic summarization task BIBREF3 , BIBREF16 , BIBREF6 . TS is different from text summarization as the focus of text summarization is to reduce the length and redundant content.
At the lexical level, lexical simplification systems often substitute difficult words using more common words, which only require a large corpus of regular text to obtain word embeddings to get words similar to the complex word BIBREF1 , BIBREF9 . Biran et al. BIBREF0 adopted an unsupervised method for learning pairs of complex and simpler synonyms from a corpus consisting of Wikipedia and Simple Wikipedia. At the sentence level, a sentence simplification model was proposed by tree transformation based on statistical machine translation (SMT) BIBREF3 . Woodsend and Lapata BIBREF17 presented a data-driven model based on a quasi-synchronous grammar, a formalism that can naturally capture structural mismatches and complex rewrite operations. Wubben et al. BIBREF18 proposed a phrase-based machine translation (PBMT) model that is trained on ordinary-simplified sentence pairs. Xu et al. BIBREF19 proposed a syntax-based machine translation model using simplification-specific objective functions and features to encourage simpler output.
Compared with SMT, neural machine translation (NMT) has shown to produce state-of-the-art results BIBREF5 , BIBREF7 . The central approach of NMT is an encoder-decoder architecture implemented by recurrent neural networks, which can represent the input sequence as a vector, and then decode that vector into an output sequence. Therefore, NMT models were used for text simplification task, and achieved good results BIBREF8 , BIBREF4 , BIBREF20 . The main limitation of the aforementioned NMT models for text simplification depended on the parallel ordinary-simplified sentence pairs. Because ordinary-simplified sentence pairs are expensive and time-consuming to build, the available largest data is EW-SEW that only have 296,402 sentence pairs. The dataset is insufficiency for NMT model if we want to NMT model can obtain the best parameters. Considering simplified data plays an important role in boosting fluency for phrase-based text simplification, and we investigate the use of simplified data for text simplification. We are the first to show that we can effectively adapt neural translation models for text simplifiation with simplified corpora.
## Simplified Corpora
We collected a simplified dataset from Simple English Wikipedia that are freely available, which has been previously used for many text simplification methods BIBREF0 , BIBREF10 , BIBREF3 . The simple English Wikipedia is pretty easy to understand than normal English Wikipedia. We downloaded all articles from Simple English Wikipedia. For these articles, we removed stubs, navigation pages and any article that consisted of a single sentence. We then split them into sentences with the Stanford CorNLP BIBREF21 , and deleted these sentences whose number of words are smaller than 10 or large than 40. After removing repeated sentences, we chose 600K sentences as the simplified data with 11.6M words, and the size of vocabulary is 82K.
## Text Simplification using Neural Machine Translation
Our work is built on attention-based NMT BIBREF5 as an encoder-decoder network with recurrent neural networks (RNN), which simultaneously conducts dynamic alignment and generation of the target simplified sentence.
The encoder uses a bidirectional RNN that consists of forward and backward RNN. Given a source sentence INLINEFORM0 , the forward RNN and backward RNN calculate forward hidden states INLINEFORM1 and backward hidden states INLINEFORM2 , respectively. The annotation vector INLINEFORM3 is obtained by concatenating INLINEFORM4 and INLINEFORM5 .
The decoder is a RNN that predicts a target simplificated sentence with Gated Recurrent Unit (GRU) BIBREF22 . Given the previously generated target (simplified) sentence INLINEFORM0 , the probability of next target word INLINEFORM1 is DISPLAYFORM0
where INLINEFORM0 is a non-linear function, INLINEFORM1 is the embedding of INLINEFORM2 , and INLINEFORM3 is a decoding state for time step INLINEFORM4 .
State INLINEFORM0 is calculated by DISPLAYFORM0
where INLINEFORM0 is the activation function GRU.
The INLINEFORM0 is the context vector computed as a weighted annotation INLINEFORM1 , computed by DISPLAYFORM0
where the weight INLINEFORM0 is computed by DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 , INLINEFORM1 and INLINEFORM2 are weight matrices. The training objective is to maximize the likelihood of the training data. Beam search is employed for decoding.
## Synthetic Simplified Sentences
We train an auxiliary system using NMT model from the simplified sentence to the ordinary sentence, which is first trained on the available parallel data. For leveraging simplified sentences to improve the quality of NMT model for text simplification, we propose to adapt the back-translation approach proposed by Sennrich et al. BIBREF11 to our scenario. More concretely, Given one sentence in simplified sentences, we use the simplified-ordinary system in translate mode with greedy decoding to translate it to the ordinary sentences, which is denoted as back-translation. This way, we obtain a synthetic parallel simplified-ordinary sentences. Both the synthetic sentences and the available parallel data are used as training data for the original NMT system.
## Evaluation
We evaluate the performance of text simplification using neural machine translation on available parallel sentences and additional simplified sentences.
Dataset. We use two simplification datasets (WikiSmall and WikiLarge). WikiSmall consists of ordinary and simplified sentences from the ordinary and simple English Wikipedias, which has been used as benchmark for evaluating text simplification BIBREF17 , BIBREF18 , BIBREF8 . The training set has 89,042 sentence pairs, and the test set has 100 pairs. WikiLarge is also from Wikipedia corpus whose training set contains 296,402 sentence pairs BIBREF19 , BIBREF20 . WikiLarge includes 8 (reference) simplifications for 2,359 sentences split into 2,000 for development and 359 for testing.
Metrics. Three metrics in text simplification are chosen in this paper. BLEU BIBREF5 is one traditional machine translation metric to assess the degree to which translated simplifications differed from reference simplifications. FKGL measures the readability of the output BIBREF23 . A small FKGL represents simpler output. SARI is a recent text-simplification metric by comparing the output against the source and reference simplifications BIBREF20 .
We evaluate the output of all systems using human evaluation. The metric is denoted as Simplicity BIBREF8 . The three non-native fluent English speakers are shown reference sentences and output sentences. They are asked whether the output sentence is much simpler (+2), somewhat simpler (+1), equally (0), somewhat more difficult (-1), and much more difficult (-2) than the reference sentence.
Methods. We use OpenNMT BIBREF24 as the implementation of the NMT system for all experiments BIBREF5 . We generally follow the default settings and training procedure described by Klein et al.(2017). We replace out-of-vocabulary words with a special UNK symbol. At prediction time, we replace UNK words with the highest probability score from the attention layer. OpenNMT system used on parallel data is the baseline system. To obtain a synthetic parallel training set, we back-translate a random sample of 100K sentences from the collected simplified corpora. OpenNMT used on parallel data and synthetic data is our model. The benchmarks are run on a Intel(R) Core(TM) i7-5930K CPU@3.50GHz, 32GB Mem, trained on 1 GPU GeForce GTX 1080 (Pascal) with CUDA v. 8.0.
We choose three statistical text simplification systems. PBMT-R is a phrase-based method with a reranking post-processing step BIBREF18 . Hybrid performs sentence splitting and deletion operations based on discourse representation structures, and then simplifies sentences with PBMT-R BIBREF25 . SBMT-SARI BIBREF19 is syntax-based translation model using PPDB paraphrase database BIBREF26 and modifies tuning function (using SARI). We choose two neural text simplification systems. NMT is a basic attention-based encoder-decoder model which uses OpenNMT framework to train with two LSTM layers, hidden states of size 500 and 500 hidden units, SGD optimizer, and a dropout rate of 0.3 BIBREF8 . Dress is an encoder-decoder model coupled with a deep reinforcement learning framework, and the parameters are chosen according to the original paper BIBREF20 . For the experiments with synthetic parallel data, we back-translate a random sample of 60 000 sentences from the collected simplified sentences into ordinary sentences. Our model is trained on synthetic data and the available parallel data, denoted as NMT+synthetic.
Results. Table 1 shows the results of all models on WikiLarge dataset. We can see that our method (NMT+synthetic) can obtain higher BLEU, lower FKGL and high SARI compared with other models, except Dress on FKGL and SBMT-SARI on SARI. It verified that including synthetic data during training is very effective, and yields an improvement over our baseline NMF by 2.11 BLEU, 1.7 FKGL and 1.07 SARI. We also substantially outperform Dress, who previously reported SOTA result. The results of our human evaluation using Simplicity are also presented in Table 1. NMT on synthetic data is significantly better than PBMT-R, Dress, and SBMT-SARI on Simplicity. It indicates that our method with simplified data is effective at creating simpler output.
Results on WikiSmall dataset are shown in Table 2. We see substantial improvements (6.37 BLEU) than NMT from adding simplified training data with synthetic ordinary sentences. Compared with statistical machine translation models (PBMT-R, Hybrid, SBMT-SARI), our method (NMT+synthetic) still have better results, but slightly worse FKGL and SARI. Similar to the results in WikiLarge, the results of our human evaluation using Simplicity outperforms the other models. In conclusion, Our method produces better results comparing with the baselines, which demonstrates the effectiveness of adding simplified training data.
## Conclusion
In this paper, we propose one simple method to use simplified corpora during training of NMT systems, with no changes to the network architecture. In the experiments on two datasets, we achieve substantial gains in all tasks, and new SOTA results, via back-translation of simplified sentences into the ordinary sentences, and treating this synthetic data as additional training data. Because we do not change the neural network architecture to integrate simplified corpora, our method can be easily applied to other Neural Text Simplification (NTS) systems. We expect that the effectiveness of our method not only varies with the quality of the NTS system used for back-translation, but also depends on the amount of available parallel and simplified corpora. In the paper, we have only utilized data from Wikipedia for simplified sentences. In the future, many other text sources are available and the impact of not only size, but also of domain should be investigated.
| 7
|
1810.05241
|
Generating Diverse Numbers of Diverse Keyphrases
|
# Generating Diverse Numbers of Diverse Keyphrases
## Abstract
Existing keyphrase generation studies suffer from the problems of generating duplicate phrases and deficient evaluation based on a fixed number of predicted phrases. We propose a recurrent generative model that generates multiple keyphrases sequentially from a text, with specific modules that promote generation diversity. We further propose two new metrics that consider a variable number of phrases. With both existing and proposed evaluation setups, our model demonstrates superior performance to baselines on three types of keyphrase generation datasets, including two newly introduced in this work: StackExchange and TextWorld ACG. In contrast to previous keyphrase generation approaches, our model generates sets of diverse keyphrases of a variable number.
## Introduction
Keyphrase generation is the task of automatically predicting keyphrases given a source text. Desired keyphrases are often multi-word units that summarize the high-level meaning and highlight certain important topics or information of the source text. Consequently, models that can successfully perform this task should be capable of not only distilling high-level information from a document, but also locating specific, important snippets therein.
To make the problem even more challenging, a keyphrase may or may not be a substring of the source text (i.e., it may be present or absent). Moreover, a given source text is usually associated with a set of multiple keyphrases. Thus, keyphrase generation is an instance of the set generation problem, where both the size of the set and the size (i.e., the number of tokens in a phrase) of each element can vary depending on the source.
Similar to summarization, keyphrase generation is often formulated as a sequence-to-sequence (Seq2Seq) generation task in most prior studies BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . Conditioned on a source text, Seq2Seq models generate phrases individually or as a longer sequence jointed by delimiting tokens. Since standard Seq2Seq models generate only one sequence at a time, thus to generate multiple phrases, a common approach is to over-generate using beam search with a large beam width. Models are then evaluated by taking a fixed number of top predicted phrases (typically 5 or 10) and comparing them against the ground truth keyphrases.
Though this approach has achieved good empirical results, we argue that it suffers from two major limitations. Firstly, models that use beam search to generate multiple keyphrases generally lack the ability to determine the dynamic number of keyphrases needed for different source texts. Meanwhile, the parallelism in beam search also fails to model the inter-relation among the generated phrases, which can often result in diminished diversity in the output. Although certain existing models take output diversity into consideration during training BIBREF1 , BIBREF2 , the effort is significantly undermined during decoding due to the reliance on over-generation and phrase ranking with beam search.
Secondly, the current evaluation setup is rather problematic, since existing studies attempt to match a fixed number of outputs against a variable number of ground truth keyphrases. Empirically, the number of keyphrases can vary drastically for different source texts, depending on a plethora of factors including the length or genre of the text, the granularity of keyphrase annotation, etc. For the several commonly used keyphrase generation datasets, for example, the average number of keyphrases per data point can range from 5.3 to 15.7, with variances sometimes as large as 64.6 (Table TABREF1 ). Therefore, using an arbitrary, fixed number INLINEFORM0 to evaluate entire datasets is not appropriate. In fact, under this evaluation setup, the F1 score for the oracle model on the KP20k dataset is 0.858 for INLINEFORM1 and 0.626 for INLINEFORM2 , which apparently poses serious normalization issues as evaluation metrics.
To overcome these problems, we propose novel decoding strategies and evaluation metrics for the keyphrase generation task. The main contributions of this work are as follows:
## Keyphrase Extraction and Generation
Traditional keyphrase extraction has been studied extensively in past decades. In most existing literature, keyphrase extraction has been formulated as a two-step process. First, lexical features such as part-of-speech tags are used to determine a list of phrase candidates by heuristic methods BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . Second, a ranking algorithm is adopted to rank the candidate list and the top ranked candidates are selected as keyphrases. A wide variety of methods were applied for ranking, such as bagged decision trees BIBREF8 , BIBREF9 , Multi-Layer Perceptron, Support Vector Machine BIBREF9 and PageRank BIBREF10 , BIBREF11 , BIBREF12 . Recently, BIBREF13 , BIBREF14 , BIBREF15 used sequence labeling models to extract keyphrases from text. Similarly, BIBREF16 used Pointer Networks to point to the start and end positions of keyphrases in a source text.
The main drawback of keyphrase extraction is that sometimes keyphrases are absent from the source text, thus an extractive model will fail predicting those keyphrases. BIBREF0 first proposed the CopyRNN, a neural generative model that both generates words from vocabulary and points to words from the source text. Recently, based on the CopyRNN architecture, BIBREF1 proposed the CorrRNN, which takes states and attention vectors from previous steps into account in both encoder and decoder to reduce duplication and improve coverage. BIBREF2 proposed semi-supervised methods by leveraging both labeled and unlabeled data for training. BIBREF3 , BIBREF2 proposed to use structure information (e.g., title of source text) to improve keyphrase generation performance. Note that none of the above works are able to generate variable number of phrases, which is one of our contributions.
## Sequence to Sequence Generation
Sequence to Sequence (Seq2Seq) learning was first introduced by BIBREF17 ; together with the soft attention mechanism of BIBREF18 , it has been widely used in natural language generation tasks. BIBREF19 , BIBREF20 used a mixture of generation and pointing to overcome the problem of large vocabulary size. BIBREF21 , BIBREF22 applied Seq2Seq models on summary generation tasks, while BIBREF23 , BIBREF24 generated questions conditioned on documents and answers from machine comprehension datasets. Seq2Seq was also applied on neural sentence simplification BIBREF25 and paraphrase generation tasks BIBREF26 .
Given a source text consisting of INLINEFORM0 words INLINEFORM1 , the encoder converts their corresponding embeddings INLINEFORM2 into a set of INLINEFORM3 real-valued vectors INLINEFORM4 with a bidirectional GRU BIBREF27 : DISPLAYFORM0
Dropout BIBREF28 is applied to both INLINEFORM0 and INLINEFORM1 for regularization.
The decoder is a uni-directional GRU, which generates a new state INLINEFORM0 at each time-step INLINEFORM1 from the word embedding INLINEFORM2 and the recurrent state INLINEFORM3 : DISPLAYFORM0
The initial state INLINEFORM0 is derived from the final encoder state INLINEFORM1 by applying a single-layer feed-forward neural net (FNN): INLINEFORM2 . Dropout is applied to both the embeddings INLINEFORM3 and the GRU states INLINEFORM4 .
When generating token INLINEFORM0 , in order to better incorporate information from the source text, an attention mechanism BIBREF18 is employed to infer the importance INLINEFORM1 of each source word INLINEFORM2 given the current decoder state INLINEFORM3 . This importance is measured by an energy function with a 2-layer FNN: DISPLAYFORM0
The output over all decoding steps INLINEFORM0 thus define a distribution over the source sequence: DISPLAYFORM0
These attention scores are then used as weights for a refined representation of the source encodings, which is then concatenated to the decoder state INLINEFORM0 to derive a generative distribution INLINEFORM1 : DISPLAYFORM0
where the output size of INLINEFORM0 equals to the target vocabulary size. Subscript INLINEFORM1 indicates the abstractive nature of INLINEFORM2 since it is a distribution over a prescribed vocabulary.
We employ the pointer softmax BIBREF19 mechanism to switch between generating a token INLINEFORM0 (from a vocabulary) and pointing (to a token in the source text). Specifically, the pointer softmax module computes a scalar switch INLINEFORM1 at each generation time-step and uses it to interpolate the abstractive distribution INLINEFORM2 over the vocabulary (see Equation EQREF16 ) and the extractive distribution INLINEFORM3 over the source text tokens: DISPLAYFORM0
where INLINEFORM0 is conditioned on both the attention-weighted source representation INLINEFORM1 and the decoder state INLINEFORM2 : DISPLAYFORM0
## Model Architecture
Given a piece of source text, our objective is to generate a variable number of multi-word phrases. To this end, we opt for the sequence-to-sequence framework (Seq2Seq) as the basis of our model, combined with attention and pointer softmax mechanisms in the decoder.
Since each data example contains one source text sequence and multiple target phrase sequences (dubbed One2Many, and each sequence can be of multi-word), two paradigms can be adopted for training Seq2Seq models. The first one BIBREF0 is to divide each One2Many data example into multiple One2One examples, and the resulting models (e.g. CopyRNN) can generate one phrase at once and must rely on beam search technique to produce more unique phrases.
To enable models to generate multiple phrases and control the number to output, we propose the second training paradigm One2Seq, in which we concatenate multiple phrases into a single sequence with a delimiter INLINEFORM0 SEP INLINEFORM1 , and this concatenated sequence is then used as the target for sequence generation during training. An overview of the model's structure is shown in Figure FIGREF8 .
## Notations
In the following subsections, we use INLINEFORM0 to denote input text tokens, INLINEFORM1 to denote token embeddings, INLINEFORM2 to denote hidden states, and INLINEFORM3 to denote output text tokens. Superscripts denote time-steps in a sequence, and subscripts INLINEFORM4 and INLINEFORM5 indicate whether a variable resides in the encoder or the decoder of the model, respectively. The absence of a superscript indicates multiplicity in the time dimension. INLINEFORM6 refers to a linear transformation and INLINEFORM7 refers to it followed by a non-linear activation function INLINEFORM8 . Angled brackets, INLINEFORM9 , denote concatenation.
## Mechanisms for Diverse Generation
There are usually multiple keyphrases for a given source text because each keyphrase represents certain aspects of the text. Therefore keyphrase diversity is desired for the keyphrase generation. Most previous keyphrase generation models generate multiple phrases by over-generation, which is highly prone to generate similar phrases due to the nature of beam search. Given our objective to generate variable numbers of keyphrases, we need to adopt new strategies for achieving better diversity in the output.
Recall that we represent variable numbers of keyphrases as delimiter-separated sequences. One particular issue we observed during error analysis is that the model tends to produce identical tokens following the delimiter token. For example, suppose a target sequence contains INLINEFORM0 delimiter tokens at time-steps INLINEFORM1 . During training, the model is rewarded for generating the same delimiter token at these time-steps, which presumably introduces much homogeneity in the corresponding decoder states INLINEFORM2 . When these states are subsequently used as inputs at the time-steps immediately following the delimiter, the decoder naturally produces highly similar distributions over the following tokens, resulting in identical tokens being decoded. To alleviate this problem, we propose two plug-in components for the sequential generation model.
We propose a mechanism called semantic coverage that focuses on the semantic representations of generated phrases. Specifically, we introduce another uni-directional recurrent model INLINEFORM0 (dubbed target encoder) which encodes decoder-generated tokens INLINEFORM1 , where INLINEFORM2 , into hidden states INLINEFORM3 . This state is then taken as an extra input to the decoder GRU, modifying Equation EQREF12 to: DISPLAYFORM0
If the target encoder were to be updated with the training signal from generation (i.e., backpropagating error from the decoder GRU to the target encoder), the resulting decoder is essentially a 2-layer GRU with residual connections. Instead, inspired by previous representation learning works BIBREF29 , BIBREF30 , BIBREF31 , we train the target encoder in an self-supervised fashion (Figure FIGREF8 ). That is, we extract target encoder's final hidden state vector INLINEFORM0 , where INLINEFORM1 is the length of target sequence, and use it as a general representation of the target phrases. We train by maximizing the mutual information between these phrase representations and the final state of the source encoder INLINEFORM2 as follows. For each phrase representation vector INLINEFORM3 , we take the enocdings INLINEFORM4 of INLINEFORM5 different source texts, where INLINEFORM6 is the encoder representation for the current source text, and the remaining INLINEFORM7 are negative samples (sampled at random) from the training data. The target encoder is trained to minimize the classification loss: DISPLAYFORM0
where INLINEFORM0 is bi-linear transformation.
The motivation here is to constrain the overall representation of generated keyphrase to be semantically close to the overall meaning of the source text. With such representations as input to the decoder, the semantic coverage mechanism can potentially help to provide useful keyphrase information and guide generation.
We also propose orthogonal regularization, which explicitly encourages the delimiter-generating decoder states to be different from each other. This is inspired by BIBREF32 , who use orthogonal regularization to encourage representations across domains to be as distinct as possible. Specifically, we stack the decoder hidden states corresponding to delimiters together to form matrix INLINEFORM0 and use the following equation as the orthogonal regularization loss: DISPLAYFORM0
where INLINEFORM0 is the matrix transpose of INLINEFORM1 , INLINEFORM2 is the identity matrix of rank INLINEFORM3 , INLINEFORM4 indicates element wise multiplication, INLINEFORM5 indicates INLINEFORM6 norm of each element in a matrix INLINEFORM7 . This loss function prefers orthogonality among the hidden states INLINEFORM8 and thus improves diversity in the tokens following the delimiters.
We adopt the widely used negative log-likelihood loss in our sequence generation model, denoted as INLINEFORM0 . The overall loss we use in our model is DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are hyper-parameters.
## Decoding Strategies
According to different task requirements, various decoding methods can be applied to generate the target sequence INLINEFORM0 . Prior studies BIBREF0 , BIBREF7 focus more on generating excessive number of phrases by leveraging beam search to proliferate the output phrases. In contrast, models trained under One2Seq paradigm are capable of determining the proper number of phrases to output. In light of previous research in psychology BIBREF33 , BIBREF34 , we name these two decoding/search strategies as Exhaustive Decoding and Self-terminating Decoding, respectively, due to their resemblance to the way humans behave in serial memory tasks. Simply speaking, the major difference lies in whether a model is capable of controlling the number of phrases to output. We describe the detailed decoding strategies used in this study as follows:
As traditional keyphrase tasks evaluate models with a fixed number of top-ranked predictions (say F-score @5 and @10), existing keyphrase generation studies have to over-generate phrases by means of beam search (commonly with a large beam size, e.g., 150 and 200 in BIBREF3 , BIBREF0 , respectively), a heuristic search algorithm that returns INLINEFORM0 approximate optimal sequences. For the One2One setting, each returned sequence is a unique phrase itself. But for One2Seq, each produced sequence contains several phrases and additional processes BIBREF2 are needed to obtain the final unique (ordered) phrase list.
It is worth noting that the time complexity of beam search is INLINEFORM0 , where INLINEFORM1 is the beam width, and INLINEFORM2 is the maximum length of generated sequences. Therefore the exhaustive decoding is generally very computationally expensive, especially for One2Seq setting where INLINEFORM3 is much larger than in One2One. It is also wasteful as we observe that less than 5% of phrases generated by One2Seq models are unique.
An innate characteristic of keyphrase tasks is that the number of keyphrases varies depending on the document and dataset genre, therefore dynamically outputting a variable number of phrases is a desirable property for keyphrase generation models. Since our proposed model is trained to generate a variable number of phrases as a single sequence joined by delimiters, we can obtain multiple phrases by simply decoding a single sequence for each given source text. The resulting model thus implicitly performs the additional task of dynamically estimating the proper size of the target phrase set: once the model believes that an adequate number of phrases have been generated, it outputs a special token INLINEFORM0 EOS INLINEFORM1 to terminate the decoding process.
One notable attribute of the self-terminating decoding strategy is that, by generating a set of phrases in a single sequence, the model conditions its current generation on all previously generated phrases. Compared to the exhaustive strategy (i.e., phrases being generated independently by beam search in parallel), our model can model the dependency among its output in a more explicit fashion. Additionally, since multiple phrases are decoded as a single sequence, decoding can be performed more efficiently than exhaustive decoding by conducting greedy search or beam search on only the top-scored sequence.
## Evaluating Keyphrase Generation
Formally, given a source text, suppose that a model predicts a list of unique keyphrases INLINEFORM0 ordered by the quality of the predictions INLINEFORM1 , and that the ground truth keyphrases for the given source text is the oracle set INLINEFORM2 . When only the top INLINEFORM3 predictions INLINEFORM4 are used for evaluation, precision, recall, and F INLINEFORM5 score are consequently conditioned on INLINEFORM6 and defined as: DISPLAYFORM0
As discussed in Section SECREF1 , the number of generated keyphrases used for evaluation can have a critical impact on the quality of the resulting evaluation metrics. Here we compare three choices of INLINEFORM0 and the implications on keyphrase evaluation for each choice:
A simple remedy is to set INLINEFORM0 as a variable number which is specific to each data example. Here we define two new metrics:
By simply extending the constant number INLINEFORM0 to different variables accordingly, both F INLINEFORM1 @ INLINEFORM2 and F INLINEFORM3 @ INLINEFORM4 are capable of reflecting the nature of variable number of phrases for each document, and a model can achieve the maximum INLINEFORM5 score of INLINEFORM6 if and only if it predicts the exact same phrases as the ground truth. Another merit of F INLINEFORM7 @ INLINEFORM8 is that it is independent from model outputs, therefore we can use it to compare existing models.
## Datasets and Experiments
In this section, we report our experiment results on multiple datasets and compare with existing models. We use INLINEFORM0 to refer to the delimiter-concatenated sequence-to-sequences model described in Section SECREF3 ; INLINEFORM1 refers to the model augmented with orthogonal regularization and semantic coverage mechanism.
To construct target sequences for training INLINEFORM0 and INLINEFORM1 , ground truth keyphrases are sorted by their order of first occurrence in the source text. Keyphrases that do not appear in the source text are appended to the end. This order may guide the attention mechanism to attend to source positions in a smoother way. Implementation details can be found in Appendix SECREF9 .
We include four non-neural extractive models and CopyRNN BIBREF0 as baselines. We use CopyRNN to denote the model reported by BIBREF0 , CopyRNN* to denote our implementation of CopyRNN based on their open sourced code. To draw fair comparison with existing study, we use the same model hyperparameter setting as used in BIBREF0 and use exhaustive decoding strategy for most experiments. KEA BIBREF4 and Maui BIBREF8 are trained on a subset of 50,000 documents from either KP20k (Table TABREF35 ) or StackEx (Table TABREF37 ) instead of all documents due to implementation limits (without fine-tuning on target dataset).
In Section SECREF42 , we apply the self-terminating decoding strategy. Since no existing model supports such decoding strategy, we only report results from our proposed models. They can be used for comparison in future studies.
## Experiments on Scientific Publications
Our first dataset consists of a collection of scientific publication datasets, namely KP20k, Inspec, Krapivin, NUS, and SemEval, that have been widely used in existing literature BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . KP20k, for example, was introduced by BIBREF0 and comprises more than half a million scientific publications. For each article, the abstract and title are used as the source text while the author keywords are used as target. The other four datasets contain much fewer articles, and thus used to test transferability of our model (without fine-tuning).
We report our model's performance on the present-keyphrase portion of the KP20k dataset in Table TABREF35 . To compare with previous works, we provide compute INLINEFORM0 and INLINEFORM1 scores. The new proposed F INLINEFORM2 @ INLINEFORM3 metric indicates consistent ranking with INLINEFORM4 for most cases. Due to its target number sensitivity, we find that its value is closer to INLINEFORM5 for KP20k and Krapivin where average target keyphrases is less and closer to INLINEFORM6 for the other three datasets.
From the result we can see that the neural-based models outperform non-neural models by large margins. Our implemented CopyRNN achieves better or comparable performance against the original model, and on NUS and SemEval the advantage is more salient.
As for the proposed models, both INLINEFORM0 and INLINEFORM1 yield comparable results to CopyRNN, indicating that One2Seq paradigm can work well as an alternative option for the keyphrase generation. INLINEFORM2 outperforms INLINEFORM3 on all metrics, suggesting the semantic coverage and orthogonal regularization help the model to generate higher quality keyphrases and achieve better generalizability. To our surprise, on the metric F INLINEFORM4 @10 for KP20k and Krapivin (average number of keyphrases is only 5), where high-recall models like CopyRNN are more favored, INLINEFORM5 is still able to outperform One2One baselines, indicating that the proposed mechanisms for diverse generation are effective.
## Experiments on The StackEx Dataset
Inspired by the StackLite tag recommendation task on Kaggle, we build a new benchmark based on the public StackExchange data. We use questions with titles as source, and user-assigned tags as target keyphrases.
Since oftentimes the questions on StackExchange contain less information than in scientific publications, there are fewer keyphrases per data point in StackEx. Furthermore, StackExchange uses a tag recommendation system that suggests topic-relevant tags to users while submitting questions; therefore, we are more likely to see general terminology such as Linux and Java. This characteristic challenges models with respect to their ability to distill major topics of a question rather than selecting specific snippets from the text.
We report our models' performance on StackEx in Table TABREF37 . Results show INLINEFORM0 performs the best; on the absent-keyphrase generation tasks, it outperforms INLINEFORM1 by a large margin.
## Generating Variable Number Keyphrases
One key advantage of our proposed model is the capability of predicting the number of keyphrases conditioned on the given source text. We thus conduct a set of experiments on KP20k and StackEx present keyphrase generation tasks, as shown in Table TABREF39 , to study such behavior. We adopt the self-terminating decoding strategy (Section SECREF28 ), and use both F INLINEFORM0 @ INLINEFORM1 and F INLINEFORM2 @ INLINEFORM3 (Section SECREF4 ) to evaluate.
In these experiments, we use beam search as in most Natural Language Generation (NLG) tasks, i.e., only use the top ranked prediction sequence as output. We compare the results with greedy search. Since no existing model is capable of generating variable number of keyphrases, in this subsection we only report performance on such setting from INLINEFORM0 and INLINEFORM1 .
From Table TABREF39 we observe that in the variable number generation setting, greedy search outperforms beam search consistently. This may because beam search tends to generate short and similar sequences. We can also see the resulting F INLINEFORM0 @ INLINEFORM1 scores are generally lower than results reported in previous subsections, this suggests an over-generation decoding strategy may still benefit from achieving higher recall.
## Ablation Study
We conduct an ablation experiment to study the effects of orthogonal regularization and semantic coverage mechanism on INLINEFORM0 . As shown in Table TABREF44 , semantic coverage provides significant boost to INLINEFORM1 's performance on all datasets. Orthogonal regularization hurts performance when is solely applied to INLINEFORM2 model. Interestingly, when both components are enabled ( INLINEFORM3 ), the model outperforms INLINEFORM4 by a noticeable margin on all datasets, this suggests the two components help keyphrase generation in a synergetic way. One future direction is to apply orthogonal regularization directly on target encoder, since the regularizer can potentially diversify target representations at phrase level, which may further encourage diverse keyphrase generation in decoder.
## Visualizing Diversified Generation
To verify our assumption that target encoding and orthogonal regularization help to boost the diversity of generated sequences, we use two metrics, one quantitative and one qualitative, to measure diversity of generation.
First, we simply calculate the average unique predictions produced by both INLINEFORM0 and INLINEFORM1 in experiments shown in Section SECREF36 . The resulting numbers are 20.38 and 89.70 for INLINEFORM2 and INLINEFORM3 respectively. Second, from the model running on the KP20k validation set, we randomly sample 2000 decoder hidden states at INLINEFORM4 steps following a delimiter ( INLINEFORM5 ) and apply an unsupervised clustering method (t-SNE BIBREF35 ) on them. From the Figure FIGREF46 we can see that hidden states sampled from INLINEFORM6 are easier to cluster while hidden states sampled from INLINEFORM7 yield one mass of vectors with no obvious distinct clusters. Results on both metrics suggest target encoding and orthogonal regularization indeed help diversifying generation of our model.
## Qualitative Analysis
To illustrate the difference of predictions between our proposed models, we show an example chosen from the KP20k validation set in Appendix SECREF10 . In this example there are 29 ground truth phrases. Neither of the models is able to generate all of the keyphrases, but it is obvious that the predictions from INLINEFORM0 all start with “test”, while predictions from INLINEFORM1 are diverse. This to some extent verifies our assumption that without the target encoder and orthogonal regularization, decoder states following delimiters are less diverse.
## Conclusion and Future Work
We propose a recurrent generative model that sequentially generates multiple keyphrases, with two extra modules that enhance generation diversity. We propose new metrics to evaluate keyphrase generation. Our model shows competitive performance on a set of keyphrase generation datasets, including one introduced in this work. In future work, we plan to investigate how target phrase order affects the generation behavior, and further explore set generation in an order invariant fashion.
## Experiment Results on KP20k Absent Subset
Generating absent keyphrases on scientific publication datasets is a rather challenging problem. Existing studies often achieve seemingly good performance by measuring recall on tens and sometimes hundreds of keyphrases produced by exhaustive decoding with a large beam size — thus completely ignoring precision.
We report the models' R@10/50 scores on the absent portion of five scientific paper datasets in Table TABREF48 to be in line with previous studies.
The absent keyphrase prediction highly prefers recall-oriented models, therefore CopyRNN with beam size of 200 is innately proper for this task setting. Howerer, from the results we observe that with the help of exhaustive decoding and diverse mechanisms, INLINEFORM0 is able to perform comparably to CopyRNN model, and it generally works better for top predictions. Even though the trend of models' performance somewhat matches what we observe on the present data, we argue that it is hard to compare different models' performance on such scale. We argue that StackEx is better testbeds for absent keyphrase generation.
## Implementation Details
Implemntation details of our proposed models are as follows. In all experiments, the word embeddings are initialized with 100-dimensional random matrices. The number of hidden units in both the encoder and decoder GRU are 150. The number of hidden units in target encoder GRU is 150. The size of vocabulary is 50,000.
The numbers of hidden units in MLPs described in Section SECREF3 are as follows. During negative sampling, we randomly sample 16 samples from the same batch, thus target encoding loss in Equation EQREF23 is a 17-way classification loss. In INLINEFORM0 , we set both the INLINEFORM1 and INLINEFORM2 in Equation EQREF27 to be 0.3. In all experiments, we use a dropout rate of 0.1.
We use Adam BIBREF36 as the step rule for optimization. The learning rate is INLINEFORM0 . The model is implemented using PyTorch BIBREF38 and OpenNMT BIBREF37 .
For exhaustive decoding, we use a beam size of 50 and a maximum sequence length of 40.
Following BIBREF0 , lowercase and stemming are performed on both the ground truth and generated keyphrases during evaluation.
We leave out 2,000 data examples as validation set for both KP20k and StackEx and use them to identify optimal checkpoints for testing. And all the scores reported in this paper are from checkpoints with best performances (F INLINEFORM0 @ INLINEFORM1 ) on validation set.
## Example Output
See Table TABREF49 .
| 19
|
1810.06743
|
Marrying Universal Dependencies and Universal Morphology
|
# Marrying Universal Dependencies and Universal Morphology
## Abstract
The Universal Dependencies (UD) and Universal Morphology (UniMorph) projects each present schemata for annotating the morphosyntactic details of language. Each project also provides corpora of annotated text in many languages - UD at the token level and UniMorph at the type level. As each corpus is built by different annotators, language-specific decisions hinder the goal of universal schemata. With compatibility of tags, each project's annotations could be used to validate the other's. Additionally, the availability of both type- and token-level resources would be a boon to tasks such as parsing and homograph disambiguation. To ease this interoperability, we present a deterministic mapping from Universal Dependencies v2 features into the UniMorph schema. We validate our approach by lookup in the UniMorph corpora and find a macro-average of 64.13% recall. We also note incompatibilities due to paucity of data on either side. Finally, we present a critical evaluation of the foundations, strengths, and weaknesses of the two annotation projects.
## Introduction
The two largest standardized, cross-lingual datasets for morphological annotation are provided by the Universal Dependencies BIBREF1 and Universal Morphology BIBREF2 , BIBREF3 projects. Each project's data are annotated according to its own cross-lingual schema, prescribing how features like gender or case should be marked. The schemata capture largely similar information, so one may want to leverage both UD's token-level treebanks and UniMorph's type-level lookup tables and unify the two resources. This would permit a leveraging of both the token-level UD treebanks and the type-level UniMorph tables of paradigms. Unfortunately, neither resource perfectly realizes its schema. On a dataset-by-dataset basis, they incorporate annotator errors, omissions, and human decisions when the schemata are underspecified; one such example is in fig:disagreement.
A dataset-by-dataset problem demands a dataset-by-dataset solution; our task is not to translate a schema, but to translate a resource. Starting from the idealized schema, we create a rule-based tool for converting UD-schema annotations to UniMorph annotations, incorporating language-specific post-edits that both correct infelicities and also increase harmony between the datasets themselves (rather than the schemata). We apply this conversion to the 31 languages with both UD and UniMorph data, and we report our method's recall, showing an improvement over the strategy which just maps corresponding schematic features to each other. Further, we show similar downstream performance for each annotation scheme in the task of morphological tagging.
This tool enables a synergistic use of UniMorph and Universal Dependencies, as well as teasing out the annotation discrepancies within and across projects. When one dataset disobeys its schema or disagrees with a related language, the flaws may not be noticed except by such a methodological dive into the resources. When the maintainers of the resources ameliorate these flaws, the resources move closer to the goal of a universal, cross-lingual inventory of features for morphological annotation.
The contributions of this work are:
## Background: Morphological Inflection
Morphological inflection is the act of altering the base form of a word (the lemma, represented in fixed-width type) to encode morphosyntactic features. As an example from English, prove takes on the form proved to indicate that the action occurred in the past. (We will represent all surface forms in quotation marks.) The process occurs in the majority of the world's widely-spoken languages, typically through meaningful affixes. The breadth of forms created by inflection creates a challenge of data sparsity for natural language processing: The likelihood of observing a particular word form diminishes.
A classic result in psycholinguistics BIBREF4 shows that inflectional morphology is a fully productive process. Indeed, it cannot be that humans simply have the equivalent of a lookup table, where they store the inflected forms for retrieval as the syntactic context requires. Instead, there needs to be a mental process that can generate properly inflected words on demand. BIBREF4 showed this insightfully through the wug-test, an experiment where she forced participants to correctly inflect out-of-vocabulary lemmata, such as the novel noun wug.
Certain features of a word do not vary depending on its context: In German or Spanish where nouns are gendered, the word for onion will always be grammatically feminine. Thus, to prepare for later discussion, we divide the morphological features of a word into two categories: the modifiable inflectional features and the fixed lexical features.
A part of speech (POS) is a coarse syntactic category (like verb) that begets a word's particular menu of lexical and inflectional features. In English, verbs express no gender, and adjectives do not reflect person or number. The part of speech dictates a set of inflectional slots to be filled by the surface forms. Completing these slots for a given lemma and part of speech gives a paradigm: a mapping from slots to surface forms. Regular English verbs have five slots in their paradigm BIBREF5 , which we illustrate for the verb prove, using simple labels for the forms in tab:ptb.
A morphosyntactic schema prescribes how language can be annotated—giving stricter categories than our simple labels for prove—and can vary in the level of detail provided. Part of speech tags are an example of a very coarse schema, ignoring details of person, gender, and number. A slightly finer-grained schema for English is the Penn Treebank tagset BIBREF6 , which includes signals for English morphology. For instance, its VBZ tag pertains to the specially inflected 3rd-person singular, present-tense verb form (e.g. proves in tab:ptb).
If the tag in a schema is detailed enough that it exactly specifies a slot in a paradigm, it is called a morphosyntactic description (MSD). These descriptions require varying amounts of detail: While the English verbal paradigm is small enough to fit on a page, the verbal paradigm of the Northeast Caucasian language Archi can have over 1500000 slots BIBREF7 .
## Two Schemata, Two Philosophies
Unlike the Penn Treebank tags, the UD and UniMorph schemata are cross-lingual and include a fuller lexicon of attribute-value pairs, such as Person: 1. Each was built according to a different set of principles. UD's schema is constructed bottom-up, adapting to include new features when they're identified in languages. UniMorph, conversely, is top-down: A cross-lingual survey of the literature of morphological phenomena guided its design. UniMorph aims to be linguistically complete, containing all known morphosyntactic attributes. Both schemata share one long-term goal: a total inventory for annotating the possible morphosyntactic features of a word.
## Universal Dependencies
The Universal Dependencies morphological schema comprises part of speech and 23 additional attributes (also called features in UD) annotating meaning or syntax, as well as language-specific attributes. In order to ensure consistent annotation, attributes are included into the general UD schema if they occur in several corpora. Language-specific attributes are used when only one corpus annotates for a specific feature.
The UD schema seeks to balance language-specific and cross-lingual concerns. It annotates for both inflectional features such as case and lexical features such as gender. Additionally, the UD schema annotates for features which can be interpreted as derivational in some languages. For example, the Czech UD guidance uses a Coll value for the Number feature to denote mass nouns (for example, "lidstvo" "humankind" from the root "lid" "people").
UD represents a confederation of datasets BIBREF8 annotated with dependency relationships (which are not the focus of this work) and morphosyntactic descriptions. Each dataset is an annotated treebank, making it a resource of token-level annotations. The schema is guided by these treebanks, with feature names chosen for relevance to native speakers. (In sec:unimorph, we will contrast this with UniMorph's treatment of morphosyntactic categories.) The UD datasets have been used in the CoNLL shared tasks BIBREF9 .
## UniMorph
In the Universal Morphological Feature Schema BIBREF10 , there are at least 212 values, spread across 23 attributes. It identifies some attributes that UD excludes like information structure and deixis, as well as providing more values for certain attributes, like 23 different noun classes endemic to Bantu languages. As it is a schema for marking morphology, its part of speech attribute does not have POS values for punctuation, symbols, or miscellany (Punct, Sym, and X in Universal Dependencies).
Like the UD schema, the decomposition of a word into its lemma and MSD is directly comparable across languages. Its features are informed by a distinction between universal categories, which are widespread and psychologically real to speakers; and comparative concepts, only used by linguistic typologists to compare languages BIBREF11 . Additionally, it strives for identity of meaning across languages, not simply similarity of terminology. As a prime example, it does not regularly label a dative case for nouns, for reasons explained in depth by BIBREF11 .
The UniMorph resources for a language contain complete paradigms extracted from Wiktionary BIBREF12 , BIBREF13 . Word types are annotated to form a database, mapping a lemma–tag pair to a surface form. The schema is explained in detail in BIBREF10 . It has been used in the SIGMORPHON shared task BIBREF14 and the CoNLL–SIGMORPHON shared tasks BIBREF15 , BIBREF16 . Several components of the UniMorph schema have been adopted by UD.
## Similarities in the annotation
While the two schemata annotate different features, their annotations often look largely similar. Consider the attested annotation of the Spanish word mandaba (I/he/she/it) commanded. tab:annotations shows that these annotations share many attributes.
Some conversions are straightforward: VERB to V, Mood=Ind to IND, Number=Sing to SG, and Person=3 to 3. One might also suggest mapping Tense=Imp to IPFV, though this crosses semantic categories: IPFV represents the imperfective aspect, whereas Tense=Imp comes from imperfect, the English name often given to Spanish's pasado continuo form. The imperfect is a verb form which combines both past tense and imperfective aspect. UniMorph chooses to split this into the atoms PST and IPFV, while UD unifies them according to the familiar name of the tense.
## UD treebanks and UniMorph tables
Prima facie, the alignment task may seem trivial. But we've yet to explore the humans in the loop. This conversion is a hard problem because we're operating on idealized schemata. We're actually annotating human decisions—and human mistakes. If both schemata were perfectly applied, their overlapping attributes could be mapped to each other simply, in a cross-lingual and totally general way. Unfortunately, the resources are imperfect realizations of their schemata. The cross-lingual, cross-resource, and within-resource problems that we'll note mean that we need a tailor-made solution for each language.
Showcasing their schemata, the Universal Dependencies and UniMorph projects each present large, annotated datasets. UD's v2.1 release BIBREF1 has 102 treebanks in 60 languages. The large resource, constructed by independent parties, evinces problems in the goal of a universal inventory of annotations. Annotators may choose to omit certain values (like the coerced gender of refrescante in fig:disagreement), and they may disagree on how a linguistic concept is encoded. (See, e.g., BIBREF11 's ( BIBREF11 ) description of the dative case.) Additionally, many of the treebanks were created by fully- or semi-automatic conversion from treebanks with less comprehensive annotation schemata than UD BIBREF0 . For instance, the Spanish word vas you go is incorrectly labeled Gender: Fem|Number: Pl because it ends in a character sequence which is common among feminine plural nouns. (Nevertheless, the part of speech field for vas is correct.)
UniMorph's development is more centralized and pipelined. Inflectional paradigms are scraped from Wiktionary, annotators map positions in the scraped data to MSDs, and the mapping is automatically applied to all of the scraped paradigms. Because annotators handle languages they are familiar with (or related ones), realization of the schema is also done on a language-by-language basis. Further, the scraping process does not capture lexical aspects that are not inflected, like noun gender in many languages. The schema permits inclusion of these details; their absence is an artifact of the data collection process. Finally, UniMorph records only exist for nouns, verbs, and adjectives, though the schema is broader than these categories.
For these reasons, we treat the corpora as imperfect realizations of the schemata. Moreover, we contend that ambiguity in the schemata leave the door open to allow for such imperfections. With no strict guidance, it's natural that annotators would take different paths. Nevertheless, modulo annotator disagreement, we assume that within a particular corpus, one word form will always be consistently annotated.
Three categories of annotation difficulty are missing values, language-specific attributes, and multiword expressions.
## A Deterministic Conversion
In our work, the goal is not simply to translate one schema into the other, but to translate one resource (the imperfect manifestation of the schema) to match the other. The differences between the schemata and discrepancies in annotation mean that the transformation of annotations from one schema to the other is not straightforward.
Two naive options for the conversion are a lookup table of MSDs and a lookup table of the individual attribute-value pairs which comprise the MSDs. The former is untenable: the table of all UD feature combinations (including null features, excluding language-specific attributes) would have 2.445e17 entries. Of course, most combinations won't exist, but this gives a sense of the table's scale. Also, it doesn't leverage the factorial nature of the annotations: constructing the table would require a massive duplication of effort. On the other hand, attribute-value lookup lacks the flexibility to show how a pair of values interacts. Neither approach would handle language- and annotator-specific tendencies in the corpora.
Our approach to converting UD MSDs to UniMorph MSDs begins with the attribute-value lookup, then amends it on a language-specific basis. Alterations informed by the MSD and the word form, like insertion, substitution, and deletion, increase the number of agreeing annotations. They are critical for work that examines the MSD monolithically instead of feature-by-feature BIBREF25 , BIBREF26 : Without exact matches, converting the individual tags becomes hollow.
Beginning our process, we relied on documentation of the two schemata to create our initial, language-agnostic mapping of individual values. This mapping has 140 pairs in it. Because the mapping was derived purely from the schemata, it is a useful approximation of how well the schemata match up. We note, however, that the mapping does not handle idiosyncrasies like the many uses of dative or features which are represented in UniMorph by argument templates: possession and ergative–absolutive argument marking. The initial step of our conversion is using this mapping to populate a proposed UniMorph MSD.
As shown in sec:results, the initial proposal is often frustratingly deficient. Thus we introduce the post-edits. To concoct these, we looked into UniMorph corpora for these languages, compared these to the conversion outputs, and then sought to bring the conversion outputs closer to the annotations in the actual UniMorph corpora. When a form and its lemma existed in both corpora, we could directly inspect how the annotations differed. Our process of iteratively refining the conversion implies a table which exactly maps any combination of UD MSD and its related values (lemma, form, etc.) to a UniMorph MSD, though we do not store the table explicitly.
Some conversion rules we've created must be applied before or after others. These sequential dependencies provide conciseness. Our post-editing procedure operates on the initial MSD hypothesis as follows:
## Experiments
We evaluate our tool on two tasks:
To be clear, our scope is limited to the schema conversion. Future work will explore NLP tasks that exploit both the created token-level UniMorph data and the existing type-level UniMorph data.
## Intrinsic evaluation
We transform all UD data to the UniMorph. We compare the simple lookup-based transformation to the one with linguistically informed post-edits on all languages with both UD and UniMorph data. We then evaluate the recall of MSDs without partial credit.
Because the UniMorph tables only possess annotations for verbs, nouns, adjectives, or some combination, we can only examine performance for these parts of speech. We consider two words to be a match if their form and lemma are present in both resources. Syncretism allows a single surface form to realize multiple MSDs (Spanish mandaba can be first- or third-person), so we define success as the computed MSD matching any of the word's UniMorph MSDs. This gives rise to an equation for recall: of the word–lemma pairs found in both resources, how many of their UniMorph-converted MSDs are present in the UniMorph tables?
Our problem here is not a learning problem, so the question is ill-posed. There is no training set, and the two resources for a given language make up a test set. The quality of our model—the conversion tool—comes from how well we encode prior knowledge about the relationship between the UD and UniMorph corpora.
## Extrinsic evaluation
If the UniMorph-converted treebanks perform differently on downstream tasks, then they convey different information. This signals a failure of the conversion process. As a downstream task, we choose morphological tagging, a critical step to leveraging morphological information on new text.
We evaluate taggers trained on the transformed UD data, choosing eight languages randomly from the intersection of UD and UniMorph resources. We report the macro-averaged F1 score of attribute-value pairs on a held-out test set, with official train/validation/test splits provided in the UD treebanks. As a reference point, we also report tagging accuracy on those languages' untransformed data.
We use the state-of-the-art morphological tagger of BIBREF0 . It is a factored conditional random field with potentials for each attribute, attribute pair, and attribute transition. The potentials are computed by neural networks, predicting the values of each attribute jointly but not monolithically. Inference with the potentials is performed approximately by loopy belief propagation. We use the authors' hyperparameters.
We note a minor implementation detail for the sake of reproducibility. The tagger exploits explicit guidance about the attribute each value pertains to. The UniMorph schema's values are globally unique, but their attributes are not explicit. For example, the UniMorph Masc denotes a masculine gender. We amend the code of BIBREF0 to incorporate attribute identifiers for each UniMorph value.
## Results
We present the intrinsic task's recall scores in tab:recall. Bear in mind that due to annotation errors in the original corpora (like the vas example from sec:resources), the optimal score is not always $100\%$ . Some shortcomings of recall come from irremediable annotation discrepancies. Largely, we are hamstrung by differences in choice of attributes to annotate. When one resource marks gender and the other marks case, we can't infer the gender of the word purely from its surface form. The resources themselves would need updating to encode the relevant morphosyntactic information. Some languages had a very low number of overlapping forms, and no tag matches or near-matches between them: Arabic, Hindi, Lithuanian, Persian, and Russian. A full list of observed, irremediable discrepancies is presented alongside the codebase.
There are three other transformations for which we note no improvement here. Because of the problem in Basque argument encoding in the UniMorph dataset—which only contains verbs—we note no improvement in recall on Basque. Irish also does not improve: UD marks gender on nouns, while UniMorph marks case. Adjectives in UD are also underspecified. The verbs, though, are already correct with the simple mapping. Finally, with Dutch, the UD annotations are impoverished compared to the UniMorph annotations, and missing attributes cannot be inferred without external knowledge.
For the extrinsic task, the performance is reasonably similar whether UniMorph or UD; see tab:tagging. A large fluctuation would suggest that the two annotations encode distinct information. On the contrary, the similarities suggest that the UniMorph-mapped MSDs have similar content. We recognize that in every case, tagging F1 increased—albeit by amounts as small as $0.16$ points. This is in part due to the information that is lost in the conversion. UniMorph's schema does not indicate the type of pronoun (demonstrative, interrogative, etc.), and when lexical information is not recorded in UniMorph, we delete it from the MSD during transformation. On the other hand, UniMorph's atomic tags have more parts to guess, but they are often related. (E.g. Ipfv always entails Pst in Spanish.) Altogether, these forces seem to have little impact on tagging performance.
## Related Work
The goal of a tagset-to-tagset mapping of morphological annotations is shared by the Interset project BIBREF28 . Interset decodes features in the source corpus to a tag interlingua, then encodes that into target corpus features. (The idea of an interlingua is drawn from machine translation, where a prevailing early mindset was to convert to a universal representation, then encode that representation's semantics in the target language. Our approach, by contrast, is a direct flight from the source to the target.) Because UniMorph corpora are noisy, the encoding from the interlingua would have to be rewritten for each target. Further, decoding the UD MSD into the interlingua cannot leverage external information like the lemma and form.
The creators of HamleDT sought to harmonize dependency annotations among treebanks, similar to our goal of harmonizing across resources BIBREF29 . The treebanks they sought to harmonize used multiple diverse annotation schemes, which the authors unified under a single scheme.
BIBREF30 present mappings into a coarse, universal part of speech for 22 languages. Working with POS tags rather than morphological tags (which have far more dimensions), their space of options to harmonize is much smaller than ours.
Our extrinsic evaluation is most in line with the paradigm of BIBREF31 (and similar work therein), who compare syntactic parser performance on UD treebanks annotated with two styles of dependency representation. Our problem differs, though, in that the dependency representations express different relationships, while our two schemata vastly overlap. As our conversion is lossy, we do not appraise the learnability of representations as they did.
In addition to using the number of extra rules as a proxy for harmony between resources, one could perform cross-lingual projection of morphological tags BIBREF32 , BIBREF33 . Our approach succeeds even without parallel corpora.
## Conclusion and Future Work
We created a tool for annotating Universal Dependencies CoNLL-U files with UniMorph annotations. Our tool is ready to use off-the-shelf today, requires no training, and is deterministic. While under-specification necessitates a lossy and imperfect conversion, ours is interpretable. Patterns of mistakes can be identified and ameliorated.
The tool allows a bridge between resources annotated in the Universal Dependencies and Universal Morphology (UniMorph) schemata. As the Universal Dependencies project provides a set of treebanks with token-level annotation, while the UniMorph project releases type-level annotated tables, the newfound compatibility opens up new experiments. A prime example of exploiting token- and type-level data is BIBREF34 . That work presents a part-of-speech (POS) dictionary built from Wiktionary, where the POS tagger is also constrained to options available in their type-level POS dictionary, improving performance. Our transformation means that datasets are prepared for similar experiments with morphological tagging. It would also be reasonable to incorporate this tool as a subroutine to UDPipe BIBREF35 and Udapi BIBREF36 . We leave open the task of converting in the opposite direction, turning UniMorph MSDs into Universal Dependencies MSDs.
Because our conversion rules are interpretable, we identify shortcomings in both resources, using each as validation for the other. We were able to find specific instances of incorrectly applied UniMorph annotation, as well as specific instances of cross-lingual inconsistency in both resources. These findings will harden both resources and better align them with their goal of universal, cross-lingual annotation.
## Acknowledgments
We thank Hajime Senuma and John Sylak-Glassman for early comments in devising the starting language-independent mapping from Universal Dependencies to UniMorph.
| 15
|
1810.09774
|
Testing the Generalization Power of Neural Network Models Across NLI Benchmarks
|
# Testing the Generalization Power of Neural Network Models Across NLI Benchmarks
## Abstract
Neural network models have been very successful in natural language inference, with the best models reaching 90% accuracy in some benchmarks. However, the success of these models turns out to be largely benchmark specific. We show that models trained on a natural language inference dataset drawn from one benchmark fail to perform well in others, even if the notion of inference assumed in these benchmarks is the same or similar. We train six high performing neural network models on different datasets and show that each one of these has problems of generalizing when we replace the original test set with a test set taken from another corpus designed for the same task. In light of these results, we argue that most of the current neural network models are not able to generalize well in the task of natural language inference. We find that using large pre-trained language models helps with transfer learning when the datasets are similar enough. Our results also highlight that the current NLI datasets do not cover the different nuances of inference extensively enough.
## Introduction
Natural Language Inference (NLI) has attracted considerable interest in the NLP community and, recently, a large number of neural network-based systems have been proposed to deal with the task. One can attempt a rough categorization of these systems into: a) sentence encoding systems, and b) other neural network systems. Both of them have been very successful, with the state of the art on the SNLI and MultiNLI datasets being 90.4%, which is our baseline with BERT BIBREF0 , and 86.7% BIBREF0 respectively. However, a big question with respect to these systems is their ability to generalize outside the specific datasets they are trained and tested on. Recently, BIBREF1 have shown that state-of-the-art NLI systems break considerably easily when, instead of tested on the original SNLI test set, they are tested on a test set which is constructed by taking premises from the training set and creating several hypotheses from them by changing at most one word within the premise. The results show a very significant drop in accuracy for three of the four systems. The system that was more difficult to break and had the least loss in accuracy was the system by BIBREF2 which utilizes external knowledge taken from WordNet BIBREF3 .
In this paper we show that NLI systems that have been very successful in specific NLI benchmarks, fail to generalize when trained on a specific NLI dataset and then these trained models are tested across test sets taken from different NLI benchmarks. The results we get are in line with BIBREF1 , showing that the generalization capability of the individual NLI systems is very limited, but, what is more, they further show the only system that was less prone to breaking in BIBREF1 , breaks too in the experiments we have conducted.
We train six different state-of-the-art models on three different NLI datasets and test these trained models on an NLI test set taken from another dataset designed for the same NLI task, namely for the task to identify for sentence pairs in the dataset if one sentence entails the other one, if they are in contradiction with each other or if they are neutral with respect to inferential relationship.
One would expect that if a model learns to correctly identify inferential relationships in one dataset, then it would also be able to do so in another dataset designed for the same task. Furthermore, two of the datasets, SNLI BIBREF4 and MultiNLI BIBREF5 , have been constructed using the same crowdsourcing approach and annotation instructions BIBREF5 , leading to datasets with the same or at least very similar definition of entailment. It is therefore reasonable to expect that transfer learning between these datasets is possible. As SICK BIBREF6 dataset has been machine-constructed, a bigger difference in performance is expected.
In this paper we show that, contrary to our expectations, most models fail to generalize across the different datasets. However, our experiments also show that BERT BIBREF0 performs much better than the other models in experiments between SNLI and MultiNLI. Nevertheless, even BERT fails when testing on SICK. In addition to the negative results, our experiments further highlight the power of pre-trained language models, like BERT, in NLI.
The negative results of this paper are significant for the NLP research community as well as to NLP practice as we would like our best models to not only to be able to perform well in a specific benchmark dataset, but rather capture the more general phenomenon this dataset is designed for. The main contribution of this paper is that it shows that most of the best performing neural network models for NLI fail in this regard. The second, and equally important, contribution is that our results highlight that the current NLI datasets do not capture the nuances of NLI extensively enough.
## Related Work
The ability of NLI systems to generalize and related skepticism has been raised in a number of recent papers. BIBREF1 show that the generalization capabilities of state-of-the-art NLI systems, in cases where some kind of external lexical knowledge is needed, drops dramatically when the SNLI test set is replaced by a test set where the premise and the hypothesis are otherwise identical except for at most one word. The results show a very significant drop in accuracy. BIBREF7 recognize the generalization problem that comes with training on datasets like SNLI, which tend to be homogeneous and with little linguistic variation. In this context, they propose to better train NLI models by making use of adversarial examples.
Multiple papers have reported hidden bias and annotation artifacts in the popular NLI datasets SNLI and MultiNLI allowing classification based on the hypothesis sentences alone BIBREF8 , BIBREF9 , BIBREF10 .
BIBREF11 evaluate the robustness of NLI models using datasets where label preserving swapping operations have been applied, reporting significant performance drops compared to the results with the original dataset. In these experiments, like in the BreakingNLI experiment, the systems that seem to be performing the better, i.e. less prone to breaking, are the ones where some kind of external knowledge is used by the model (KIM by BIBREF2 is one of those systems).
On a theoretical and methodological level, there is discussion on the nature of various NLI datasets, as well as the definition of what counts as NLI and what does not. For example, BIBREF12 , BIBREF13 present an overview of the most standard datasets for NLI and show that the definitions of inference in each of them are actually quite different, capturing only fragments of what seems to be a more general phenomenon.
BIBREF4 show that a simple LSTM model trained on the SNLI data fails when tested on SICK. However, their experiment is limited to this single architecture and dataset pair. BIBREF5 show that different models that perform well on SNLI have lower accuracy on MultiNLI. However in their experiments they did not systematically test transfer learning between the two datasets, but instead used separate systems where the training and test data were drawn from the same corpora.
## Experimental Setup
In this section we describe the datasets and model architectures included in the experiments.
## Data
We chose three different datasets for the experiments: SNLI, MultiNLI and SICK. All of them have been designed for NLI involving three-way classification with the labels entailment, neutral and contradiction. We did not include any datasets with two-way classification, e.g. SciTail BIBREF14 . As SICK is a relatively small dataset with approximately only 10k sentence pairs, we did not use it as training data in any experiment. We also trained the models with a combined SNLI + MultiNLI training set.
For all the datasets we report the baseline performance where the training and test data are drawn from the same corpus. We then take these trained models and test them on a test set taken from another NLI corpus. For the case where the models are trained with SNLI + MultiNLI we report the baseline using the SNLI test data. All the experimental combinations are listed in Table 1 . Examples from the selected datasets are provided in Table 2 . To be more precise, we vary three things: training dataset, model and testing dataset. We should qualify this though, since the three datasets we look at, can also be grouped by text domain/genre and type of data collection, with MultiNLI and SNLI using the same data collection style, and SNLI and SICK using roughly the same domain/genre. Hopefully, our set up will let us determine which of these factors matters the most.
We describe the source datasets in more detail below.
The Stanford Natural Language Inference (SNLI) corpus BIBREF4 is a dataset of 570k human-written sentence pairs manually labeled with the labels entailment, contradiction, and neutral. The source for the premise sentences in SNLI were image captions taken from the Flickr30k corpus BIBREF15 .
The Multi-Genre Natural Language Inference (MultiNLI) corpus BIBREF5 consisting of 433k human-written sentence pairs labeled with entailment, contradiction and neutral. MultiNLI contains sentence pairs from ten distinct genres of both written and spoken English. Only five genres are included in the training set. The development and test sets have been divided into matched and mismatched, where the former includes only sentences from the same genres as the training data, and the latter includes sentences from the remaining genres not present in the training data.
We used the matched development set (MultiNLI-m) for the experiments. The MultiNLI dataset was annotated using very similar instructions as for the SNLI dataset. Therefore we can assume that the definitions of entailment, contradiction and neutral is the same in these two datasets.
SICK BIBREF6 is a dataset that was originally constructed to test compositional distributional semantics (DS) models. The dataset contains 9,840 examples pertaining to logical inference (negation, conjunction, disjunction, apposition, relative clauses, etc.). The dataset was automatically constructed taking pairs of sentences from a random subset of the 8K ImageFlickr data set BIBREF15 and the SemEval 2012 STS MSRVideo Description dataset BIBREF16 .
## Model and Training Details
We perform experiments with six high-performing models covering the sentence encoding models, cross-sentence attention models as well as fine-tuned pre-trained language models.
For sentence encoding models, we chose a simple one-layer bidirectional LSTM with max pooling (BiLSTM-max) with the hidden size of 600D per direction, used e.g. in InferSent BIBREF17 , and HBMP BIBREF18 . For the other models, we have chosen ESIM BIBREF19 , which includes cross-sentence attention, and KIM BIBREF2 , which has cross-sentence attention and utilizes external knowledge. We also selected two model involving a pre-trained language model, namely ESIM + ELMo BIBREF20 and BERT BIBREF0 . KIM is particularly interesting in this context as it performed significantly better than other models in the Breaking NLI experiment conducted by BIBREF1 . The success of pre-trained language models in multiple NLP tasks make ESIM + ELMo and BERT interesting additions to this experiment. Table 3 lists the different models used in the experiments.
For BiLSTM-max we used the Adam optimizer BIBREF21 , a learning rate of 5e-4 and batch size of 64. The learning rate was decreased by the factor of 0.2 after each epoch if the model did not improve. Dropout of 0.1 was used between the layers of the multi-layer perceptron classifier, except before the last layer.The BiLSTM-max models were initialized with pre-trained GloVe 840B word embeddings of size 300 dimensions BIBREF22 , which were fine-tuned during training. Our BiLSMT-max model was implemented in PyTorch.
For HBMP, ESIM, KIM and BERT we used the original implementations with the default settings and hyperparameter values as described in BIBREF18 , BIBREF19 , BIBREF2 and BIBREF0 respectively. For BERT we used the uncased 768-dimensional model (BERT-base). For ESIM + ELMo we used the AllenNLP BIBREF23 PyTorch implementation with the default settings and hyperparameter values.
## Experimental Results
Table 4 contains all the experimental results.
Our experiments show that, while all of the six models perform well when the test set is drawn from the same corpus as the training and development set, accuracy is significantly lower when we test these trained models on a test set drawn from a separate NLI corpus, the average difference in accuracy being 24.9 points across all experiments.
Accuracy drops the most when a model is tested on SICK. The difference in this case is between 19.0-29.0 points when trained on MultiNLI, between 31.6-33.7 points when trained on SNLI and between 31.1-33.0 when trained on SNLI + MultiNLI. This was expected, as the method of constructing the sentence pairs was different, and hence there is too much difference in the kind of sentence pairs included in the training and test sets for transfer learning to work. However, the drop was more dramatic than expected.
The most surprising result was that the accuracy of all models drops significantly even when the models were trained on MultiNLI and tested on SNLI (3.6-11.1 points). This is surprising as both of these datasets have been constructed with a similar data collection method using the same definition of entailment, contradiction and neutral. The sentences included in SNLI are also much simpler compared to those in MultiNLI, as they are taken from the Flickr image captions. This might also explain why the difference in accuracy for all of the six models is lowest when the models are trained on MultiNLI and tested on SNLI. It is also very surprising that the model with the biggest difference in accuracy was ESIM + ELMo which includes a pre-trained ELMo language model. BERT performed significantly better than the other models in this experiment having an accuracy of 80.4% and only 3.6 point difference in accuracy.
The poor performance of most of the models with the MultiNLI-SNLI dataset pair is also very surprising given that neural network models do not seem to suffer a lot from introduction of new genres to the test set which were not included in the training set, as can be seen from the small difference in test accuracies for the matched and mismatched test sets (see e.g BIBREF5 ). In a sense SNLI could be seen as a separate genre not included in MultiNLI. This raises the question if the SNLI and MultiNLI have e.g. different kinds of annotation artifacts, which makes transfer learning between these datasets more difficult.
All the models, except BERT, perform almost equally poorly across all the experiments. Both BiLSTM-max and HBMP have an average drop in accuracy of 24.4 points, while the average for KIM is 25.5 and for ESIM + ELMo 25.6. ESIM has the highest average difference of 27.0 points. In contrast to the findings of BIBREF1 , utilizing external knowledge did not improve the model's generalization capability, as KIM performed equally poorly across all dataset combinations.
Also including a pretrained ELMo language model did not improve the results significantly. The overall performance of BERT was significantly better than the other models, having the lowest average difference in accuracy of 22.5 points. Our baselines for SNLI (90.4%) and SNLI + MultiNLI (90.6%) outperform the previous state-of-the-art accuracy for SNLI (90.1%) by BIBREF24 .
To understand better the types of errors made by neural network models in NLI we looked at some example failure-pairs for selected models. Tables 5 and 6 contain some randomly selected failure-pairs for two models: BERT and HBMP, and for three set-ups: SNLI $\rightarrow $ SICK, SNLI $\rightarrow $ MultiNLI and MultiNLI $\rightarrow $ SICK. We chose BERT as the current the state of the art NLI model. HBMP was selected as a high performing model in the sentence encoding model type. Although the listed sentence pairs represent just a small sample of the errors made by these models, they do include some interesting examples. First, it seems that SICK has a more narrow notion of contradiction – corresponding more to logical contradiction – compared to the contradiction in SNLI and MultiNLI, where especially in SNLI the sentences are contradictory if they describe a different state of affairs. This is evident in the sentence pair: A young child is running outside over the fallen leaves and A young child is lying down on a gravel road that is covered with dead leaves, which is predicted by BERT to be contradiction although the gold label is neutral. Another interesting example is the sentence pair: A boat pear with people boarding and disembarking some boats. and people are boarding and disembarking some boats, which is incorrectly predicted by BERT to be contradiction although it has been labeled as entailment. Here the two sentences describe the same event from different points of view: the first one describing a boat pear with some people on it and the second one describing the people directly. Interestingly the added information about the boat pear seems to confuse the model.
## Discussion and Conclusion
In this paper we have shown that neural network models for NLI fail to generalize across different NLI benchmarks. We experimented with six state-of-the-art models covering sentence encoding approaches, cross-sentence attention models and pre-trained and fine-tuned language models. For all the systems, the accuracy drops between 3.6-33.7 points (the average drop being 24.9 points), when testing with a test set drawn from a separate corpus from that of the training data, as compared to when the test and training data are splits from the same corpus. Our findings, together with the previous negative findings, indicate that the state-of-the-art models fail to capture the semantics of NLI in a way that will enable them to generalize across different NLI situations.
The results highlight two issues to be taken into consideration: a) using datasets involving a fraction of what NLI is, will fail when tested in datasets that are testing for a slightly different definition of inference. This is evident when we move from the SNLI to the SICK dataset. b) NLI is to some extent genre/context dependent. Training on SNLI and testing on MultiNLI gives worse results than vice versa. This is particularly evident in the case of BERT. These results highlight that training on multiple genres helps. However, this help is still not enough given that, even in the case of training on MultiNLI (multi genre) and training on SNLI (single genre and same definition of inference with MultiNLI), accuracy drops significantly.
We also found that involving a large pre-trained language model helps with transfer learning when the datasets are similar enough, as is the case with SNLI and MultiNLI. Our results further corroborate the power of pre-trained and fine-tuned language models like BERT in NLI. However, not even BERT is able to generalize from SNLI and MultiNLI to SICK, possibly due to the difference between what kind of inference relations are contained in these datasets.
Our findings motivate us to look for novel neural network architectures and approaches that better capture the semantics on natural language inference beyond individual datasets. However, there seems to be a need to start with better constructed datasets, i.e. datasets that will not only capture fractions of what NLI is in reality. Better NLI systems need to be able to be more versatile on the types of inference they can recognize. Otherwise, we would be stuck with systems that can cover only some aspects of NLI. On a theoretical level, and in connection to the previous point, we need a better understanding of the range of phenomena NLI must be able to cover and focus our future endeavours for dataset construction towards this direction. In order to do this a more systematic study is needed on the different kinds of entailment relations NLI datasets need to include. Our future work will include a more systematic and broad-coverage analysis of the types of errors the models make and in what kinds of sentence-pairs they make successful predictions.
## Acknowledgments
The first author is supported by the FoTran project, funded by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 771113).
The first author also gratefully acknowledges the support of the Academy of Finland through project 314062 from the ICT 2023 call on Computation, Machine Learning and Artificial Intelligence.
The second author is supported by grant 2014-39 from the Swedish Research Council, which funds the Centre for Linguistic Theory and Studies in Probability (CLASP) in the Department of Philosophy, Linguistics, and Theory of Science at the University of Gothenburg.
| 8
|
1810.12196
|
ReviewQA: a relational aspect-based opinion reading dataset
|
# ReviewQA: a relational aspect-based opinion reading dataset
## Abstract
Deep reading models for question-answering have demonstrated promising performance over the last couple of years. However current systems tend to learn how to cleverly extract a span of the source document, based on its similarity with the question, instead of seeking for the appropriate answer. Indeed, a reading machine should be able to detect relevant passages in a document regarding a question, but more importantly, it should be able to reason over the important pieces of the document in order to produce an answer when it is required. To motivate this purpose, we present ReviewQA, a question-answering dataset based on hotel reviews. The questions of this dataset are linked to a set of relational understanding competencies that we expect a model to master. Indeed, each question comes with an associated type that characterizes the required competency. With this framework, it is possible to benchmark the main families of models and to get an overview of what are the strengths and the weaknesses of a given model on the set of tasks evaluated in this dataset. Our corpus contains more than 500.000 questions in natural language over 100.000 hotel reviews. Our setup is projective, the answer of a question does not need to be extracted from a document, like in most of the recent datasets, but selected among a set of candidates that contains all the possible answers to the questions of the dataset. Finally, we present several baselines over this dataset.
## Introduction
A large majority of the human knowledge is recorded through text documents. That is why ability for a system to automatically infer information from text without any structured data has become a major challenge. Answering questions about a given document is a relevant proxy task that has been proposed as a way to evaluate the reading ability of a given model. In this configuration, a text document such as a news article, a document from Wikipedia or any type of text is presented to a machine with an associated set of questions. The system is then expected to answer these questions and evaluated by its accuracy on this task. The machine reading framework is very general and we can imagine a large panel of questions that can possibly handle most of the standard natural language processing tasks. For example, the task of named entities recognition can be formulated as a machine reading one where your document is the sentence and the question would be 'What are the named entities mentioned in this sentence?'. These natural language interactions are an important objective for reading systems.
Recently, many datasets have been proposed to build and evaluate reading models BIBREF0 , BIBREF1 . From cloze style questions BIBREF2 to open questions BIBREF3 , from synthetic data BIBREF4 to human written articles BIBREF5 , many styles of documents and questions have been proposed to challenge reading models. The correct answer to the questions proposed in most of these datasets is a span of text of the source document, which can be restricted to a single word in several cases. It means that the answer should explicitly be present in the source document and that the model should be able to locate it.
Different models have already shown superhuman performance on several of these datasets and particularly on the SQuAD dataset composed of Wikipedia articles BIBREF6 , BIBREF7 . However, some limits of such models have been highlighted when they encounter perturbations into the input documents BIBREF8 . Indeed almost all of the state of the art models on the SQuAD dataset suffer from a lack of robustness against adversarial examples. Once the model is trained, a meaningless sentence added at the end of the text document can completely disturb the reading system. Conversely, these adversarial examples do not seem to fool a human reader who will be capable of answering the questions as well as without this perturbation. One possible explanation of this phenomenon is that computers are good at extracting patterns in the document that match the representation of the question. If multiple spans of the documents look similar to the questions, the reader might not be able to decide which one is relevant. Moreover, Wikipedia articles tend to be written with the same standard writing style, factual, unambiguous. Such writing style tends to favor the pattern matching between the questions and the documents. This format of documents/questions has certainly influenced the design of the comprehension models that have been proposed so far. Most of them are composed of stacked attention layers that match question and document representations.
Following concepts proposed in the 20 bAbI tasks BIBREF4 or in the visual question-answering dataset CLEVR BIBREF9 , we think that the challenge, limited to the detection of relevant passages in a document, is only the first step in building systems that truly understand text. The second step is the ability of reasoning with the relevant information extracted from a document. To set up this challenge, we propose to leverage on a hotel reviews corpus that requires reasoning skills to answer natural language questions. The reviews we used have been extracted from TripAdvisor and originally proposed in BIBREF10 , BIBREF11 . In the original data, each review comes with a set of rated aspects among the seventh available: Business service, Check in / Front Desk, Cleanliness, Location, Room, Sleep Quality, Value and for all the reviews an Overall rating. In this articles we propose to exploit these data to create a dataset of question-answering that will challenge 8 competencies of the reader.
Our contributions can be summarized as follow:
## Machine comprehension datasets
ReviewQA is proposed as a novel dataset regarding the collection of the existing ones. Indeed a large panel of available datasets, that evaluate models on different types of documents, can only be valuable for designing efficient models and learning protocols. In this following part, we describe several of these datasets.
SQuAD: The Standford Question Answering Dataset (SQuAD) introduced in BIBREF0 is a large dataset of natural questions over the 500 most popular articles of Wikipedia. All the questions have been crowdsourced and answers are spans of text extracted from source documents. This dataset has been very popular these last two years and the performance of the architectures that have been proposed have rapidly increased until several models surpass the human score. Indeed, in the original paper human performance has been measured at 82.304 points for the exact match metric and at the time we are writing this paper four models have already a higher score. In another hand BIBREF8 has shown that these models suffer from a lack of robustness against adversarial examples that are meaningless from a human point of view. This suggests the need for a more challenging dataset that will allow developing strongest reasoning architectures.
NewsQA: NewsQA BIBREF1 is a dataset very similar to SQuAD. It contains 120.000 human generated questions over 12.000 articles form CNN originally introduced in BIBREF5 . It has been designed to be more challenging than SQuAD with questions that might require to extract multiple spans of text or not be answerable.
WikiHop and MedHop: These are two recent datasets introduced in BIBREF13 . Unlike SQuAD and NewsQA, important facts are spread out across multiple documents and, in order to answer a question, it is necessary to jump over a set of passages to collect the required information. The relevant passages are not explicitly mentioned in the data so this dataset measures the ability that a model has to navigate across multiple documents. The questions come with a set of candidates which are all present in the text.
MS Marco: This dataset has been released in BIBREF14 . The documents come from the internet and the questions are real user queries asked through the bing search engine. The dataset contains around 100.000 queries and each of them comes with a set of approximatively 10 relevant passages. Like in SQuAD, several models are already doing superhuman performances on this dataset.
Facebook bAbI tasks: This is a set of 20 toy tasks proposed in BIBREF4 and designed to measure text understanding. Each task requires a certain capability to be completed like induction, deduction and more. Documents are synthetic stories, composed of few sentences that describe a set of actions. This dataset was one of the first attempt to introduce a general set of prerequisite capabilities required for the reading task. Although it has been a very challenging framework, beneficial to the emergence of the attention mechanism inside the reading architectures, a Gated end-to-end memory network BIBREF15 now succeed in almost all of the 20 tasks. One of the possible reason is that the data are synthetic data, without noise or ambiguity. We propose a comparable framework with understanding and reasoning tasks based on user-generated comments that are much more realistic and that required language competencies to be understood.
CLEVR: Beyond textual question-answering, Visual Question-Answering (VQA) has been largely studied during the last couple of years. More recently, the problem of relational reasoning has been introduced through this dataset BIBREF9 . The main original idea was to introduce relational reasoning questions over object shapes and placements. This dataset has already motivated the development of original deep models. To the best of our knowledge, no natural language question-answering corpus has been designed to investigate such capabilities. As we will present in the following of this paper, we think sentiment analysis is particularly suited for this task and we will introduce a novel machine reading corpus with such capability requirements.
## Attention-based models for aspect-based sentiment analysis
Sentiment analysis is one of the historical tasks of Natural Language Processing. It is an important challenge for companies, restaurants, hotels that aim to analyze customer satisfaction regarding products and quality of services. Given a text document, the objective is to predict its overall polarity. Generally, it can be positive, negative or neutral. This analysis gives a quick overview of a general sentiment over a set of documents, but this framework tends to be restrictive. Indeed, one document tends to express multiple opinions of different aspects. For instance, in the sentence: The fish was very good but the service was terrible, there is not a general dominant sentiment, and a finer analysis is needed. The task of aspect-based sentiment analysis aims to predict a polarity of a sentence regarding a given aspect. In the previous example a positive polarity should be associated to the aspect food, and on the contrary, a negative sentiment is expressed regarding the quality of the service.
The idea of using models originally designed for question-answering, for the sentiment analysis task has been introduced in BIBREF16 , BIBREF17 . In these papers, several adaptations of the end-to-end memory network (MemN2N) BIBREF18 are used to predict the polarity of a review regarding a given aspect. In that configuration, the review is encoded into the memory cells and the controller, usually initialized with a representation of the question, is initialized with a representation of the aspect. The analysis of the attention between the values of the controller and the document has shown interesting results, by highlighting relevant part of a document regarding an aspect.
## ReviewQA dataset
We think that evaluating the task of sentiment analysis through the setup of question-answering is a relevant playground for machine reading research. Indeed natural language questions about the different aspects of the targeted venues are typical kind of questions we want to be able to ask to a system. In this context, we introduce a set of reasoning questions types over the relationships between aspects. We propose ReviewQA, a dataset of natural language questions over hotel reviews. These questions are divided into 8 groups, regarding the competency required to be answered. In this section, we describe each task and the process followed to generate this dataset.
## Original data
We used a set of reviews extracted from the TripAdvisor website and originally proposed in BIBREF10 and BIBREF11 . This corpus is available at http://www.cs.virginia.edu/~hw5x/Data/LARA/TripAdvisor/TripAdvisorJson.tar.bz2. Each review comes with the name of the associated hotel, a title, an overall rating, a comment and a list of rated aspects. From 0 to 7 aspects, among value, room, location, cleanliness, check-in/front desk, service, business service, can possibly be rated in a review. Figure FIGREF8 displays a review extracted from this dataset.
## Relational reasoning competencies
Objective: Starting with the original corpus, we aim at building a machine reading task where natural language questions will challenge the model on its understanding of the reviews. Indeed learning relational reasoning competencies over natural language documents is a major challenge of the current reading models. These original raw data allow us to generate relational questions that can possibly require a global understanding of the comment and reasoning skills to be treated. For example, asking a question like What is the best aspect rated in this comment ? is not an easy question that can be answered without a deep understanding of the review. It is necessary to capture all the aspects mentioned in the text, to predict their rating and finally to select the best one. The tasks and the dataset we propose are publicly available at http://www.europe.naverlabs.com/Blog/ReviewQA-A-novel-relational-aspect-based-opinion-dataset-for-machine-reading
We introduce a list of 8 different competencies that a reading system should master in order to process reviews and text documents in general. These 8 tasks require different competencies and a different level of understanding of the document to be well answered. For instance, detecting if an aspect is mentioned in a review will require less understanding of the review than predicting explicitly the rating of this aspect. Table TABREF10 presents the 8 tasks we have introduced in this dataset with an example of a question that corresponds to each task. We also provide the expected type of the answer (Yes/No question, rating question...). It can be an additional tool to analyze the errors of the readers.
## Construction of the dataset
We sample 100.000 reviews from the original corpus. Figure FIGREF12 presents the distribution of the number of words of the reviews in the dataset. We explicitly favor reviews which contain an important number of words. In average, a review contains 200 words. Indeed these long reviews are most likely to contain challenging relations between different aspects. A short review which deals with only a few aspects is more likely to not be very relevant to the challenge we want to propose in this dataset. Figure FIGREF14 displays the distribution of the ratings per aspects in the 100.000 reviews we based our dataset. We can see that the average values of these ratings tend to be quite high. It could have introduced bias if it was not the case for all the aspects. For example, we do not want that the model learns that in general, the service is rated better than the location and them answer without looking at the document. Since this situation is the same for all the aspects, the relational tasks introduced in this dataset remains extremely relevant.
Then we randomly select 6 tasks for each review (the same task can be selected multiple times) and randomly select a natural language question that corresponds to this task. The questions are human-generated patterns that we have crowdsourced in order to produce a dataset as rich as possible. To this end, we have generated several patterns that correspond to the capabilities we wanted to express in a given question and we have crowdsourced rephrasing of these patterns.
The final dataset we propose is composed of more than 500.000 questions about 100.000 reviews. Table TABREF13 shows the repartition of the documents and queries into the train and test set. Each review contains a maximum of 6 questions. Sometimes less when it is not possible to generate all. For example, if only two or three aspects are mentioned in a review, we will be able to generate only a little set of relational questions. Figure FIGREF15 depicts the repartition of the answers in the generated dataset. A majority of the tasks we introduced, even if they possibly require a high level of understanding of the document and the question, are binary questions. It means that in the generated dataset the answers yes and no tend to be more present than the others. To balance in a better way the distribution of the answers, we chose to affect a higher probability of sampling to the task 5, 6, 7.1, 8. Indeed, these tasks are not binary questions and required an aspect name as the answer. Figure FIGREF17 represents the repartition of question types in our dataset. Finally, figure FIGREF15 shows the repartition of the answers in the dataset.
## Paraphrase augmentation using backtranslation
In order to generate more paraphrases of the questions, we used a backtranslation method to enrich them. The idea is to use a translation model that will translate our human-generated questions into another language, and then translate them back to English. This double translation will introduce rewordings of the questions that we will be able to integrate into this dataset. This approach has been used in BIBREF7 to perform data augmentation on the training set. For this purpose, we have trained a fairseq BIBREF19 model to translate sentences from English to French and for French to English. In order to preserve the quality of the sentences we have so far, we only keep the most probable translation of each original sentence. Indeed a beam search is used during the translation to predict the most probable translations which mean that we each translation comes with an associated probability. By selecting only the first translations, we almost double the number of questions without degrading the quality of the questions proposed in the dataset.
## Models
In this section, we present the performance of four different models on our dataset: a logistic regression and three neural models. The first one is a basic LSTM BIBREF20 , the second a MemN2N BIBREF18 and the third one is a model of our own design. This fourth model reuses the encoding layers of the R-net BIBREF12 and we modify the final layers with a projection layer that will be able to select the answer among the set of candidates instead of pointing the answerer directly into the source document.
Logistic regression: To produce the representation of the input, we concatenate the Bag-Of-Words representation of the document with the Bag-Of-Words representation of the question. It produces an array of size INLINEFORM0 where INLINEFORM1 is the vocabulary size. Then we use a logistic regression to select the most probable answer among the INLINEFORM2 possibilities.
LSTM: We start with a concatenation of the sequence of indexes of the document with the sequence of indexes of the question. Them we feed an LSTM network with this vector and use the final state as the representation of the input. Finally, we apply a logistic regression over this representation to produce the final decision.
End-to-end memory networks: This architecture is based on two different memory cells (input and output) that contain a representation of the document. A controller, initialized with the encoding of the question, is used to calculate an attention between this controller and the representation of the document in the input memory. This attention is them used to re-weight the representation of the document in the output memory. This response from the output memory is them utilized to update the controller. After that, either a matrix is used to project this representation into the answer space either the controller is used to go through an over hop of memory. This architecture allows the model to sequentially look into the initial document seeking for important information regarding the current state of its controller. This model achieves very good performances on the 20 bAbI tasks dataset.
Deep projective reader: This is a model of our own design, largely inspired by the efficient R-net reader BIBREF12 . The overall architecture is composed of 4 stacked layers: an encoding layer, a question/document attention, a self-attention layer and a projection layer. The following paragraphs briefly describe the overall utility of each of these layers.
Encoding: The sentence is tokenized by words. Each token is represented by the concatenation of its embedding vector and the final state of a bidirectional recurrent network over the characters of this word. Finally, another bidirectional RNN on the top of this representation produce the encoding of the document and the question.
Question/document attention: We apply a question/document attention layer that matches the representation of the question with each token of the document individually to output an attention that gives more weight to the important tokens of the document regarding the question.
Self-attention layer: The previous layer has built a question-aware representation of the document. One problem with such representation is that form the moment each token has only a good knowledge of its closest neighbors. To tackle this problem, BIBREF12 have proposed to use a self-attention layer that matches each individual token with all the other tokens of the document. Doing that, each token is now aware of a larger context.
Output layer: A bidirectional RNN is applied on the top of the last layer and we use its final state as the representation of the input. We use a projection matrix to project this representation into the answer space and select the most probable one
## Training details
We propose to train these models on the entire set of tasks and them to measure the overall performance and the accuracy of each individual task. In all the models, we use the Adam optimizer BIBREF21 with a learning rate of 0.01 and the batch size is set to 64. All the parameter are initialized from a Gaussian distribution with mean 0 and a standard deviation of 0.01. The dimension of the word embeddings in the projective deep reading model and the LSTM model is 300 and we use Glove pre-trained vectors ( BIBREF22 ). We use a MemN2N with 5 memory hops and a linear start of 5 epochs. The reviews are split by sentence and each memory block corresponds to one sentence. Each sentence is represented by its bag-of-word representation augmented with temporal encoding as it is suggested in BIBREF18 .
## Model performance
Table TABREF19 displays the performance of the 4 baselines on the ReviewQA's test set. These results are the performance achieved by our own implementation of these 4 models. According to our results, the simple LSTM network and the MemN2N perform very poorly on this dataset. Especially on the most advanced reasoning tasks. Indeed, the task 5 which corresponds to the prediction of the exact rating of an aspect seems to be very challenging for these model. Maybe the tokenization by sentence to create the memory blocks of the MemN2N, which is appropriated in the case of the bAbI tasks, is not a good representation of the documents when it has to handle human generated comments. However, the logistic regression achieves reasonable performance on these tasks, and do not suffer from catastrophic performance on any tasks. Its worst result comes on task 6 and one of the reason is probably that this architecture is not designed to predict a list of answers. On the contrary, the deep projective reader achieves encouraging on this dataset. It outperforms all the other baselines, with very good scores on the first fourth tasks. The question/document and document/document attention layers proposed in BIBREF12 seem once again to produce rich encodings of the inputs which are relevant for our projection layer.
## Conclusion
In this paper, we formalize the sentiment analysis task through the framework of machine reading and release ReviewQA, a relational question-answering corpus. This dataset allows evaluating a set of relational reasoning skills through natural language questions. It is composed of a large panel of human-generated questions. Moreover, we propose to augment the dataset with backtranslated reformulations of these questions. Finally, we evaluate 4 models on this dataset, including a projective model of our own design that seems to be a strong baseline for this dataset. We expect that this large dataset will encourage the research community to develop reasoning models and evaluate their models on this set of tasks.
## Acknowledgment
We thank Vassilina Nikoulina and Stéphane Clinchant for the help regarding the backtranslation rewording of the questions.
| 13
|
1811.01734
|
Transductive Learning with String Kernels for Cross-Domain Text Classification
|
# Transductive Learning with String Kernels for Cross-Domain Text Classification
## Abstract
For many text classification tasks, there is a major problem posed by the lack of labeled data in a target domain. Although classifiers for a target domain can be trained on labeled text data from a related source domain, the accuracy of such classifiers is usually lower in the cross-domain setting. Recently, string kernels have obtained state-of-the-art results in various text classification tasks such as native language identification or automatic essay scoring. Moreover, classifiers based on string kernels have been found to be robust to the distribution gap between different domains. In this paper, we formally describe an algorithm composed of two simple yet effective transductive learning approaches to further improve the results of string kernels in cross-domain settings. By adapting string kernels to the test set without using the ground-truth test labels, we report significantly better accuracy rates in cross-domain English polarity classification.
## Introduction
Domain shift is a fundamental problem in machine learning, that has attracted a lot of attention in the natural language processing and vision communities BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 . To understand and address this problem, generated by the lack of labeled data in a target domain, researchers have studied the behavior of machine learning methods in cross-domain settings BIBREF2 , BIBREF11 , BIBREF10 and came up with various domain adaptation techniques BIBREF12 , BIBREF5 , BIBREF6 , BIBREF9 . In cross-domain classification, a classifier is trained on data from a source domain and tested on data from a (different) target domain. The accuracy of machine learning methods is usually lower in the cross-domain setting, due to the distribution gap between different domains. However, researchers proposed several domain adaptation techniques by using the unlabeled test data to obtain better performance BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 , BIBREF7 . Interestingly, some recent works BIBREF10 , BIBREF17 indicate that string kernels can yield robust results in the cross-domain setting without any domain adaptation. In fact, methods based on string kernels have demonstrated impressive results in various text classification tasks ranging from native language identification BIBREF18 , BIBREF19 , BIBREF20 , BIBREF21 and authorship identification BIBREF22 to dialect identification BIBREF23 , BIBREF17 , BIBREF24 , sentiment analysis BIBREF10 , BIBREF25 and automatic essay scoring BIBREF26 . As long as a labeled training set is available, string kernels can reach state-of-the-art results in various languages including English BIBREF19 , BIBREF10 , BIBREF26 , Arabic BIBREF27 , BIBREF20 , BIBREF17 , BIBREF24 , Chinese BIBREF25 and Norwegian BIBREF20 . Different from all these recent approaches, we use unlabeled data from the test set in a transductive setting in order to significantly increase the performance of string kernels. In our recent work BIBREF28 , we proposed two transductive learning approaches combined into a unified framework that improves the results of string kernels in two different tasks. In this paper, we provide a formal and detailed description of our transductive algorithm and present results in cross-domain English polarity classification.
The paper is organized as follows. Related work on cross-domain text classification and string kernels is presented in Section SECREF2 . Section SECREF3 presents our approach to obtain domain adapted string kernels. The transductive transfer learning method is described in Section SECREF4 . The polarity classification experiments are presented in Section SECREF5 . Finally, we draw conclusions and discuss future work in Section SECREF6 .
## Related Work
## Cross-Domain Classification
Transfer learning (or domain adaptation) aims at building effective classifiers for a target domain when the only available labeled training data belongs to a different (source) domain. Domain adaptation techniques can be roughly divided into graph-based methods BIBREF1 , BIBREF29 , BIBREF9 , BIBREF30 , probabilistic models BIBREF3 , BIBREF4 , knowledge-based models BIBREF14 , BIBREF31 , BIBREF11 and joint optimization frameworks BIBREF12 . The transfer learning methods from the literature show promising results in a variety of real-world applications, such as image classification BIBREF12 , text classification BIBREF13 , BIBREF16 , BIBREF3 , polarity classification BIBREF1 , BIBREF29 , BIBREF4 , BIBREF6 , BIBREF30 and others BIBREF32 .
General transfer learning approaches. Long et al. BIBREF12 proposed a novel transfer learning framework to model distribution adaptation and label propagation in a unified way, based on the structural risk minimization principle and the regularization theory. Shu et al. BIBREF5 proposed a method that bridges the distribution gap between the source domain and the target domain through affinity learning, by exploiting the existence of a subset of data points in the target domain that are distributed similarly to the data points in the source domain. In BIBREF7 , deep learning is employed to jointly optimize the representation, the cross-domain transformation and the target label inference in an end-to-end fashion. More recently, Sun et al. BIBREF8 proposed an unsupervised domain adaptation method that minimizes the domain shift by aligning the second-order statistics of source and target distributions, without requiring any target labels. Chang et al. BIBREF9 proposed a framework based on using a parallel corpus to calibrate domain-specific kernels into a unified kernel for leveraging graph-based label propagation between domains.
Cross-domain text classification. Joachims BIBREF13 introduced the Transductive Support Vector Machines (TSVM) framework for text classification, which takes into account a particular test set and tries to minimize the error rate for those particular test samples. Ifrim et al. BIBREF14 presented a transductive learning approach for text classification based on combining latent variable models for decomposing the topic-word space into topic-concept and concept-word spaces, and explicit knowledge models with named concepts for populating latent variables. Guo et al. BIBREF16 proposed a transductive subspace representation learning method to address domain adaptation for cross-lingual text classification. Zhuang et al. BIBREF3 presented a probabilistic model, by which both the shared and distinct concepts in different domains can be learned by the Expectation-Maximization process which optimizes the data likelihood. In BIBREF33 , an algorithm to adapt a classification model by iteratively learning domain-specific features from the unlabeled test data is described.
Cross-domain polarity classification. In recent years, cross-domain sentiment (polarity) classification has gained popularity due to the advances in domain adaptation on one side, and to the abundance of documents from various domains available on the Web, expressing positive or negative opinion, on the other side. Some of the general domain adaptation frameworks have been applied to polarity classification BIBREF3 , BIBREF33 , BIBREF9 , but there are some approaches that have been specifically designed for the cross-domain sentiment classification task BIBREF0 , BIBREF34 , BIBREF1 , BIBREF29 , BIBREF11 , BIBREF4 , BIBREF6 , BIBREF10 , BIBREF30 . To the best of our knowledge, Blitzer et al. BIBREF0 were the first to report results on cross-domain classification proposing the structural correspondence learning (SCL) method, and its variant based on mutual information (SCL-MI). Pan et al. BIBREF1 proposed a spectral feature alignment (SFA) algorithm to align domain-specific words from different domains into unified clusters, using domain-independent words as a bridge. Bollegala et al. BIBREF31 used a cross-domain lexicon creation to generate a sentiment-sensitive thesaurus (SST) that groups different words expressing the same sentiment, using unigram and bigram features as BIBREF0 , BIBREF1 . Luo et al. BIBREF4 proposed a cross-domain sentiment classification framework based on a probabilistic model of the author's emotion state when writing. An Expectation-Maximization algorithm is then employed to solve the maximum likelihood problem and to obtain a latent emotion distribution of the author. Franco-Salvador et al. BIBREF11 combined various recent and knowledge-based approaches using a meta-learning scheme (KE-Meta). They performed cross-domain polarity classification without employing any domain adaptation technique. More recently, Fernández et al. BIBREF6 introduced the Distributional Correspondence Indexing (DCI) method for domain adaptation in sentiment classification. The approach builds term representations in a vector space common to both domains where each dimension reflects its distributional correspondence to a highly predictive term that behaves similarly across domains. A graph-based approach for sentiment classification that models the relatedness of different domains based on shared users and keywords is proposed in BIBREF30 .
## String Kernels
In recent years, methods based on string kernels have demonstrated remarkable performance in various text classification tasks BIBREF35 , BIBREF36 , BIBREF22 , BIBREF19 , BIBREF10 , BIBREF17 , BIBREF26 . String kernels represent a way of using information at the character level by measuring the similarity of strings through character n-grams. Lodhi et al. BIBREF35 used string kernels for document categorization, obtaining very good results. String kernels were also successfully used in authorship identification BIBREF22 . More recently, various combinations of string kernels reached state-of-the-art accuracy rates in native language identification BIBREF19 and Arabic dialect identification BIBREF17 . Interestingly, string kernels have been used in cross-domain settings without any domain adaptation, obtaining impressive results. For instance, Ionescu et al. BIBREF19 have employed string kernels in a cross-corpus (and implicitly cross-domain) native language identification experiment, improving the state-of-the-art accuracy by a remarkable INLINEFORM0 . Giménez-Pérez et al. BIBREF10 have used string kernels for single-source and multi-source polarity classification. Remarkably, they obtain state-of-the-art performance without using knowledge from the target domain, which indicates that string kernels provide robust results in the cross-domain setting without any domain adaptation. Ionescu et al. BIBREF17 obtained the best performance in the Arabic Dialect Identification Shared Task of the 2017 VarDial Evaluation Campaign BIBREF37 , with an improvement of INLINEFORM1 over the second-best method. It is important to note that the training and the test speech samples prepared for the shared task were recorded in different setups BIBREF37 , or in other words, the training and the test sets are drawn from different distributions. Different from all these recent approaches BIBREF19 , BIBREF10 , BIBREF17 , we use unlabeled data from the target domain to significantly increase the performance of string kernels in cross-domain text classification, particularly in English polarity classification.
## Transductive String Kernels
String kernels. Kernel functions BIBREF38 capture the intuitive notion of similarity between objects in a specific domain. For example, in text mining, string kernels can be used to measure the pairwise similarity between text samples, simply based on character n-grams. Various string kernel functions have been proposed to date BIBREF35 , BIBREF38 , BIBREF19 . Perhaps one of the most recently introduced string kernels is the histogram intersection string kernel BIBREF19 . For two strings over an alphabet INLINEFORM0 , INLINEFORM1 , the intersection string kernel is formally defined as follows: DISPLAYFORM0
where INLINEFORM0 is the number of occurrences of n-gram INLINEFORM1 as a substring in INLINEFORM2 , and INLINEFORM3 is the length of INLINEFORM4 . The spectrum string kernel or the presence bits string kernel can be defined in a similar fashion BIBREF19 .
Transductive string kernels. We present a simple and straightforward approach to produce a transductive similarity measure suitable for strings. We take the following steps to derive transductive string kernels. For a given kernel (similarity) function INLINEFORM0 , we first build the full kernel matrix INLINEFORM1 , by including the pairwise similarities of samples from both the train and the test sets. For a training set INLINEFORM2 of INLINEFORM3 samples and a test set INLINEFORM4 of INLINEFORM5 samples, such that INLINEFORM6 , each component in the full kernel matrix is defined as follows: DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are samples from the set INLINEFORM2 , for all INLINEFORM3 . We then normalize the kernel matrix by dividing each component by the square root of the product of the two corresponding diagonal components: DISPLAYFORM0
We transform the normalized kernel matrix into a radial basis function (RBF) kernel matrix as follows: DISPLAYFORM0
Each row in the RBF kernel matrix INLINEFORM0 is now interpreted as a feature vector. In other words, each sample INLINEFORM1 is represented by a feature vector that contains the similarity between the respective sample INLINEFORM2 and all the samples in INLINEFORM3 . Since INLINEFORM4 includes the test samples as well, the feature vector is inherently adapted to the test set. Indeed, it is easy to see that the features will be different if we choose to apply the string kernel approach on a set of test samples INLINEFORM5 , such that INLINEFORM6 . It is important to note that through the features, the subsequent classifier will have some information about the test samples at training time. More specifically, the feature vector conveys information about how similar is every test sample to every training sample. We next consider the linear kernel, which is given by the scalar product between the new feature vectors. To obtain the final linear kernel matrix, we simply need to compute the product between the RBF kernel matrix and its transpose: DISPLAYFORM0
In this way, the samples from the test set, which are included in INLINEFORM0 , are used to obtain new (transductive) string kernels that are adapted to the test set at hand.
[!tpb] Transductive Kernel Algorithm
Input:
INLINEFORM0 – the training set of INLINEFORM1 training samples and associated class labels;
INLINEFORM0 – the set of INLINEFORM1 test samples;
INLINEFORM0 – a kernel function;
INLINEFORM0 – the number of test samples to be added in the second round of training;
INLINEFORM0 – a binary kernel classifier.
Domain-Adapted Kernel Matrix Computation Steps:
INLINEFORM0 INLINEFORM1 ; INLINEFORM2 ; INLINEFORM3 ; INLINEFORM4
INLINEFORM0 INLINEFORM1 INLINEFORM2
INLINEFORM0 INLINEFORM1 INLINEFORM2
INLINEFORM0
INLINEFORM0
Transductive Kernel Classifier Steps:
INLINEFORM0
INLINEFORM0
INLINEFORM0
INLINEFORM0 INLINEFORM1
INLINEFORM0
INLINEFORM0
INLINEFORM0 INLINEFORM1 the dual weights of INLINEFORM2 trained on INLINEFORM3 with the labels INLINEFORM4
INLINEFORM0
INLINEFORM0 ; INLINEFORM1
INLINEFORM0 INLINEFORM1
INLINEFORM0
INLINEFORM0 INLINEFORM1 sort INLINEFORM2 in descending order and return the sorted indexes
INLINEFORM0
INLINEFORM0
INLINEFORM0
INLINEFORM0
INLINEFORM0
Output:
INLINEFORM0 – the set of predicted labels for the test samples in INLINEFORM1 .
## Transductive Kernel Classifier
We next present a simple yet effective approach for adapting a one-versus-all kernel classifier trained on a source domain to a different target domain. Our transductive kernel classifier (TKC) approach is composed of two learning iterations. Our entire framework is formally described in Algorithm SECREF3 .
Notations. We use the following notations in the algorithm. Sets, arrays and matrices are written in capital letters. All collection types are considered to be indexed starting from position 1. The elements of a set INLINEFORM0 are denoted by INLINEFORM1 , the elements of an array INLINEFORM2 are alternatively denoted by INLINEFORM3 or INLINEFORM4 , and the elements of a matrix INLINEFORM5 are denoted by INLINEFORM6 or INLINEFORM7 when convenient. The sequence INLINEFORM8 is denoted by INLINEFORM9 . We use sequences to index arrays or matrices as well. For example, for an array INLINEFORM10 and two integers INLINEFORM11 and INLINEFORM12 , INLINEFORM13 denotes the sub-array INLINEFORM14 . In a similar manner, INLINEFORM15 denotes a sub-matrix of the matrix INLINEFORM16 , while INLINEFORM17 returns the INLINEFORM18 -th row of M and INLINEFORM19 returns the INLINEFORM20 -th column of M. The zero matrix of INLINEFORM21 components is denoted by INLINEFORM22 , and the square zero matrix is denoted by INLINEFORM23 . The identity matrix is denoted by INLINEFORM24 .
Algorithm description. In steps 8-17, we compute the domain-adapted string kernel matrix, as described in the previous section. In the first learning iteration (when INLINEFORM0 ), we train several classifiers to distinguish each individual class from the rest, according to the one-versus-all (OVA) scheme. In step 27, the kernel classifier INLINEFORM1 is trained to distinguish a class from the others, assigning a dual weight to each training sample from the source domain. The returned column vector of dual weights is denoted by INLINEFORM2 and the bias value is denoted by INLINEFORM3 . The vector of weights INLINEFORM4 contains INLINEFORM5 values, such that the weight INLINEFORM6 corresponds to the training sample INLINEFORM7 . When the test kernel matrix INLINEFORM8 of INLINEFORM9 components is multiplied with the vector INLINEFORM10 in step 28, the result is a column vector of INLINEFORM11 positive or negative scores. Afterwards (step 34), the test samples are sorted in order to maximize the probability of correctly predicted labels. For each test sample INLINEFORM12 , we consider the score INLINEFORM13 (step 32) produced by the classifier for the chosen class INLINEFORM14 (step 31), which is selected according to the OVA scheme. The sorting is based on the hypothesis that if the classifier associates a higher score to a test sample, it means that the classifier is more confident about the predicted label for the respective test sample. Before the second learning iteration, a number of INLINEFORM15 test samples from the top of the sorted list are added to the training set (steps 35-39) for another round of training. As the classifier is more confident about the predicted labels INLINEFORM16 of the added test samples, the chance of including noisy examples (with wrong labels) is minimized. On the other hand, the classifier has the opportunity to learn some useful domain-specific patterns of the test domain. We believe that, at least in the cross-domain setting, the added test samples bring more useful information than noise. We would like to stress out that the ground-truth test labels are never used in our transductive algorithm. Although the test samples are required beforehand, their labels are not necessary. Hence, our approach is suitable in situations where unlabeled data from the target domain can be collected cheaply, and such situations appear very often in practice, considering the great amount of data available on the Web.
## Polarity Classification
Data set. For the cross-domain polarity classification experiments, we use the second version of Multi-Domain Sentiment Dataset BIBREF0 . The data set contains Amazon product reviews of four different domains: Books (B), DVDs (D), Electronics (E) and Kitchen appliances (K). Reviews contain star ratings (from 1 to 5) which are converted into binary labels as follows: reviews rated with more than 3 stars are labeled as positive, and those with less than 3 stars as negative. In each domain, there are 1000 positive and 1000 negative reviews.
Baselines. We compare our approach with several methods BIBREF1 , BIBREF31 , BIBREF11 , BIBREF8 , BIBREF10 , BIBREF39 in two cross-domain settings. Using string kernels, Giménez-Pérez et al. BIBREF10 reported better performance than SST BIBREF31 and KE-Meta BIBREF11 in the multi-source domain setting. In addition, we compare our approach with SFA BIBREF1 , CORAL BIBREF8 and TR-TrAdaBoost BIBREF39 in the single-source setting.
Evaluation procedure and parameters. We follow the same evaluation methodology of Giménez-Pérez et al. BIBREF10 , to ensure a fair comparison. Furthermore, we use the same kernels, namely the presence bits string kernel ( INLINEFORM0 ) and the intersection string kernel ( INLINEFORM1 ), and the same range of character n-grams (5-8). To compute the string kernels, we used the open-source code provided by Ionescu et al. BIBREF19 , BIBREF40 . For the transductive kernel classifier, we select INLINEFORM2 unlabeled test samples to be included in the training set for the second round of training. We choose Kernel Ridge Regression BIBREF38 as classifier and set its regularization parameter to INLINEFORM3 in all our experiments. Although Giménez-Pérez et al. BIBREF10 used a different classifier, namely Kernel Discriminant Analysis, we observed that Kernel Ridge Regression produces similar results ( INLINEFORM4 ) when we employ the same string kernels. As Giménez-Pérez et al. BIBREF10 , we evaluate our approach in two cross-domain settings. In the multi-source setting, we train the models on all domains, except the one used for testing. In the single-source setting, we train the models on one of the four domains and we independently test the models on the remaining three domains.
Results in multi-source setting. The results for the multi-source cross-domain polarity classification setting are presented in Table TABREF8 . Both the transductive presence bits string kernel ( INLINEFORM0 ) and the transductive intersection kernel ( INLINEFORM1 ) obtain better results than their original counterparts. Moreover, according to the McNemar's test BIBREF41 , the results on the DVDs, the Electronics and the Kitchen target domains are significantly better than the best baseline string kernel, with a confidence level of INLINEFORM2 . When we employ the transductive kernel classifier (TKC), we obtain even better results. On all domains, the accuracy rates yielded by the transductive classifier are more than INLINEFORM3 better than the best baseline. For example, on the Books domain the accuracy of the transductive classifier based on the presence bits kernel ( INLINEFORM4 ) is INLINEFORM5 above the best baseline ( INLINEFORM6 ) represented by the intersection string kernel. Remarkably, the improvements brought by our transductive string kernel approach are statistically significant in all domains.
Results in single-source setting. The results for the single-source cross-domain polarity classification setting are presented in Table TABREF9 . We considered all possible combinations of source and target domains in this experiment, and we improve the results in each and every case. Without exception, the accuracy rates reached by the transductive string kernels are significantly better than the best baseline string kernel BIBREF10 , according to the McNemar's test performed at a confidence level of INLINEFORM0 . The highest improvements (above INLINEFORM1 ) are obtained when the source domain contains Books reviews and the target domain contains Kitchen reviews. As in the multi-source setting, we obtain much better results when the transductive classifier is employed for the learning task. In all cases, the accuracy rates of the transductive classifier are more than INLINEFORM2 better than the best baseline string kernel. Remarkably, in four cases (E INLINEFORM3 B, E INLINEFORM4 D, B INLINEFORM5 K and D INLINEFORM6 K) our improvements are greater than INLINEFORM7 . The improvements brought by our transductive classifier based on string kernels are statistically significant in each and every case. In comparison with SFA BIBREF1 , we obtain better results in all but one case (K INLINEFORM8 D). Remarkably, we surpass the other state-of-the-art approaches BIBREF8 , BIBREF39 in all cases.
## Conclusion
In this paper, we presented two domain adaptation approaches that can be used together to improve the results of string kernels in cross-domain settings. We provided empirical evidence indicating that our framework can be successfully applied in cross-domain text classification, particularly in cross-domain English polarity classification. Indeed, the polarity classification experiments demonstrate that our framework achieves better accuracy rates than other state-of-the-art methods BIBREF1 , BIBREF31 , BIBREF11 , BIBREF8 , BIBREF10 , BIBREF39 . By using the same parameters across all the experiments, we showed that our transductive transfer learning framework can bring significant improvements without having to fine-tune the parameters for each individual setting. Although the framework described in this paper can be generally applied to any kernel method, we focused our work only on string kernel approaches used in text classification. In future work, we aim to combine the proposed transductive transfer learning framework with different kinds of kernels and classifiers, and employ it for other cross-domain tasks.
| 8
|
1811.08048
|
QuaRel: A Dataset and Models for Answering Questions about Qualitative Relationships
|
# QuaRel: A Dataset and Models for Answering Questions about Qualitative Relationships
## Abstract
Many natural language questions require recognizing and reasoning with qualitative relationships (e.g., in science, economics, and medicine), but are challenging to answer with corpus-based methods. Qualitative modeling provides tools that support such reasoning, but the semantic parsing task of mapping questions into those models has formidable challenges. We present QuaRel, a dataset of diverse story questions involving qualitative relationships that characterize these challenges, and techniques that begin to address them. The dataset has 2771 questions relating 19 different types of quantities. For example,"Jenny observes that the robot vacuum cleaner moves slower on the living room carpet than on the bedroom carpet. Which carpet has more friction?"We contribute (1) a simple and flexible conceptual framework for representing these kinds of questions; (2) the QuaRel dataset, including logical forms, exemplifying the parsing challenges; and (3) two novel models for this task, built as extensions of type-constrained semantic parsing. The first of these models (called QuaSP+) significantly outperforms off-the-shelf tools on QuaRel. The second (QuaSP+Zero) demonstrates zero-shot capability, i.e., the ability to handle new qualitative relationships without requiring additional training data, something not possible with previous models. This work thus makes inroads into answering complex, qualitative questions that require reasoning, and scaling to new relationships at low cost. The dataset and models are available at http://data.allenai.org/quarel.
## Introduction
Many natural language tasks require recognizing and reasoning with qualitative relationships. For example, we may read about temperatures rising (climate science), a drug dose being increased (medicine), or the supply of goods being reduced (economics), and want to reason about the effects. Qualitative story problems, of the kind found in elementary exams (e.g., Figure FIGREF1 ), form a natural example of many of these linguistic and reasoning challenges, and is the target of this work.
Understanding and answering such questions is particularly challenging. Corpus-based methods perform poorly in this setting, as the questions ask about novel scenarios rather than facts that can be looked up. Similarly, word association methods struggle, as a single word change (e.g., “more” to “less”) can flip the answer. Rather, the task appears to require knowledge of the underlying qualitative relations (e.g., “more friction implies less speed”).
Qualitative modeling BIBREF0 , BIBREF1 , BIBREF2 provides a means for encoding and reasoning about such relationships. Relationships are expressed in a natural, qualitative way (e.g., if X increases, then so will Y), rather than requiring numeric equations, and inference allows complex questions to be answered. However, the semantic parsing task of mapping real world questions into these models is formidable and presents unique challenges. These challenges must be solved if natural questions involving qualitative relationships are to be reliably answered.
We make three contributions: (1) a simple and flexible conceptual framework for formally representing these kinds of questions, in particular ones that express qualitative comparisons between two scenarios; (2) a challenging new dataset (QuaRel), including logical forms, exemplifying the parsing challenges; and (3) two novel models that extend type-constrained semantic parsing to address these challenges.
Our first model, QuaSP+, addresses the problem of tracking different “worlds” in questions, resulting in significantly higher scores than with off-the-shelf tools (Section SECREF36 ). The second model, QuaSP+Zero, demonstrates zero-shot capability, i.e., the ability to handle new qualitative relationships on unseen properties, without requiring additional training data, something not possible with previous models (Section SECREF44 ). Together these contributions make inroads into answering complex, qualitative questions by linking language and reasoning, and offer a new dataset and models to spur further progress by the community.
## Related Work
There has been rapid progress in question-answering (QA), spanning a wide variety of tasks and phenomena, including factoid QA BIBREF3 , entailment BIBREF4 , sentiment BIBREF5 , and ellipsis and coreference BIBREF6 . Our contribution here is the first dataset specifically targeted at qualitative relationships, an important category of language that has been less explored. While questions requiring reasoning about qualitative relations sometimes appear in other datasets, e.g., BIBREF7 , our dataset specifically focuses on them so their challenges can be studied.
For answering such questions, we treat the problem as mapping language to a structured formalism (semantic parsing) where simple qualitative reasoning can occur. Semantic parsing has a long history BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , using datasets about geography BIBREF8 , travel booking BIBREF12 , factoid QA over knowledge bases BIBREF10 , Wikipedia tables BIBREF13 , and many more. Our contributions to this line of research are: a dataset that features phenomena under-represented in prior datasets, namely (1) highly diverse language describing open-domain qualitative problems, and (2) the need to reason over entities that have no explicit formal representation; and methods for adapting existing semantic parsers to address these phenomena.
For the target formalism itself, we draw on the extensive body of work on qualitative reasoning BIBREF0 , BIBREF1 , BIBREF2 to create a logical form language that can express the required qualitative knowledge, yet is sufficiently constrained that parsing into it is feasible, described in more detail in Section SECREF3 .
There has been some work connecting language with qualitative reasoning, although mainly focused on extracting qualitative models themselves from text rather than question interpretation, e.g., BIBREF14 , BIBREF15 . Recent work by BIBREF16 crouse2018learning also includes interpreting questions that require identifying qualitative processes in text, in constrast to our setting of interpreting NL story questions that involve qualitative comparisons.
Answering story problems has received attention in the domain of arithmetic, where simple algebra story questions (e.g., “Sue had 5 cookies, then gave 2 to Joe...”) are mapped to a system of equations, e.g., BIBREF17 , BIBREF18 , BIBREF19 , BIBREF20 . This task is loosely analogous to ours (we instead map to qualitative relations) except that in arithmetic the entities to relate are often identifiable (namely, the numbers). Our qualitative story questions lack this structure, adding an extra challenge.
The QuaRel dataset shares some structure with the Winograd Schema Challenge BIBREF21 , being 2-way multiple choice questions invoking both commonsense and coreference. However, they test different aspects of commonsense: Winograd uses coreference resolution to test commonsense understanding of scenarios, while QuaRel tests reasoning about qualitative relationships requiring tracking of coreferent “worlds.”
Finally, crowdsourcing datasets has become a driving force in AI, producing significant progress, e.g., BIBREF3 , BIBREF22 , BIBREF23 . However, for semantic parsing tasks, one obstacle has been the difficulty in crowdsourcing target logical forms for questions. Here, we show how those logical forms can be obtained indirectly from workers without training the workers in the formalism, loosely similar to BIBREF24 .
## Knowledge Representation
We first describe our framework for representing questions and the knowledge to answer them. Our dataset, described later, includes logical forms expressed in this language.
## Qualitative Background Knowledge
We use a simple representation of qualitative relationships, leveraging prior work in qualitative reasoning BIBREF0 . Let INLINEFORM0 be the set of properties relevant to the question set's domain (e.g., smoothness, friction, speed). Let INLINEFORM1 be a set of qualitative values for property INLINEFORM2 (e.g., fast, slow). For the background knowledge about the domain itself (a qualitative model), following BIBREF0 Forbus1984QualitativePT, we use the following predicates: [vskip=1mm,leftmargin=5mm] q+(property1, property2)
q-(property1, property2) q+ denotes that property1 and property2 are qualitatively proportional, e.g., if property1 goes up, property2 will too, while q- denotes inverse proportionality, e.g., [vskip=1mm,leftmargin=5mm] # If friction goes up, speed goes down.
q-(friction, speed). We also introduce the predicate: [vskip=1mm,leftmargin=5mm] higher-than( INLINEFORM0 , INLINEFORM1 , property INLINEFORM2 ) where INLINEFORM3 , allowing an ordering of property values to be specified, e.g., higher-than(fast, slow, speed). For our purposes here, we simplify to use just two property values, low and high, for all properties. (The parser learns mappings from words to these values, described later).
Given these primitives, compact theories can be authored for a particular domain by choosing relevant properties INLINEFORM0 , and specifying qualitative relationships (q+,q-) and ordinal values (higher-than) for them. For example, a simple theory about friction is sketched graphically in Figure FIGREF3 . Our observation is that these theories are relatively small, simple, and easy to author. Rather, the primary challenge is in mapping the complex and varied language of questions into a form that interfaces with this representation.
This language can be extended to include additional primitives from qualitative modeling, e.g., i+(x,y) (“the rate of change of x is qualitatively proportional to y”). That is, the techniques we present are not specific to our particular qualitative modeling subset. The only requirement is that, given a set of absolute values or qualitative relationships from a question, the theory can compute an answer.
## Representing Questions
A key feature of our representation is the conceptualization of questions as describing events happening in two worlds, world1 and world2, that are being compared. That comparison may be between two different entities, or the same entity at different time points. E.g., in Figure FIGREF1 the two worlds being compared are the car on wood, and the car on carpet. The tags world1 and world2 denote these different situations, and semantic parsing (Section SECREF5 ) requires learning to correctly associate these tags with parts of the question describing those situations. This abstracts away irrelevant details of the worlds, while still keeping track of which world is which.
We define the following two predicates to express qualitative information in questions: [vskip=1mm,leftmargin=5mm] qrel(property, direction, world)
qval(property, value, world) where property ( INLINEFORM0 ) INLINEFORM1 P, value INLINEFORM2 INLINEFORM3 , direction INLINEFORM4 {higher, lower}, and world INLINEFORM5 {world1, world2}. qrel() denotes the relative assertion that property is higher/lower in world compared with the other world, which is left implicit, e.g., from Figure FIGREF1 : [vskip=1mm,leftmargin=5mm] # The car rolls further on wood.
qrel(distance, higher, world1) where world1 is a tag for the “car on wood” situation (hence world2 becomes a tag for the opposite “car on carpet” situation). qval() denotes that property has an absolute value in world, e.g., [vskip=1mm,leftmargin=5mm] # The car's speed is slow on carpet.
qval(speed, low, world2)
## Logical Forms for Questions
Despite the wide variation in language, the space of logical forms (LFs) for the questions that we consider is relatively compact. In each question, the question body establishes a scenario and each answer option then probes an implication. We thus express a question's LF as a tuple: [vskip=1mm,leftmargin=5mm] (setup, answer-A, answer-B) where setup is the predicate(s) describing the scenario, and answer-* are the predicate(s) being queried for. If answer-A follows from setup, as inferred by the reasoner, then the answer is (A); similarly for (B). For readability we will write this as [vskip=1mm,leftmargin=5mm] setup INLINEFORM0 answer-A ; answer-B We consider two styles of LF, covering a large range of questions. The first is: [vskip=1mm,leftmargin=5mm] (1) qrel( INLINEFORM1 ) INLINEFORM2
qrel( INLINEFORM0 ) ; qrel( INLINEFORM1 ) which deals with relative values of properties between worlds, and applies when the question setup includes a comparative. An example of this is in Figure FIGREF1 . The second is: [vskip=1mm,leftmargin=5mm] (2) qval( INLINEFORM2 ), qval( INLINEFORM3 ) INLINEFORM4
qrel( INLINEFORM0 ) ; qrel( INLINEFORM1 ) which deals with absolute values of properties, and applies when the setup uses absolute terms instead of comparatives. An example is the first question in Figure FIGREF4 , shown simplified below, whose LF looks as follows (colors showing approximate correspondences): [vskip=1mm,leftmargin=5mm] # Does a bar stool orangeslide faster along the redbar surface with tealdecorative raised bumps or the magentasmooth wooden bluefloor? (A) redbar (B) bluefloor
qval(tealsmoothness, low, redworld1),
qval(magentasmoothness, high, blueworld2) INLINEFORM0
qrel(orangespeed, higher, redworld1) ;
qrel(orangespeed, higher, blueworld2)
## Inference
A small set of rules for qualitative reasoning connects these predicates together. For example, (in logic) if the value of P is higher in world1 than the value of P in world2 and q+(P,Q) then the value of Q will be higher in world1 than the value of Q in world2. Given a question's logical form, a qualitative model, and these rules, a Prolog-style inference engine determines which answer option follows from the premise.
## The QuaRel Dataset
QuaRel is a crowdsourced dataset of 2771 multiple-choice story questions, including their logical forms. The size of the dataset is similar to several other datasets with annotated logical forms used for semantic parsing BIBREF8 , BIBREF25 , BIBREF24 . As the space of LFs is constrained, the dataset is sufficient for a rich exploration of this space.
We crowdsourced multiple-choice questions in two parts, encouraging workers to be imaginative and varied in their use of language. First, workers were given a seed qualitative relation q+/-( INLINEFORM0 ) in the domain, expressed in English (e.g., “If a surface has more friction, then an object will travel slower”), and asked to enter two objects, people, or situations to compare. They then created a question, guided by a large number of examples, and were encouraged to be imaginative and use their own words. The results are a remarkable variety of situations and phrasings (Figure FIGREF4 ).
Second, the LFs were elicited using a novel technique of reverse-engineering them from a set of follow-up questions, without exposing workers to the underlying formalism. This is possible because of the constrained space of LFs. Referring to LF templates (1) and (2) earlier (Section SECREF13 ), these questions are as follows:
From this information, we can deduce the target LF ( INLINEFORM0 is the complement of INLINEFORM1 , INLINEFORM2 , we arbitrarily set INLINEFORM3 =world1, hence all other variables can be inferred). Three independent workers answer these follow-up questions to ensure reliable results.
We also had a human answer the questions in the dev partition (in principle, they should all be answerable). The human scored 96.4%, the few failures caused by occasional annotation errors or ambiguities in the question set itself, suggesting high fidelity of the content.
About half of the dataset are questions about friction, relating five different properties (friction, heat, distance, speed, smoothness). These questions form a meaningful, connected subset of the dataset which we denote QuaRel INLINEFORM0 . The remaining questions involve a wide variety of 14 additional properties and their relations, such as “exercise intensity vs. sweat” or “distance vs. brightness”.
Figure FIGREF4 shows typical examples of questions in QuaRel, and Table TABREF26 provides summary statistics. In particular, the vocabulary is highly varied (5226 unique words), given the dataset size. Figure FIGREF27 shows some examples of the varied phrases used to describe smoothness.
## Baseline Systems
We use four systems to evaluate the difficulty of this dataset. (We subsequently present two new models, extending the baseline neural semantic parser, in Sections SECREF36 and SECREF44 ). The first two are an information retrieval system and a word-association method, following the designs of BIBREF26 Clark2016CombiningRS. These are naive baselines that do not parse the question, but nevertheless may find some signal in a large corpus of text that helps guess the correct answer. The third is a CCG-style rule-based semantic parser written specifically for friction questions (the QuaRel INLINEFORM0 subset), but prior to data being collected. The last is a state-of-the-art neural semantic parser. We briefly describe each in turn.
## Baseline Experiments
We ran the above systems on the QuaRel dataset. QuaSP was trained on the training set, using the model with highest parse accuracy on the dev set (similarly BiLSTM used highest answer accuracy on the dev set) . The results are shown in Table TABREF34 . The 95% confidence interval is +/- 4% on the full test set. The human score is the sanity check on the dev set (Section SECREF4 ).
As Table TABREF34 shows, the QuaSP model performs better than other baseline approaches which are only slightly above random. QuaSP scores 56.1% (61.7% on the friction subset), indicating the challenges of this dataset.
For the rule-based system, we observe that it is unable to parse the majority (66%) of questions (hence scoring 0.5 for those questions, reflecting a random guess), due to the varied and unexpected vocabulary present in the dataset. For example, Figure FIGREF27 shows some of the ways that the notion of “smoother/rougher” is expressed in questions, many of which are not covered by the hand-written CCG grammar. This reflects the typical brittleness of hand-built systems.
For QuaSP, we also analyzed the parse accuracies, shown in Table TABREF35 , the score reflecting the percentage of times it produced exactly the right logical form. The random baseline for parse accuracy is near zero given the large space of logical forms, while the model parse accuracies are relatively high, much better than a random baseline.
Further analysis of the predicted LFs indicates that the neural model does well at predicting the properties ( INLINEFORM0 25% of errors on dev set), but struggles to predict the worlds in the LFs reliably ( INLINEFORM1 70% of errors on dev set). This helps explain why non-trivial parse accuracy does not necessarily translate into correspondingly higher answer accuracy: If only the world assignment is wrong, the answer will flip and give a score of zero, rather than the average 0.5.
## New Models
We now present two new models, both extensions of the neural baseline QuaSP. The first, QuaSP+, addresses the leading cause of failure just described, namely the problem of identifying the two worlds being compared, and significantly outperforms all the baseline systems. The second, QuaSP+Zero, addresses the scaling problem, namely the costly requirement of needing many training examples each time a new qualitative property is introduced. It does this by instead using only a small amount of lexical information about the new property, thus achieving “zero shot” performance, i.e., handling properties unseen in the training examples BIBREF34 , a capability not present in the baseline systems. We present the models and results for each.
## QuaSP+: A Model Incorporating World Tracking
We define the world tracking problem as identifying and tracking references to different “worlds” being compared in text, i.e., correctly mapping phrases to world identifiers, a critical aspect of the semantic parsing task. There are three reasons why this is challenging. First, unlike properties, the worlds being compared in questions are distinct in almost every question, and thus there is no obvious, learnable mapping from phrases to worlds. For example, while a property (like speed) has learnable ways to refer to it (“faster”, “moves rapidly”, “speeds”, “barely moves”), worlds are different in each question (e.g., “on a road”, “countertop”, “while cutting grass”) and thus learning to identify them is hard. Second, different phrases may be used to refer to the same world in the same question (see Figure FIGREF43 ), further complicating the task. Finally, even if the model could learn to identify worlds in other ways, e.g., by syntactic position in the question, there is the problem of selecting world1 or world2 consistently throughout the parse, so that the equivalent phrasings are assigned the same world.
This problem of mapping phrases to world identifiers is similar to the task of entity linking BIBREF35 . In prior semantic parsing work, entity linking is relatively straightforward: simple string-matching heuristics are often sufficient BIBREF36 , BIBREF37 , or an external entity linking system can be used BIBREF38 , BIBREF39 . In QuaRel, however, because the phrases denoting world1 and world2 are different in almost every question, and the word “world” is never used, such methods cannot be applied.
To address this, we have developed QuaSP+, a new model that extends QuaSP by adding an extra initial step to identify and delexicalize world references in the question. In this delexicalization process, potentially new linguistic descriptions of worlds are replaced by canonical tokens, creating the opportunity for the model to generalize across questions. For example, the world mentions in the question: [vskip=1mm,leftmargin=5mm] “A ball rolls further on wood than carpet because the (A) carpet is smoother (B) wood is smoother” are delexicalized to: [vskip=1mm,leftmargin=5mm] “A ball rolls further on World1 than World2 because the (A) World2 is smoother (B) World1 is smoother” This approach is analogous to BIBREF40 Herzig2018DecouplingSA, who delexicalized words to POS tags to avoid memorization. Similar delexicalized features have also been employed in Open Information Extraction BIBREF41 , so the Open IE system could learn a general model of how relations are expressed. In our case, however, delexicalizing to World1 and World2 is itself a significant challenge, because identifying phrases referring to worlds is substantially more complex than (say) identifying parts of speech.
To perform this delexicalization step, we use the world annotations included as part of the training dataset (Section SECREF4 ) to train a separate tagger to identify “world mentions” (text spans) in the question using BIO tags (BiLSTM encoder followed by a CRF). The spans are then sorted into World1 and World2 using the following algorithm:
If one span is a substring of another, they are are grouped together. Remaining spans are singleton groups.
The two groups containing the longest spans are labeled as the two worlds being compared.
Any additional spans are assigned to one of these two groups based on closest edit distance (or ignored if zero overlap).
The group appearing first in the question is labeled World1, the other World2.
The result is a question in which world mentions are canonicalized. The semantic parser QuaSP is then trained using these questions. We call the combined system (delexicalization plus semantic parser) QuaSP+.
The results for QuaSP+ are included in Table TABREF34 . Most importantly, QuaSP+ significantly outperforms the baselines by over 12% absolute. Similarly, the parse accuracies are significantly improved from 32.2% to 43.8% (Table TABREF35 ). This suggests that this delexicalization technique is an effective way of making progress on this dataset, and more generally on problems where multiple situations are being compared, a common characteristic of qualitative problems.
## QuaSP+Zero: A Model for the Zero-Shot Task
While our delexicalization procedure demonstrates a way of addressing the world tracking problem, the approach still relies on annotated data; if we were to add new qualitative relations, new training data would be needed, which is a significant scalability obstacle. To address this, we define the zero-shot problem as being able to answer questions involving a new predicate p given training data only about other predicates P different from p. For example, if we add a new property (e.g., heat) to the qualitative model (e.g., adding q+(friction, heat); “more friction implies more heat”), we want to answer questions involving heat without creating new annotated training questions, and instead only use minimal extra information about the new property. A parser that achieved good zero-shot performance, i.e., worked well for new properties unseen at training time, would be a substantial advance, allowing a new qualitative model to link to questions with minimal effort.
QuaRel provides an environment in which methods for this zero-shot theory extension can be devised and evaluated. To do this, we consider the following experimental setting: All questions mentioning a particular property are removed, the parser is trained on the remainder, and then tested on those withheld questions, i.e., questions mentioning a property unseen in the training data.
We present and evaluate a model that we have developed for this, called QuaSP+Zero, that modifies the QuaSP+ parser as follows: During decoding, at points where the parser is selecting which property to include in the LF (e.g., Figure FIGREF31 ), it does not just consider the question tokens, but also the relationship between those tokens and the properties INLINEFORM0 used in the qualitative model. For example, a question token such as “longer” can act as a cue for (the property) length, even if unseen in the training data, because “longer” and a lexical form of length (e.g.,“length”) are similar. This approach follows the entity-linking approach used by BIBREF11 Krishnamurthy2017NeuralSP, where the similarity between question tokens and (words associated with) entities - called the entity linking score - help decide which entities to include in the LF during parsing. Here, we modify their entity linking score INLINEFORM1 , linking question tokens INLINEFORM2 and property “entities” INLINEFORM3 , to be: INLINEFORM4
where INLINEFORM0 is a diagonal matrix connecting the embedding of the question token INLINEFORM1 and words INLINEFORM2 associated with the property INLINEFORM3 . For INLINEFORM4 , we provide a small list of words for each property (such as “speed”, “velocity”, and “fast” for the speed property), a small-cost requirement.
The results with QuaSP+Zero are in Table TABREF45 , shown in detail on the QuaRel INLINEFORM0 subset and (due to space constraints) summarized for the full QuaRel. We can measure overall performance of QuaSP+Zero by averaging each of the zero-shot test sets (weighted by the number of questions in each set), resulting in an overall parse accuracy of 38.9% and answer accuracy 61.0% on QuaRel INLINEFORM1 , and 25.7% (parse) and 59.5% (answer) on QuaRel, both significantly better than random. These initial results are encouraging, suggesting that it may be possible to parse into modified qualitative models that include new relations, with minimal annotation effort, significantly opening up qualitative reasoning methods for QA.
## Summary and Conclusion
Our goal is to answer questions that involve qualitative relationships, an important genre of task that involves both language and knowledge, but also one that presents significant challenges for semantic parsing. To this end we have developed a simple and flexible formalism for representing these questions; constructed QuaRel, the first dataset of qualitative story questions that exemplifies these challenges; and presented two new models that adapt existing parsing techniques to this task. The first model, QuaSP+, illustrates how delexicalization can help with world tracking (identifying different “worlds” in questions), resulting in state-of-the-art performance on QuaRel. The second model, QuaSP+Zero, illustrates how zero-shot learning can be achieved (i.e., adding new qualitative relationships without requiring new training examples) by using an entity-linking approach applied to properties - a capability not present in previous models.
There are several directions in which this work can be expanded. First, quantitative property values (e.g., “10 mph”) are currently not handled well, as their mapping to “low” or “high” is context-dependent. Second, some questions do not fit our two question templates (Section SECREF13 ), e.g., where two property values are a single answer option (e.g., “....(A) one floor is smooth and the other floor is rough”). Finally, some questions include an additional level of indirection, requiring an inference step to map to qualitative relations. For example, “Which surface would be best for a race? (A) gravel (B) blacktop” requires the additional commonsense inference that “best for a race” implies “higher speed”.
Given the ubiquity of qualitative comparisons in natural text, recognizing and reasoning with qualitative relationships is likely to remain an important task for AI. This work makes inroads into this task, and contributes a dataset and models to encourage progress by others. The dataset and models are publicly available at http://data.allenai.org/quarel.
| 14
|
1812.06705
|
Conditional BERT Contextual Augmentation
|
# Conditional BERT Contextual Augmentation
## Abstract
We propose a novel data augmentation method for labeled sentences called conditional BERT contextual augmentation. Data augmentation methods are often applied to prevent overfitting and improve generalization of deep neural network models. Recently proposed contextual augmentation augments labeled sentences by randomly replacing words with more varied substitutions predicted by language model. BERT demonstrates that a deep bidirectional language model is more powerful than either an unidirectional language model or the shallow concatenation of a forward and backward model. We retrofit BERT to conditional BERT by introducing a new conditional masked language model\footnote{The term"conditional masked language model"appeared once in original BERT paper, which indicates context-conditional, is equivalent to term"masked language model". In our paper,"conditional masked language model"indicates we apply extra label-conditional constraint to the"masked language model".} task. The well trained conditional BERT can be applied to enhance contextual augmentation. Experiments on six various different text classification tasks show that our method can be easily applied to both convolutional or recurrent neural networks classifier to obtain obvious improvement.
## Introduction
Deep neural network-based models are easy to overfit and result in losing their generalization due to limited size of training data. In order to address the issue, data augmentation methods are often applied to generate more training samples. Recent years have witnessed great success in applying data augmentation in the field of speech area BIBREF0 , BIBREF1 and computer vision BIBREF2 , BIBREF3 , BIBREF4 . Data augmentation in these areas can be easily performed by transformations like resizing, mirroring, random cropping, and color shifting. However, applying these universal transformations to texts is largely randomized and uncontrollable, which makes it impossible to ensure the semantic invariance and label correctness. For example, given a movie review “The actors is good", by mirroring we get “doog si srotca ehT", or by random cropping we get “actors is", both of which are meaningless.
Existing data augmentation methods for text are often loss of generality, which are developed with handcrafted rules or pipelines for specific domains. A general approach for text data augmentation is replacement-based method, which generates new sentences by replacing the words in the sentences with relevant words (e.g. synonyms). However, words with synonyms from a handcrafted lexical database likes WordNet BIBREF5 are very limited , and the replacement-based augmentation with synonyms can only produce limited diverse patterns from the original texts. To address the limitation of replacement-based methods, Kobayashi BIBREF6 proposed contextual augmentation for labeled sentences by offering a wide range of substitute words, which are predicted by a label-conditional bidirectional language model according to the context. But contextual augmentation suffers from two shortages: the bidirectional language model is simply shallow concatenation of a forward and backward model, and the usage of LSTM models restricts their prediction ability to a short range.
BERT, which stands for Bidirectional Encoder Representations from Transformers, pre-trained deep bidirectional representations by jointly conditioning on both left and right context in all layers. BERT addressed the unidirectional constraint by proposing a “masked language model" (MLM) objective by masking some percentage of the input tokens at random, and predicting the masked words based on its context. This is very similar to how contextual augmentation predict the replacement words. But BERT was proposed to pre-train text representations, so MLM task is performed in an unsupervised way, taking no label variance into consideration.
This paper focuses on the replacement-based methods, by proposing a novel data augmentation method called conditional BERT contextual augmentation. The method applies contextual augmentation by conditional BERT, which is fine-tuned on BERT. We adopt BERT as our pre-trained language model with two reasons. First, BERT is based on Transformer. Transformer provides us with a more structured memory for handling long-term dependencies in text. Second, BERT, as a deep bidirectional model, is strictly more powerful than the shallow concatenation of a left-to-right and right-to left model. So we apply BERT to contextual augmentation for labeled sentences, by offering a wider range of substitute words predicted by the masked language model task. However, the masked language model predicts the masked word based only on its context, so the predicted word maybe incompatible with the annotated labels of the original sentences. In order to address this issue, we introduce a new fine-tuning objective: the "conditional masked language model"(C-MLM). The conditional masked language model randomly masks some of the tokens from an input, and the objective is to predict a label-compatible word based on both its context and sentence label. Unlike Kobayashi's work, the C-MLM objective allows a deep bidirectional representations by jointly conditioning on both left and right context in all layers. In order to evaluate how well our augmentation method improves performance of deep neural network models, following Kobayashi BIBREF6 , we experiment it on two most common neural network structures, LSTM-RNN and CNN, on text classification tasks. Through the experiments on six various different text classification tasks, we demonstrate that the proposed conditional BERT model augments sentence better than baselines, and conditional BERT contextual augmentation method can be easily applied to both convolutional or recurrent neural networks classifier. We further explore our conditional MLM task’s connection with style transfer task and demonstrate that our conditional BERT can also be applied to style transfer too.
Our contributions are concluded as follows:
To our best knowledge, this is the first attempt to alter BERT to a conditional BERT or apply BERT on text generation tasks.
## Fine-tuning on Pre-trained Language Model
Language model pre-training has attracted wide attention and fine-tuning on pre-trained language model has shown to be effective for improving many downstream natural language processing tasks. Dai BIBREF7 pre-trained unlabeled data to improve Sequence Learning with recurrent networks. Howard BIBREF8 proposed a general transfer learning method, Universal Language Model Fine-tuning (ULMFiT), with the key techniques for fine-tuning a language model. Radford BIBREF9 proposed that by generative pre-training of a language model on a diverse corpus of unlabeled text, large gains on a diverse range of tasks could be realized. Radford BIBREF9 achieved large improvements on many sentence-level tasks from the GLUE benchmark BIBREF10 . BERT BIBREF11 obtained new state-of-the-art results on a broad range of diverse tasks. BERT pre-trained deep bidirectional representations which jointly conditioned on both left and right context in all layers, following by discriminative fine-tuning on each specific task. Unlike previous works fine-tuning pre-trained language model to perform discriminative tasks, we aim to apply pre-trained BERT on generative tasks by perform the masked language model(MLM) task. To generate sentences that are compatible with given labels, we retrofit BERT to conditional BERT, by introducing a conditional masked language model task and fine-tuning BERT on the task.
## Text Data Augmentation
Text data augmentation has been extensively studied in natural language processing. Sample-based methods includes downsampling from the majority classes and oversampling from the minority class, both of which perform weakly in practice. Generation-based methods employ deep generative models such as GANs BIBREF12 or VAEs BIBREF13 , BIBREF14 , trying to generate sentences from a continuous space with desired attributes of sentiment and tense. However, sentences generated in these methods are very hard to guarantee the quality both in label compatibility and sentence readability. In some specific areas BIBREF15 , BIBREF16 , BIBREF17 . word replacement augmentation was applied. Wang BIBREF18 proposed the use of neighboring words in continuous representations to create new instances for every word in a tweet to augment the training dataset. Zhang BIBREF19 extracted all replaceable words from the given text and randomly choose $r$ of them to be replaced, then substituted the replaceable words with synonyms from WordNet BIBREF5 . Kolomiyets BIBREF20 replaced only the headwords under a task-specific assumption that temporal trigger words usually occur as headwords. Kolomiyets BIBREF20 selected substitute words with top- $K$ scores given by the Latent Words LM BIBREF21 , which is a LM based on fixed length contexts. Fadaee BIBREF22 focused on the rare word problem in machine translation, replacing words in a source sentence with only rare words. A word in the translated sentence is also replaced using a word alignment method and a rightward LM. The work most similar to our research is Kobayashi BIBREF6 . Kobayashi used a fill-in-the-blank context for data augmentation by replacing every words in the sentence with language model. In order to prevent the generated words from reversing the information related to the labels of the sentences, Kobayashi BIBREF6 introduced a conditional constraint to control the replacement of words. Unlike previous works, we adopt a deep bidirectional language model to apply replacement, and the attention mechanism within our model allows a more structured memory for handling long-term dependencies in text, which resulting in more general and robust improvement on various downstream tasks.
## Preliminary: Masked Language Model Task
In general, the language model(LM) models the probability of generating natural language sentences or documents. Given a sequence $\textbf {\textit {S}}$ of N tokens, $<t_1,t_2,...,t_N>$ , a forward language model allows us to predict the probability of the sequence as:
$$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_1, t_2,..., t_{i-1}).$$ (Eq. 8)
Similarly, a backward language model allows us to predict the probability of the sentence as:
$$p(t_1,t_2,...,t_N) = \prod _{i=1}^{N}p(t_i|t_{i+1}, t_{i+2},..., t_N).$$ (Eq. 9)
Traditionally, a bidirectional language model a shallow concatenation of independently trained forward and backward LMs.
In order to train a deep bidirectional language model, BERT proposed Masked Language Model (MLM) task, which was also referred to Cloze Task BIBREF23 . MLM task randomly masks some percentage of the input tokens, and then predicts only those masked tokens according to their context. Given a masked token ${t_i}$ , the context is the tokens surrounding token ${t_i}$ in the sequence $\textbf {\textit {S}}$ , i.e. cloze sentence ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ . The final hidden vectors corresponding to the mask tokens are fed into an output softmax over the vocabulary to produce words with a probability distribution ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . MLM task only predicts the masked words rather than reconstructing the entire input, which suggests that more pre-training steps are required for the model to converge. Pre-trained BERT can augment sentences through MLM task, by predicting new words in masked positions according to their context.
## Conditional BERT
As shown in Fig 1 , our conditional BERT shares the same model architecture with the original BERT. The differences are the input representation and training procedure.
The input embeddings of BERT are the sum of the token embeddings, the segmentation embeddings and the position embeddings. For the segmentation embeddings in BERT, a learned sentence A embedding is added to every token of the first sentence, and if a second sentence exists, a sentence B embedding will be added to every token of the second sentence. However, the segmentation embeddings has no connection to the actual annotated labels of a sentence, like sense, sentiment or subjectivity, so predicted word is not always compatible with annotated labels. For example, given a positive movie remark “this actor is good", we have the word “good" masked. Through the Masked Language Model task by BERT, the predicted word in the masked position has potential to be negative word likes "bad" or "boring". Such new generated sentences by substituting masked words are implausible with respect to their original labels, which will be harmful if added to the corpus to apply augmentation. In order to address this issue, we propose a new task: “conditional masked language model".
The conditional masked language model randomly masks some of the tokens from the labeled sentence, and the objective is to predict the original vocabulary index of the masked word based on both its context and its label. Given a masked token ${t_i}$ , the context ${\textbf {\textit {S}}\backslash \lbrace t_i \rbrace }$ and label ${y}$ are both considered, aiming to calculate ${p(\cdot |y,\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ , instead of calculating ${p(\cdot |\textbf {\textit {S}}\backslash \lbrace t_i \rbrace )}$ . Unlike MLM pre-training, the conditional MLM objective allows the representation to fuse the context information and the label information, which allows us to further train a label-conditional deep bidirectional representations.
To perform conditional MLM task, we fine-tune on pre-trained BERT. We alter the segmentation embeddings to label embeddings, which are learned corresponding to their annotated labels on labeled datasets. Note that the BERT are designed with segmentation embedding being embedding A or embedding B, so when a downstream task dataset with more than two labels, we have to adapt the size of embedding to label size compatible. We train conditional BERT using conditional MLM task on labeled dataset. After the model has converged, it is expected to be able to predict words in masked position both considering the context and the label.
## Conditional BERT Contextual Augmentation
After the conditional BERT is well-trained, we utilize it to augment sentences. Given a labeled sentence from the corpus, we randomly mask a few words in the sentence. Through conditional BERT, various words compatibly with the label of the sentence are predicted by conditional BERT. After substituting the masked words with predicted words, a new sentences is generated, which shares similar context and same label with original sentence. Then new sentences are added to original corpus. We elaborate the entire process in algorithm "Conditional BERT Contextual Augmentation" .
Conditional BERT contextual augmentation algorithm. Fine-tuning on the pre-trained BERT , we retrofit BERT to conditional BERT using conditional MLM task on labeled dataset. After the model converged, we utilize it to augment sentences. New sentences are added into dataset to augment the dataset. [1] Alter the segmentation embeddings to label embeddings Fine-tune the pre-trained BERT using conditional MLM task on labeled dataset D until convergence each iteration i=1,2,...,M Sample a sentence $s$ from D Randomly mask $k$ words Using fine-tuned conditional BERT to predict label-compatible words on masked positions to generate a new sentence $S^{\prime }$ Add new sentences into dataset $D$ to get augmented dataset $D^{\prime }$ Perform downstream task on augmented dataset $D^{\prime }$
## Experiment
In this section, we present conditional BERT parameter settings and, following Kobayashi BIBREF6 , we apply different augmentation methods on two types of neural models through six text classification tasks. The pre-trained BERT model we used in our experiment is BERT $_{BASE}$ , with number of layers (i.e., Transformer blocks) $L = 12$ , the hidden size $ H = 768$ , and the number of self-attention heads $A = 12$ , total parameters $= 110M$ . Detailed pre-train parameters setting can be found in original paper BIBREF11 . For each task, we perform the following steps independently. First, we evaluate the augmentation ability of original BERT model pre-trained on MLM task. We use pre-trained BERT to augment dataset, by predicted masked words only condition on context for each sentence. Second, we fine-tune the original BERT model to a conditional BERT. Well-trained conditional BERT augments each sentence in dataset by predicted masked words condition on both context and label. Third, we compare the performance of the two methods with Kobayashi's BIBREF6 contextual augmentation results. Note that the original BERT’s segmentation embeddings layer is compatible with two-label dataset. When the task-specific dataset is with more than two different labels, we should re-train a label size compatible label embeddings layer instead of directly fine-tuning the pre-trained one.
## Datasets
Six benchmark classification datasets are listed in table 1 . Following Kim BIBREF24 , for a dataset without validation data, we use 10% of its training set for the validation set. Summary statistics of six classification datasets are shown in table 1.
SST BIBREF25 SST (Stanford Sentiment Treebank) is a dataset for sentiment classification on movie reviews, which are annotated with five labels (SST5: very positive, positive, neutral, negative, or very negative) or two labels (SST2: positive or negative).
Subj BIBREF26 Subj (Subjectivity dataset) is annotated with whether a sentence is subjective or objective.
MPQA BIBREF27 MPQA Opinion Corpus is an opinion polarity detection dataset of short phrases rather than sentences, which contains news articles from a wide variety of news sources manually annotated for opinions and other private states (i.e., beliefs, emotions, sentiments, speculations, etc.).
RT BIBREF28 RT is another movie review sentiment dataset contains a collection of short review excerpts from Rotten Tomatoes collected by Bo Pang and Lillian Lee.
TREC BIBREF29 TREC is a dataset for classification of the six question types (whether the question is about person, location, numeric information, etc.).
## Text classification
We evaluate the performance improvement brought by conditional BERT contextual augmentation on sentence classification tasks, so we need to prepare two common sentence classifiers beforehand. For comparison, following Kobayashi BIBREF6 , we adopt two typical classifier architectures: CNN or LSTM-RNN. The CNN-based classifier BIBREF24 has convolutional filters of size 3, 4, 5 and word embeddings. All outputs of each filter are concatenated before applied with a max-pooling over time, then fed into a two-layer feed-forward network with ReLU, followed by the softmax function. An RNN-based classifier has a single layer LSTM and word embeddings, whose output is fed into an output affine layer with the softmax function. For both the architectures, dropout BIBREF30 and Adam optimization BIBREF31 are applied during training. The train process is finish by early stopping with validation at each epoch.
Sentence classifier hyper-parameters including learning rate, embedding dimension, unit or filter size, and dropout ratio, are selected using grid-search for each task-specific dataset. We refer to Kobayashi's implementation in the released code. For BERT, all hyper-parameters are kept the same as Devlin BIBREF11 , codes in Tensorflow and PyTorch are all available on github and pre-trained BERT model can also be downloaded. The number of conditional BERT training epochs ranges in [1-50] and number of masked words ranges in [1-2].
We compare the performance improvements obtained by our proposed method with the following baseline methods, “w/" means “with":
w/synonym: Words are randomly replaced with synonyms from WordNet BIBREF5 .
w/context: Proposed by Kobayashi BIBREF6 , which used a bidirectional language model to apply contextual augmentation, each word was replaced with a probability.
w/context+label: Kobayashi’s contextual augmentation method BIBREF6 in a label-conditional LM architecture.
Table 2 lists the accuracies of the all methods on two classifier architectures. The results show that, for various datasets on different classifier architectures, our conditional BERT contextual augmentation improves the model performances most. BERT can also augments sentences to some extent, but not as much as conditional BERT does. For we masked words randomly, the masked words may be label-sensitive or label-insensitive. If label-insensitive words are masked, words predicted through BERT may not be compatible with original labels. The improvement over all benchmark datasets also shows that conditional BERT is a general augmentation method for multi-labels sentence classification tasks.
We also explore the effect of number of training steps to the performance of conditional BERT data augmentation. The fine-tuning epoch setting ranges in [1-50], we list the fine-tuning epoch of conditional BERT to outperform BERT for various benchmarks in table 3 . The results show that our conditional BERT contextual augmentation can achieve obvious performance improvement after only a few fine-tuning epochs, which is very convenient to apply to downstream tasks.
## Connection to Style Transfer
In this section, we further deep into the connection to style transfer and apply our well trained conditional BERT to style transfer task. Style transfer is defined as the task of rephrasing the text to contain specific stylistic properties without changing the intent or affect within the context BIBREF32 . Our conditional MLM task changes words in the text condition on given label without changing the context. View from this point, the two tasks are very close. So in order to apply conditional BERT to style transfer task, given a specific stylistic sentence, we break it into two steps: first, we find the words relevant to the style; second, we mask the style-relevant words, then use conditional BERT to predict new substitutes with sentence context and target style property. In order to find style-relevant words in a sentence, we refer to Xu BIBREF33 , which proposed an attention-based method to extract the contribution of each word to the sentence sentimental label. For example, given a positive movie remark “This movie is funny and interesting", we filter out the words contributes largely to the label and mask them. Then through our conditional BERT contextual augmentation method, we fill in the masked position by predicting words conditioning on opposite label and sentence context, resulting in “This movie is boring and dull". The words “boring" and “dull" contribute to the new sentence being labeled as negative style. We sample some sentences from dataset SST2, transferring them to the opposite label, as listed in table 4 .
## Conclusions and Future Work
In this paper, we fine-tune BERT to conditional BERT by introducing a novel conditional MLM task. After being well trained, the conditional BERT can be applied to data augmentation for sentence classification tasks. Experiment results show that our model outperforms several baseline methods obviously. Furthermore, we demonstrate that our conditional BERT can also be applied to style transfer task. In the future, (1)We will explore how to perform text data augmentation on imbalanced datasets with pre-trained language model, (2) we believe the idea of conditional BERT contextual augmentation is universal and will be applied to paragraph or document level data augmentation.
| 11
|
1901.00570
|
Event detection in Twitter: A keyword volume approach
|
# Event detection in Twitter: A keyword volume approach
## Abstract
Event detection using social media streams needs a set of informative features with strong signals that need minimal preprocessing and are highly associated with events of interest. Identifying these informative features as keywords from Twitter is challenging, as people use informal language to express their thoughts and feelings. This informality includes acronyms, misspelled words, synonyms, transliteration and ambiguous terms. In this paper, we propose an efficient method to select the keywords frequently used in Twitter that are mostly associated with events of interest such as protests. The volume of these keywords is tracked in real time to identify the events of interest in a binary classification scheme. We use keywords within word-pairs to capture the context. The proposed method is to binarize vectors of daily counts for each word-pair by applying a spike detection temporal filter, then use the Jaccard metric to measure the similarity of the binary vector for each word-pair with the binary vector describing event occurrence. The top n word-pairs are used as features to classify any day to be an event or non-event day. The selected features are tested using multiple classifiers such as Naive Bayes, SVM, Logistic Regression, KNN and decision trees. They all produced AUC ROC scores up to 0.91 and F1 scores up to 0.79. The experiment is performed using the English language in multiple cities such as Melbourne, Sydney and Brisbane as well as the Indonesian language in Jakarta. The two experiments, comprising different languages and locations, yielded similar results.
## Introduction
Event detection is important for emergency services to react rapidly and minimize damage. For example, terrorist attacks, protests, or bushfires may require the presence of ambulances, firefighters, and police as soon as possible to save people. This research aims to detect events as soon as they occur and are reported via some Twitter user. The event detection process requires to know the keywords associated with each event and to assess the minimal count of each word to decide confidently that an event has occurred. In this research, we propose a novel method of spike matching to identify keywords, and use probabilistic classification to assess the probability of having an event given the volume of each word.
Event detection and prediction from social networks have been studied frequently in recent years. Most of the predictive frameworks use textual content such as likes, shares, and retweets, as features. The text is used as features either by tracking the temporal patterns of keywords, clustering words into topics, or by evaluating sentiment scores and polarity. The main challenge in keyword-based models is to determine which words to use in the first place, especially as people use words in a non-standard way, particularly on Twitter.
In this research, we aim for detecting large events as soon as they happen with near-live sensitivity. For example, When spontaneous protests occur just after recent news such as increasing taxes or decreasing budget, we need to have indicators to raise the flag of a happening protest. Identifying these indicators requires to select a set of words that are mostly associated with the events of interest such as protests. We then track the volume of these words and evaluate the probability of an event occurring given the current volume of each of the tracked features. The main challenge is to find this set of features that allow such probabilistic classification.
Using text as features in Twitter is challenging because of the informal nature of the tweets, the limited length of the tweet, platform-specific language, and multilingual nature of Twitter BIBREF0 , BIBREF1 , BIBREF2 . The main challenges for text analysis in Twitter are listed below:
We approached the first and second challenges by using a Bayesian approach to learn which terms were associated with events, regardless of whether they are standard language, acronyms, or even a made-up word, so long as they match the events of interest. The third and fourth challenges are approached by using word-pairs, where we extract all the pairs of co-occurring words within each tweet. This allows us to recognize the context of the word ('Messi','strike' ) is different than ('labour','strike').
According to the distributional semantic hypothesis, event-related words are likely to be used on the day of an event more frequently than any normal day before or after the event. This will form a spike in the keyword count magnitude along the timeline as illustrated in Figure FIGREF6 . To find the words most associated with events, we search for the words that achieve the highest number of spikes matching the days of events. We use the Jaccard similarity metric as it values the spikes matching events and penalizes spikes with no event and penalizes events without spikes. Separate words can be noisy due to the misuse of the term by people, especially in big data environments. So, we rather used the word-pairs as textual features in order to capture the context of the word. For example, this can differentiate between the multiple usages of the word “strike” within the contexts of “lightning strike”, “football strike” and “labour strike”
In this paper, we propose a method to find the best word-pairs to represent the events of interest. These word-pairs can be used for time series analysis to predict future events as indicated in Figure FIGREF1 . They can also be used as seeds for topic modelling, or to find related posts and word-pairs using dynamic query expansion. The proposed framework uses a temporal filter to identify the spikes within the word-pair signal to binarize the word-pair time series vector BIBREF3 . The binary vector of the word-pair is compared to the protest days vector using Jaccard similarity index BIBREF4 , BIBREF5 , where the word-pairs with highest similarity scores are the most associated word-pairs with protest days. This feature selection method is built upon the assumption that people discuss an event on the day of that event more than on any day before or after the event. This implies that word-pairs related to the event will form a spike on this specific day. Some of the spiking word-pairs are related to the nature of the event itself, such as “taxi protest” or “fair education”. These word-pairs will appear once or twice along the time frame. Meanwhile, more generic word-pairs such as “human rights” or “labour strike” will spike more frequently in the days of events regardless the protest nature.
To test our method, we developed two experiments using all the tweets in Melbourne and Sydney over a period of 640 days. The total number of tweets exceeded 4 million tweets per day, with a total word-pair count of 12 million different word-pairs per day, forming 6 billion word-pairs over the entire timeframe. The selected word-pairs from in each city are used as features to classify if there will be an event or not on a specific day in that city. We classified events from the extracted word-pairs using 9 classifiers including Naive Bayes, Decision Trees, KNN, SVM, and logistic regression.
In Section 2, we describe the event detection methods. Section 3 states the known statistical methods used for data association and feature selection. Section 4 describes the proposed feature selection method. Section 5 describes model training and prediction. Section 6 describes the experiment design, the data and the results. Section 7 summarizes the paper, discuss the research conclusion and explains future work.
## Event Detection Methods
Analyzing social networks for event detection is approached from multiple perspectives depending on the research objective. This can be predicting election results, a contest winner, or predicting peoples' reaction to a government decision through protest. The main perspectives to analyze the social networks are (1) content analysis, where the textual content of each post is analyzed using natural language processing to identify the topic or the sentiment of the authors. (2) Network structure analysis, where the relation between the users are described in a tree structure for the follower-followee patterns, or in a graph structure for friendship and interaction patterns. These patterns can be used to know the political preference of people prior to elections. (3) Behavioural analysis of each user including sentiment, response, likes, retweets, location, to identify responses toward specific events. This might be useful to identify users with terrorist intentions. In this section, we will focus on textual content-based models, where text analysis and understanding can be achieved using keywords, topic modelling or sentiment analysis.
## Keyword-based approaches
Keyword-based approaches focus on sequence analysis of the time series for each keyword. They also consider different forms for each keyword, including n-gram, skip-gram, and word-pairs BIBREF6 . The keyword-based approaches use the concept of the distributional semantics to group semantically-related words as synonyms to be used as a single feature BIBREF7 . In this approach, keywords are usually associated with events by correlation, entropy or distance metrics. Also, Hossny et al. proposed using SVD with K-Means to strengthen keyword signals, by grouping words having similar temporal patterns, then mapping them into one central word that has minimum distance to the other members of the cluster BIBREF8 .
Sayyadi et al. used co-occurring keywords in documents such as news articles to build a network of keywords. This network is used as a graph to feed a community detection algorithm in order to identify and classify events BIBREF9 . Takeshi et al. created a probabilistic spatio-temporal model to identify natural disasters events such as earthquakes and typhoons using multiple tweet-based features such as words counts per tweet, event-related keywords, and tweet context. They considered each Twitter user as a social sensor and applied both of the Kalman filter and particle filter for location estimation. This model could detect 96% of Japanese earthquakes BIBREF10 . Zhou et al. developed a named entity recognition model to find location names within tweets and use them as keyword-features for event detection, then estimated the impact of the detected events qualitatively BIBREF11 .
Weng et al. introduced “Event Detection by Clustering of Wavelet-based Signals” (EDCow). This model used wavelets to analyze the frequency of word signals, then calculated the autocorrelations of each word signal in order to filter outlier words. The remaining words were clustered using a modularity-based graph partitioning technique to form events BIBREF12 . Ning et al. proposed a model to identify evidence-based precursors and forecasts of future events. They used as a set of news articles to develop a nested multiple instance learning model to predict events across multiple countries. This model can identify the news articles that can be used as precursors for a protest BIBREF13 .
## Topic modelling approaches
Topic modelling approaches focus on clustering related words according to their meaning, and indexing them using some similarity metric such as cosine similarity or Euclidean distance. The most recognized techniques are (1) Latent Semantic Indexing (LSI), where the observation matrix is decomposed using singular value decomposition and the data are clustered using K-Means BIBREF7 ,(2) Latent Dirichlet Allocation (LDA), where the words are clustered using Gaussian mixture models (GMM) according to the likelihood of term co-occurrence within the same context BIBREF14 , (3) Word2Vec, which uses a very large corpus to compute continuous vector representations, where we can apply standard vector operations to map one vector to another BIBREF15 .
Cheng et al. suggested using space-time scan statistics to detect events by looking for clusters within data across both time and space, regardless of the content of each individual tweet BIBREF16 . The clusters emerging during spatio-temporal relevant events are used as an indicator of a currently occurring event, as people tweet more often about event topics and news. Ritter et al. proposed a framework that uses the calendar date, cause and event type to describe any event in a way similar to the way Twitter users mention the important events. This framework used temporal resolution, POS tagging, an event tagger, and named entity recognition. Once features are extracted, the association between the combination of features and the events is measured in order to know what are the most important features and how significant the event will be BIBREF17 .
Zhou et al. introduced a graphical model to capture the information in the social data including time, content, and location, calling it location-time constrained topic (LTT). They measure the similarity between the tweets using KL divergence to assess media content uncertainty. Then, they measure the similarity between users using a “longest common subsequence” (LCS) metric. They aggregate the two measurements by augmenting weights as a measure for message similarity. They used the similarity between streaming posts in a social network to detect social events BIBREF18 .
Ifrim et al. presented another approach for topic detection that combines aggressive pre-processing of data with hierarchical clustering of tweets. The framework analyzes different factors affecting the quality of topic modelling results BIBREF19 , along with real-time data streams of live tweets to produce topic streams in close to real-time rate.
Xing et al. presented the mutually generative Latent Dirichlet Allocation model (MGE-LDA) that uses hashtags and topics, as they both are generated mutually by each other in tweets. This process models the relationship between topics and hashtags in tweets, and uses them both as features for event discovery BIBREF20 . Azzam et al. used deep learning and cosine similarity to understand short text posts in communities of question answering BIBREF21 , BIBREF22 . Also, Hossny et al. used inductive logic programming to understand short sentences from news for translation purposes BIBREF23
## Sentiment analysis approaches
The third approach is to identify sentiment through the context of the post, which is another application for distributional semantics requiring a huge amount of training data to build the required understanding of the context. Sentiment analysis approaches focus on recognizing the feelings of the crowd and use the score of each feeling as a feature to calculate the probability of social events occurring. The sentiment can represent the emotion, attitude, or opinion of the user towards the subject of the post. One approach to identify sentiment is to find smiley faces such as emoticons and emojis within a tweet or a post. Another approach is to use a sentiment labelled dictionary such as SentiWordNet to assess the sentiment associated with each word.
Generally, sentiment analysis has not been used solely to predict civil unrest, especially as it still faces the challenges of sarcasm and understanding negation in ill-formed sentences. Meanwhile, it is used as an extra feature in combination with features from other approaches such as keywords and topic modelling. Paul et al. proposed a framework to predict the results of the presidential election in the United States in 2017. The proposed framework applied topic modelling to identify related topics in news, then used the topics as seeds for Word2Vec and LSTM to generate a set of enriched keywords. The generated keywords will be used to classify politics-related tweets, which are used to evaluate the sentiment towards each candidate. The sentiment score trend is used to predict the winning candidate BIBREF24 .
## Feature Selection Methods
Keywords can be selected as features as a single term or a word-pair or a skip-grams, which can be used for classification using multiple methods such as mutual information, TF-IDF, INLINEFORM0 , or traditional statistical methods such as ANOVA or correlation. Our problem faces two challenges: the first is the huge number of word-pairs extracted from all tweets for the whole time frame concurrently, which make some techniques such as TF-IDF and INLINEFORM1 computationally unfeasible as they require the technique to be distributable on parallel processors on a cluster. The second challenge is the temporal nature of the data which require some techniques that can capture the distributional semantics of terms along with the ground truth vector. In this section, we describe briefly a set of data association methods used to find the best word-pairs to identify the event days.
Pearson correlation measures the linear dependency of the response variable on the independent variable with the maximum dependency of 1 and no dependency of zero. This technique needs to satisfy multiple assumptions to assess the dependency properly. These assumptions require the signals of the variables to be normally distributed, homoskedastic, stationary and have no outliers BIBREF25 , BIBREF26 . In social network and human-authored tweets, we cannot guarantee that the word-pairs signals throughout the timeframe will satisfy the required assumptions. Another drawback for Pearson correlation is that zero score does not necessarily imply no correlation, while no correlation implies zero score.
Spearman is a rank-based metric that evaluates the linear association between the rank variables for each of the independent and the response variables. It simply evaluates the linear correlation between the ranked variables of the original variables. Spearman correlation assumes the monotonicity of the variables but it relaxes the Pearson correlation requirements of the signal to be normal, homoskedastic and stationary. Although the text signals in the social network posts do not satisfy the monotonicity assumption, Spearman correlation can select some word-pairs to be used as predictive features for classification. Spearman correlation has the same drawback of Pearson correlation that zero score does not necessarily imply no correlation while no correlation implies zero score.
Distance correlation is introduced by Szekely et al . (2007) to measure the nonlinear association between two variables BIBREF27 . Distance correlation measures the statistical distance between probability distributions by dividing the Brownian covariance (distance covariance) between X and Y by the product of the distance standard deviations BIBREF28 , BIBREF29 .
TF-IDF is the short of term frequency-inverse document frequency technique that is used for word selection for classification problems. The concept of this technique is to give the words that occur frequently within a specific class high weight as a feature and to penalize the words that occur frequently among multiple classes. for example; the term “Shakespeare” is considered a useful feature to classify English literature documents as it occurs frequently in English literature and rarely occurs in any other kind of documents. Meanwhile, the term “act” will occur frequently in English literature, but it also occurs frequently in the other types of document, so this term will be weighted for its frequent appearance and it will be penalized for its publicity among the classes by what we call inverse-document-frequency BIBREF30 .
Mutual information is a metric for the amount of information one variable can tell the other one. MI evaluates how similar are the joint distributions of the two variables with the product of the marginal distributions of each individual variable, which makes MI more general than correlation as it is not limited by the real cardinal values, it can also be applied to binary, ordinal and nominal values BIBREF31 . As mutual information uses the similarity of the distribution, it is not concerned with pairing the individual observations of X and Y as much as it cares about the whole statistical distribution of X and Y. This makes MI very useful for clustering purposes rather than classification purposes BIBREF32 .
Cosine similarity metric calculates the cosine of the angle between two vectors. The cosine metric evaluates the direction similarity of the vectors rather the magnitude similarity. The cosine similarity score equals to 1 if the two vectors have the angle of zero between the directions of two vectors, and the score is set to zero when the two vectors are perpendicular BIBREF33 . if the two vectors are oriented to opposite directions, the similarity score is -1. Cosine similarity metric is usually used in the positive space, which makes the scores limited within the interval of [0,1].
Jaccard index or coefficient is a metric to evaluate the similarity of two sets by comparing their members to identify the common elements versus the distinct ones. The main advantage of Jaccard similarity is it ignores the default value or the null assumption in the two vectors and it only considers the non-default correct matches compared to the mismatches. This consideration makes the metric immune to the data imbalance. Jaccard index is similar to cosine-similarity as it retains the sparsity property and it also allows the discrimination of the collinear vectors.
## Spike Matching Method:
The proposed model extracts the word-pairs having a high association with event days according to the distributional semantic hypothesis and use them for training the model that will be used later for the binary classification task BIBREF34 as illustrated in figure FIGREF10 . The first step is the data preparation where we load all the tweets for each day, then we exclude the tweets having URLs or unrelated topics, then we clean each tweet by removing the hashtags, non-Latin script and stopping words. Then we lemmatize and stem each word in each tweet using Lancaster stemmer. Finally, we extract the word-pairs in each tweet. The word-pair is the list of n words co-occurring together within the same tweet.
The second step is to count the frequency of each word-pair per each day, which are used as features to classify the day as either event or no-event day. The formulation is a matrix with rows as word-pairs and columns as days and values are daily counts of each word-pair. The third step is to binarize the event count vector (ground truth) as well as the vector of each word-pair. Binarizing the event vector is done by checking if the count of events in each day is larger than zero. The binarization of the word-pair count vectors is done by applying a temporal filter to the time series in order to identify the spikes as explained in equation EQREF11 , where the days with spikes are set to ones and days without spike are set to zeros BIBREF35 , BIBREF36 . DISPLAYFORM0
Where x is the count of the word-pair, INLINEFORM0 is the time variable, INLINEFORM1 is the time difference, the threshold is the minimum height of the spike. Afterwards, we compare the binary vector for each word-pair with the ground truth binary vector using the Jaccard similarity index as stated in equation EQREF12 BIBREF4 , BIBREF5 . The word-pairs are then sorted descendingly according to the similarity score. The word-pairs with the highest scores are used as a feature for training the model in the fourth step. DISPLAYFORM0
where WP is the word pair vector, GT is the ground truth vector
## Training and Prediction
Once we identify the best word-pairs to be used as features for classification, we split the time series vector of each word-pair into a training vector and a testing vector. then we use the list of the training vectors of the selected word-pairs to train the model as explained in subsection SECREF13 and use the list of testing vectors for the same word-pairs to classify any day to event/nonevent day SECREF16 .
## Training the model:
The third step is to train the model using the set of features generated in the first step. We selected the Naive Bayes classifier to be our classification technique for the following reasons: (1) the high bias of the NB classifier reduces the possibility of over-fitting, and our problem has a high probability of over-fitting due to the high number of features and the low number of observations, (2) the response variable is binary, so we do not need to regress the variable real value as much as we need to know the event-class, and (3) The counts of the word-pairs as independent variables are limited between 0 and 100 occurrences per each day, which make the probabilistic approaches more effective than distance based approaches.
The training process aims to calculate three priori probabilities to be used later in calculating the posterior probabilities: (1) the probability of each word-pair count in a specific day given the status of the day as “event” or “non-event”. (2) the priori conditional probability of each word-pair given event status INLINEFORM0 . (3) the probability of each event class as well as the probability of each word-pair as stated in equations EQREF15 and EQREF15 . DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 is the word-pair, INLINEFORM1 is any class for event occurrence and word-pair is the vector of counts for the word-pairs extracted from tweets
## Predicting Civil Unrest
Once the priori probabilities are calculated using the training data, we use them to calculate the posterior probability of both classes of event-days and non-event-days given the values of the word-pairs using the equation EQREF17 . DISPLAYFORM0
where INLINEFORM0 is the word-pair, INLINEFORM1 INLINEFORM2 As the word-pairs are assumed to be independent and previously known from the training step.
## Experiments and Results
The experiments are designed to detect civil unrest events in Melbourne on any specific day. In this experiment, we used all the tweets posted from Melbourne within a time frame of 640 days between December 2015 and September 2017. This time frame will be split into 500 days for model training and 140 days for model testing on multiple folds. The tweet location is specified using (1) longitude and latitude meta-tag, (2) tweet location meta-tag, (3) the profile location meta-tag, and (4) The time zone meta-tag. The total number of tweets exceeded 4 million tweets daily. Firstly, we cleaned the data from noisy signals, performed stemming and lemmatization then extracted the word-pairs from each tweet and count each word-pair per each day. Example 1 illustrates how each tweet is cleaned, prepared and vectorized before being used for training the model. The steps are explained below:
As explained in example 1, each word-pair will be transformed from a vector of integer values into a vector of binary values and denoted as INLINEFORM0 . INLINEFORM1 will be used to calculate the Jaccard similarity index of the binary vector with the events binary vector. Each word-pair will have a similarity score according to the number of word-pair spikes matching the event days. This method uses the concept of distributional semantic, where the co-occurring signals are likely to be semantically associated BIBREF34 .
Example 1: Original Tweet: Protesters may be unmasked in wake of Coburg clash https://t.co/djjVIfzO3e (News) #melbourne #victoria Cleaned Tweet: protest unmask wake coburg clash news List of two-words-word-pairs: [`protest', `unmask'], [`protest', `wake'], [`protest', `Coburg'], ..., [`unmask', `wake'], [`unmask', `coburg'],..., [`clash', `news'] [`protest', `unmask'] training : INLINEFORM0 [`protest', `unmask'] testing : INLINEFORM1 Assuming a time frame of 20 days word-pair: [2,3,3,4,5,3,2,3,8,3,3,1,3,9,3,1,2,4,5,1] Spikes ( INLINEFORM2 ): [0,0,0,0,1,0,0,0,1,0,0,0,0,1,0,0,0,0,1,0] Events( INLINEFORM3 ): [0,0,0,0,1,0,0,0,0,0,0,0,0,1,0,0,1,0,1,0] INLINEFORM4
Once we selected the most informative word-pairs as features, we will use the raw values to train the Naive Bayes classifier. The classifier is trained using 500 days selected randomly along the whole timeframe, then it is used to predict the other 140 days. To ensure the robustness of our experiment, We applied 10-folds cross-validation, where we performed the same experiment 10 times using 10 different folds of randomly selected training and testing data. The prediction achieved an average area under the ROC curve of 90%, which statistically significant and achieved F-score of 91%, which is immune to data imbalance as listed in table TABREF18 . Figure FIGREF25 shows the ROC curves for the results of a single fold of Naive Bayes classification that uses the features extracted by each selection methods. The classification results of the proposed method outperformed the benchmarks and state of the art developed by Cui et al. (2017), Nguyen et al. (2017), Willer et al. (2016), and Adedoyin-Olowe et al. (2016) as illustrated in the table TABREF33 BIBREF12 , BIBREF38 , BIBREF39 , BIBREF40 , BIBREF41 , BIBREF42 .
The same experiment has been applied to Sydney, Brisbane and Perth in Australia on a time frame of 640 days with 500 days training data and 140 days testing data and the results were similar to Melbourne results with average AUC of 0.91 and average F-Score of 0.79. To ensure that the proposed method is language independent, we used the same method to classify civil unrest days in Jakarta using the Indonesian language, the classification scores were lower than the average scores for English language by 0.05 taking into consideration that we did not apply any NLP pre-processing to the Indonesian tweets such as stemming and lemmatization.
To verify the robustness of this feature selection method, we tested the selected features using multiple classifiers such as KNN, SVM, naive Bayes and decision trees. The results emphasized that the word-pairs selected using the spike-matching method achieve better AUC scores than the other correlation methods as listed in table TABREF19
## Conclusions
In this paper, we proposed a framework to detect civil unrest events by tracking each word-pair volume in twitter. The main challenge with this model is to identify the word-pairs that are highly associated with the events with predictive power. We used temporal filtering to detect the spike within the time series vector and used Jaccard similarity to calculate the scores of each word-pair according to its similarity with the binary vector of event days. These scores are used to rank the word-pairs as features for prediction.
Once the word-pairs are identified, we trained a Naive Bayes classifier to identify any day in a specific region to be an event or non-event days. We performed the experiment on both Melbourne and Sydney regions in Australia, and we achieved a classification accuracy of 87% with the precision of 77%, Recall of 82 %, area under the ROC curve of 91% and F-Score of 79%. The results are all achieved after 10-folds randomized cross-validation as listed in table TABREF32 .
The main contributions of this paper are (1) to overcome twitter challenges of acronyms, short text, ambiguity and synonyms, (2) to identify the set of word-pairs to be used as features for live event detection, (3) to build an end-to-end framework that can detect the events lively according to the word counts. This work can be applied to similar problems, where specific tweets can be associated with life events such as disease outbreak or stock market fluctuation. This work can be extended to predict future events with one day in advance, where we will use the same method for feature selection in addition to to time series analysis of the historical patterns of the word-pairs.
## Acknowledgments
This research was fully supported by the School of Mathematical Sciences at the University of Adelaide. All the data, computation and technical framework were supported by Data-To-Decision-Collaborative-Research-Center (D2DCRC).
| 13
|
1901.01010
|
A Joint Model for Multimodal Document Quality Assessment
|
# A Joint Model for Multimodal Document Quality Assessment
## Abstract
The quality of a document is affected by various factors, including grammaticality, readability, stylistics, and expertise depth, making the task of document quality assessment a complex one. In this paper, we explore this task in the context of assessing the quality of Wikipedia articles and academic papers. Observing that the visual rendering of a document can capture implicit quality indicators that are not present in the document text --- such as images, font choices, and visual layout --- we propose a joint model that combines the text content with a visual rendering of the document for document quality assessment. Experimental results over two datasets reveal that textual and visual features are complementary, achieving state-of-the-art results.
## Introduction
The task of document quality assessment is to automatically assess a document according to some predefined inventory of quality labels. This can take many forms, including essay scoring (quality = language quality, coherence, and relevance to a topic), job application filtering (quality = suitability for role + visual/presentational quality of the application), or answer selection in community question answering (quality = actionability + relevance of the answer to the question). In the case of this paper, we focus on document quality assessment in two contexts: Wikipedia document quality classification, and whether a paper submitted to a conference was accepted or not.
Automatic quality assessment has obvious benefits in terms of time savings and tractability in contexts where the volume of documents is large. In the case of dynamic documents (possibly with multiple authors), such as in the case of Wikipedia, it is particularly pertinent, as any edit potentially has implications for the quality label of that document (and around 10 English Wikipedia documents are edited per second). Furthermore, when the quality assessment task is decentralized (as in the case of Wikipedia and academic paper assessment), quality criteria are often applied inconsistently by different people, where an automatic document quality assessment system could potentially reduce inconsistencies and enable immediate author feedback.
Current studies on document quality assessment mainly focus on textual features. For example, BIBREF0 examine features such as the article length and the number of headings to predict the quality class of a Wikipedia article. In contrast to these studies, in this paper, we propose to combine text features with visual features, based on a visual rendering of the document. Figure 1 illustrates our intuition, relative to Wikipedia articles. Without being able to read the text, we can tell that the article in Figure 1 has higher quality than Figure 1 , as it has a detailed infobox, extensive references, and a variety of images. Based on this intuition, we aim to answer the following question: can we achieve better accuracy on document quality assessment by complementing textual features with visual features?
Our visual model is based on fine-tuning an Inception V3 model BIBREF1 over visual renderings of documents, while our textual model is based on a hierarchical biLSTM. We further combine the two into a joint model. We perform experiments on two datasets: a Wikipedia dataset novel to this paper, and an arXiv dataset provided by BIBREF2 split into three sub-parts based on subject category. Experimental results on the visual renderings of documents show that implicit quality indicators, such as images and visual layout, can be captured by an image classifier, at a level comparable to a text classifier. When we combine the two models, we achieve state-of-the-art results over 3/4 of our datasets.
This paper makes the following contributions:
All code and data associated with this research will be released on publication.
## Related Work
A variety of approaches have been proposed for document quality assessment across different domains: Wikipedia article quality assessment, academic paper rating, content quality assessment in community question answering (cQA), and essay scoring. Among these approaches, some use hand-crafted features while others use neural networks to learn features from documents. For each domain, we first briefly describe feature-based approaches and then review neural network-based approaches. Wikipedia article quality assessment: Quality assessment of Wikipedia articles is a task that assigns a quality class label to a given Wikipedia article, mirroring the quality assessment process that the Wikipedia community carries out manually. Many approaches have been proposed that use features from the article itself, meta-data features (e.g., the editors, and Wikipedia article revision history), or a combination of the two. Article-internal features capture information such as whether an article is properly organized, with supporting evidence, and with appropriate terminology. For example, BIBREF3 use writing styles represented by binarized character trigram features to identify featured articles. BIBREF4 and BIBREF0 explore the number of headings, images, and references in the article. BIBREF5 use nine readability scores, such as the percentage of difficult words in the document, to measure the quality of the article. Meta-data features, which are indirect indicators of article quality, are usually extracted from revision history, and the interaction between editors and articles. For example, one heuristic that has been proposed is that higher-quality articles have more edits BIBREF6 , BIBREF7 . BIBREF8 use the percentage of registered editors and the total number of editors of an article. Article–editor dependencies have also been explored. For example, BIBREF9 use the authority of editors to measure the quality of Wikipedia articles, where the authority of editors is determined by the articles they edit. Deep learning approaches to predicting Wikipedia article quality have also been proposed. For example, BIBREF10 use a version of doc2vec BIBREF11 to represent articles, and feed the document embeddings into a four hidden layer neural network. BIBREF12 first obtain sentence representations by averaging words within a sentence, and then apply a biLSTM BIBREF13 to learn a document-level representation, which is combined with hand-crafted features as side information. BIBREF14 exploit two stacked biLSTMs to learn document representations.
Academic paper rating: Academic paper rating is a relatively new task in NLP/AI, with the basic formulation being to automatically predict whether to accept or reject a paper. BIBREF2 explore hand-crafted features, such as the length of the title, whether specific words (such as outperform, state-of-the-art, and novel) appear in the abstract, and an embedded representation of the abstract as input to different downstream learners, such as logistic regression, decision tree, and random forest. BIBREF15 exploit a modularized hierarchical convolutional neural network (CNN), where each paper section is treated as a module. For each paper section, they train an attention-based CNN, and an attentive pooling layer is applied to the concatenated representation of each section, which is then fed into a softmax layer.
Content quality assessment in cQA: Automatic quality assessment in cQA is the task of determining whether an answer is of high quality, selected as the best answer, or ranked higher than other answers. To measure answer content quality in cQA, researchers have exploited various features from different sources, such as the answer content itself, the answerer's profile, interactions among users, and usage of the content. The most common feature used is the answer length BIBREF16 , BIBREF17 , with other features including: syntactic and semantic features, such as readability scores. BIBREF18 ; similarity between the question and the answer at lexical, syntactic, and semantic levels BIBREF18 , BIBREF19 , BIBREF20 ; or user data (e.g., a user's status points or the number of answers written by the user). There have also been approaches using neural networks. For example, BIBREF21 combine CNN-learned representations with hand-crafted features to predict answer quality. BIBREF22 use a 2-dimensional CNN to learn the semantic relevance of an answer to the question, and apply an LSTM to the answer sequence to model thread context. BIBREF23 and BIBREF24 model the problem similarly to machine translation quality estimation, treating answers as competing translation hypotheses and the question as the reference translation, and apply neural machine translation to the problem. Essay scoring: Automated essay scoring is the task of assigning a score to an essay, usually in the context of assessing the language ability of a language learner. The quality of an essay is affected by the following four primary dimensions: topic relevance, organization and coherence, word usage and sentence complexity, and grammar and mechanics. To measure whether an essay is relevant to its “prompt” (the description of the essay topic), lexical and semantic overlap is commonly used BIBREF25 , BIBREF26 . BIBREF27 explore word features, such as the number of verb formation errors, average word frequency, and average word length, to measure word usage and lexical complexity. BIBREF28 use sentence structure features to measure sentence variety. The effects of grammatical and mechanic errors on the quality of an essay are measured via word and part-of-speech $n$ -gram features and “mechanics” features BIBREF29 (e.g., spelling, capitalization, and punctuation), respectively. BIBREF30 , BIBREF31 , and BIBREF32 use an LSTM to obtain an essay representation, which is used as the basis for classification. Similarly, BIBREF33 utilize a CNN to obtain sentence representation and an LSTM to obtain essay representation, with an attention layer at both the sentence and essay levels.
## The Proposed Joint Model
We treat document quality assessment as a classification problem, i.e., given a document, we predict its quality class (e.g., whether an academic paper should be accepted or rejected). The proposed model is a joint model that integrates visual features learned through Inception V3 with textual features learned through a biLSTM. In this section, we present the details of the visual and textual embeddings, and finally describe how we combine the two. We return to discuss hyper-parameter settings and the experimental configuration in the Experiments section.
## Visual Embedding Learning
A wide range of models have been proposed to tackle the image classification task, such as VGG BIBREF34 , ResNet BIBREF35 , Inception V3 BIBREF1 , and Xception BIBREF36 . However, to the best of our knowledge, there is no existing work that has proposed to use visual renderings of documents to assess document quality. In this paper, we use Inception V3 pretrained on ImageNet (“Inception” hereafter) to obtain visual embeddings of documents, noting that any image classifier could be applied to our task. The input to Inception is a visual rendering (screenshot) of a document, and the output is a visual embedding, which we will later integrate with our textual embedding.
Based on the observation that it is difficult to decide what types of convolution to apply to each layer (such as 3 $\times $ 3 or 5 $\times $ 5), the basic Inception model applies multiple convolution filters in parallel and concatenates the resulting features, which are fed into the next layer. This has the benefit of capturing both local features through smaller convolutions and abstracted features through larger convolutions. Inception is a hybrid of multiple Inception models of different architectures. To reduce computational cost, Inception also modifies the basic model by applying a 1 $\times $ 1 convolution to the input and factorizing larger convolutions into smaller ones.
## Textual Embedding Learning
We adopt a bi-directional LSTM model to generate textual embeddings for document quality assessment, following the method of BIBREF12 (“biLSTM” hereafter). The input to biLSTM is a textual document, and the output is a textual embedding, which will later integrate with the visual embedding.
For biLSTM, each word is represented as a word embedding BIBREF37 , and an average-pooling layer is applied to the word embeddings to obtain the sentence embedding, which is fed into a bi-directional LSTM to generate the document embedding from the sentence embeddings. Then a max-pooling layer is applied to select the most salient features from the component sentences.
## The Joint Model
The proposed joint model (“Joint” hereafter) combines the visual and textual embeddings (output of Inception and biLSTM) via a simple feed-forward layer and softmax over the document label set, as shown in Figure 2 . We optimize our model based on cross-entropy loss.
## Experiments
In this section, we first describe the two datasets used in our experiments: (1) Wikipedia, and (2) arXiv. Then, we report the experimental details and results.
## Datasets
The Wikipedia dataset consists of articles from English Wikipedia, with quality class labels assigned by the Wikipedia community. Wikipedia articles are labelled with one of six quality classes, in descending order of quality: Featured Article (“FA”), Good Article (“GA”), B-class Article (“B”), C-class Article (“C”), Start Article (“Start”), and Stub Article (“Stub”). A description of the criteria associated with the different classes can be found in the Wikipedia grading scheme page. The quality class of a Wikipedia article is assigned by Wikipedia reviewers or any registered user, who can discuss through the article's talk page to reach consensus. We constructed the dataset by first crawling all articles from each quality class repository, e.g., we get FA articles by crawling pages from the FA repository: https://en.wikipedia.org/wiki/Category:Featured_articles. This resulted in around 5K FA, 28K GA, 212K B, 533K C, 2.6M Start, and 3.2M Stub articles.
We randomly sampled 5,000 articles from each quality class and removed all redirect pages, resulting in a dataset of 29,794 articles. As the wikitext contained in each document contains markup relating to the document category such as {Featured Article} or {geo-stub}, which reveals the label, we remove such information. We additionally randomly partitioned this dataset into training, development, and test splits based on a ratio of 8:1:1. Details of the dataset are summarized in Table 1 .
We generate a visual representation of each document via a 1,000 $\times $ 2,000-pixel screenshot of the article via a PhantomJS script over the rendered version of the article, ensuring that the screenshot and wikitext versions of the article are the same version. Any direct indicators of document quality (such as the FA indicator, which is a bronze star icon in the top right corner of the webpage) are removed from the screenshot.
The arXiv dataset BIBREF2 consists of three subsets of academic articles under the arXiv repository of Computer Science (cs), from the three subject areas of: Artificial Intelligence (cs.ai), Computation and Language (cs.cl), and Machine Learning (cs.lg). In line with the original dataset formulation BIBREF2 , a paper is considered to have been accepted (i.e. is positively labeled) if it matches a paper in the DBLP database or is otherwise accepted by any of the following conferences: ACL, EMNLP, NAACL, EACL, TACL, NIPS, ICML, ICLR, or AAAI. Failing this, it is considered to be rejected (noting that some of the papers may not have been submitted to one of these conferences). The median numbers of pages for papers in cs.ai, cs.cl, and cs.lg are 11, 10, and 12, respectively. To make sure each page in the PDF file has the same size in the screenshot, we crop the PDF file of a paper to the first 12; we pad the PDF file with blank pages if a PDF file has less than 12 pages, using the PyPDF2 Python package. We then use ImageMagick to convert the 12-page PDF file to a single 1,000 $\times $ 2,000 pixel screenshot. Table 2 details this dataset, where the “Accepted” column denotes the percentage of positive instances (accepted papers) in each subset.
## Experimental Setting
As discussed above, our model has two main components — biLSTM and Inception— which generate textual and visual representations, respectively. For the biLSTM component, the documents are preprocessed as described in BIBREF12 , where an article is divided into sentences and tokenized using NLTK BIBREF38 . Words appearing more than 20 times are retained when building the vocabulary. All other words are replaced by the special UNK token. We use the pre-trained GloVe BIBREF39 50-dimensional word embeddings to represent words. For words not in GloVe, word embeddings are randomly initialized based on sampling from a uniform distribution $U(-1, 1)$ . All word embeddings are updated in the training process. We set the LSTM hidden layer size to 256. The concatenation of the forward and backward LSTMs thus gives us 512 dimensions for the document embedding. A dropout layer is applied at the sentence and document level, respectively, with a probability of 0.5.
For Inception, we adopt data augmentation techniques in the training with a “nearest” filling mode, a zoom range of 0.1, a width shift range of 0.1, and a height shift range of 0.1. As the original screenshots have the size of 1,000 $\times 2$ ,000 pixels, they are resized to 500 $\times $ 500 to feed into Inception, where the input shape is (500, 500, 3). A dropout layer is applied with a probability of 0.5. Then, a GlobalAveragePooling2D layer is applied, which produces a 2,048 dimensional representation.
For the Joint model, we get a representation of 2,560 dimensions by concatenating the 512 dimensional representation from the biLSTM with the 2,048 dimensional representation from Inception. The dropout layer is applied to the two components with a probability of 0.5. For biLSTM, we use a mini-batch size of 128 and a learning rate of 0.001. For both Inception and joint model, we use a mini-batch size of 16 and a learning rate of 0.0001. All hyper-parameters were set empirically over the development data, and the models were optimized using the Adam optimizer BIBREF40 .
In the training phase, the weights in Inception are initialized by parameters pretrained on ImageNet, and the weights in biLSTM are randomly initialized (except for the word embeddings). We train each model for 50 epochs. However, to prevent overfitting, we adopt early stopping, where we stop training the model if the performance on the development set does not improve for 20 epochs. For evaluation, we use (micro-)accuracy, following previous studies BIBREF5 , BIBREF2 .
## Baseline Approaches
We compare our models against the following five baselines:
Majority: the model labels all test samples with the majority class of the training data.
Benchmark: a benchmark method from the literature. In the case of Wikipedia, this is BIBREF5 , who use structural features and readability scores as features to build a random forest classifier; for arXiv, this is BIBREF2 , who use hand-crafted features, such as the number of references and TF-IDF weighted bag-of-words in abstract, to build a classifier based on the best of logistic regression, multi-layer perception, and AdaBoost.
Doc2Vec: doc2vec BIBREF11 to learn document embeddings with a dimension of 500, and a 4-layer feed-forward classification model on top of this, with 2000, 1000, 500, and 200 dimensions, respectively.
biLSTM: first derive a sentence representation by averaging across words in a sentence, then feed the sentence representation into a biLSTM and a maxpooling layer over output sequence to learn a document level representation with a dimension of 512, which is used to predict document quality.
Inception $_{\text{fixed}}$ : the frozen Inception model, where only parameters in the last layer are fine-tuned during training.
The hyper-parameters of Benchmark, Doc2Vec, and biLSTM are based on the corresponding papers except that: (1) we fine-tune the feed forward layer of Doc2Vec on the development set and train the model 300 epochs on Wikipedia and 50 epochs on arXiv; (2) we do not use hand-crafted features for biLSTM as we want the baselines to be comparable to our models, and the main focus of this paper is not to explore the effects of hand-crafted features (e.g., see BIBREF12 ).
## Experimental Results
Table 3 shows the performance of the different models over our two datasets, in the form of the average accuracy on the test set (along with the standard deviation) over 10 runs, with different random initializations.
On Wikipedia, we observe that the performance of biLSTM, Inception, and Joint is much better than that of all four baselines. Inception achieves 2.9% higher accuracy than biLSTM. The performance of Joint achieves an accuracy of 59.4%, which is 5.3% higher than using textual features alone (biLSTM) and 2.4% higher than using visual features alone (Inception). Based on a one-tailed Wilcoxon signed-rank test, the performance of Joint is statistically significant ( $p<0.05$ ). This shows that the textual and visual features complement each other, achieving state-of-the-art results in combination.
For arXiv, baseline methods Majority, Benchmark, and Inception $_{\text{fixed}}$ outperform biLSTM over cs.ai, in large part because of the class imbalance in this dataset (90% of papers are rejected). Surprisingly, Inception $_{\text{fixed}}$ is better than Majority and Benchmark over the arXiv cs.lg subset, which verifies the usefulness of visual features, even when only the last layer is fine-tuned. Table 3 also shows that Inception and biLSTM achieve similar performance on arXiv, showing that textual and visual representations are equally discriminative: Inception and biLSTM are indistinguishable over cs.cl; biLSTM achieves 1.8% higher accuracy over cs.lg, while Inception achieves 1.3% higher accuracy over cs.ai. Once again, the Joint model achieves the highest accuracy on cs.ai and cs.cl by combining textual and visual representations (at a level of statistical significance for cs.ai). This, again, confirms that textual and visual features complement each other, and together they achieve state-of-the-art results. On arXiv cs.lg, Joint achieves a 0.6% higher accuracy than Inception by combining visual features and textual features, but biLSTM achieves the highest accuracy. One characteristic of cs.lg documents is that they tend to contain more equations than the other two arXiv datasets, and preliminary analysis suggests that the biLSTM is picking up on a correlation between the volume/style of mathematical presentation and the quality of the document.
## Analysis
In this section, we first analyze the performance of Inception and Joint. We also analyze the performance of different models on different quality classes. The high-level representations learned by different models are also visualized and discussed. As the Wikipedia test set is larger and more balanced than that of arXiv, our analysis will focus on Wikipedia.
## Inception
To better understand the performance of Inception, we generated the gradient-based class activation map BIBREF41 , by maximizing the outputs of each class in the penultimate layer, as shown in Figure 3 . From Figure 3 and Figure 3 , we can see that Inception identifies the two most important regions (one at the top corresponding to the table of contents, and the other at the bottom, capturing both document length and references) that contribute to the FA class prediction, and a region in the upper half of the image that contributes to the GA class prediction (capturing the length of the article body). From Figure 3 and Figure 3 , we can see that the most important regions in terms of B and C class prediction capture images (down the left and right of the page, in the case of B and C), and document length/references. From Figure 3 and Figure 3 , we can see that Inception finds that images in the top right corner are the strongest predictor of Start class prediction, and (the lack of) images/the link bar down the left side of the document are the most important for Stub class prediction.
## Joint
Table 4 shows the confusion matrix of Joint on Wikipedia. We can see that more than 50% of documents for each quality class are correctly classified, except for the C class where more documents are misclassified into B. Analysis shows that when misclassified, documents are usually misclassified into adjacent quality classes, which can be explained by the Wikipedia grading scheme, where the criteria for adjacent quality classes are more similar.
We also provide a breakdown of precision (“ $\mathcal {P}$ ”), recall (“ $\mathcal {R}$ ”), and F1 score (“ $\mathcal {F}_{\beta =1}$ ”) for biLSTM, Inception, and Joint across the quality classes in Table 5 . We can see that Joint achieves the highest accuracy in 11 out of 18 cases. It is also worth noting that all models achieve higher scores for FA, GA, and Stub articles than B, C and Start articles. This can be explained in part by the fact that FA and GA articles must pass an official review based on structured criteria, and in part by the fact that Stub articles are usually very short, which is discriminative for Inception, and Joint. All models perform worst on the B and C quality classes. It is difficult to differentiate B articles from C articles even for Wikipedia contributors. As evidence of this, when we crawled a new dataset including talk pages with quality class votes from Wikipedia contributors, we found that among articles with three or more quality labels, over 20% percent of B and C articles have inconsistent votes from Wikipedia contributors, whereas for FA and GA articles the number is only 0.7%.
We further visualize the learned document representations of biLSTM, Inception, and Joint in the form of a t-SNE plot BIBREF42 in Figure 4 . The degree of separation between Start and Stub achieved by Inception is much greater than for biLSTM, with the separation between Start and Stub achieved by Joint being the clearest among the three models. Inception and Joint are better than biLSTM at separating Start and C. Joint achieves slightly better performance than Inception in separating GA and FA. We can also see that it is difficult for all models to separate B and C, which is consistent with the findings of Tables 4 and 5 .
## Conclusions
We proposed to use visual renderings of documents to capture implicit document quality indicators, such as font choices, images, and visual layout, which are not captured in textual content. We applied neural network models to capture visual features given visual renderings of documents. Experimental results show that we achieve a 2.9% higher accuracy than state-of-the-art approaches based on textual features over Wikipedia, and performance competitive with or surpassing state-of-the-art approaches over arXiv. We further proposed a joint model, combining textual and visual representations, to predict the quality of a document. Experimental results show that our joint model outperforms the visual-only model in all cases, and the text-only model on Wikipedia and two subsets of arXiv. These results underline the feasibility of assessing document quality via visual features, and the complementarity of visual and textual document representations for quality assessment.
| 15
|
1901.02262
|
Multi-style Generative Reading Comprehension
|
# Multi-style Generative Reading Comprehension
## Abstract
This study tackles generative reading comprehension (RC), which consists of answering questions based on textual evidence and natural language generation (NLG). We propose a multi-style abstractive summarization model for question answering, called Masque. The proposed model has two key characteristics. First, unlike most studies on RC that have focused on extracting an answer span from the provided passages, our model instead focuses on generating a summary from the question and multiple passages. This serves to cover various answer styles required for real-world applications. Second, whereas previous studies built a specific model for each answer style because of the difficulty of acquiring one general model, our approach learns multi-style answers within a model to improve the NLG capability for all styles involved. This also enables our model to give an answer in the target style. Experiments show that our model achieves state-of-the-art performance on the Q&A task and the Q&A + NLG task of MS MARCO 2.1 and the summary task of NarrativeQA. We observe that the transfer of the style-independent NLG capability to the target style is the key to its success.
## Introduction
Question answering has been a long-standing research problem. Recently, reading comprehension (RC), a challenge to answer a question given textual evidence provided in a document set, has received much attention. Here, current mainstream studies have treated RC as a process of extracting an answer span from one passage BIBREF0 , BIBREF1 or multiple passages BIBREF2 , which is usually done by predicting the start and end positions of the answer BIBREF3 , BIBREF4 .
The demand for answering questions in natural language is increasing rapidly, and this has led to the development of smart devices such as Siri and Alexa. However, in comparison with answer span extraction, the natural language generation (NLG) ability for RC has been less studied. While datasets such as MS MARCO BIBREF5 have been proposed for providing abstractive answers in natural language, the state-of-the-art methods BIBREF6 , BIBREF7 are based on answer span extraction, even for the datasets. Generative models such as S-Net BIBREF8 suffer from a dearth of training data to cover open-domain questions.
Moreover, to satisfy various information needs, intelligent agents should be capable of answering one question in multiple styles, such as concise phrases that do not contain the context of the question and well-formed sentences that make sense even without the context of the question. These capabilities complement each other; however, the methods used in previous studies cannot utilize and control different answer styles within a model.
In this study, we propose a generative model, called Masque, for multi-passage RC. On the MS MARCO 2.1 dataset, Masque achieves state-of-the-art performance on the dataset's two tasks, Q&A and NLG, with different answer styles. The main contributions of this study are that our model enables the following two abilities.
## Problem Formulation
The task considered in this paper, is defined as:
Problem 1 Given a question with $J$ words $x^q = \lbrace x^q_1, \ldots , x^q_J\rbrace $ , a set of $K$ passages, where each $k$ -th passage is composed of $L$ words $x^{p_k} = \lbrace x^{p_k}_1, \ldots , x^{p_k}_{L}\rbrace $ , and an answer style $s$ , an RC system outputs an answer $y = \lbrace y_1, \ldots , y_T \rbrace $ conditioned on the style.
In short, for inference, given a set of 3-tuples $(x^q, \lbrace x^{p_k}\rbrace , s)$ , the system predicts $P(y)$ . The training data is a set of 6-tuples: $(x^q, \lbrace x^{p_k}\rbrace , s, y, a, \lbrace r^{p_k}\rbrace )$ , where $a$ is 1 if the question is answerable with the provided passages and 0 otherwise, and $r^{p_k}$ is 1 if the $k$ -th passage is required to formulate the answer and 0 otherwise.
## Proposed Model
Our proposed model, Masque, is based on multi-source abstractive summarization; the answer our model generates can be viewed as a summary from the question and multiple passages. It is also style-controllable; one model can generate the answer with the target style.
Masque directly models the conditional probability $p(y|x^q, \lbrace x^{p_k}\rbrace , s)$ . In addition to multi-style learning, it considers passage ranking and answer possibility classification together as multi-task learning in order to improve accuracy. Figure 2 shows the model architecture. It consists of the following modules.
1 The question-passages reader (§ "Question-Passages Reader" ) models interactions between the question and passages.
2 The passage ranker (§ "Passage Ranker" ) finds relevant passages to the question.
3 The answer possibility classifier (§ "Answer Possibility Classifier" ) identifies answerable questions.
4 The answer sentence decoder (§ "Answer Sentence Decoder" ) outputs a sequence of words conditioned on the style.
## Question-Passages Reader
Given a question and passages, the question-passages reader matches them so that the interactions among the question (passage) words conditioned on the passages (question) can be captured.
Let $x^q$ and $x^{p_k}$ represent one-hot vectors of words in the question and $k$ -th passage. First, this layer projects each of the one-hot vectors (of size $V$ ) into a $d_\mathrm {word}$ -dimensional continuous vector space with a pre-trained weight matrix $W^e \in \mathbb {R}^{d_\mathrm {word} \times V}$ such as GloVe BIBREF15 . Next, it uses contextualized word representations, ELMo BIBREF16 , which is a character-level two-layer bidirectional language model pre-trained on a large-scale corpus. ELMo representations allow our model to use morphological clues to form robust representations for out-of-vocabulary words unseen in training. Then, the concatenation of the word and contextualized embedding vectors is passed to a two-layer highway network BIBREF17 that is shared for the question and passages.
This layer uses a stack of Transformer blocks, which are shared for the question and passages, on top of the embeddings provided by the word embedding layer. The input of the first block is immediately mapped to a $d$ -dimensional vector by a linear transformation. The outputs of this layer are sequences of $d$ -dimensional vectors: $E^{p_k} \in \mathbb {R}^{d \times L}$ for the $k$ -th passage and $E^q \in \mathbb {R}^{d \times J}$ for the question.
It consists of two sub-layers: a self-attention layer and a position-wise feed-forward network. For the self-attention layer, we adopt the multi-head attention mechanism defined in BIBREF12 . The feed-forward network consists of two linear transformations with a GELU BIBREF18 activation in between, following OpenAI GPT BIBREF19 . Each sub-layer is placed inside a residual block BIBREF20 . For an input $x$ and a given sub-layer function $f$ , the output is $\mathrm {LayerNorm}(f(x)+x)$ , where $\mathrm {LayerNorm}$ indicates the layer normalization proposed in BIBREF21 . To facilitate these residual connections, all sub-layers produce outputs of dimension $d$ . Note that our model does not use any position embeddings because ELMo gives the positional information of the words in each sequence.
This layer fuses information from the passages to the question as well as from the question to the passages in a dual mechanism.
It first computes a similarity matrix $U^{p_k} \in \mathbb {R}^{L{\times }J}$ between the question and $k$ -th passage, as is done in BIBREF22 , where
$$U^{p_k}_{lj} = {w^a}^\top [ E^{p_k}_l; E^q_j; E^{p_k}_l \odot E^q_j ]$$ (Eq. 15)
indicates the similarity between the $l$ -th word of the $k$ -th passage and the $j$ -th question word. $w^a \in \mathbb {R}^{3d}$ are learnable parameters. The $\odot $ operator denotes the Hadamard product, and the $[;]$ operator means vector concatenation across the rows. Next, it obtains the row and column normalized similarity matrices $A^{p_k} = \mathrm {softmax}_j({U^{p_k}}^\top ) \in \mathbb {R}^{J\times L}$ and $B^{p_k} = \mathrm {softmax}_{l}(U^{p_k}) \in \mathbb {R}^{L \times J}$ . We use DCN BIBREF23 as the dual attention mechanism to obtain question-to-passage representations $G^{q \rightarrow p_k} \in \mathbb {R}^{5d \times L}$ :
$$\nonumber [E^{p_k}; \bar{A}^{p_k}; \bar{\bar{A}}^{p_k}; E^{p_k} \odot \bar{A}^{p_k}; E^{p_k} \odot \bar{\bar{A}}^{p_k}]$$ (Eq. 16)
and passage-to-question ones $G^{p \rightarrow q} \in \mathbb {R}^{5d \times J}$ :
$$\begin{split}
\nonumber & [ E^{q} ; \max _k(\bar{B}^{p_k}); \max _k(\bar{\bar{B}}^{p_k}); \\
&\hspace{10.0pt} E^{q} \odot \max _k(\bar{B}^{p_k}); E^{q} \odot \max _k(\bar{\bar{B}}^{p_k}) ] \mathrm {\ \ where}
\end{split}\\
\nonumber &\bar{A}^{p_k} = E^q A^{p_k}\in \mathbb {R}^{d \times L}, \ \bar{B}^{p_k} = E^{p_k} B^{p_k} \in \mathbb {R}^{d \times J} \\
\nonumber &\bar{\bar{A}}^{p_k} = \bar{B}^{p_k} A^{p_k} \in \mathbb {R}^{d \times L}, \ \bar{\bar{B}}^{p_k} = \bar{A}^{p_k} B^{p_k} \in \mathbb {R}^{d \times J}.$$ (Eq. 17)
This layer uses a stack of Transformer encoder blocks for question representations and obtains $M^q \in \mathbb {R}^{d \times J}$ from $G^{p \rightarrow q}$ . It also uses an another stack for passage representations and obtains $M^{p_k} \in \mathbb {R}^{d \times L}$ from $G^{q \rightarrow p_k}$ for each $k$ -th passage. The outputs of this layer, $M^q$ and $\lbrace M^{p_k}\rbrace $ , are passed on to the answer sentence decoder; the $\lbrace M^{p_k}\rbrace $ are also passed on to the passage ranker and answer possibility classifier.
## Passage Ranker
The passage ranker maps the output of the modeling layer, $\lbrace M^{p_k}\rbrace $ , to the relevance score of each passage. To obtain a fixed-dimensional pooled representation of each passage sequence, this layer takes the output for the first passage word, $M^{p_k}_1$ , which corresponds to the beginning-of-sentence token. It calculates the relevance of each $k$ -th passage to the question as:
$$\beta ^{p_k} = \mathrm {sigmoid}({w^r}^\top M^{p_k}_1),$$ (Eq. 20)
where $w^r \in \mathbb {R}^{d}$ are learnable parameters.
## Answer Possibility Classifier
The answer possibility classifier maps the output of the modeling layer, $\lbrace M^{p_k}\rbrace $ , to the probability of the answer possibility. The classifier takes the output for the first word, $M^{p_k}_1$ , for all passages and concatenates them to obtain a fixed-dimensional representation. It calculates the answer possibility to the question as:
$$P(a) = \mathrm {sigmoid}({w^c}^\top [M^{p_1}_1; \ldots ; M^{p_K}_1]),$$ (Eq. 22)
where $w^c \in \mathbb {R}^{Kd}$ are learnable parameters.
## Answer Sentence Decoder
Given the outputs provided by the reader, the decoder generates a sequence of answer words one element at a time. It is auto-regressive BIBREF24 , consuming the previously generated words as additional input at each decoding step.
Let $y = \lbrace y_1, \ldots , y_{T}\rbrace $ represent one-hot vectors of words in the answer. This layer has the same components as the word embedding layer of the question-passages reader, except that it uses a unidirectional ELMo in order to ensure that the predictions for position $t$ depend only on the known outputs at positions less than $t$ .
Moreover, to be able to make use of multiple answer styles within a single system, our model introduces an artificial token corresponding to the target style at the beginning of the answer sentence ( $y_1$ ), like BIBREF14 . At test time, the user can specify the first token to control the answer styles. This modification does not require any changes to the model architecture. Note that introducing the tokens on the decoder side prevents the passage ranker and answer possibility classifier from depending on the answer style.
This layer uses a stack of Transformer decoder blocks on top of the embeddings provided by the word embedding layer. The input is immediately mapped to a $d$ -dimensional vector by a linear transformation, and the output of this layer is a sequence of $d$ -dimensional vectors: $\lbrace s_1, \ldots , s_T\rbrace $ .
In addition to the encoder block, this block consists of second and third sub-layers after the self-attention block and before the feed-forward network, as shown in Figure 2 . As in BIBREF12 , the self-attention sub-layer uses a sub-sequent mask to prevent positions from attending to subsequent positions. The second and third sub-layers perform the multi-head attention over $M^q$ and $M^{p_\mathrm {all}}$ , respectively. The $M^{p_\mathrm {all}}$ is the concatenated outputs of the encoder stack for the passages,
$$M^{p_\mathrm {all}} = [M^{p_1}, \ldots , M^{p_K}] \in \mathbb {R}^{d \times KL}.$$ (Eq. 27)
The $[,]$ operator means vector concatenation across the columns. This attention for the concatenated passages enables our model to produce attention weights that are comparable between passages.
Our extended mechanism allows both words to be generated from a fixed vocabulary and words to be copied from both the question and multiple passages. Figure 3 shows the overview.
Let the extended vocabulary, $V_\mathrm {ext}$ , be the union of the common words (a small subset of the full vocabulary, $V$ , defined by the reader-side word embedding matrix) and all words appearing in the input question and passages. $P^v$ denotes the probability distribution of the $t$ -th answer word, $y_t$ , over the extended vocabulary. It is defined as:
$$P^v(y_t) =\mathrm {softmax}({W^2}^\top (W^1 s_t + b^1)),$$ (Eq. 31)
where the output embedding $W^2 \in \mathbb {R}^{d_\mathrm {word} \times V_\mathrm {ext}}$ is tied with the corresponding part of the input embedding BIBREF25 , and $W^1 \in \mathbb {R}^{d_\mathrm {word} \times d}$ and $b^1 \in \mathbb {R}^{d_\mathrm {word}}$ are learnable parameters. $P^v(y_t)$ is zero if $y_t$ is an out-of-vocabulary word for $V$ .
The copy mechanism used in the original pointer-generator is based on the attention weights of a single-layer attentional RNN decoder BIBREF9 . The attention weights in our decoder stack are the intermediate outputs in multi-head attentions and are not suitable for the copy mechanism. Therefore, our model also uses additive attentions for the question and multiple passages on top of the decoder stack.
The layer takes $s_t$ as the query and outputs $\alpha ^q_t \in \mathbb {R}^J$ ( $\alpha ^p_t \in \mathbb {R}^{KL}$ ) as the attention weights and $c^q_t \in \mathbb {R}^d$ ( $c^p_t \in \mathbb {R}^d$ ) as the context vectors for the question (passages):
$$e^q_j &= {w^q}^\top \tanh (W^{qm} M_j^q + W^{qs} s_t +b^q), \\
\alpha ^q_t &= \mathrm {softmax}(e^q), \\
c^q_t &= \textstyle \sum _j \alpha ^q_{tj} M_j^q, \\
e^{p_k}_l &= {w^p}^\top \tanh (W^{pm} M_l^{p_k} + W^{ps} s_t +b^p), \\
\alpha ^p_t &= \mathrm {softmax}([e^{p_1}; \ldots ; e^{p_K}]), \\
c^p_t &= \textstyle \sum _{l} \alpha ^p_{tl} M^{p_\mathrm {all}}_{l},$$ (Eq. 33)
where $w^q$ , $w^p \in \mathbb {R}^d$ , $W^{qm}$ , $W^{qs}$ , $W^{pm}$ , $W^{ps} \in \mathbb {R}^{d \times d}$ , and $b^q$ , $b^p \in \mathbb {R}^d$ are learnable parameters.
$P^q$ and $P^p$ are the copy distributions over the extended vocabulary, defined as:
$$P^q(y_t) &= \textstyle \sum _{j: x^q_j = y_t} \alpha ^q_{tj}, \\
P^p(y_t) &= \textstyle \sum _{l: x^{p_{k(l)}}_{l} = y_t} \alpha ^p_{tl},$$ (Eq. 34)
where $k(l)$ means the passage index corresponding to the $l$ -th word in the concatenated passages.
The final distribution of the $t$ -th answer word, $y_t$ , is defined as a mixture of the three distributions:
$$P(y_t) = \lambda ^v P^v(y_t) + \lambda ^q P^q(y_t) + \lambda ^p P^p(y_t),$$ (Eq. 36)
where the mixture weights are given by
$$\lambda ^v, \lambda ^q, \lambda ^p = \mathrm {softmax}(W^m [s_t; c^q_t; c^p_t] + b^m).$$ (Eq. 37)
$W^m \in \mathbb {R}^{3 \times 3d}$ , $b^m \in \mathbb {R}^3$ are learnable parameters.
In order not to use words in irrelevant passages, our model introduces the concept of combined attention BIBREF26 . While the original technique combines the word and sentence level attentions, our model combines the passage-level relevance $\beta ^{p_k}$ and word-level attentions $\alpha ^p_t$ by using simple scalar multiplication and re-normalization. The updated word attention is:
$$\alpha ^p_{tl} & := \frac{\alpha ^p_{tl} \beta ^{p_{k(l)} }}{\sum _{l^{\prime }} \alpha ^p_{tl^{\prime }} \beta ^{p_{k(l^{\prime })}}}.$$ (Eq. 39)
## Loss Function
We define the training loss as the sum of losses in
$$L(\theta ) = L_\mathrm {dec} + \gamma _\mathrm {rank} L_\mathrm {rank} + \gamma _\mathrm {cls} L_\mathrm {cls}$$ (Eq. 41)
where $\theta $ is the set of all learnable parameters, and $\gamma _\mathrm {rank}$ and $\gamma _\mathrm {cls}$ are balancing parameters.
The loss of the decoder, $L_\mathrm {dec}$ , is the negative log likelihood of the whole target answer sentence averaged over $N_\mathrm {able}$ answerable examples:
$$L_\mathrm {dec} = - \frac{1}{N_\mathrm {able}}\sum _{(a,y)\in \mathcal {D}} \frac{a}{T} \sum _t \log P(y_{t}),$$ (Eq. 42)
where $\mathcal {D}$ is the training dataset.
The losses of the passage ranker, $L_\mathrm {rank}$ , and the answer possibility classifier, $L_\mathrm {cls}$ , are the binary cross entropy between the true and predicted values averaged over all $N$ examples:
$$L_\mathrm {rank} = - \frac{1}{NK} \sum _k \sum _{r^{p_k}\in \mathcal {D}}
\biggl (
\begin{split}
&r^{p_k} \log \beta ^{p_k} + \\
&(1-r^{p_k}) \log (1-\beta ^{p_k})
\end{split}
\biggr ),\\
L_\mathrm {cls} = - \frac{1}{N} \sum _{a \in \mathcal {D}}
\biggl (
\begin{split}
&a \log P(a) + \\
&(1-a) \log (1-P(a))
\end{split}
\biggr ).$$ (Eq. 43)
## Setup
We conducted experiments on the two tasks of MS MARCO 2.1 BIBREF5 . The answer styles considered in the experiments corresponded to the two tasks. The NLG task requires a well-formed answer that is an abstractive summary of the question and ten passages, averaging 16.6 words. The Q&A task also requires an abstractive answer but prefers a more concise answer than the NLG task, averaging 13.1 words, where many of the answers do not contain the context of the question. For instance, for the question “tablespoon in cup”, the answer in the Q&A task will be “16”, and the answer in the NLG task will be “There are 16 tablespoons in a cup.” In addition to the ALL dataset, we prepared two subsets (Table 1 ). The ANS set consists of answerable questions, and the WFA set consists of the answerable questions and well-formed answers, where WFA $\subset $ ANS $\subset $ ALL.
We trained our model on a machine with eight NVIDIA P100 GPUs. Our model was jointly trained with the two answer styles in the ALL set for a total of eight epochs with a batch size of 80. The training took roughly six days. The ensemble model consists of six training runs with the identical architecture and hyperparameters. The hidden size $d$ was 304, and the number of attention heads was 8. The inner state size of the feed-forward networks was 256. The numbers of shared encoding blocks, modeling blocks for question, modeling blocks for passages, and decoder blocks were 3, 2, 5, and 8, respectively. We used the pre-trained uncased 300-dimensional GloVe BIBREF15 and the original 512-dimensional ELMo BIBREF16 . We used the spaCy tokenizer, and all words were lowercased except the input for ELMo. The number of common words in $V_\mathrm {ext}$ was 5,000.
We used the Adam optimization BIBREF27 with $\beta _1 = 0.9$ , $\beta _2 = 0.999$ , and $\epsilon = 10^{-8}$ . Weights were initialized using $N(0, 0.02)$ , except that the biases of all the linear transformations were initialized with zero vectors. The learning rate was increased linearly from zero to $2.5 \times 10^{-4}$ in the first 2,000 steps and annealed to 0 using a cosine schedule. All parameter gradients were clipped to a maximum norm of 1. An exponential moving average was applied to all trainable variables with a decay rate 0.9995. The balancing factors of joint learning, $\lambda _\mathrm {rank}$ and $\lambda _\mathrm {cls}$ , were set to 0.5 and 0.1.
We used a modified version of the L $_2$ regularization proposed in BIBREF28 , with $w = 0.01$ . We additionally used a dropout BIBREF29 rate of 0.3 for all highway networks and residual and scaled dot-product attention operations in the multi-head attention mechanism. We also used one-sided label smoothing BIBREF30 for the passage relevance and answer possibility labels. We smoothed only the positive labels to 0.9.
## Results
Table 2 shows that our ensemble model, controlled with the NLG and Q&A styles, achieved state-of-the-art performance on the NLG and Q&A tasks in terms of Rouge-L. In particular, for the NLG task, our single model outperformed competing models in terms of both Rouge-L and Bleu-1. The capability of creating abstractive summaries from the question and passages contributed to its improvements over the state-of-the-art extractive approaches BIBREF6 , BIBREF7 .
Table 3 shows the results of the ablation test for our model (controlled with the NLG style) on the well-formed answers of the WFA dev. set. Our model, which was trained with the ALL set consisting of the two styles, outperformed the model trained with the WFA set consisting of the single style. Multi-style learning allowed our model to improve NLG performance by also using non-sentence answers.
Table 3 shows that our model outperformed the model that used RNNs and self-attentions instead of Transformer blocks as in MCAN BIBREF11 . Our deep Transformer decoder captured the interaction among the question, the passages, and the answer better than a single-layer LSTM decoder.
Table 3 shows that our model (jointly trained with the passage ranker and answer possibility classifier) outperformed the model that did not use the ranker and classifier. The joint learning has a regularization effect on the question-passages reader.
We also confirmed that the gold passage ranker, which can predict passage relevances perfectly, improves RC performance significantly. Passage re-ranking will be a key to developing a system that can outperform humans.
Table 4 shows the passage re-ranking performance for the ten given passages on the ANS dev. set. Our ranker improved the initial ranking provided by Bing by a significant margin. Also, the ranker shares the question-passages reader with the answer decoder, and this sharing contributed to the improvements over the ranker trained without the answer decoder. This result is similar to those reported in BIBREF33 . Moreover, the joint learning with the answer possibility classifier and multiple answer styles, which enables our model to learn from a larger number of data, improved the re-ranking.
Figure 4 shows the precision-recall curve of answer possibility classification on the ALL dev. set, where the positive class is the answerable data. Our model identified the answerable questions well. The maximum $F_1$ score was 0.7893. This is the first report on answer possibility classification with MS MARCO 2.1.
Figure 5 shows the lengths of the answers generated by our model, which are broken down by answer style and query type. The generated answers were relatively shorter than the reference answers but well controlled with the target style in every query type.
Also, we should note that our model does not guarantee the consistency in terms of meaning across the answer styles. We randomly selected 100 questions and compared the answers our model generated with the NLG and Q&A styles. The consistency ratio was 0.81, where major errors were due to copying words from different parts of the passages and generating different words, especially yes/no, from a fixed vocabulary.
Appendix "Reading Comprehension Examples generated by Masque from MS MARCO 2.1" shows examples of generated answers. We found (d) style errors; (e) yes/no classification errors; (f) copy errors with respect to numerical values; and (c,e) grammatical errors that were originally contained in the inputs.
## Conclusion
We believe our study makes two contributions to the study of multi-passage RC with NLG. Our model enables 1) multi-source abstractive summarization based RC and 2) style-controllable RC. The key strength of our model is its high accuracy of generating abstractive summaries from the question and passages; our model achieved state-of-the-art performance in terms of Rouge-L on the Q&A and NLG tasks of MS MARCO 2.1 that have different answer styles BIBREF5 .
The styles considered in this paper are only related to the context of the question in the answer sentence; our model will be promising for controlling other styles such as length and speaking styles. Future work will involve exploring the potential of hybrid models combining extractive and abstractive approaches and improving the passage re-ranking and answerable question identification.
| 11
|
1901.03438
|
Grammatical Analysis of Pretrained Sentence Encoders with Acceptability Judgments
|
# Grammatical Analysis of Pretrained Sentence Encoders with Acceptability Judgments
## Abstract
Recent pretrained sentence encoders achieve state of the art results on language understanding tasks, but does this mean they have implicit knowledge of syntactic structures? We introduce a grammatically annotated development set for the Corpus of Linguistic Acceptability (CoLA; Warstadt et al., 2018), which we use to investigate the grammatical knowledge of three pretrained encoders, including the popular OpenAI Transformer (Radford et al., 2018) and BERT (Devlin et al., 2018). We fine-tune these encoders to do acceptability classification over CoLA and compare the models' performance on the annotated analysis set. Some phenomena, e.g. modification by adjuncts, are easy to learn for all models, while others, e.g. long-distance movement, are learned effectively only by models with strong overall performance, and others still, e.g. morphological agreement, are hardly learned by any model.
## Introduction
The effectiveness and ubiquity of pretrained sentence embeddings for natural language understanding has grown dramatically in recent years. Recent sentence encoders like OpenAI's Generative Pretrained Transformer BIBREF3 and BERT BIBREF2 achieve the state of the art on the GLUE benchmark BIBREF4 . Among the GLUE tasks, these state-of-the-art systems make their greatest gains on the acceptability task with the Corpus of Linguistic Acceptability BIBREF0 . CoLA contains example sentences from linguistics publications labeled by experts for grammatical acceptability, and written to show subtle grammatical features. Because minimal syntactic differences can separate acceptable sentences from unacceptable ones (What did Bo write a book about? / *What was a book about written by Bo?), and acceptability classifiers are more reliable when trained on GPT and BERT than on recurrent models, it stands to reason that GPT and BERT have better implicit knowledge of syntactic features relevant to acceptability.
Our goal in this paper is to develop an evaluation dataset that can locate which syntactic features that a model successfully learns by identifying the syntactic domains of CoLA in which it performs the best. Using this evaluation set, we compare the syntactic knowledge of GPT and BERT in detail, and investigate the strengths of these models over the baseline BiLSTM model published by warstadt2018neural. The analysis set includes expert annotations labeling the entire CoLA development set for the presence of 63 fine-grained syntactic features.
We identify many specific syntactic features that make sentences harder to classify, and many that have little effect. For instance, sentences involving unusual or marked argument structures are no harder than the average sentence, while sentences with long distance dependencies are hard to learn. We also find features of sentences that accentuate or minimize the differences between models. Specifically, the transformer models seem to learn long-distance dependencies much better than the recurrent model, yet have no advantage on sentences with morphological violations.
## Analysis Set
We introduce a grammatically annotated version of the entire CoLA development set to facilitate detailed error analysis of acceptability classifiers. These 1043 sentences are expert-labeled for the presence of 63 minor grammatical features organized into 15 major features. Each minor feature belongs to a single major feature. A sentence belongs to a major feature if it belongs to one or more of the relevant minor features. The Appendix includes descriptions of each feature along with examples and the criteria used for annotation.
The 63 minor features and 15 major features are illustrated in Table TABREF5 . Considering minor features, an average of 4.31 features is present per sentence (SD=2.59). The average feature is present in 71.3 sentences (SD=54.7). Turning to major features, the average sentence belongs to 3.22 major features (SD=1.66), and the average major feature is present in 224 sentences (SD=112). Every sentence is labeled with at least one feature.
## Annotation
The sentences were annotated manually by one of the authors, who is a PhD student with extensive training in formal linguistics. The features were developed in a trial stage, in which the annotator performed a similar annotation with different annotation schema for several hundred sentences from CoLA not belonging to the development set.
## Feature Descriptions
Here we briefly summarize the feature set in order of the major features. Many of these constructions are well-studied in syntax, and further background can be found in textbooks such as adger2003core and sportiche2013introduction.
This major feature contains only one minor feature, simple, including sentences with a syntactically simplex subject and predicate.
These three features correspond to predicative phrases, including copular constructions, small clauses (I saw Bo jump), and resultatives/depictives (Bo wiped the table clean).
These six features mark various kinds of optional modifiers. This includes modifiers of NPs (The boy with blue eyes gasped) or VPs (The cat meowed all morning), and temporal (Bo swam yesterday) or locative (Bo jumped on the bed).
These five features identify syntactically selected arguments, differentiating, for example, obliques (I gave a book to Bo), PP arguments of NPs and VPs (Bo voted for Jones), and expletives (It seems that Bo left).
These four features mark VPs with unusual argument structures, including added arguments (I baked Bo a cake) or dropped arguments (Bo knows), and the passive (I was applauded).
This contains only one feature for imperative clauses (Stop it!).
These are two minor features, one for bound reflexives (Bo loves himself), and one for other bound pronouns (Bo thinks he won).
These five features apply to sentences with question-like properties. They mark whether the interrogative is an embedded clause (I know who you are), a matrix clause (Who are you?), or a relative clause (Bo saw the guy who left); whether it contains an island out of which extraction is unacceptable (*What was a picture of hanging on the wall?); or whether there is pied-piping or a multi-word wh-expressions (With whom did you eat?).
These six features apply to various complement clauses (CPs), including subject CPs (That Bo won is odd); CP arguments of VPs or NPs/APs (The fact that Bo won); CPs missing a complementizer (I think Bo's crazy); or non-finite CPs (This is ready for you to eat).
These four minor features mark the presence of auxiliary or modal verbs (I can win), negation, or “pseudo-auxiliaries” (I have to win).
These five features mark various infinitival embedded VPs, including control VPs (Bo wants to win); raising VPs (Bo seemed to fly); VP arguments of NPs or APs (Bo is eager to eat); and VPs with extraction (e.g. This is easy to read ts ).
These seven features mark complex NPs and APs, including ones with PP arguments (Bo is fond of Mo), or CP/VP arguments; noun-noun compounds (Bo ate mud pie); modified NPs, and NPs derived from verbs (Baking is fun).
These seven features mark various unrelated syntactic constructions, including dislocated phrases (The boy left who was here earlier); movement related to focus or information structure (This I've gotta see this); coordination, subordinate clauses, and ellipsis (I can't); or sentence-level adjuncts (Apparently, it's raining).
These four features mark various determiners, including quantifiers, partitives (two of the boys), negative polarity items (I *do/don't have any pie), and comparative constructions.
These three features apply only to unacceptable sentences, and only ones which are ungrammatical due to a semantic or morphological violation, or the presence or absence of a single salient word.
## Correlations
We wish to emphasize that these features are overlapping and in many cases are correlated, thus not all results from using this analysis set will be independent. We analyzed the pairwise Matthews Correlation Coefficient BIBREF17 of the 63 minor features (giving 1953 pairs), and of the 15 major features (giving 105 pairs). MCC is a special case of Pearson's INLINEFORM0 for Boolean variables. These results are summarized in Table TABREF25 . Regarding the minor features, 60 pairs had a correlation of 0.2 or greater, 17 had a correlation of 0.4 or greater, and 6 had a correlation of 0.6 or greater. None had an anti-correlation of greater magnitude than -0.17. Turning to the major features, 6 pairs had a correlation of 0.2 or greater, and 2 had an anti-correlation of greater magnitude than -0.2.
We can see at least three reasons for these observed correlations. First, some correlations can be attributed to overlapping feature definitions. For instance, expletive arguments (e.g. There are birds singing) are, by definition, non-canonical arguments, and thus are a subset of add arg. However, some added arguments, such as benefactives (Bo baked Mo a cake), are not expletives. Second, some correlations can be attributed to grammatical properties of the relevant constructions. For instance, question and aux are correlated because main-clause questions in English require subject-aux inversion and in many cases the insertion of auxiliary do (Do lions meow?). Third, some correlations may be a consequence of the sources sampled in CoLA and the phenomena they focus on. For instance, the unusually high correlation of Emb-Q and ellipsis/anaphor can be attributed to BIBREF18 , which is an article about the sluicing construction involving ellipsis of an embedded interrogative (e.g. I saw someone, but I don't know who).
Finally, two strongest anti-correlations between major features are between simple and the two features related to argument structure, argument types and arg altern. This follows from the definition of simple, which excludes any sentence containing a large number or unusual configuration of arguments.
## Models Evaluated
We train MLP acceptability classifiers for CoLA on top of three sentence encoders: (1) the CoLA baseline encoder with ELMo-style embeddings, (2) OpenAI GPT, and (3) BERT. We use publicly available sentence encoders with pretrained weights.
## Overall CoLA Results
The overall performance of the three sentence encoders is shown in Table TABREF33 . Performance on CoLA is measured using MCC BIBREF14 . We present the best single restart for each encoder, the mean over restarts for an encoder, and the result of ensembling the restarts for a given encoder, i.e. taking the majority classification for a given sentence, or the majority label of acceptable if tied. For BERT results, we exclude 5 out of the 20 restarts because they were degenerate (MCC=0).
Across the board, BERT outperforms GPT, which outperforms the CoLA baseline. However, BERT and GPT are much closer in performance than they are to CoLA baseline. While ensemble performance exceeded the average for BERT and GPT, it did not outperform the best single model.
## Analysis Set Results
The results for the major features and minor features are shown in Figures FIGREF26 and FIGREF35 , respectively. For each feature, we measure the MCC of the sentences including that feature. We plot the mean of these results across the different restarts for each model, and error bars mark the mean INLINEFORM0 standard deviation. For the Violations features, MCC is technically undefined because these features only contain unacceptable sentences. We report MCC in these cases by including for each feature a single acceptable example that is correctly classified by all models.
Comparison across features reveals that the presence of certain features has a large effect on performance, and we comment on some overall patterns below. Within a given feature, the effect of model type is overwhelmingly stable, and resembles the overall difference in performance. However, we observe several interactions, i.e. specific features where the relative performance of models does not track their overall relative performance.
Among the major features (Figure FIGREF26 ), performance is universally highest on the simple sentences, and is higher than each model's overall performance. Though these sentences are simple, we notice that the proportion of ungrammatical ones is on par with the entire dataset. Otherwise we find that a model's performance on sentences of a given feature is on par with or lower than its overall performance, reflecting the fact that features mark the presence of unusual or complex syntactic structure.
Performance is also high (and close to overall performance) on sentences with marked argument structures (Argument Types and Arg(ument) Alt(ernation)). While these models are still worse than human (overall) performance on these sentences, this result indicates that argument structure is relatively easy to learn.
Comparing different kinds of embedded content, we observe higher performance on sentences with embedded clauses (major feature=Comp Clause) embedded VPs (major feature=to-VP) than on sentences with embedded interrogatives (minor features=Emb-Q, Rel Clause). An exception to this trend is the minor feature No C-izer, which labels complement clauses without a complementizer (e.g. I think that you're crazy). Low performance on these sentences compared to most other features in Comp Clause might indicate that complementizers are an important syntactic cue for these models.
As the major feature Question shows, the difficulty of sentences with question-like syntax applies beyond just embedded questions. Excluding polar questions, sentences with question-like syntax almost always involve extraction of a wh-word, creating a long-distance dependency between the wh-word and its extraction site, which may be difficult for models to recognize.
The most challenging features are all related to Violations. Low performance on Infl/Agr Violations, which marks morphological violations (He washed yourself, This is happy), is especially striking because a relatively high proportion (29%) of these sentences are Simple. These models are likely to be deficient in encoding morphological features is that they are word level models, and do not have direct access sub-word information like inflectional endings, which indicates that these features are difficult to learn effectively purely from lexical distributions.
Finally, unusual performance on some features is due to small samples, and have a high standard deviation, suggesting the result is unreliable. This includes CP Subj, Frag/Paren, imperative, NPI/FCI, and Comparative.
Comparing within-feature performance of the three encoders to their overall performance, we find they have differing strengths and weaknesses. BERT stands out over other models in Deep Embed, which includes challenging sentences with doubly-embedded, as well as in several features involving extraction (i.e. long-distance dependencies) such as VP+Extract and Info-Struc. The transformer models show evidence of learning long-distance dependencies better than the CoLA baseline. They outperform the CoLA baseline by an especially wide margin on Bind:Refl, which all involves establishing a dependency between a reflexive and its antecedent (Bo tries to love himself). They also have a large advantage in dislocation, in which expressions are separated from their dependents (Bo practiced on the train an important presentation). The advantage of BERT and GPT may be due in part to their use of the transformer architecture. Unlike the BiLSTM used by the CoLA baseline, the transformer uses a self-attention mechanism that associates all pairs of words regardless of distance.
In some cases models showed surprisingly good or bad performance, revealing possible idiosyncrasies of the sentence embeddings they output. For instance, the CoLA baseline performs on par with the others on the major feature adjunct, especially considering the minor feature Particle (Bo looked the word up).
Furthermore, all models struggle equally with sentences in Violation, indicating that the advantages of the transformer models over the CoLA baseline does not extend to the detection of morphological violations (Infl/Agr Violation) or single word anomalies (Extra/Missing Expr).
## Length Analysis
For comparison, we analyze the effect of sentence length on acceptability classifier performance. The results are shown in Figure FIGREF39 . The results for the CoLA baseline are inconsistent, but do drop off as sentence length increases. For BERT and GPT, performance decreases very steadily with length. Exceptions are extremely short sentences (length 1-3), which may be challenging due to insufficient information; and extremely long sentences, where we see a small (but somewhat unreliable) boost in BERT's performance. BERT and GPT are generally quite close in performance, except on the longest sentences, where BERT's performance is considerably better.
## Conclusion
Using a new grammatically annotated analysis set, we identify several syntactic phenomena that are predictive of good or bad performance of current state of the art sentence encoders on CoLA. We also use these results to develop hypotheses about why BERT is successful, and why transformer models outperform sequence models.
Our findings can guide future work on sentence embeddings. A current weakness of all sentence encoders we investigate, including BERT, is the identification of morphological violations. Future engineering work should investigate whether switching to a character-level model can mitigate this problem. Additionally, transformer models appear to have an advantage over sequence models with long-distance dependencies, but still struggle with these constructions relative to more local phenomena. It stands to reason that this performance gap might be widened by training larger or deeper transformer models, or training on longer or more complex sentences. This analysis set can be used by engineers interested in evaluating the syntactic knowledge of their encoders.
Finally, these findings suggest possible controlled experiments that could confirm whether there is a causal relation between the presence of the syntactic features we single out as interesting and model performance. Our results are purely correlational, and do not mark whether a particular construction is crucial for the acceptability of the sentence. Future experiments following ettinger2018assessing and kann2019verb can semi-automatically generate datasets manipulating, for example, length of long-distance dependencies, inflectional violations, or the presence of interrogatives, while controlling for factors like sentence length and word choice, in order determine the extent to which these features impact the quality of sentence embeddings.
## Acknowledgments
We would like to thank Jason Phang and Thibault Févry for sharing GPT and BERT model predictions on CoLA, and Alex Wang for feedback.
## Simple
These are sentences with transitive or intransitive verbs appearing with their default syntax and argument structure. All arguments are noun phrases (DPs), and there are no modifiers or adjuncts on DPs or the VP.
. Included J̇ohn owns the book. (37) Park Square has a festive air. (131) *Herself likes Mary's mother. (456)
. Excluded Ḃill has eaten cake. I gave Joe a book.
## Pred (Predicates)
These are sentences including the verb be used predicatively. Also, sentences where the object of the verb is itself a predicate, which applies to the subject. Not included are auxiliary uses of be or other predicate phrases that are not linked to a subject by a verb.
. Included J̇ohn is eager. (27) He turned into a frog. (150) To please John is easy. (315)
. Excluded Ṫhere is a bench to sit on. (309) John broke the geode open. The cake was eaten.
These sentences involve predication of a non-subject argument by another non-subject argument, without the presence of a copula. Some of these cases may be analyzed as small clauses. BIBREF35
. Included J̇ohn called the president a fool. (234) John considers himself proud of Mary. (464) They want them arrested. (856) the election of John president surprised me. (1001)
Modifiers that act as predicates of an argument. Resultatives express a resulting state of that argument, and depictives describe that argument during the matrix event. See BIBREF24 .
. Included Ṙesultative Ṭhe table was wiped by John clean. (625) The horse kicked me black and blue. (898) . Depictive J̇ohn left singing. (971) In which car was the man seen? (398)
. Excluded Ḣe turned into a frog. (150)
## Adjunct
Particles are lone prepositions associated with verbs. When they appear with transitive verbs they may immediately follow the verb or the object. Verb-particle pairs may have a non-compositional (idiomatic) meaning. See [pp. 69-70]carnie2013syntax and [pp. 16-17]kim2008syntax.
. Included Ṭhe argument was summed by the coach up. (615) Some sentences go on and on and on. (785) *He let the cats which were whining out. (71)
Adjuncts modifying verb phrases. Adjuncts are (usually) optional, and they do not change the category of the expression they modify. See BIBREF33 .
. Included ṖP-adjuncts, e.g. locative, temporal, instrumental, beneficiary Ṅobody who hates to eat anything should work in a delicatessen. (121) Felicia kicked the ball off the bench. (127) . Adverbs Ṁary beautifully plays the violin. (40) John often meets Mary. (65) . Purpose VPs Ẇe need another run to win. (769) .
0.5em. Excluded ṖP arguments Ṣue gave to Bill a book. (42) Everything you like is on the table. (736) . S-adjuncts J̇ohn lost the race, unfortunately.
These are adjuncts modifying noun phrases. Adjuncts are (usually) optional, and they do not change the category of the expression they modify. Single-word prenominal adjectives are excluded, as are relative clauses (this has another category). . Included ṖP-adjuncts Ṭom's dog with one eye attacked Frank's with three legs. (676) They were going to meet sometime on Sunday, but the faculty didn't know when. (565) . Phrasal adjectives Ȧs a statesman, scarcely could he do anything worth mentioning. (292) . Verbal modifiers Ṫhe horse raced past the barn fell. (900)
. Excluded Ṗrenominal Adjectives İt was the policeman met that several young students in the park last night. (227) . Relative Clauses NP arguments
These are adjuncts of VPs and NPs that specify a time or modify tense or aspect or frequency of an event. Adjuncts are (usually) optional, and they do not change the category of the expression they modify. . Included Ṡhort adverbials (never, today, now, always) Ẉhich hat did Mike quip that she never wore? (95) . PPs Ḟiona might be here by 5 o'clock. (426) . When İ inquired when could we leave. (520)
These are adjuncts of VPs and NPs that specify a location of an event or a part of an event, or of an individual. Adjuncts are (usually) optional, and they do not change the category of the expression they modify. . Included Ṡhort adverbials PPs Ṫhe bed was slept in. (298) *Anson demonized up the Khyber (479) Some people consider dogs in my neighborhood dangerous. (802) Mary saw the boy walking toward the railroad station. (73) . Where İ found the place where we can relax. (307)
. Excluded Ŀocative arguments Ṣam gave the ball out of the basket. (129) Jessica loaded boxes on the wagon. (164) I went to Rome.
These are adjuncts of VPs and NPs not described by some other category (with the exception of (6-7)), i.e. not temporal, locative, or relative clauses. Adjuncts are (usually) optional, and they do not change the category of the expression they modify.
. Included Ḃeneficiary Ị know which book José didn't read for class, and which book Lilly did it for him. (58) . Instrument Ŀee saw the student with a telescope. (770) . Comitative J̇oan ate dinner with someone but I don't know who. (544) . VP adjuncts Ẇhich article did Terry file papers without reading? (431) . Purpose Ẇe need another run to win. (769)
## Argument Types
Oblique arguments of verbs are individual-denoting arguments (DPs or PPs) which act as the third argument of verb, i.e. not a subject or (direct) object. They may or may not be marked by a preposition. Obliques are only found in VPs that have three or more individual arguments. Arguments are selected for by the verb, and they are (generally) not optional, though in some cases they may be omitted where they are understood or implicitly existentially quantified over. See [p.40]kim2008syntax.
. Included Ṗrepositional Ṣue gave to Bill a book. (42) Mary has always preferred lemons to limes. (70) *Janet broke Bill on the finger. (141) . Benefactives Ṁartha carved the baby a toy out of wood. (139) . Double object Ṡusan told her a story. (875) Locative arguments Ȧnn may spend her vacation in Italy. (289) . High-arity Passives Ṃary was given by John the book. (626)
. Excluded Ṅon-DP arguments Ẇe want John to win (28) . 3rd argments where not all three arguments are DPs Ẇe want John to win (28)
Prepositional Phrase arguments of VPs are individual-denoting arguments of a verb which are marked by a proposition. They may or may not be obliques. Arguments are selected for by the verb, and they are (generally) not optional, though in some cases they may be omitted where they are understood or implicitly existentially quantified over.
. Included Ḋative Ṣue gave to Bill a book. (42) . Conative (at) C̣arla slid at the book. (179) . Idiosyncratic prepositional verbs İ wonder who to place my trust in. (711) She voted for herself. (743) . Locative J̇ohn was found in the office. (283) . PP predicates Ėverything you like is on the table. (736)
. Excluded ṖP adjuncts Particles Arguments of deverbal expressions ṭhe putter of books left. (892) . By-phrase Ṫed was bitten by the spider. (613)
Prepositional Phrase arguments of NPs or APs are individual-denoting arguments of a noun or adjective which are marked by a proposition. Arguments are selected for by the head, and they are (generally) not optional, though in some cases they may be omitted where they are understood or implicitly existentially quantified over.
. Included Ṙelational adjectives Ṁany people were fond of Pat. (936) *I was already aware of fact. (824) . Relational nouns Ẇe admired the pictures of us in the album. (759) They found the book on the atom. (780) . Arguments of deverbal nouns ṭhe putter of books left. (892)
Prepositional arguments introduced with by. Usually, this is the (semantic) subject of a passive verb, but in rare cases it may be the subject of a nominalized verb. Arguments are usually selected for by the head, and they are generally not optional. In this case, the argument introduced with by is semantically selected for by the verb, but it is syntactically optional. See [p.190]adger2003core and []collins2005smuggling.
. Included Ṗassives Ṫed was bitten by the spider. (613) . Subjects of deverbal nouns ṫhe attempt by John to leave surprised me. (1003)
Expletives, or “dummy” arguments, are semantically inert arguments. The most common expletives in English are it and there, although not all occurrences of these items are expletives. Arguments are usually selected for by the head, and they are generally not optional. In this case, the expletive occupies a syntactic argument slot, but it is not semantically selected by the verb, and there is often a syntactic variation without the expletive. See [p.170-172]adger2003core and [p.82-83]kim2008syntax.
. Included Ṫhere—inserted, existential Ṭhere loved Sandy. (939) There is a nurse available. (466) . It—cleft, inserted İt was a brand new car that he bought. (347) It bothers me that John coughs. (314) It is nice to go abroad. (47) . Environmental it K̇erry remarked it was late. (821) Poor Bill, it had started to rain and he had no umbrella. (116) You've really lived it up. (160)
. Excluded J̇ohn counted on Bill to get there on time. (996) I bought it to read. (1026)
## Arg Altern (Argument Alternations)
These are verbs with 3 or more arguments of any kind. Arity refers to the number of arguments that a head (or function) selects for. Arguments are usually selected for by the head, and they are generally not optional. They may be DPs, PPs, CPs, VPs, APs or other categories.
. Included Ḋitransitive [̣Sue] gave [to Bill] [a book]. (42) [Martha] carved [the baby] [a toy] out of wood. (139) . VP arguments [̣We] believed [John] [to be a fountain in the park]. (274) [We] made [them] [be rude]. (260) . Particles He] let [the cats which were whining] [out]. (71) . Passives with by-phrase [̣A good friend] is remained [to me] [by him]. (237) . Expletives [̣We] expect [there] [to will rain]. (282) [There] is [a seat] [available]. (934) [It] bothers [me] [that he is here]. (1009) . Small clause John] considers [Bill] [silly]. (1039)
. Excluded Ṙesults, depictives John] broke [the geode] [open].
These are VPs where a canonical argument of the verb is missing. This can be difficult to determine, but in many cases the missing argument is understood with existential quantification or generically, or contextually salient. See [p.106-109]sportiche2013introduction.
. Included Ṁiddle voice/causative inchoative Ṭhe problem perceives easily. (66) . Passive Ṫhe car was driven. (296) . Null complement anaphora J̇ean persuaded Robert. (380) Nobody told Susan. (883) . Dropped argument Ḳim put in the box. (253) The guests dined. (835) I wrote to Bill. (1030) . Transitive adjective J̇ohn is eager. (27) We pulled free. (144) . Transitive noun İ sensed his eagerness. (155) . Expletive insertion Ịt loved Sandy. (949)
. Excluded Ṫed was bitten by the spider. (613)
These are VPs in which a non-canonical argument of the verb has been added. These cases are clearer to identify where the additional argument is a DP. In general, PPs which mark locations, times, beneficiaries, or purposes should be analyzed as adjuncts, while PPs marking causes can be considered arguments. See []pylkkanen2008introducing.
. Included Ėxtra argument Ḷinda winked her lip. (202) Sharon fainted from hunger. (204) I shaved myself. (526) . Causative Ị squeaked the door. (207) . Expletive insertion Ṫhere is a monster in Loch Ness. (928) It annoys people that dogs bark. (943) . Benefactive Ṁartha carved the baby a toy out of wood. (139)
The passive voice is marked by the demotion of the subject (either complete omission or to a by-phrase) and the verb appearing as a past participle. In the stereotypical construction there is an auxiliary be verb, though this may be absent. See [p.175-190]kim2008syntax, collins2005smuggling, and [p.311-333]sag2003syntactic.
. Included V̇erbs Ṫhe earth was believed to be round. (157) . Psuedopassive Ṫhe bed was slept in. (298) . Past participle adjuncts Ṫhe horse raced past the barn fell. (900)
## Imperative
The imperative mood is marked by the absence of the a subject and the bare form of the verb, and expresses a command, request, or other directive speech act.
. Included Ẉash you! (224) Somebody just left - guess who. (528)
## Binding
These are cases in which a reflexive (non-possessive) pronoun, usually bound by an antecedent. See [p.163-186]sportiche2013introduction and [p.203-226]sag2003syntactic.
. Included Ọurselves like ourselves. (742) Which pictures of himself does John like? (386)
These are cases in which a non-reflexive pronoun appears along with its antecedent. This includes donkey anaphora, quantificational binding, and bound possessives, among other bound pronouns. See [p.163-186]sportiche2013introduction and [p.203-226]sag2003syntactic.
. Included Ḃound possessor Ṫhe children admire their mother. (382) . Quantificational binding Ėverybody gets on well with a certain relative, but often only his therapist knows which one. (562) . Bound pronoun Ẉe gave us to the cause. (747)
## Question
These are sentences in which the matrix clause is interrogative (either a wh- or polar question). See [pp.282-213]adger2003core, [pp.193-222]kim2008syntax, and [p.315-350]carnie2013syntax.
. Included Ẇh-question Ẇho always drinks milk? (684) . Polar question Ḋid Athena help us? (486)
These are embedded interrogative clauses appearing as arguments of verbs, nouns, and adjectives. Not including relative clauses and free relatives. See [p.297]adger2003core.
. Included U̇nder VP İ forgot how good beer tastes. (235) *What did you ask who saw? (508) . Under NP Ṫhat is the reason why he resigned. (313) . Under AP Ṫhey claimed they had settled on something, but it wasn't clear what they had settled on. (529) . Free relative Ẇhat the water did to the bottle was fill it. (33)
. Excluded Relative clauses, free relatives
These are phrasal Wh-phrases, in which the wh-word moves along with other expressions, including prepositions (pied-piping) or nouns in the case of determiner wh-words such as how many and which.
. Included Ṗied-piping Ṭhe ship sank, but I don't know with what. (541) . Other phrasal wh-phrases İ know which book Mag read, and which book Bob read my report that you hadn't. (61) How sane is Peter? (88)
Relative clauses are noun modifiers appearing with a relativizer (either that or a wh-word) and an associated gap. See [p.223-244]kim2008syntax.
. Included Ṫhough he may hate those that criticize Carter, it doesn't matter. (332) *The book what inspired them was very long. (686) Everything you like is on the table. (736)
. Excluded Ṭhe more you would want, the less you would eat. (6)
This is wh-movement out of an extraction island, or near-island. Islands include, for example, complex NPs, adjuncts, embedded questions, coordination. A near-island is an extraction that closely resembles an island violation, such as extraction out of an embedded clause, or across-the-board extraction. See [pp.323-333]adger2003core and [pp.332-334]carnie2013syntax.
. Included Ėmbedded question *What did you ask who Medea gave? (493) . Adjunct Ẉhat did you leave before they did? (598) . Parasitic gaps Ẇhich topic did you choose without getting his approval? (311) . Complex NP Ẇho did you get an accurate description of? (483)
## Comp Clause (Complement Clauses)
These are complement clauses acting as the (syntactic) subject of verbs. See [pp.90-91]kim2008syntax.
. Included Ṫhat dogs bark annoys people. (942) The socks are ready for for you to put on to be planned. (112)
. Excluded Ėxpletive insertion İt bothers me that John coughs. (314)
These are complement clauses acting as (non-subject) arguments of verbs. See [pp.84-90]kim2008syntax.
. Included İ can't believe Fred won't, either. (50) I saw that gas can explode. (222) It bothers me that John coughs. (314) Clefts İt was a brand new car that he bought. (347)
These are complement clauses acting as an argument of a noun or adjective. See [pp.91-94]kim2008syntax.
. Included U̇nder NP Ḋo you believe the claim that somebody was looking for something? (99) . Under AP Ṭhe children are fond that they have ice cream. (842)
These are complement clauses with a non-finite matrix verb. Often, the complementizer is for, or there is no complementizer. See [pp.252-253,256-260]adger2003core.
. Included Ḟor complementizer İ would prefer for John to leave. (990) . No Complementizer Ṁary intended John to go abroad. (48) . Ungrammatical Ḣeidi thinks that Andy to eat salmon flavored candy bars. (363) . V-ing Ȯnly Churchill remembered Churchill giving the Blood, Sweat and Tears speech. (469)
These are complement clauses with no overt complementizer.
. Included Ċomplement clause İ'm sure we even got these tickets! (325) He announced he would marry the woman he loved most, but none of his relatives could figure out who. (572) . Relative clause Ṫhe Peter we all like was at the party (484)
These are sentences with three or nested verbs, where VP is not an aux or modal, i.e. with the following syntax: [S ...[ VP ...[ VP ...[ VP ...] ...] ...] ...]
. Included Ėmbedded VPs Ṁax seemed to be trying to force Ted to leave the room, and Walt, Ira. (657) . Embedded clauses İ threw away a book that Sandy thought we had read. (713)
## Aux (Auxiliaries)
Any occurrence of negation in a sentence, including sentential negation, negative quantifiers, and negative adverbs.
. Included Ṡentential İ can't remember the name of somebody who had misgivings. (123) . Quantifier Ṅo writer, and no playwright, meets in Vienna. (124) . Adverb Ṫhey realised that never had Sir Thomas been so offended. (409)
Modal verbs (may, might, can, could, will, would, shall, should, must). See [pp.152-155]kim2008syntax.
. Included J̇ohn can kick the ball. (280) As a statesman, scarcely could he do anything worth mentioning. (292)
. Excluded Ṗseudo-modals Ṡandy was trying to work out which students would be able to solve a certain problem. (600)
Auxiliary verbs (e.g. be, have, do). See [pp.149-174]kim2008syntax.
. Included Ṫhey love to play golf, but I do not. (290) The car was driven. (296) he had spent five thousand dollars. (301)
. Excluded Ṗseudo-auxiliaries Ṣally asked if somebody was going to fail math class, but I can't remember who. (589) The cat got bitten. (926)
These are predicates acting as near-auxiliary (e.g. get-passive) or near-modals (e.g. willing)
. Included Ṅear-auxiliaries Ṃary came to be introduced by the bartender and I also came to be. (55) *Sally asked if somebody was going to fail math class, but I can't remember who. (589) The cat got bitten. (926) . Near-modals Ċlinton is anxious to find out which budget dilemmas Panetta would be willing to tackle in a certain way, but he won't say in which. (593) Sandy was trying to work out which students would be able to solve a certain problem. (600)
## to-VP (Infinitival VPs)
These are VPs with control verbs, where one argument is a non-finite to-VP without a covert subject co-indexed with an argument of the matrix verb. See [pp.252,266-291]adger2003core, [pp.203-222]sportiche2013introduction, and [pp.125-148]kim2008syntax.
. Included İntransitive subject control Ịt tries to leave the country. (275) . Transitive subject control J̇ohn promised Bill to leave. (977) . Transitive object control İ want her to dance. (379) John considers Bill to be silly. (1040)
. Excluded V̇P args of NP/AP Ṫhis violin is difficult to play sonatas on. (114) . Purpose Ṫhere is a bench to sit on. (309) . Subject VPs Ṫo please John is easy. (315) . Argument present participles Ṁedea denied poisoning the phoenix. (490) . Raising Ȧnson believed himself to be handsome. (499)
These are VPs with raising predicates, where one argument is a non-finite to-VP without a covert subject co-indexed with an argument of the matrix verb. Unlike control verbs, the coindexed argument is not a semantic argument of the raising predicate. See [pp.260-266]adger2003core, [pp.203-222]sportiche2013introduction, and [pp.125-148]kim2008syntax.
. Included Ṡubject raising U̇nder the bed seems to be a fun place to hide. (277) . Object raising Ȧnson believed himself to be handsome. (499) . Raising adjective J̇ohn is likely to leave. (370)
These are embedded infinitival VPs containing a (non-subject) gap that is filled by an argument in the upper clause. Examples are purpose-VPs and tough-movement. See [pp.246-252]kim2008syntax.
. Included Ṫough-movement Ḍrowning cats, which is against the law, are hard to rescue. (79) . Infinitival relatives F̣ed knows which politician her to vote for. (302) . Purpose ṫhe one with a red cover takes a very long time to read. (352) . Other non-finite VPs with extraction Ȧs a statesman, scarcely could he do anything worth mentioning. (292)
These are non-finite VP arguments of nouns and adjectives.
. Included Ṙaising adjectives J̇ohn is likely to leave. (370) . Control adjectives Ṫhe administration has issued a statement that it is willing to meet a student group, but I'm not sure which one. (604) . Control nouns Ȧs a teacher, you have to deal simultaneously with the administration's pressure on you to succeed, and the children's to be a nice guy. (673) . Purpose VPs ṫhere is nothing to do. (983)
These are miscellaneous non-finite VPs.
. Included İ saw that gas can explode. (222) Gerunds/Present participles Ṣtudents studying English reads Conrad's Heart of Darkness while at university. (262) Knowing the country well, he took a short cut. (411) John became deadly afraid of flying. (440) . Subject VPs Ṫo please John is easy. (315) . Nominalized VPs Ẉhat Mary did Bill was give a book. (473)
. Excluded ṫo-VPs acting as complements or modifiers of verbs, nouns, or adjectives
## N, Adj (Nouns and Adjectives)
These are nouns and adjectives derived from verbs.
. Included Ḋeverbal nouns ṭhe election of John president surprised me. (1001) . “Light” verbs Ṫhe birds give the worm a tug. (815) . Gerunds İf only Superman would stop flying planes! (773) . Event-wh Ẇhat the water did to the bottle was fill it. (33) . Deverbal adjectives Ḣis or her least known work. (95)
Relational nouns are NPs with an obligatory (or existentially closed) argument. A particular relation holds between the members of the extension of NP and the argument. The argument must be a DP possessor or a PP. See [pp.82-83]kim2008syntax.
. Included Ṅouns with of-arguments J̇ohn has a fear of dogs. (353) . Nouns with other PP-arguments Ḣenri wants to buy which books about cooking? (442) . Measure nouns İ bought three quarts of wine and two of Clorox. (667) . Possessed relational nouns J̣ohn's mother likes himself. (484)
. Excluded Ṅouns with PP modifiers Ṡome people consider dogs in my neighborhood dangerous. (802)
Transitive (non-relational) nouns take a VP or CP argument. See [pp.82-83]kim2008syntax.
. Included V̇P argument ṫhe attempt by John to leave surprised me. (1003) . CP argument Ẉhich report that John was incompetent did he submit? (69) . QP argument Ṫhat is the reason why he resigned. (313)
These are complex NPs, including coordinated nouns and nouns with modifiers (excluding prenominal adjectives).
. Included Ṁodified NPs Ṭhe madrigals which Henry plays the lute and sings sound lousy. (84) John bought a book on the table. (233) . NPs with coordination Ṭhe soundly and furry cat slept. (871) The love of my life and mother of my children would never do such a thing. (806)
Noun-noun compounds are NPs consisting of two constituent nouns.
. Included İt was the peasant girl who got it. (320) A felon was elected to the city council. (938)
These are adjectives that take an obligatory (or existentially closed) argument. A particular relation holds between the members of the extension of the modified NP and the argument. The argument must be a DP or PP. See [pp.80-82]kim2008syntax.
. Included Ȯf-arguments Ṫhe chickens seem fond of the farmer. (254) . Other PP arguments Ṫhis week will be a difficult one for us. (241) John made Bill mad at himself. (1035)
A transitive (non-relational) adjective. I.e. an adjectives that takes a VP or CP argument. See [pp.80-82]kim2008syntax.
. Included V̇P argument J̇ohn is likely to leave. (370) . CP argument J̇ohn is aware of it that Bill is here. (1013) . QP argument Ṫhe administration has issued a statement that it is willing to meet a student group, but I'm not sure which one. (604)
## S-Syntax (Sentence-Level Syntax)
These are expressions with non-canonical word order. See, for example, [p.76]sportiche2013introduction.
. Includes Ṗarticle shift Ṃickey looked up it. (24) . Preposed modifiers Ȯut of the box jumped a little white rabbit. (215) *Because she's so pleasant, as for Mary I really like her. (331) . Quantifier float Ṫhe men will all leave. (43) . Preposed argument Ẇith no job would John be happy. (333) . Relative clause extraposition Ẇhich book's, author did you meet who you liked? (731) . Misplaced phrases Ṁary was given by John the book. (626)
This includes topicalization and focus constructions. See [pp.258-269]kim2008syntax and [pp.68-75]sportiche2013introduction.
. Included Ṫopicalization Ṁost elections are quickly forgotten, but the election of 2000, everyone will remember for a long time. (807) . Clefts İt was a brand new car that he bought. (347) . Pseudo-clefts Ẇhat John promised is to be gentle. (441)
. Excluded Ṫhere-insertion Passive
These are parentheticals or fragmentary expressions. . Included Ṗarenthetical Ṁary asked me if, in St. Louis, John could rent a house cheap. (704) . Fragments Ṫhe soup cooks, thickens. (448) . Tag question Ġeorge has spent a lot of money, hasn't he? (291)
Coordinations and disjunctions are expressions joined with and, but, or, etc. See [pp.61-68]sportiche2013introduction.
. Included ḊP coordination Ḋave, Dan, Erin, Jaime, and Alina left. (341) . Right Node Raising K̇im gave a dollar to Bobbie and a dime to Jean. (435) . Clausal coordination Ṡhe talked to Harry, but I don't know who else. (575) . Or, nor Ṇo writer, nor any playwright, meets in Vienna. (125) . Pseudo-coordination İ want to try and buy some whiskey. (432) . Juxtaposed clauses Ŀights go out at ten. There will be no talking afterwards. (779)
This includes subordinate clauses, especially with subordinating conjunctions, and conditionals.
. Included Ċonditional İf I can, I will work on it. (56) . Subordinate clause Ẉhat did you leave before they did? (598) *Because Steve's of a spider's eye had been stolen, I borrowed Fred's diagram of a snake's fang. (677) . Correlative Ạs you eat the most, you want the least. (5)
This includes VP or NP ellipsis, or anaphora standing for VPs or NPs (not DPs). See [pp.55-61]sportiche2013introduction.
. Included V̇P Ellipsis İf I can, I will work on it. (56) Mary likes to tour art galleries, but Bill hates to. (287) . VP Anaphor İ saw Bill while you did so Mary. (472) . NP Ellipsis Ṫom's dog with one eye attacked Fred's. (679) . NP anaphor ṫhe one with a red cover takes a very long time to read. (352) . Sluicing Ṁost columnists claim that a senior White House official has been briefing them, and the newspaper today reveals which one. (557) . Gapping Ḃill ate the peaches, but Harry the grapes. (646)
These are adjuncts modifying sentences, sentence-level adverbs, subordinate clauses.
. Included Ṡentence-level adverbs Ṡuddenly, there arrived two inspectors from the INS. (447) . Subordinate clauses Ṫhe storm arrived while we ate lunch. (852)
## Determiner
These are quantificational DPs, i.e. the determiner is a quantifier.
. Included Q̇uantifiers Ẹvery student, and he wears socks, is a swinger. (118) We need another run to win. (769) . Partitive Ṇeither of students failed. (265)
These are quantifiers that take PP arguments, and measure nouns. See [pp.109-118]kim2008syntax.
. Included Q̇uantifiers with PP arguments Ṇeither of students failed. (265) . Numerals Ȯne of Korea's most famous poets wrote these lines. (294) . Measure nouns İ bought three quarts of wine and two of Clorox. (667)
These are negative polarity items (any, ever, etc.) and free choice items (any). See kadmon1993any.
. Included ṄPI Ėverybody around here who ever buys anything on credit talks in his sleep. (122) I didn't have a red cent. (350) . FCI Ȧny owl hunts mice. (387)
These are comparative constructions. See BIBREF22 .
. Included Ċorrelative Ṫhe angrier Mary got, the more she looked at pictures. (9) They may grow as high as bamboo. (337) I know you like the back of my hand. (775)
## Violations
These are sentences that include a semantic violation, including type mismatches, violations of selectional restrictions, polarity violations, definiteness violations.
. Included V̇olation of selectional restrictions ṃany information was provided. (218) *It tries to leave the country. (275) . Aspectual violations J̣ohn is tall on several occasions. (540) . Definiteness violations Ịt is the problem that he is here. (1018) . Polarity violations Ȧny man didn't eat dinner. (388)
These are sentences that include a violation in inflectional morphology, including tense-aspect marking, or agreement.
. Included Ċase Ụs love they. (46) . Agreement Ṣtudents studying English reads Conrad's Heart of Darkness while at university. (262) . Gender Ṣally kissed himself. (339) . Tense/Aspect Ḳim alienated cats and beating his dog. (429)
These are sentences with a violation that can be identified with the presence or absence of a single word.
. Included Ṁissing word J̣ohn put under the bathtub. (247) *I noticed the. (788) . Extra word Ẹveryone hopes everyone to sleep. (467) *He can will go (510)
| 26
|
1901.04899
|
Conversational Intent Understanding for Passengers in Autonomous Vehicles
|
# Conversational Intent Understanding for Passengers in Autonomous Vehicles
## Abstract
Understanding passenger intents and extracting relevant slots are important building blocks towards developing a contextual dialogue system responsible for handling certain vehicle-passenger interactions in autonomous vehicles (AV). When the passengers give instructions to AMIE (Automated-vehicle Multimodal In-cabin Experience), the agent should parse such commands properly and trigger the appropriate functionality of the AV system. In our AMIE scenarios, we describe usages and support various natural commands for interacting with the vehicle. We collected a multimodal in-cabin data-set with multi-turn dialogues between the passengers and AMIE using a Wizard-of-Oz scheme. We explored various recent Recurrent Neural Networks (RNN) based techniques and built our own hierarchical models to recognize passenger intents along with relevant slots associated with the action to be performed in AV scenarios. Our experimental results achieved F1-score of 0.91 on utterance-level intent recognition and 0.96 on slot extraction models.
## Introduction
Understanding passenger intents and extracting relevant slots are important building blocks towards developing a contextual dialogue system responsible for handling certain vehicle-passenger interactions in autonomous vehicles (AV). When the passengers give instructions to AMIE (Automated-vehicle Multimodal In-cabin Experience), the agent should parse such commands properly and trigger the appropriate functionality of the AV system. In our AMIE scenarios, we describe usages and support various natural commands for interacting with the vehicle. We collected a multimodal in-cabin data-set with multi-turn dialogues between the passengers and AMIE using a Wizard-of-Oz scheme. We explored various recent Recurrent Neural Networks (RNN) based techniques and built our own hierarchical models to recognize passenger intents along with relevant slots associated with the action to be performed in AV scenarios. Our experimental results achieved F1-score of 0.91 on utterance-level intent recognition and 0.96 on slot extraction models.
## Methodology
Our AV in-cabin data-set includes 30 hours of multimodal data collected from 30 passengers (15 female, 15 male) in 20 rides/sessions. 10 types of passenger intents are identified and annotated as: Set/Change Destination, Set/Change Route (including turn-by-turn instructions), Go Faster, Go Slower, Stop, Park, Pull Over, Drop Off, Open Door, and Other (turn music/radio on/off, open/close window/trunk, change AC/temp, show map, etc.). Relevant slots are identified and annotated as: Location, Position/Direction, Object, Time-Guidance, Person, Gesture/Gaze (this, that, over there, etc.), and None. In addition to utterance-level intent types and their slots, word-level intent keywords are annotated as Intent as well. We obtained 1260 unique utterances having commands to AMIE from our in-cabin data-set. We expanded this data-set via Amazon Mechanical Turk and ended up with 3347 utterances having intents. The annotations for intents and slots are obtained on the transcribed utterances by majority voting of 3 annotators.
For slot filling and intent keywords extraction tasks, we experimented with seq2seq LSTMs and GRUs, and also Bidirectional LSTM/GRUs. The passenger utterance is fed into a Bi-LSTM network via an embedding layer as a sequence of words, which are transformed into word vectors. We also experimented with GloVe, word2vec, and fastText as pre-trained word embeddings. To prevent overfitting, a dropout layer is used for regularization. Best performing results are obtained with Bi-LSTMs and GloVe embeddings (6B tokens, 400K vocab size, dim 100).
For utterance-level intent detection, we experimented with mainly 5 models: (1) Hybrid: RNN + Rule-based, (2) Separate: Seq2one Bi-LSTM + Attention, (3) Joint: Seq2seq Bi-LSTM for slots/intent keywords & utterance-level intents, (4) Hierarchical + Separate, (5) Hierarchical + Joint. For (1), we extract intent keywords/slots (Bi-LSTM) and map them into utterance-level intent types (rule-based via term frequencies for each intent). For (2), we feed the whole utterance as input sequence and intent-type as single target. For (3), we experiment with the joint learning models BIBREF0 , BIBREF1 , BIBREF2 where we jointly train word-level intent keywords/slots and utterance-level intents (adding <BOU>/<EOU> terms to the start/end of utterances with intent types). For (4) and (5), we experiment with the hierarchical models BIBREF3 , BIBREF4 , BIBREF5 where we extract intent keywords/slots first, and then only feed the predicted keywords/slots as a sequence into (2) and (3), respectively.
## Experimental Results
The slot extraction and intent keywords extraction results are given in Table TABREF1 and Table TABREF2 , respectively. Table TABREF3 summarizes the results of various approaches we investigated for utterance-level intent understanding. Table TABREF4 shows the intent-wise detection results for our AMIE scenarios with the best performing utterance-level intent recognizer.
## Conclusion
After exploring various recent Recurrent Neural Networks (RNN) based techniques, we built our own hierarchical joint models to recognize passenger intents along with relevant slots associated with the action to be performed in AV scenarios. Our experimental results outperformed certain competitive baselines and achieved overall F1-scores of 0.91 for utterance-level intent recognition and 0.96 for slot extraction tasks.
| 4
|
1901.09755
|
Language Independent Sequence Labelling for Opinion Target Extraction
|
# Language Independent Sequence Labelling for Opinion Target Extraction
## Abstract
In this research note we present a language independent system to model Opinion Target Extraction (OTE) as a sequence labelling task. The system consists of a combination of clustering features implemented on top of a simple set of shallow local features. Experiments on the well known Aspect Based Sentiment Analysis (ABSA) benchmarks show that our approach is very competitive across languages, obtaining best results for six languages in seven different datasets. Furthermore, the results provide further insights into the behaviour of clustering features for sequence labelling tasks. The system and models generated in this work are available for public use and to facilitate reproducibility of results.
## Introduction
Opinion Mining and Sentiment Analysis (OMSA) are crucial for determining opinion trends and attitudes about commercial products, companies reputation management, brand monitoring, or to track attitudes by mining social media, etc. Furthermore, given the explosion of information produced and shared via the Internet, especially in social media, it is simply not possible to keep up with the constant flow of new information by manual methods.
Early approaches to OMSA were based on document classification, where the task was to determine the polarity (positive, negative, neutral) of a given document or review BIBREF0 , BIBREF1 . A well known benchmark for polarity classification at document level is that of BIBREF2 . Later on, a finer-grained OMSA was deemed necessary. This was motivated by the fact that in a given review more than one opinion about a variety of aspects or attributes of a given product is usually conveyed. Thus, Aspect Based Sentiment Analysis (ABSA) was defined as a task which consisted of identifying several components of a given opinion: the opinion holder, the target, the opinion expression (the textual expression conveying polarity) and the aspects or features. Aspects are mostly domain-dependent. In restaurant reviews, relevant aspects would include “food quality”, “price”, “service”, “restaurant ambience”, etc. Similarly, if the reviews were about consumer electronics such as laptops, then aspects would include “size”, “battery life”, “hard drive capacity”, etc.
In the review shown by Figure FIGREF1 there are three different opinions about two different aspects (categories) of the restaurant, namely, the first two opinions are about the quality of the food and the third one about the general ambience of the place. Furthermore, there are just two opinion targets because the target of the third opinion, the restaurant itself, remains implicit. Finally, each aspect is assigned a polarity; in this case all three opinion aspects are negative.
In this work we focus on Opinion Target Extraction, which we model as a sequence labelling task. In order to do so, we convert an annotated review such as the one in Figure FIGREF1 into the BIO scheme for learning sequence labelling models BIBREF3 . Example (1) shows the review in BIO format. Tokens in the review are tagged depending on whether they are at the beginning (B-target), inside (I-target) or outside (O) of the opinion target expression. Note that the third opinion target in Figure FIGREF1 is implicit.
We learn language independent models which consist of a set of local, shallow features complemented with semantic distributional features based on clusters obtained from a variety of data sources. We show that our approach, despite the lack of hand-engineered, language-specific features, obtains state-of-the-art results in 7 datasets for 6 languages on the ABSA benchmarks BIBREF4 , BIBREF5 , BIBREF6 .
The main contribution of this research note is providing an extension or addendum to previous work on sequence labelling BIBREF7 by reporting additional experimental results as well as further insights on the performance of our model across languages on a different NLP task such as Opinion Target Extraction (OTE). Thus, we empirically demonstrate the validity and strong performance of our approach for six languages in seven different datasets of the restaurant domain. Every experiment and result presented in this note is novel.
In this sense, we show that our approach is not only competitive across languages and domains for Named Entity Recognition, as shown by BIBREF7 , but that it can be straightforwardly adapted to different tasks and domains such as OTE. Furthermore, we release the system and every model trained for public use and to facilitate reproducibility of results.
## Background
Early approaches to Opinion Target Extraction (OTE) were unsupervised, although later on the vast majority of works have been based on supervised and deep learning models. To the best of our knowledge, the first work on OTE was published by BIBREF8 . They created a new task which consisted of generating overviews of the main product features from a collection of customer reviews on consumer electronics. They addressed such task using an unsupervised algorithm based on association mining. Other early unsupervised approaches include BIBREF9 which used a dependency parser to obtain more opinion targets, and BIBREF10 which aimed at extracting opinion targets in newswire via Semantic Role Labelling. From a supervised perspective, BIBREF11 presented an approach which learned the opinion target candidates and a combination of dependency and part-of-speech (POS) paths connecting such pairs. Their results improved the baseline provided by BIBREF8 . Another influential work was BIBREF12 , an unsupervised algorithm called Double Propagation which roughly consists of incrementally augmenting a set of seeds via dependency parsing.
Closer to our work, BIBREF13 , BIBREF14 and BIBREF15 approached OTE as a sequence labelling task, modelling the opinion targets using the BIO scheme. The first approach implemented HMM whereas the last two proposed CRFs to solve the problem. In all three cases, their systems included extensive human-designed and linguistically motivated features, such as POS tags, lemmas, dependencies, constituent parsing structure, lexical patterns and semantic features extracted from WordNet BIBREF16 .
Quite frequently these works used a third party dataset, or a subset of the original one, or created their own annotated data for their experiments. The result was that it was difficult to draw precise conclusions about the advantages or disadvantages of the proposed methods. In this context, the Aspect Based Sentiment Analysis (ABSA) tasks at SemEval BIBREF4 , BIBREF5 , BIBREF6 provided standard training and evaluation data thereby helping to establish a clear benchmark for the OTE task.
Finally, it should be noted that there is a closely related task, namely, the SemEval 2016 task on Stance Detection. Stance detection is related to ABSA, but there is a significant difference. In ABSA the task is to determine whether a piece of text is positive, negative, or neutral with respect to an aspect and a given target (which in Stance Detection is called “author's favorability” towards a given target). However, in Stance Detection the text may express opinion or sentiment about some other target, not mentioned in the given text, and the targets are predefined, whereas in ABSA the targets are open-ended.
## ABSA Tasks at SemEval
Three ABSA editions were held within the SemEval Evaluation Exercises between 2014 and 2016. The ABSA 2014 and 2015 tasks consisted of English reviews only, whereas in the 2016 task 7 more languages were added. Additionally, reviews from four domains were collected for the various sub-tasks across the three editions, namely, Consumer Electronics, Telecommunications, Museums and Restaurant reviews. In any case, the only constant in each of the ABSA editions was the inclusion, for the Opinion Target Extraction (OTE) sub-task, of restaurant reviews for every language. Thus, for the experiments presented in this paper we decided to focus on the restaurant domain across 6 languages and the three different ABSA editions. Similarly, this section will be focused on reviewing the OTE results for the restaurant domain.
The ABSA task consisted of identifying, for each opinion, the opinion target, the aspect referred to by the opinion and the aspect's polarity. Figure FIGREF1 illustrates the original annotation of a restaurant review in the ABSA 2016 dataset. It should be noted that, out of the three opinion components, only the targets are explicitly represented in the text, which means that OTE can be independently modelled as a sequence labelling problem as shown by Example (1). It is particularly important to notice that the opinion expressions (“dry”, “greasy”, “loud and rude”) are not annotated.
Following previous approaches, the first competitive systems for OTE at ABSA were supervised. Among the participants (for English) in the three editions, one team BIBREF17 , BIBREF18 was particularly successful. For ABSA 2014 and 2015 they developed a CRF system with extensive handcrafted linguistic features: POS, head word, dependency relations, WordNet relations, gazetteers and Name Lists based on applying the Double Propagation algorithm BIBREF12 on an initial list of 551 seeds. Interestingly, they also introduced word representation features based on Brown and K-mean clusters. For ABSA 2016, they improved their system by using the output of a Recurrent Neural Network (RNN) to provide additional features. The RNN is trained on the following input features: word embeddings, Name Lists and word clusters BIBREF19 . They were the best system in 2014 and 2016. In 2015 they obtained the second best result, in which the best system, a preliminary version of the one presented in this note, was submitted by the EliXa team BIBREF20 .
From 2015 onwards most works have been based on deep learning. BIBREF21 applied RNNs on top of a variety of pre-trained word embeddings, while BIBREF22 presented an architecture in which a RNN based tagger is stacked on top of the features generated by a Convolutional Neural Network (CNN). These systems were evaluated on the 2014 and 2015 datasets, respectively, but they did not go beyond the state-of-the-art.
BIBREF23 presented a 7 layer deep CNN combining word embeddings trained on a INLINEFORM0 5 billion word corpus extracted from Amazon BIBREF24 , POS tag features and manually developed linguistic patterns based on syntactic analysis and SenticNet BIBREF25 a concept-level knowledge based build for Sentiment Analysis applications. They only evaluate their system on the English 2014 ABSA data, obtaining best results up to date on that benchmark.
More recently, BIBREF26 proposed a coupled multi-layer attention (CMLA) network where each layer consists of a couple of attentions with tensor operators. Unlike previous approaches, their system does not use complex linguistic-based features designed for one specific language. However, whereas previous successful approaches modelled OTE as an independent task, in the CMLA model the attentions interactively learn both the opinion targets and the opinion expressions. As opinion expressions are not available in the original ABSA datasets, they had to manually annotate the ABSA training and testing data with the required opinion expressions. Although BIBREF26 did not release the datasets with the annotated opinion expressions, Figure FIGREF5 illustrates what these annotations would look like. Thus, two new attributes (pfrom and pto) annotate the opinion expressions for each of the three opinions (“dry”, “greasy” and “loud and rude”, respectively). Using this new manual information to train their CMLA network they reported the best results so far for ABSA 2014 and 2015 (English only).
Finally, BIBREF27 develop a multi-task learning framework consisting of two LSTMs equipped with extended memories and neural memory operations. As BIBREF26 , they use opinion expressions annotations for a joint modelling of opinion targets and expressions. However, unlike BIBREF26 they do not manually annotate the opinion expressions. Instead they manually add sentiment lexicons and rules based on dependency parsing in order to find the opinion words required to train their system. Using this hand-engineered system, they report state of the art results only for English on the ABSA 2016 dataset. They do not provide evaluation results on the 2014 and 2015 restaurant datasets.
With respect to other languages, the IIT-T team presented systems for 4 out of the 7 languages in ABSA 2016, obtaining the best score for French and Dutch, second in Spanish but with very poor results for English, well below the baseline. The GTI team BIBREF28 implemented a CRF system using POS, lemmas and bigrams as features. They obtained the best result for Spanish and rather modest results for English.
Summarizing, the most successful systems for OTE have been based on supervised approaches with rather elaborate, complex and linguistically inspired features. BIBREF23 obtains best results on the ABSA 2014 data by means of a CNN with word embeddings trained on 5 billion words from Amazon, POS features, manual patterns based on syntactic analysis and SenticNet. More recently, the CMLA deep learning model has established new state-of-the-art results for the 2015 dataset, whereas BIBREF27 provide the state of the art for the 2016 benchmark. Thus, there is not currently a multilingual system that obtains competitive results across (at least) several of the languages included in ABSA.
As usual, most of the work has been done for English, with the large majority of the previous systems providing results only for one of the three English ABSA editions and without exploring the multilingual aspect. This could be due to the complex and language-specific systems that performed best for English BIBREF23 , or perhaps because the CMLA approach of BIBREF26 would require, in addition to the opinion targets, the gold standard annotations of the opinion expressions for each of the 6 languages other than English in the ABSA datasets.
## Methodology
The work presented in this research note requires the following resources: (i) Aspect Based Sentiment Analysis (ABSA) data for training and testing; (ii) large unlabelled corpora to obtain semantic distributional features from clustering lexicons; and (iii) a sequence labelling system. In this section we will describe each of the resources used.
## ABSA Datasets
Table TABREF7 shows the ABSA datasets from the restaurants domain for English, Spanish, French, Dutch, Russian and Turkish. From left to right each row displays the number of tokens, number of targets and the number of multiword targets for each training and test set. For English, it should be noted that the size of the 2015 set is less than half with respect to the 2014 dataset in terms of tokens, and only one third in number of targets. The French, Spanish and Dutch datasets are quite similar in terms of tokens although the number of targets in the Dutch dataset is comparatively smaller, possibly due to the tendency to construct compound terms in that language. The Russian dataset is the largest whereas the Turkish set is by far the smallest one.
Additionally, we think it is also interesting to note the low number of targets that are multiwords. To provide a couple of examples, for Spanish only the %35.59 of the targets are multiwords whereas for Dutch the percentage goes down to %25.68. If we compare these numbers with the CoNLL 2002 data for Named Entity Recognition (NER), a classic sequence labelling task, we find that in the ABSA data there is less than half the number of multiword targets than the number of multiword entities that can be found in the CoNLL Spanish and Dutch data (%35.59 vs %74.33 for Spanish and %25.68 vs %44.96 for Dutch).
## Unlabelled Corpora
Apart from the manually annotated data, we also leveraged large, publicly available, unlabelled data to train the clusters: (i) Brown 1000 clusters and (ii) Clark and Word2vec clusters in the 100-800 range.
In order to induce clusters from the restaurant domain we used the Yelp Academic Dataset, from which three versions were created. First, the full dataset, containing 225M tokens. Second, a subset consisting of filtering out those categories that do not correspond directly to food related reviews BIBREF29 . Thus, out of the 720 categories contained in the Yelp Academic Dataset, we kept the reviews from 173 of them. This Yelp food dataset contained 117M tokens in 997,721 reviews. Finally, we removed two more categories (Hotels and Hotels & Travel) from the Yelp food dataset to create the Yelp food-hotels subset containing around 102M tokens. For the rest of the languages we used their corresponding Wikipedia dumps. The pre-processing and tokenization is performed with the IXA pipes tools BIBREF30 .
The number of words used for each dataset, language and cluster type are described in Table TABREF9 . For example, the first row reads “Yelp Academic Dataset containing 225M words was used; after pre-processing, 156M words were taken to induce Brown clusters, whereas Clark and Word2vec clusters were trained on the whole corpus”. As explained in BIBREF7 , we pre-process the corpus before training Brown clusters, resulting in a smaller dataset than the original. Additionally, due to efficiency reasons, when the corpus is too large we use the pre-processed version to induce the Clark clusters.
## System
We use the sequence labeller implemented within IXA pipes BIBREF7 . It learns supervised models based on the Perceptron algorithm BIBREF31 . To avoid duplication of efforts, it uses the Apache OpenNLP project implementation customized with its own features. By design, the sequence labeller aims to establish a simple and shallow feature set, avoiding any linguistic motivated features, with the objective of removing any reliance on costly extra gold annotations and/or cascading errors across annotations.
The system consists of: (i) Local, shallow features based mostly on orthographic, word shape and n-gram features plus their context; and (ii) three types of simple clustering features, based on unigram matching: (i) Brown BIBREF32 clusters, taking the 4th, 8th, 12th and 20th node in the path; (ii) Clark BIBREF33 clusters and, (iii) Word2vec BIBREF34 clusters, based on K-means applied over the extracted word vectors using the skip-gram algorithm.
The clustering features look for the cluster class of the incoming token in one or more of the clustering lexicons induced following the three methods listed above. If found, then the class is added as feature (“not found” otherwise). As we work on a 5 token window, for each token and clustering lexicon at least 5 features are generated. For Brown, the number of features generated depend on the number of nodes found in the path for each token and clustering lexicon used.
Figure FIGREF13 depicts how our system relates, via clusters, unseen words with those words that have been seen as targets during the training process. Thus, the tokens `french-onions' and `salmon' would be annotated as opinion targets because they occur in the same clusters as seen words which in the training data are labeled as targets.
The word representation features are combined and stacked using the clustering lexicons induced over the different data sources listed in Table TABREF9 . In other words, stacking means adding various clustering features of the same type obtained from different data sources (for example, using clusters trained on Yelp and on Wikipedia); combining refers to combining different types of clustering features obtained from the same data source (e.g., using features from Brown and Clark clustering lexicons).
To choose the best combination of clustering features we tried, via 5-fold cross validation on the training set, every possible permutation of the available Clark and Word2vec clustering lexicons obtained from the data sources. Once the best combination of Clark and Word2vec clustering lexicons per data source was found, we tried to combine them with the Brown clusters. The result is a rather simple but very competitive system that has proven to be highly successful in the most popular Named Entity Recognition and Classification (NER) benchmarks, both in out-of-domain and in-domain evaluations. Furthermore, it was demonstrated that the system also performed robustly across languages without any language-specific tuning. Details of the system's implementation, including detailed description of the local and clustering features, can be found in BIBREF7 , including a section on how to combine the clustering features.
A preliminary version of this system BIBREF20 was the winner of the OTE sub-task in the ABSA 2015 edition (English only). In the next section we show that this system obtains state-of-the-art results not only across domains and languages for NER, but also for other tasks such as Opinion Target Extraction. The results reported are obtained using the official ABSA evaluation scripts BIBREF4 , BIBREF5 , BIBREF6 .
## Experimental Results
In this section we report on the experiments performed using the system and data described above. First we will present the English results for the three ABSA editions as well as a comparison with previous work. After that we will do the same for 5 additional languages included in the ABSA 2016 edition: Dutch, French, Russian, Spanish and Turkish. The local and clustering features, as described in Section SECREF11 , are the same for every language and evaluation setting. The only change is the clustering lexicons used for the different languages. As stated in section SECREF11 , the best cluster combination is chosen via 5-fold cross validation (CV) on the training data. We first try every permutation with the Clark and Word2vec clusters. Once the best combination is obtained, we then try with the Brown clusters obtaining thus the final model for each language and dataset.
## English
Table TABREF16 provides detailed results on the Opinion Target Extraction (OTE) task for English. We show in bold our best model (ALL) chosen via 5-fold CV on the training data. Moreover, we also show the results of the best models using only one type of clustering feature, namely, the best Brown, Clark and Word2vec models, respectively.
The first noteworthy issue is that the same model obtains the best results on the three English datasets. Second, it is also interesting to note the huge gains obtained by the clustering features, between 6-7 points in F1 score across the three ABSA datasets. Third, the results show that the combination of clustering features induced from different data sources is crucial. Fourth, the clustering features improve the recall by 12-15 points in the 2015 and 2016 data, and around 7 points for 2014. Finally, while in 2014 the precision also increases, in the 2015 setting it degrades almost by 4 points in F1 score.
Table TABREF17 compares our results with previous work. MIN refers to the multi-task learning framework consisting of two LSTMs equipped with extended memories and neural memory operations with manually developed rules for detecting opinion expressions BIBREF27 . CNN-SenticNet is the 7 layer CNN with Amazon word embeddings, POS, linguistic rules based on syntax patterns and SenticNet BIBREF23 .
LSTM is a Long Short Term Memory neural network built on top of word embeddings as proposed by BIBREF21 . WDEmb BIBREF35 uses word and dependency path, linear context and dependency context embedding features the input to a CRF. RNCRF is a joint model with CRF and a recursive neural network whereas CMLA is the Coupled Multilayer Attentions model described in section SECREF4 , both systems proposed by BIBREF26 . DLIREC-NLANGP is the winning system at ABSA 2014 and 2016 BIBREF17 , BIBREF18 , BIBREF19 while the penultimate row refers to our own system for all the three benchmarks (details in Table TABREF16 ).
The results of Table TABREF17 show that our system, despite its simplicity, is highly competitive, obtaining the best results on the 2015 and 2016 datasets and a competitive performance on the 2014 benchmark. In particular, we outperform much more complex and language-specific approaches tuned via language-specific features, such as that of DLIREC-NLANGP. Furthermore, while the deep learning approaches (enriched with human-engineered linguistic features) obtain comparable or better results on the 2014 data, that is not the case for the 2015 and 2016 benchmarks, where our system outperforms also the MIN and CMLA models (systems which require manually added rules and gold-standard opinion expressions to obtain their best results, as explained in section SECREF4 ). In this sense, this means that our system obtains better results than MIN and CMLA by learning the targets independently instead of jointly learning the target and those expressions that convey the polarity of the opinion, namely, the opinion expression.
There seems to be also a correlation between the size of the datasets and performance, given that the results on the 2014 data are much higher than those obtained using the 2015 and 2016 datasets. This might be due to the fact that the 2014 training set is substantially larger, as detailed in Table TABREF7 . In fact, the smaller datasets seem to affect more the deep learning approaches (LSTM, WDEmb, RNCRF) where only the MIN and CMLA models obtain similar results to ours, albeit using manually added language-specific annotations.
Finally, it would have been interesting to compare MIN, CNN-SenticNet and CMLA with our system on the three ABSA benchmarks, but their systems are not publicly available.
## Multilingual
We trained our system for 5 other languages on the ABSA 2016 datasets, using the same strategy as for English. We choose the best Clark-Word2vec combination (with and without Brown clusters) via 5-cross validation on the training data. The features are exactly the same as those used for English, the only change is the data on which the clusters are trained. Table TABREF19 reports on the detailed results obtained for each of the languages. In bold we show the best model chosen via 5-fold CV. Moreover, we also show the best models using only one of each of the clustering features.
The first difference with respect to the English results is that the Brown clustering features are, in three out of five settings, detrimental to performance. Second, that combining clustering features is only beneficial for Spanish. Third, the overall results are in general lower than those obtained in the 2016 English data. Finally, the difference between the best results and the results using the Local features is lower than for English, even though the Local results are similar to those obtained with the English datasets (except for Turkish, but this is due to the significantly smaller size of the data, as shown in Table TABREF7 ).
We believe that all these four issues are caused, at least partially, by the lack of domain-specific clustering features used for the multilingual experiments. In other words, while for the English experiments we leveraged the Yelp dataset to train the clustering algorithms, in the multilingual setting we first tried with already available clusters induced from the Wikipedia. Thus, it is to be expected that the gains obtained by clustering features obtained from domain-specific data such as Yelp would be superior to those achieved by the clusters trained on out-of-domain data.
In spite of this, Table TABREF20 shows that our system outperforms the best previous approaches across the five languages. In some cases, such as Turkish and Russian, the best previous scores were the baselines provided by the ABSA organizers, but for Dutch, French and Spanish our system is significantly better than current state-of-the-art. In particular, and despite using the same system for every language, we improve over GTI's submission, which implemented a CRF system with linguistic features specific to Spanish BIBREF28 .
## Discussion and Error Analysis
Considering the simplicity of our approach, we obtain best results for 6 languages and 7 different settings in the Opinion Target Extraction (OTE) benchmark for the restaurant domain using the ABSA 2014-2016 datasets.
These results are obtained without linguistic or manually-engineered features, relying on injecting external knowledge from the combination of clustering features to obtain a robust system across languages, outperforming other more complex and language-specific systems. Furthermore, the feature set used is the same for every setting, reducing human intervention to a minimum and establishing a clear methodology for a fast and easy creation of competitive OTE multilingual taggers.
The results also confirm the behaviour of these clustering algorithms to provide features for sequence labelling tasks such as OTE and Named Entity Recognition (NER), as previously discussed in BIBREF7 . Thus, in every evaluation setting the best results using Brown clusters as features were obtained when data close to the application domain and text genre, even if relatively small, was used to train the Brown algorithm. This can be clearly seen if we compare the English with the multilingual results. For English, the models including Brown clusters improve the Local features over 3-5 points in F1 score, whereas for Spanish, Dutch and Russian, they worsen performance. The reason is that for English the Yelp dataset is used whereas for the rest of languages the clusters are induced using the Wikipedia, effectively an out-of-domain corpus. The exception is Turkish, for which a 6 point gain in F1 score is obtained, but we believe that is probably due to the small size of the training data used for training the Local model.
In contrast, Word2vec clusters clearly benefit from larger amounts of data, as illustrated by the best English Word2vec model being the one trained using the Wikipedia, and not the Yelp dataset, which is closer to the application domain. Finally, the Clark algorithm seems to be the most versatile as it consistently outperforms the other two clustering methods in 4 out of the 8 evaluation settings presented.
Summarizing: (i) Brown clusters perform better when leveraged from source data close to the application domain, even if small in size; (ii) Clark clusters are the most robust of the three with respect to the size and domain of the data used; and (iii) for Word2vec size is the crucial factor. The larger the source data the better the performance. Thus, instead of choosing over one clustering type or the other, our system provides a method to effectively combining them, depending on the data sources available, to obtain robust and language independent sequence labelling systems.
Finally, results show that our models are particularly competitive when the amount of training data available is small, allowing us to compete with more complex systems including also manually-engineered features, as shown especially by the English results on the 2015 and 2016 data.
## Error Analysis
We will now discuss the shortcomings and most common errors performed by our system for the OTE task. By looking at the overall results in terms of precision and recall, it is possible to see the following patterns: With respect to the Local models, precision is consistently better than recall or, in other words, the coverage of the Local models is quite low. Tables TABREF16 and TABREF19 show that adding clustering features to the Local models allows to improve the recall for every evaluation setting, although with different outcomes. Overall, precision suffers, except for French. Furthermore, in three cases (English 2014, 2016 and Russian) precision is lower than recall, whereas the remaining 5 evaluations show that, despite large improvements in F1 score, most errors in our system are caused by false negatives, as it can be seen in Table TABREF23 .
Table TABREF25 displays the top 5 most common false positives and false negative errors for English, Spanish and French. By inspecting our system's output, and both the test and training sets, we found out that there were three main sources of errors: (a) errors caused by ambiguity in the use of certain source forms that may or may not refer to an opinion target; (b) span errors, where the target has only been partially annotated; and (c) unknown targets, which the system was unable to annotate by generalizing on the training data or clusters.
With respect to type (a), it is useful to look at the most common errors for all three languages, namely, `place', `food' and `restaurant', which are also among the top 5 most frequent targets in the gold standard sets. By looking at Examples (1-3) we would say that in all three cases `place' should be annotated as opinion target. However, (2) is a false positive (FP), (3) is a false negative (FN) and (1) is an example from the training set in which `place' is annotated as target. This is the case with many instances of `place' for which there seems to be some inconsistency in the actual annotation of the training and test set examples.
Example (1): Avoid this place!
Example (2): this place is a keeper!
Example (3): it is great place to watch sporting events.
For other frequent type (a) errors, ambiguity is the main problem. Thus, in Spanish the use of `comida' and `restaurante' is highly ambiguous and causes many FPs and FNs because sometimes it is actually an opinion target whereas in many other other cases it is just referring to the meal or the restaurant themselves without expressing any opinion about them. The same phenomenon occurs for “food” and “restaurant” in English and for `cuisine' and `restaurant' in French.
Span type (b) errors are typically caused by long opinion targets such as “filet mignon on top of spinach and mashed potatoes” for which our system annotates “filet” and “spinach” as separate targets, or “chicken curry and chicken tikka masala” which is wrongly tagged as one target. These cases are difficult because on the surface they look similar but the first one refers to one dish only, hence one target, whereas the second one refers to two separate dishes for which two different opinion targets should be annotated. Of course, these cases are particularly hurtful because they count as both FP and FN.
Finally, type (c) errors are usually caused by lack of generalization of our system to deal with unknown targets. Example (4-7) contain various mentions to the “Ray's” restaurant, which is in the top 5 errors for the English 2016 test set.
Example (4): After 12 years in Seattle Ray's rates as the place we always go back to.
Example (5): We were only in Seattle for one night and I'm so glad we picked Rays for dinner!
Example (6): I love Dungeness crabs and at Ray's you can get them served in about 6 different ways!
Example (7): Imagine my happy surprise upon finding that the views are only the third-best thing about Ray's!
Example (8): Ray's is something of a Seattle institution
Examples (4), (5) and (7) are FNs, (6) is a FP caused by wrongly identifying the target as “Ray's you”, whereas (8) is not event annotated in the gold standard or by our system, although it should had been.
## Concluding Remarks
In this research note we provide additional empirical experimentation to BIBREF7 , reporting best results for Opinion Target Extraction for 6 languages and 7 datasets using the same set of simple, shallow and language independent features. Furthermore, the results provide some interesting insights with respect to the use of clusters to inject external knowledge via semi-supervised features.
First, Brown clusters are particularly beneficial when trained on domain-related data. This seems to be the case in the multilingual setting, where the Brown clusters (trained on out-of-domain Wikipedia data) worsen the system's performance for every language except for Turkish.
Second, the results also show that Clark and Word2vec improve results in general, even if induced on out-of-domain data. Thirdly, for best performance it is convenient to combine clusters obtained from diverse data sources, both from in- and out-of-domain corpora.
Finally, the results indicate that, even when the amount of training data is small, such as in the 2015 and 2016 English benchmarks, our system's performance remains competitive thanks to the combination of clustering features. This, together with the lack of linguistic features, facilitates the easy and fast development of systems for new domains or languages. These considerations thus confirm the hypotheses stated in BIBREF7 with respect to the use of clustering features to obtain robust sequence taggers across languages and tasks.
The system and models for every language and dataset are available as part of the ixa-pipe-opinion module for public use and reproducibility of results.
## Acknowledgments
First, we would like to thank the anonymous reviewers for their comments to improve the paper. We would also like to thank Iñaki San Vicente for his help obtaining the Yelp data. This work has been supported by the Spanish Ministry of Economy and Competitiveness (MINECO/FEDER, UE), under the projects TUNER (TIN2015-65308-C5-1-R) and CROSSTEXT (TIN2015-72646-EXP).
| 14
|
1902.06843
|
Fusing Visual, Textual and Connectivity Clues for Studying Mental Health
|
# Fusing Visual, Textual and Connectivity Clues for Studying Mental Health
## Abstract
With ubiquity of social media platforms, millions of people are sharing their online persona by expressing their thoughts, moods, emotions, feelings, and even their daily struggles with mental health issues voluntarily and publicly on social media. Unlike the most existing efforts which study depression by analyzing textual content, we examine and exploit multimodal big data to discern depressive behavior using a wide variety of features including individual-level demographics. By developing a multimodal framework and employing statistical techniques for fusing heterogeneous sets of features obtained by processing visual, textual and user interaction data, we significantly enhance the current state-of-the-art approaches for identifying depressed individuals on Twitter (improving the average F1-Score by 5 percent) as well as facilitate demographic inference from social media for broader applications. Besides providing insights into the relationship between demographics and mental health, our research assists in the design of a new breed of demographic-aware health interventions.
0pt*0*0
0pt*0*0
0pt*0*0 0.95
1]Amir Hossein Yazdavar 1]Mohammad Saeid Mahdavinejad 2]Goonmeet Bajaj
3]William Romine 1]Amirhassan Monadjemi 1]Krishnaprasad Thirunarayan
1]Amit Sheth 4]Jyotishman Pathak [1]Department of Computer Science & Engineering, Wright State University, OH, USA [2]Ohio State University, Columbus, OH, USA [3]Department of Biological Science, Wright State University, OH, USA [4] Division of Health Informatics, Weill Cornell University, New York, NY, USA
[1] yazdavar.2@wright.edu
With ubiquity of social media platforms, millions of people are sharing their online persona by expressing their thoughts, moods, emotions, feelings, and even their daily struggles with mental health issues voluntarily and publicly on social media. Unlike the most existing efforts which study depression by analyzing textual content, we examine and exploit multimodal big data to discern depressive behavior using a wide variety of features including individual-level demographics. By developing a multimodal framework and employing statistical techniques for fusing heterogeneous sets of features obtained by processing visual, textual and user interaction data, we significantly enhance the current state-of-the-art approaches for identifying depressed individuals on Twitter (improving the average F1-Score by 5 percent) as well as facilitate demographic inference from social media for broader applications. Besides providing insights into the relationship between demographics and mental health, our research assists in the design of a new breed of demographic-aware health interventions.
## Introduction
Depression is a highly prevalent public health challenge and a major cause of disability worldwide. Depression affects 6.7% (i.e., about 16 million) Americans each year . According to the World Mental Health Survey conducted in 17 countries, on average, about 5% of people reported having an episode of depression in 2011 BIBREF0 . Untreated or under-treated clinical depression can lead to suicide and other chronic risky behaviors such as drug or alcohol addiction.
Global efforts to curb clinical depression involve identifying depression through survey-based methods employing phone or online questionnaires. These approaches suffer from under-representation as well as sampling bias (with very small group of respondents.) In contrast, the widespread adoption of social media where people voluntarily and publicly express their thoughts, moods, emotions, and feelings, and even share their daily struggles with mental health problems has not been adequately tapped into studying mental illnesses, such as depression. The visual and textual content shared on different social media platforms like Twitter offer new opportunities for a deeper understanding of self-expressed depression both at an individual as well as community-level. Previous research efforts have suggested that language style, sentiment, users' activities, and engagement expressed in social media posts can predict the likelihood of depression BIBREF1 , BIBREF2 . However, except for a few attempts BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , these investigations have seldom studied extraction of emotional state from visual content of images in posted/profile images. Visual content can express users' emotions more vividly, and psychologists noted that imagery is an effective medium for communicating difficult emotions.
According to eMarketer, photos accounted for 75% of content posted on Facebook worldwide and they are the most engaging type of content on Facebook (87%). Indeed, "a picture is worth a thousand words" and now "photos are worth a million likes." Similarly, on Twitter, the tweets with image links get twice as much attention as those without , and video-linked tweets drive up engagement . The ease and naturalness of expression through visual imagery can serve to glean depression-indicators in vulnerable individuals who often seek social support through social media BIBREF7 . Further, as psychologist Carl Rogers highlights, we often pursue and promote our Ideal-Self . In this regard, the choice of profile image can be a proxy for the online persona BIBREF8 , providing a window into an individual's mental health status. For instance, choosing emaciated legs of girls covered with several cuts as profile image portrays negative self-view BIBREF9 .
Inferring demographic information like gender and age can be crucial for stratifying our understanding of population-level epidemiology of mental health disorders. Relying on electronic health records data, previous studies explored gender differences in depressive behavior from different angles including prevalence, age at onset, comorbidities, as well as biological and psychosocial factors. For instance, women have been diagnosed with depression twice as often as men BIBREF10 and national psychiatric morbidity survey in Britain has shown higher risk of depression in women BIBREF11 . On the other hand, suicide rates for men are three to five times higher compared to that of the women BIBREF12 .
Although depression can affect anyone at any age, signs and triggers of depression vary for different age groups . Depression triggers for children include parental depression, domestic violence, and loss of a pet, friend or family member. For teenagers (ages 12-18), depression may arise from hormonal imbalance, sexuality concerns and rejection by peers. Young adults (ages 19-29) may develop depression due to life transitions, poverty, trauma, and work issues. Adult (ages 30-60) depression triggers include caring simultaneously for children and aging parents, financial burden, work and relationship issues. Senior adults develop depression from common late-life issues, social isolation, major life loses such as the death of a spouse, financial stress and other chronic health problems (e.g., cardiac disease, dementia). Therefore, inferring demographic information while studying depressive behavior from passively sensed social data, can shed better light on the population-level epidemiology of depression.
The recent advancements in deep neural networks, specifically for image analysis task, can lead to determining demographic features such as age and gender BIBREF13 . We show that by determining and integrating heterogeneous set of features from different modalities – aesthetic features from posted images (colorfulness, hue variance, sharpness, brightness, blurriness, naturalness), choice of profile picture (for gender, age, and facial expression), the screen name, the language features from both textual content and profile's description (n-gram, emotion, sentiment), and finally sociability from ego-network, and user engagement – we can reliably detect likely depressed individuals in a data set of 8,770 human-annotated Twitter users.
We address and derive answers to the following research questions: 1) How well do the content of posted images (colors, aesthetic and facial presentation) reflect depressive behavior? 2) Does the choice of profile picture show any psychological traits of depressed online persona? Are they reliable enough to represent the demographic information such as age and gender? 3) Are there any underlying common themes among depressed individuals generated using multimodal content that can be used to detect depression reliably?
## Related Work
Mental Health Analysis using Social Media:
Several efforts have attempted to automatically detect depression from social media content utilizing machine/deep learning and natural language processing approaches. Conducting a retrospective study over tweets, BIBREF14 characterizes depression based on factors such as language, emotion, style, ego-network, and user engagement. They built a classifier to predict the likelihood of depression in a post BIBREF14 , BIBREF15 or in an individual BIBREF1 , BIBREF16 , BIBREF17 , BIBREF18 . Moreover, there have been significant advances due to the shared task BIBREF19 focusing on methods for identifying depressed users on Twitter at the Computational Linguistics and Clinical Psychology Workshop (CLP 2015). A corpus of nearly 1,800 Twitter users was built for evaluation, and the best models employed topic modeling BIBREF20 , Linguistic Inquiry and Word Count (LIWC) features, and other metadata BIBREF21 . More recently, a neural network architecture introduced by BIBREF22 combined posts into a representation of user's activities for detecting depressed users. Another active line of research has focused on capturing suicide and self-harm signals BIBREF23 , BIBREF24 , BIBREF25 , BIBREF26 , BIBREF2 , BIBREF27 . Moreover, the CLP 2016 BIBREF28 defined a shared task on detecting the severity of the mental health from forum posts. All of these studies derive discriminative features to classify depression in user-generated content at message-level, individual-level or community-level. Recent emergence of photo-sharing platforms such as Instagram, has attracted researchers attention to study people's behavior from their visual narratives – ranging from mining their emotions BIBREF29 , and happiness trend BIBREF30 , to studying medical concerns BIBREF31 . Researchers show that people use Instagram to engage in social exchange and storytelling about their difficult experiences BIBREF4 . The role of visual imagery as a mechanism of self-disclosure by relating visual attributes to mental health disclosures on Instagram was highlighted by BIBREF3 , BIBREF5 where individual Instagram profiles were utilized to build a prediction framework for identifying markers of depression. The importance of data modality to understand user behavior on social media was highlighted by BIBREF32 . More recently, a deep neural network sequence modeling approach that marries audio and text data modalities to analyze question-answer style interviews between an individual and an agent has been developed to study mental health BIBREF32 . Similarly, a multimodal depressive dictionary learning was proposed to detect depressed users on Twitter BIBREF33 . They provide a sparse user representations by defining a feature set consisting of social network features, user profile features, visual features, emotional features BIBREF34 , topic-level features, and domain-specific features. Particularly, our choice of multi-model prediction framework is intended to improve upon the prior works involving use of images in multimodal depression analysis BIBREF33 and prior works on studying Instagram photos BIBREF6 , BIBREF35 .
Demographic information inference on Social Media:
There is a growing interest in understanding online user's demographic information due to its numerous applications in healthcare BIBREF36 , BIBREF37 . A supervised model developed by BIBREF38 for determining users' gender by employing features such as screen-name, full-name, profile description and content on external resources (e.g., personal blog). Employing features including emoticons, acronyms, slangs, punctuations, capitalization, sentence length and included links/images, along with online behaviors such as number of friends, post time, and commenting activity, a supervised model was built for predicting user's age group BIBREF39 . Utilizing users life stage information such as secondary school student, college student, and employee, BIBREF40 builds age inference model for Dutch Twitter users. Similarly, relying on profile descriptions while devising a set of rules and patterns, a novel model introduced for extracting age for Twitter users BIBREF41 . They also parse description for occupation by consulting the SOC2010 list of occupations and validating it through social surveys. A novel age inference model was developed while relying on homophily interaction information and content for predicting age of Twitter users BIBREF42 . The limitations of textual content for predicting age and gender was highlighted by BIBREF43 . They distinguish language use based on social gender, age identity, biological sex and chronological age by collecting crowdsourced signals using a game in which players (crowd) guess the biological sex and age of a user based only on their tweets. Their findings indicate how linguistic markers can misguide (e.g., a heart represented as <3 can be misinterpreted as feminine when the writer is male.) Estimating age and gender from facial images by training a convolutional neural networks (CNN) for face recognition is an active line of research BIBREF44 , BIBREF13 , BIBREF45 .
## Dataset
Self-disclosure clues have been extensively utilized for creating ground-truth data for numerous social media analytic studies e.g., for predicting demographics BIBREF36 , BIBREF41 , and user's depressive behavior BIBREF46 , BIBREF47 , BIBREF48 . For instance, vulnerable individuals may employ depressive-indicative terms in their Twitter profile descriptions. Others may share their age and gender, e.g., "16 years old suicidal girl"(see Figure FIGREF15 ). We employ a huge dataset of 45,000 self-reported depressed users introduced in BIBREF46 where a lexicon of depression symptoms consisting of 1500 depression-indicative terms was created with the help of psychologist clinician and employed for collecting self-declared depressed individual's profiles. A subset of 8,770 users (24 million time-stamped tweets) containing 3981 depressed and 4789 control users (that do not show any depressive behavior) were verified by two human judges BIBREF46 . This dataset INLINEFORM0 contains the metadata values of each user such as profile descriptions, followers_count, created_at, and profile_image_url.
Age Enabled Ground-truth Dataset: We extract user's age by applying regular expression patterns to profile descriptions (such as "17 years old, self-harm, anxiety, depression") BIBREF41 . We compile "age prefixes" and "age suffixes", and use three age-extraction rules: 1. I am X years old 2. Born in X 3. X years old, where X is a "date" or age (e.g., 1994). We selected a subset of 1061 users among INLINEFORM0 as gold standard dataset INLINEFORM1 who disclose their age. From these 1061 users, 822 belong to depressed class and 239 belong to control class. From 3981 depressed users, 20.6% disclose their age in contrast with only 4% (239/4789) among control group. So self-disclosure of age is more prevalent among vulnerable users. Figure FIGREF18 depicts the age distribution in INLINEFORM2 . The general trend, consistent with the results in BIBREF42 , BIBREF49 , is biased toward young people. Indeed, according to Pew, 47% of Twitter users are younger than 30 years old BIBREF50 . Similar data collection procedure with comparable distribution have been used in many prior efforts BIBREF51 , BIBREF49 , BIBREF42 . We discuss our approach to mitigate the impact of the bias in Section 4.1. The median age is 17 for depressed class versus 19 for control class suggesting either likely depressed-user population is younger, or depressed youngsters are more likely to disclose their age for connecting to their peers (social homophily.) BIBREF51
Gender Enabled Ground-truth Dataset: We selected a subset of 1464 users INLINEFORM0 from INLINEFORM1 who disclose their gender in their profile description. From 1464 users 64% belonged to the depressed group, and the rest (36%) to the control group. 23% of the likely depressed users disclose their gender which is considerably higher (12%) than that for the control class. Once again, gender disclosure varies among the two gender groups. For statistical significance, we performed chi-square test (null hypothesis: gender and depression are two independent variables). Figure FIGREF19 illustrates gender association with each of the two classes. Blue circles (positive residuals, see Figure FIGREF19 -A,D) show positive association among corresponding row and column variables while red circles (negative residuals, see Figure FIGREF19 -B,C) imply a repulsion. Our findings are consistent with the medical literature BIBREF10 as according to BIBREF52 more women than men were given a diagnosis of depression. In particular, the female-to-male ratio is 2.1 and 1.9 for Major Depressive Disorder and Dysthymic Disorder respectively. Our findings from Twitter data indicate there is a strong association (Chi-square: 32.75, p-value:1.04e-08) between being female and showing depressive behavior on Twitter.
## Data Modality Analysis
We now provide an in-depth analysis of visual and textual content of vulnerable users.
Visual Content Analysis: We show that the visual content in images from posts as well as profiles provide valuable psychological cues for understanding a user's depression status. Profile/posted images can surface self-stigmatization BIBREF53 . Additionally, as opposed to typical computer vision framework for object recognition that often relies on thousands of predetermined low-level features, what matters more for assessing user's online behavior is the emotions reflected in facial expressions BIBREF54 , attributes contributing to the computational aesthetics BIBREF55 , and sentimental quotes they may subscribe to (Figure FIGREF15 ) BIBREF8 .
Facial Presence:
For capturing facial presence, we rely on BIBREF56 's approach that uses multilevel convolutional coarse-to-fine network cascade to tackle facial landmark localization. We identify facial presentation, emotion from facial expression, and demographic features from profile/posted images . Table TABREF21 illustrates facial presentation differences in both profile and posted images (media) for depressed and control users in INLINEFORM0 . With control class showing significantly higher in both profile and media (8%, 9% respectively) compared to that for the depressed class. In contrast with age and gender disclosure, vulnerable users are less likely to disclose their facial identity, possibly due to lack of confidence or fear of stigma.
Facial Expression:
Following BIBREF8 's approach, we adopt Ekman's model of six emotions: anger, disgust, fear, joy, sadness and surprise, and use the Face++ API to automatically capture them from the shared images. Positive emotions are joy and surprise, and negative emotions are anger, disgust, fear, and sadness. In general, for each user u in INLINEFORM0 , we process profile/shared images for both the depressed and the control groups with at least one face from the shared images (Table TABREF23 ). For the photos that contain multiple faces, we measure the average emotion.
Figure FIGREF27 illustrates the inter-correlation of these features. Additionally, we observe that emotions gleaned from facial expressions correlated with emotional signals captured from textual content utilizing LIWC. This indicates visual imagery can be harnessed as a complementary channel for measuring online emotional signals.
General Image Features:
The importance of interpretable computational aesthetic features for studying users' online behavior has been highlighted by several efforts BIBREF55 , BIBREF8 , BIBREF57 . Color, as a pillar of the human vision system, has a strong association with conceptual ideas like emotion BIBREF58 , BIBREF59 . We measured the normalized red, green, blue and the mean of original colors, and brightness and contrast relative to variations of luminance. We represent images in Hue-Saturation-Value color space that seems intuitive for humans, and measure mean and variance for saturation and hue. Saturation is defined as the difference in the intensities of the different light wavelengths that compose the color. Although hue is not interpretable, high saturation indicates vividness and chromatic purity which are more appealing to the human eye BIBREF8 . Colorfulness is measured as a difference against gray background BIBREF60 . Naturalness is a measure of the degree of correspondence between images and the human perception of reality BIBREF60 . In color reproduction, naturalness is measured from the mental recollection of the colors of familiar objects. Additionally, there is a tendency among vulnerable users to share sentimental quotes bearing negative emotions. We performed optical character recognition (OCR) with python-tesseract to extract text and their sentiment score. As illustrated in Table TABREF26 , vulnerable users tend to use less colorful (higher grayscale) profile as well as shared images to convey their negative feelings, and share images that are less natural (Figure FIGREF15 ). With respect to the aesthetic quality of images (saturation, brightness, and hue), depressed users use images that are less appealing to the human eye. We employ independent t-test, while adopting Bonferroni Correction as a conservative approach to adjust the confidence intervals. Overall, we have 223 features, and choose Bonferroni-corrected INLINEFORM0 level of INLINEFORM1 (*** INLINEFORM2 , ** INLINEFORM3 ).
** alpha= 0.05, *** alpha = 0.05/223
Demographics Inference & Language Cues: LIWC has been used extensively for examining the latent dimensions of self-expression for analyzing personality BIBREF61 , depressive behavior, demographic differences BIBREF43 , BIBREF40 , etc. Several studies highlight that females employ more first-person singular pronouns BIBREF62 , and deictic language BIBREF63 , while males tend to use more articles BIBREF64 which characterizes concrete thinking, and formal, informational and affirmation words BIBREF65 . For age analysis, the salient findings include older individuals using more future tense verbs BIBREF62 triggering a shift in focus while aging. They also show positive emotions BIBREF66 and employ fewer self-references (i.e. 'I', 'me') with greater first person plural BIBREF62 . Depressed users employ first person pronouns more frequently BIBREF67 , repeatedly use negative emotions and anger words. We analyzed psycholinguistic cues and language style to study the association between depressive behavior as well as demographics. Particularly, we adopt Levinson's adult development grouping that partitions users in INLINEFORM0 into 5 age groups: (14,19],(19,23], (23,34],(34,46], and (46,60]. Then, we apply LIWC for characterizing linguistic styles for each age group for users in INLINEFORM1 .
Qualitative Language Analysis: The recent LIWC version summarizes textual content in terms of language variables such as analytical thinking, clout, authenticity, and emotional tone. It also measures other linguistic dimensions such as descriptors categories (e.g., percent of target words gleaned by dictionary, or longer than six letters - Sixltr) and informal language markers (e.g., swear words, netspeak), and other linguistic aspects (e.g., 1st person singular pronouns.)
Thinking Style:
Measuring people's natural ways of trying to analyze, and organize complex events have strong association with analytical thinking. LIWC relates higher analytic thinking to more formal and logical reasoning whereas a lower value indicates focus on narratives. Also, cognitive processing measures problem solving in mind. Words such as "think," "realize," and "know" indicates the degree of "certainty" in communications. Critical thinking ability relates to education BIBREF68 , and is impacted by different stages of cognitive development at different ages . It has been shown that older people communicate with greater cognitive complexity while comprehending nuances and subtle differences BIBREF62 . We observe a similar pattern in our data (Table TABREF40 .) A recent study highlights how depression affects brain and thinking at molecular level using a rat model BIBREF69 . Depression can promote cognitive dysfunction including difficulty in concentrating and making decisions. We observed a notable differences in the ability to think analytically in depressed and control users in different age groups (see Figure FIGREF39 - A, F and Table TABREF40 ). Overall, vulnerable younger users are not logical thinkers based on their relative analytical score and cognitive processing ability.
Authenticity:
Authenticity measures the degree of honesty. Authenticity is often assessed by measuring present tense verbs, 1st person singular pronouns (I, me, my), and by examining the linguistic manifestations of false stories BIBREF70 . Liars use fewer self-references and fewer complex words. Psychologists often see a child's first successfull lie as a mental growth. There is a decreasing trend of the Authenticity with aging (see Figure FIGREF39 -B.) Authenticity for depressed youngsters is strikingly higher than their control peers. It decreases with age (Figure FIGREF39 -B.)
Clout:
People with high clout speak more confidently and with certainty, employing more social words with fewer negations (e.g., no, not) and swear words. In general, midlife is relatively stable w.r.t. relationships and work. A recent study shows that age 60 to be best for self-esteem BIBREF71 as people take on managerial roles at work and maintain a satisfying relationship with their spouse. We see the same pattern in our data (see Figure FIGREF39 -C and Table TABREF40 ). Unsurprisingly, lack of confidence (the 6th PHQ-9 symptom) is a distinguishable characteristic of vulnerable users, leading to their lower clout scores, especially among depressed users before middle age (34 years old).
Self-references:
First person singular words are often seen as indicating interpersonal involvement and their high usage is associated with negative affective states implying nervousness and depression BIBREF66 . Consistent with prior studies, frequency of first person singular for depressed people is significantly higher compared to that of control class. Similarly to BIBREF66 , youngsters tend to use more first-person (e.g. I) and second person singular (e.g. you) pronouns (Figure FIGREF39 -G).
Informal Language Markers; Swear, Netspeak:
Several studies highlighted the use of profanity by young adults has significantly increased over the last decade BIBREF72 . We observed the same pattern in both the depressed and the control classes (Table TABREF40 ), although it's rate is higher for depressed users BIBREF1 . Psychologists have also shown that swearing can indicate that an individual is not a fragmented member of a society. Depressed youngsters, showing higher rate of interpersonal involvement and relationships, have a higher rate of cursing (Figure FIGREF39 -E). Also, Netspeak lexicon measures the frequency of terms such as lol and thx.
Sexual, Body:
Sexual lexicon contains terms like "horny", "love" and "incest", and body terms like "ache", "heart", and "cough". Both start with a higher rate for depressed users while decreasing gradually while growing up, possibly due to changes in sexual desire as we age (Figure FIGREF39 -H,I and Table TABREF40 .)
Quantitative Language Analysis:
We employ one-way ANOVA to compare the impact of various factors and validate our findings above. Table TABREF40 illustrates our findings, with a degree of freedom (df) of 1055. The null hypothesis is that the sample means' for each age group are similar for each of the LIWC features.
*** alpha = 0.001, ** alpha = 0.01, * alpha = 0.05
## Demographic Prediction
We leverage both the visual and textual content for predicting age and gender.
Prediction with Textual Content:
We employ BIBREF73 's weighted lexicon of terms that uses the dataset of 75,394 Facebook users who shared their status, age and gender. The predictive power of this lexica was evaluated on Twitter, blog, and Facebook, showing promising results BIBREF73 . Utilizing these two weighted lexicon of terms, we are predicting the demographic information (age or gender) of INLINEFORM0 (denoted by INLINEFORM1 ) using following equation: INLINEFORM2
where INLINEFORM0 is the lexicon weight of the term, and INLINEFORM1 represents the frequency of the term in the user generated INLINEFORM2 , and INLINEFORM3 measures total word count in INLINEFORM4 . As our data is biased toward young people, we report age prediction performance for each age group separately (Table TABREF42 ). Moreover, to measure the average accuracy of this model, we build a balanced dataset (keeping all the users above 23 -416 users), and then randomly sampling the same number of users from the age ranges (11,19] and (19,23]. The average accuracy of this model is 0.63 for depressed users and 0.64 for control class. Table TABREF44 illustrates the performance of gender prediction for each class. The average accuracy is 0.82 on INLINEFORM5 ground-truth dataset.
Prediction with Visual Imagery:
Inspired by BIBREF56 's approach for facial landmark localization, we use their pretrained CNN consisting of convolutional layers, including unshared and fully-connected layers, to predict gender and age from both the profile and shared images. We evaluate the performance for gender and age prediction task on INLINEFORM0 and INLINEFORM1 respectively as shown in Table TABREF42 and Table TABREF44 .
Demographic Prediction Analysis:
We delve deeper into the benefits and drawbacks of each data modality for demographic information prediction. This is crucial as the differences between language cues between age groups above age 35 tend to become smaller (see Figure FIGREF39 -A,B,C) and making the prediction harder for older people BIBREF74 . In this case, the other data modality (e.g., visual content) can play integral role as a complementary source for age inference. For gender prediction (see Table TABREF44 ), on average, the profile image-based predictor provides a more accurate prediction for both the depressed and control class (0.92 and 0.90) compared to content-based predictor (0.82). For age prediction (see Table TABREF42 ), textual content-based predictor (on average 0.60) outperforms both of the visual-based predictors (on average profile:0.51, Media:0.53).
However, not every user provides facial identity on his account (see Table TABREF21 ). We studied facial presentation for each age-group to examine any association between age-group, facial presentation and depressive behavior (see Table TABREF43 ). We can see youngsters in both depressed and control class are not likely to present their face on profile image. Less than 3% of vulnerable users between 11-19 years reveal their facial identity. Although content-based gender predictor was not as accurate as image-based one, it is adequate for population-level analysis.
## Multi-modal Prediction Framework
We use the above findings for predicting depressive behavior. Our model exploits early fusion BIBREF32 technique in feature space and requires modeling each user INLINEFORM0 in INLINEFORM1 as vector concatenation of individual modality features. As opposed to computationally expensive late fusion scheme where each modality requires a separate supervised modeling, this model reduces the learning effort and shows promising results BIBREF75 . To develop a generalizable model that avoids overfitting, we perform feature selection using statistical tests and all relevant ensemble learning models. It adds randomness to the data by creating shuffled copies of all features (shadow feature), and then trains Random Forest classifier on the extended data. Iteratively, it checks whether the actual feature has a higher Z-score than its shadow feature (See Algorithm SECREF6 and Figure FIGREF45 ) BIBREF76 .
Main each Feature INLINEFORM0 INLINEFORM1
RndForrest( INLINEFORM0 ) Calculate Imp INLINEFORM1 INLINEFORM2 Generate next hypothesis , INLINEFORM3 Once all hypothesis generated Perform Statistical Test INLINEFORM4 //Binomial Distribution INLINEFORM5 Feature is important Feature is important
Ensemble Feature Selection
Next, we adopt an ensemble learning method that integrates the predictive power of multiple learners with two main advantages; its interpretability with respect to the contributions of each feature and its high predictive power. For prediction we have INLINEFORM0 where INLINEFORM1 is a weak learner and INLINEFORM2 denotes the final prediction.
In particular, we optimize the loss function: INLINEFORM0 where INLINEFORM1 incorporates INLINEFORM2 and INLINEFORM3 regularization. In each iteration, the new INLINEFORM4 is obtained by fitting weak learner to the negative gradient of loss function. Particularly, by estimating the loss function with Taylor expansion : INLINEFORM5 where its first expression is constant, the second and the third expressions are first ( INLINEFORM6 ) and second order derivatives ( INLINEFORM7 ) of the loss. INLINEFORM8
For exploring the weak learners, assume INLINEFORM0 has k leaf nodes, INLINEFORM1 be subset of users from INLINEFORM2 belongs to the node INLINEFORM3 , and INLINEFORM4 denotes the prediction for node INLINEFORM5 . Then, for each user INLINEFORM6 belonging to INLINEFORM7 , INLINEFORM8 and INLINEFORM9 INLINEFORM10
Next, for each leaf node INLINEFORM0 , deriving w.r.t INLINEFORM1 : INLINEFORM2
and by substituting weights: INLINEFORM0
which represents the loss for fixed weak learners with INLINEFORM0 nodes. The trees are built sequentially such that each subsequent tree aims to reduce the errors of its predecessor tree. Although, the weak learners have high bias, the ensemble model produces a strong learner that effectively integrate the weak learners by reducing bias and variance (the ultimate goal of supervised models) BIBREF77 . Table TABREF48 illustrates our multimodal framework outperform the baselines for identifying depressed users in terms of average specificity, sensitivity, F-Measure, and accuracy in 10-fold cross-validation setting on INLINEFORM1 dataset. Figure FIGREF47 shows how the likelihood of being classified into the depressed class varies with each feature addition to the model for a sample user in the dataset. The prediction bar (the black bar) shows that the log-odds of prediction is 0.31, that is, the likelihood of this person being a depressed user is 57% (1 / (1 + exp(-0.3))). The figure also sheds light on the impact of each contributing feature. The waterfall charts represent how the probability of being depressed changes with the addition of each feature variable. For instance, the "Analytic thinking" of this user is considered high 48.43 (Median:36.95, Mean: 40.18) and this decreases the chance of this person being classified into the depressed group by the log-odds of -1.41. Depressed users have significantly lower "Analytic thinking" score compared to control class. Moreover, the 40.46 "Clout" score is a low value (Median: 62.22, Mean: 57.17) and it decreases the chance of being classified as depressed. With respect to the visual features, for instance, the mean and the median of 'shared_colorfulness' is 112.03 and 113 respectively. The value of 136.71 would be high; thus, it decreases the chance of being depressed for this specific user by log-odds of -0.54. Moreover, the 'profile_naturalness' of 0.46 is considered high compared to 0.36 as the mean for the depressed class which justifies pull down of the log-odds by INLINEFORM2 . For network features, for instance, 'two_hop_neighborhood' for depressed users (Mean : 84) are less than that of control users (Mean: 154), and is reflected in pulling down the log-odds by -0.27.
Baselines:
To test the efficacy of our multi-modal framework for detecting depressed users, we compare it against existing content, content-network, and image-based models (based on the aforementioned general image feature, facial presence, and facial expressions.)
| 7
|
1902.09666
|
Predicting the Type and Target of Offensive Posts in Social Media
|
# Predicting the Type and Target of Offensive Posts in Social Media
## Abstract
As offensive content has become pervasive in social media, there has been much research in identifying potentially offensive messages. However, previous work on this topic did not consider the problem as a whole, but rather focused on detecting very specific types of offensive content, e.g., hate speech, cyberbulling, or cyber-aggression. In contrast, here we target several different kinds of offensive content. In particular, we model the task hierarchically, identifying the type and the target of offensive messages in social media. For this purpose, we complied the Offensive Language Identification Dataset (OLID), a new dataset with tweets annotated for offensive content using a fine-grained three-layer annotation scheme, which we make publicly available. We discuss the main similarities and differences between OLID and pre-existing datasets for hate speech identification, aggression detection, and similar tasks. We further experiment with and we compare the performance of different machine learning models on OLID.
## Introduction
Offensive content has become pervasive in social media and a reason of concern for government organizations, online communities, and social media platforms. One of the most common strategies to tackle the problem is to train systems capable of recognizing offensive content, which then can be deleted or set aside for human moderation. In the last few years, there have been several studies published on the application of computational methods to deal with this problem. Most prior work focuses on a different aspect of offensive language such as abusive language BIBREF0 , BIBREF1 , (cyber-)aggression BIBREF2 , (cyber-)bullying BIBREF3 , BIBREF4 , toxic comments INLINEFORM0 , hate speech BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , and offensive language BIBREF11 . Prior work has focused on these aspects of offensive language in Twitter BIBREF3 , BIBREF7 , BIBREF8 , BIBREF11 , Wikipedia comments, and Facebook posts BIBREF2 .
Recently, Waseem et. al. ( BIBREF12 ) acknowledged the similarities among prior work and discussed the need for a typology that differentiates between whether the (abusive) language is directed towards a specific individual or entity or towards a generalized group and whether the abusive content is explicit or implicit. Wiegand et al. ( BIBREF11 ) followed this trend as well on German tweets. In their evaluation, they have a task to detect offensive vs not offensive tweets and a second task for distinguishing between the offensive tweets as profanity, insult, or abuse. However, no prior work has explored the target of the offensive language, which is important in many scenarios, e.g., when studying hate speech with respect to a specific target.
Therefore, we expand on these ideas by proposing a a hierarchical three-level annotation model that encompasses:
Using this annotation model, we create a new large publicly available dataset of English tweets. The key contributions of this paper are as follows:
## Related Work
Different abusive and offense language identification sub-tasks have been explored in the past few years including aggression identification, bullying detection, hate speech, toxic comments, and offensive language.
Aggression identification: The TRAC shared task on Aggression Identification BIBREF2 provided participants with a dataset containing 15,000 annotated Facebook posts and comments in English and Hindi for training and validation. For testing, two different sets, one from Facebook and one from Twitter were provided. Systems were trained to discriminate between three classes: non-aggressive, covertly aggressive, and overtly aggressive.
Bullying detection: Several studies have been published on bullying detection. One of them is the one by xu2012learning which apply sentiment analysis to detect bullying in tweets. xu2012learning use topic models to to identify relevant topics in bullying. Another related study is the one by dadvar2013improving which use user-related features such as the frequency of profanity in previous messages to improve bullying detection.
Hate speech identification: It is perhaps the most widespread abusive language detection sub-task. There have been several studies published on this sub-task such as kwok2013locate and djuric2015hate who build a binary classifier to distinguish between `clean' comments and comments containing hate speech and profanity. More recently, Davidson et al. davidson2017automated presented the hate speech detection dataset containing over 24,000 English tweets labeled as non offensive, hate speech, and profanity.
Offensive language: The GermEval BIBREF11 shared task focused on Offensive language identification in German tweets. A dataset of over 8,500 annotated tweets was provided for a course-grained binary classification task in which systems were trained to discriminate between offensive and non-offensive tweets and a second task where the organizers broke down the offensive class into three classes: profanity, insult, and abuse.
Toxic comments: The Toxic Comment Classification Challenge was an open competition at Kaggle which provided participants with comments from Wikipedia labeled in six classes: toxic, severe toxic, obscene, threat, insult, identity hate.
While each of these sub-tasks tackle a particular type of abuse or offense, they share similar properties and the hierarchical annotation model proposed in this paper aims to capture this. Considering that, for example, an insult targeted at an individual is commonly known as cyberbulling and that insults targeted at a group are known as hate speech, we pose that OLID's hierarchical annotation model makes it a useful resource for various offensive language identification sub-tasks.
## Hierarchically Modelling Offensive Content
In the OLID dataset, we use a hierarchical annotation model split into three levels to distinguish between whether language is offensive or not (A), and type (B) and target (C) of the offensive language. Each level is described in more detail in the following subsections and examples are shown in Table TABREF10 .
## Level A: Offensive language Detection
Level A discriminates between offensive (OFF) and non-offensive (NOT) tweets.
Not Offensive (NOT): Posts that do not contain offense or profanity;
Offensive (OFF): We label a post as offensive if it contains any form of non-acceptable language (profanity) or a targeted offense, which can be veiled or direct. This category includes insults, threats, and posts containing profane language or swear words.
## Level B: Categorization of Offensive Language
Level B categorizes the type of offense and two labels are used: targeted (TIN) and untargeted (INT) insults and threats.
Targeted Insult (TIN): Posts which contain an insult/threat to an individual, group, or others (see next layer);
Untargeted (UNT): Posts containing non-targeted profanity and swearing. Posts with general profanity are not targeted, but they contain non-acceptable language.
## Level C: Offensive Language Target Identification
Level C categorizes the targets of insults and threats as individual (IND), group (GRP), and other (OTH).
Individual (IND): Posts targeting an individual. It can be a a famous person, a named individual or an unnamed participant in the conversation. Insults and threats targeted at individuals are often defined as cyberbulling.
Group (GRP): The target of these offensive posts is a group of people considered as a unity due to the same ethnicity, gender or sexual orientation, political affiliation, religious belief, or other common characteristic. Many of the insults and threats targeted at a group correspond to what is commonly understood as hate speech.
Other (OTH): The target of these offensive posts does not belong to any of the previous two categories (e.g. an organization, a situation, an event, or an issue).
## Data Collection
The data included in OLID has been collected from Twitter. We retrieved the data using the Twitter API by searching for keywords and constructions that are often included in offensive messages, such as `she is' or `to:BreitBartNews'. We carried out a first round of trial annotation of 300 instances with six experts. The goal of the trial annotation was to 1) evaluate the proposed tagset; 2) evaluate the data retrieval method; and 3) create a gold standard with instances that could be used as test questions in the training and test setting annotation which was carried out using crowdsourcing. The breakdown of keywords and their offensive content in the trial data of 300 tweets is shown in Table TABREF14 . We included a left (@NewYorker) and far-right (@BreitBartNews) news accounts because there tends to be political offense in the comments. One of the best offensive keywords was tweets that were flagged as not being safe by the Twitter `safe' filter (the `-' indicates `not safe'). The vast majority of content on Twitter is not offensive so we tried different strategies to keep a reasonable number of tweets in the offensive class amounting to around 30% of the dataset including excluding some keywords that were not high in offensive content such as `they are` and `to:NewYorker`. Although `he is' is lower in offensive content we kept it as a keyword to avoid gender bias. In addition to the keywords in the trial set, we searched for more political keywords which tend to be higher in offensive content, and sampled our dataset such that 50% of the the tweets come from political keywords and 50% come from non-political keywords. In addition to the keywords `gun control', and `to:BreitbartNews', political keywords used to collect these tweets are `MAGA', `antifa', `conservative' and `liberal'. We computed Fliess' INLINEFORM0 on the trial set for the five annotators on 21 of the tweets. INLINEFORM1 is .83 for Layer A (OFF vs NOT) indicating high agreement. As to normalization and anonymization, no user metadata or Twitter IDs have been stored, and URLs and Twitter mentions have been substituted to placeholders. We follow prior work in related areas (burnap2015cyber,davidson2017automated) and annotate our data using crowdsourcing using the platform Figure Eight. We ensure data quality by: 1) we only received annotations from individuals who were experienced in the platform; and 2) we used test questions to discard annotations of individuals who did not reach a certain threshold. Each instance in the dataset was annotated by multiple annotators and inter-annotator agreement has been calculated. We first acquired two annotations for each instance. In case of 100% agreement, we considered these as acceptable annotations, and in case of disagreement, we requested more annotations until the agreement was above 66%. After the crowdsourcing annotation, we used expert adjudication to guarantee the quality of the annotation. The breakdown of the data into training and testing for the labels from each level is shown in Table TABREF15 .
## Experiments and Evaluation
We assess our dataset using traditional and deep learning methods. Our simplest model is a linear SVM trained on word unigrams. SVMs have produced state-of-the-art results for many text classification tasks BIBREF13 . We also train a bidirectional Long Short-Term-Memory (BiLSTM) model, which we adapted from the sentiment analysis system of sentimentSystem,rasooli2018cross and altered to predict offensive labels instead. It consists of (1) an input embedding layer, (2) a bidirectional LSTM layer, (3) an average pooling layer of input features. The concatenation of the LSTM's and average pool layer is passed through a dense layer and the output is passed through a softmax function. We set two input channels for the input embedding layers: pre-trained FastText embeddings BIBREF14 , as well as updatable embeddings learned by the model during training. Finally, we also apply a Convolutional Neural Network (CNN) model based on the architecture of BIBREF15 , using the same multi-channel inputs as the above BiLSTM.
Our models are trained on the training data, and evaluated by predicting the labels for the held-out test set. The distribution is described in Table TABREF15 . We evaluate and compare the models using the macro-averaged F1-score as the label distribution is highly imbalanced. Per-class Precision (P), Recall (R), and F1-score (F1), also with other averaged metrics are also reported. The models are compared against baselines of predicting all labels as the majority or minority classes.
## Offensive Language Detection
The performance on discriminating between offensive (OFF) and non-offensive (NOT) posts is reported in Table TABREF18 . We can see that all systems perform significantly better than chance, with the neural models being substantially better than the SVM. The CNN outperforms the RNN model, achieving a macro-F1 score of 0.80.
## Categorization of Offensive Language
In this experiment, the two systems were trained to discriminate between insults and threats (TIN) and untargeted (UNT) offenses, which generally refer to profanity. The results are shown in Table TABREF19 .
The CNN system achieved higher performance in this experiment compared to the BiLSTM, with a macro-F1 score of 0.69. All systems performed better at identifying target and threats (TIN) than untargeted offenses (UNT).
## Offensive Language Target Identification
The results of the offensive target identification experiment are reported in Table TABREF20 . Here the systems were trained to distinguish between three targets: a group (GRP), an individual (IND), or others (OTH). All three models achieved similar results far surpassing the random baselines, with a slight performance edge for the neural models.
The performance of all systems for the OTH class is 0. This poor performances can be explained by two main factors. First, unlike the two other classes, OTH is a heterogeneous collection of targets. It includes offensive tweets targeted at organizations, situations, events, etc. making it more challenging for systems to learn discriminative properties of this class. Second, this class contains fewer training instances than the other two. There are only 395 instances in OTH, and 1,075 in GRP, and 2,407 in IND.
## Conclusion and Future Work
This paper presents OLID, a new dataset with annotation of type and target of offensive language. OLID is the official dataset of the shared task SemEval 2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval) BIBREF16 . In OffensEval, each annotation level in OLID is an independent sub-task. The dataset contains 14,100 tweets and is released freely to the research community. To the best of our knowledge, this is the first dataset to contain annotation of type and target of offenses in social media and it opens several new avenues for research in this area. We present baseline experiments using SVMs and neural networks to identify the offensive tweets, discriminate between insults, threats, and profanity, and finally to identify the target of the offensive messages. The results show that this is a challenging task. A CNN-based sentence classifier achieved the best results in all three sub-tasks.
In future work, we would like to make a cross-corpus comparison of OLID and datasets annotated for similar tasks such as aggression identification BIBREF2 and hate speech detection BIBREF8 . This comparison is, however, far from trivial as the annotation of OLID is different.
## Acknowledgments
The research presented in this paper was partially supported by an ERAS fellowship awarded by the University of Wolverhampton.
| 13
|
1903.00058
|
Non-Parametric Adaptation for Neural Machine Translation
|
# Non-Parametric Adaptation for Neural Machine Translation
## Abstract
Neural Networks trained with gradient descent are known to be susceptible to catastrophic forgetting caused by parameter shift during the training process. In the context of Neural Machine Translation (NMT) this results in poor performance on heterogeneous datasets and on sub-tasks like rare phrase translation. On the other hand, non-parametric approaches are immune to forgetting, perfectly complementing the generalization ability of NMT. However, attempts to combine non-parametric or retrieval based approaches with NMT have only been successful on narrow domains, possibly due to over-reliance on sentence level retrieval. We propose a novel n-gram level retrieval approach that relies on local phrase level similarities, allowing us to retrieve neighbors that are useful for translation even when overall sentence similarity is low. We complement this with an expressive neural network, allowing our model to extract information from the noisy retrieved context. We evaluate our semi-parametric NMT approach on a heterogeneous dataset composed of WMT, IWSLT, JRC-Acquis and OpenSubtitles, and demonstrate gains on all 4 evaluation sets. The semi-parametric nature of our approach opens the door for non-parametric domain adaptation, demonstrating strong inference-time adaptation performance on new domains without the need for any parameter updates.
## Introduction
Over the last few years, neural sequence to sequence models BIBREF0 , BIBREF1 , BIBREF2 have revolutionized the field of machine translation by significantly improving translation quality over their phrase based counterparts BIBREF3 , BIBREF4 , BIBREF5 . With more gains arising from continued research on new neural network architectures and accompanying training techniques BIBREF6 , BIBREF7 , BIBREF8 , NMT researchers, both in industry and academia, have doubled down on their ability to train high capacity models on large corpora with gradient based optimization.
However, despite huge improvements in overall translation quality NMT has shown some glaring weaknesses, including idiom processing, and rare word or phrase translation BIBREF9 , BIBREF10 , BIBREF11 - tasks that should be easy if the model could retain learned information from individual training examples. NMT has also been shown to perform poorly when dealing with multi-domain data BIBREF12 . This `catastrophic forgetting' problem has been well-studied in traditional neural network literature, caused by parameter shift during the training process BIBREF13 , BIBREF14 . Non-parametric methods, on the other hand, are resistant to forgetting but are prone to over-fitting due to their reliance on individual training examples. We focus on a non-parametric extension to NMT, hoping to combine the generalization ability of neural networks with the eidetic memory of non-parametric methods. Given a translation query, we rely on an external retrieval mechanism to find similar source-target instances in the training corpus, which are then utilized by the model.
There has been some work on semi-parametric NMT BIBREF15 , BIBREF16 , BIBREF17 , but its effectiveness has been confined to narrow domain datasets. Existing approaches have relied on sentence level similarity metrics for retrieval, which works well for domains with high train-test overlap, but fails to retrieve useful candidates for broad domains. Even if we could find training instances with overlapping phrases it's likely that the information in most retrieved source-target pairs is noise for the purpose of translating the current query.
To retrieve useful candidates when sentence similarity is low, we use n-gram retrieval instead of sentence retrieval. This results in neighbors which have high local overlap with the source sentence, even if they are significantly different in terms of overall sentence similarity. This is intuitively similar to utilizing information from a phrase table BIBREF18 within NMT BIBREF19 , without losing the global context lost when constructing the phrase table. We also propose another simple extension using dense vectors for n-gram retrieval which allows us to exploit similarities beyond lexical overlap.
To effectively extract the signal from the noisy retrieved neighbors, we develop an extension of the approach proposed in BIBREF17 . While BIBREF17 encode the retrieved targets without any context, we incorporate information from the current and retrieved sources while encoding the retrieved target, in order to distinguish useful information from noise.
We evaluate our semi-parametric NMT approach on two tasks.
## Semi-parametric NMT
Standard approaches for Neural Machine Translation rely on seq2seq architectures BIBREF0 , BIBREF1 , where given a source sequence INLINEFORM0 and a target sequence INLINEFORM1 , the goal is to model the probability distribution, INLINEFORM2 .
Semi-parametric NMT BIBREF19 , BIBREF15 approaches this learning problem with a different formulation, by modeling INLINEFORM0 instead, where INLINEFORM1 is the set of sentence pairs where the source sentence is a neighbor of INLINEFORM2 , retrieved from the training corpus using some similarity metric. This relies on a two step approach - the retrieval stage finds training instances, INLINEFORM3 , similar to the source sentence INLINEFORM4 , and the translation stage generates the target sequence INLINEFORM5 given INLINEFORM6 and INLINEFORM7 . We follow this setup, proposing improvements to both stages in order to enhance the applicability of semi-parametric NMT to more general translation tasks.
## Retrieval Approaches
Existing approaches have proposed using off the shelf search engines for the retrieval stage. However, our objective differs from traditional information retrieval, since the goal of retrieval in semi-parametric NMT is to find neighbors which might improve translation performance, which might not correlate with maximizing sentence similarity.
Our baseline strategy relies on a sentence level similarity score, similar to those used for standard information retrieval tasks BIBREF24 . We compare this against finer-grained n-gram retrieval using the same similarity metric. We also propose a dense vector based n-gram retrieval strategy, using representations extracted from a pre-trained NMT model.
Our baseline approach relies on a simple inverse document frequency (IDF) based similarity score. We define the IDF score of any token, INLINEFORM0 , as INLINEFORM1 , where INLINEFORM2 is the number of sentence pairs in training corpus and INLINEFORM3 is the number of sentences INLINEFORM4 occurs in. Let any two sentence pairs in the corpus be INLINEFORM5 and INLINEFORM6 . Then we define the similarity between INLINEFORM7 and INLINEFORM8 by, DISPLAYFORM0
For every sentence in the training, dev and test corpora, we find the INLINEFORM0 most similar training sentence pairs and provide them as context to NMT.
Motivated by phrase based SMT, we retrieve neighbors which have high local, sub-sentence level overlap with the source sentence. We adapt our approach to retrieve n-grams instead of sentences. We note that the similarity metric defined above for sentences is equally applicable for n-gram retrieval.
Let INLINEFORM0 be a sentence. Then the set of all possible n-grams of X, for a given INLINEFORM1 , can be defined as INLINEFORM2 (also including padding at the end). To reduce the number of n-grams used to represent every sentence, we define the reduced set of n-grams for X to be INLINEFORM3 .
We represent every sentence by their reduced n-gram set. For every n-gram in INLINEFORM0 , we find the closest n-gram in the training set using the IDF similarity defined above. For each retrieved n-gram we find the corresponding sentence (In case an n-gram is present in multiple sentences, we choose one randomly). The set of neighbors of INLINEFORM1 is then the set of all sentences in the training corpus that contain an n-gram that maximizes the n-gram similarity with any n-gram in INLINEFORM2 .
To capture phrases of different lengths we use multiple n-gram widths, INLINEFORM0 . In case a sentence has already been added to the retrieved set, we find the next most similar sentence to avoid having duplicates. The number of neighbors retrieved for each source sentence is proportional to its length.
We also extend our n-gram retrieval strategy with dense vector based n-gram representations. The objective behind using a dense vector based approach is to incorporate information relevant to the translation task in the retrieval stage. We use a pre-trained Transformer Base BIBREF6 encoder trained on WMT to generate sub-word level dense representations for the sentence. The representation for each n-gram is now defined to be the mean of the representations of all its constituent sub-words. We use the INLINEFORM0 distance of n-gram representations as the retrieval criterion. Note that we use a sub-word level decomposition of sentences for dense retrieval, as compared to word-level for IDF based retrieval (i.e., n-grams are composed of sub-words instead of words).
Following the approach described for IDF based n-gram retrieval, we use multiple values of INLINEFORM0 , and remove duplicate neighbors while creating the retrieved set.
## NMT with Context Retrieval
To incorporate the retrieved neighbors, INLINEFORM0 , within the NMT model, we first encode them using Transformer layers, as described in subsection UID12 . This encoded memory is then used within the decoder via an attention mechanism, as described in subsection UID15 .
We now describe how each retrieved translation pair, INLINEFORM0 , is encoded. This architecture is illustrated in Figure FIGREF9 .
We first encode the retrieved source, INLINEFORM0 , in a Transformer layer. Apart from self-attention, we incorporate information from the encoder representation of the current source, INLINEFORM1 , using decoder style cross-attention.
The retrieved target, INLINEFORM0 , is encoded in a similar manner, attending the encoded representation of INLINEFORM1 generated in the previous step.
The encoded representations for all targets, INLINEFORM0 , are then concatenated along the time axis to form the Conditional Source Target Memory (CSTM).
We use gated multi-source attention to combine the context from the source encoder representations and the CSTM. This is similar to the gated attention employed by BIBREF17 . We use a Transformer based decoder that attends to both, the encoder outputs and the CSTM, in every cross-attention layer. The rest of the decoder architecture remains unchanged.
Let the context vectors obtained by applying multi-head attention to the source and memory, with query INLINEFORM0 be INLINEFORM1 and INLINEFORM2 respectively. Then the gated context vector, INLINEFORM3 , is given by, DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 is the scalar gating variable at time-step t, and INLINEFORM1 and INLINEFORM2 are learned parameters. These steps are illustrated in Figure FIGREF10 .
## Data and Evaluation
We compare the performance of a standard Transformer Base model and our semi-parametric NMT approach on an English-French translation task. We create a new heterogeneous dataset, constructed from a combination of the WMT training set (36M pairs), the IWSLT bilingual corpus (237k pairs), JRC-Acquis (797k pairs) and OpenSubtitles (33M pairs). For WMT, we use newstest 13 for validation and newstest 14 for test. For IWSLT, we use a combination of the test corpora from 2012-14 for validation and test 2015 for eval. For OpenSubtitles and JRC-Acquis, we create our own splits for validation and test, since no benchmark split is publicly available. After deduping, the JRC-Acquis test and validation set contain 6574 and 5121 sentence pairs respectively. The OpenSubtitles test and validation sets contain 3975 and 3488 pairs. For multi-domain training, the validation set is a concatenation of the four individual validation sets.
All datasets are tokenized with the Moses tokenizer BIBREF25 and mixed without any sampling. We use a shared vocabulary Sentence-Piece Model BIBREF26 for sub-word tokenization, with a vocabulary size of 32000 tokens. We train each model for 1M steps, and choose the best checkpoint from the last 5 checkpoints based on validation performance. BLEU scores are computed with tokenized true-cased output and references with multi-bleu.perl from Moses.
For IDF based sentence retrieval, for each sentence in the training, dev and test corpus, we use INLINEFORM0 neighbors per example during both, training and evaluation. For the N-Gram level retrieval strategies, we used INLINEFORM1 neighbors during training, and neighbors corresponding to all n-grams during decoding. This was meant to limit memory requirements and enable the model to fit on P100s during training. We used n-gram width, INLINEFORM2 , for both IDF and dense vector based n-gram retrieval approaches. For scalability reasons, we restricted the retrieval set to the in-domain training corpus, i.e. neighbors for all train, dev and test sentences in the JRC-Acquis corpus were retrieved from the JRC-Acquis training split, and similarly for the other datasets.
## Hyper-parameters and Optimization
For our baseline model we use the standard Transformer Base model BIBREF6 . For the semi-parametric model, all our hyper-parameters for attention (8 attention heads), model dimensions (512) and hidden dimensions (2048), including those used in the CSTM memory are equivalent to Transformer Base.
The Transformer baselines are trained on 16 GPUs, with the learning rate, warm-up schedule and batching scheme described in BIBREF6 . The semi-parametric models were trained on 32 GPUs with each replica split over 2 GPUs, one to train the translation model and the other for computing the CSTM. We used a conservative learning rate schedule (3, 40K) BIBREF8 to train the semi-parametric models.
We apply a dropout rate BIBREF27 of 0.1 to all inputs, residuals, attentions and ReLU connections in both models. We use Adam BIBREF28 to train all models, and apply label smoothing with an uncertainty of 0.1 BIBREF29 . In addition to the transformer layers, layer normalization BIBREF30 was applied to the output of the CSTM. All models are implemented in Tensorflow-Lingvo BIBREF31 .
## Results
We compare the test performance of a multi-domain Transformer Base and our semi-parametric model using dense vector based n-gram retrieval and CSTM in Table TABREF21 . Apart from significantly improving performance by more than 10 BLEU points on JRC-Acquis, 2-3 BLEU on OpenSubtitles and IWSLT, we notice a moderate gain of 0.5 BLEU points on WMT 14.
## Comparison of retrieval strategies
We compare the performance of all 3 retrieval strategies in Table TABREF21 . The semi-parametric model with sentence level retrieval out-performs the seq2seq model by a huge margin on JRC-Acquis and OpenSubtitles. A sample from the JRC-Acquis dataset where the semi-parametric approach improves significantly over the neural approach is included in Table TABREF22 . We notice that there is a lot of overlap between the source sentence and the retrieved source, resulting in the semi-parametric model copying large chunks from the retrieved target. However, its performance is noticeably worse on WMT and IWSLT. Based on a manual inspection of the retrieved candidates, we attribute these losses to retrieval failures. For broad domain datasets like WMT and IWSLT sentence retrieval fails to find good candidates.
Switching to n-gram level retrieval brings the WMT performance close to the seq2seq approach, and IWSLT performance to 2 BLEU points above the baseline model. Representative examples from IWSLT and WMT where n-gram retrieval improves over sentence level retrieval can be seen in Tables TABREF24 and TABREF25 . Despite the majority of the retrieved neighbor having nothing in common with the source sentence, n-gram retrieval is able to find neighbors that contain local overlaps.
Using dense n-gram retrieval allows us to move beyond lexical overlap and retrieve semantically similar n-grams even when the actual tokens are different. As a result, dense n-gram retrieval improves performance over all our models on all 4 datasets. An illustrative example from WMT is included in Table TABREF26 .
## Memory Ablation Experiments
We report the performance of the various memory ablations in Table TABREF27 . We first remove the retrieved sources, INLINEFORM0 , from the CSTM, resulting in an architecture where the encoding of a retrieved target, INLINEFORM1 , only incorporates information from the source INLINEFORM2 , represented by the row CTM in the table. This results in a clear drop in performance on all datasets. We ablate further by removing the attention to the original source INLINEFORM3 , resulting in a slightly smaller drop in performance (represented by TM). These experiments indicate that incorporating context from the sources significantly contributes to performance, by allowing the model to distinguish between relevant context and noise.
## Non-Parametric Adaptation
Using a semi-parametric formulation for MT opens up the possibility of non-parametric adaptation. The biggest advantage of this approach is the possibility of training a single massively customizable model which can be adapted to any new dataset or document at inference time, by just updating the retrieval dataset.
We evaluate our model's performance on non-parametric adaptation and compare it against a fully fine-tuned model. In this setting, we train a baseline model and a dense n-gram based semi-parametric model on the WMT training corpus. We only retrieve and train on examples from the WMT corpus during training. We use the same hyper-parameters and training approaches used for the multi-domain experiments, as in Section SECREF3 .
The baseline model is then fine-tuned independently on JRC-Acquis, OpenSubtitles and IWSLT. The semi-parametric model is adapted non-parametrically to these three datasets, without any parameter updates. Adaptation is achieved via the retrieval mechanism - while evaluating, we retrieve similar examples from their respective training datasets. To quantify headroom, we also fine-tune our semi-parametric model on each of these datasets.
The results for non-parametric adaptation experiments are documented in Table TABREF30 . We notice that the non-parametric adaptation strategy significantly out-performs the base model on all 4 datasets. More importantly, the we find that our approach is capable of adapting to both, JRC-Acquis and OpenSubtitles, via just the retrieval apparatus, and out-performs the fully fine-tuned model indicating that non-parametric adaptation might be a reasonable approach when adapting to a lot of narrow domains or documents.
In-domain fine-tuning on top of non-parametric adaptation further improves by 2 BLEU points on all datasets, increasing the gap even further with the seq2seq adapted models.
## Related Work
Tools incorporating information from individual translation pairs, or translation memories BIBREF32 , BIBREF33 , have been widely utilized by human translators in the industry. There have been a few efforts attempting to combine non-parametric methods with NMT BIBREF15 , BIBREF16 , BIBREF17 , but the key difference of our approach is the introduction of local, sub-sentence level similarity in the retrieval process, via n-gram level retrieval. Combined with our architectural improvements, motivated by the target encoder and gated attention from BIBREF17 and the extended transformer model from BIBREF34 , our semi-parametric NMT model is able to out-perform purely neural models in broad multi-domain settings.
Some works have proposed using phrase tables or the outputs of Phrase based MT within NMT BIBREF19 , BIBREF35 , BIBREF36 . While this reduces the noise present within the retrieved translation pairs, it requires training and maintaining a separate SMT system which might introduce errors of its own.
Another class of methods requires fine-tuning the entire NMT model to every instance at inference time, using retrieved examples BIBREF37 , BIBREF38 , but these approaches require running expensive gradient descent steps before every translation.
Beyond NMT, there have been a few other attempts to incorporate non-parametric approaches into neural generative models BIBREF39 , BIBREF40 , BIBREF41 . This strong trend towards combining neural generative models with non-parametric methods is an attempt to counter the weaknesses of neural networks, especially their failure to remember information from individual training instances and the diversity problem of seq2seq models BIBREF42 , BIBREF43 .
While our approach relies purely on retrieval from the training corpus, there has been quite a lot of work, especially on Question Answering, that attempts to find additional signals to perform the supervised task in the presence of external knowledge sources BIBREF44 , BIBREF45 . Retrieving information from unsupervised corpora by utilizing multilingual representations BIBREF46 might be another interesting extension of this work.
## Conclusions and Future Work
We make two major technical contributions in this work which enable us to improve the quality of semi-parametric NMT on broad domain datasets. First, we propose using n-gram retrieval, with standard Inverse Document Frequency similarity and with dense vector representations, that takes into account local sentence similarities that are critical to translation. As a result we are able to retrieve useful candidates even for broad domain tasks with little train-test overlap. Second, we propose a novel architecture to encode retrieved source-target pairs, allowing the model to distinguish useful information from noise by encoding the retrieved targets in context of the current translation task.
We demonstrate, for the first time, that semi-parametric methods can beat neural models by significant margins on multi-domain Machine Translation. By successfully training semi-parametric neural models on a broad domain dataset (WMT), we also open the door for non-parametric adaptation, showing huge improvements on new domains without any parameter updates.
While we constrain this work to retrieved context, our architecture can be utilized to incorporate information from other sources of context, including documents, bilingual dictionaries etc. Using dense representations for retrieval also allows extending semi-parametric neural methods to other input modalities, including images and speech.
With this work, we hope to motivate further investigation into semi-parametric neural models for and beyond Neural Machine Translation.
## Acknowledgments
We would like to thank Naveen Arivazhagan, Macduff Hughes, Dmitry Lepikhin, Mia Chen, Yuan Cao, Ciprian Chelba, Zhifeng Chen, Melvin Johnson and other members of the Google Brain and Google Translate teams for their useful inputs and discussions. We would also like to thank the entire Lingvo development team for their foundational contributions to this project.
| 13
|
1903.03467
|
Filling Gender&Number Gaps in Neural Machine Translation with Black-box Context Injection
|
# Filling Gender&Number Gaps in Neural Machine Translation with Black-box Context Injection
## Abstract
When translating from a language that does not morphologically mark information such as gender and number into a language that does, translation systems must"guess"this missing information, often leading to incorrect translations in the given context. We propose a black-box approach for injecting the missing information to a pre-trained neural machine translation system, allowing to control the morphological variations in the generated translations without changing the underlying model or training data. We evaluate our method on an English to Hebrew translation task, and show that it is effective in injecting the gender and number information and that supplying the correct information improves the translation accuracy in up to 2.3 BLEU on a female-speaker test set for a state-of-the-art online black-box system. Finally, we perform a fine-grained syntactic analysis of the generated translations that shows the effectiveness of our method.
## Introduction
A common way for marking information about gender, number, and case in language is morphology, or the structure of a given word in the language. However, different languages mark such information in different ways – for example, in some languages gender may be marked on the head word of a syntactic dependency relation, while in other languages it is marked on the dependent, on both, or on none of them BIBREF0 . This morphological diversity creates a challenge for machine translation, as there are ambiguous cases where more than one correct translation exists for the same source sentence. For example, while the English sentence “I love language” is ambiguous with respect to the gender of the speaker, Hebrew marks verbs for the gender of their subject and does not allow gender-neutral translation. This allows two possible Hebrew translations – one in a masculine and the other in a feminine form. As a consequence, a sentence-level translator (either human or machine) must commit to the gender of the speaker, adding information that is not present in the source. Without additional context, this choice must be done arbitrarily by relying on language conventions, world knowledge or statistical (stereotypical) knowledge.
Indeed, the English sentence “I work as a doctor” is translated into Hebrew by Google Translate using the masculine verb form oved, indicating a male speaker, while “I work as a nurse” is translated with the feminine form ovedet, indicating a female speaker (verified on March 2019). While this is still an issue, there have been recent efforts to reduce it for specific language pairs.
We present a simple black-box method to influence the interpretation chosen by an NMT system in these ambiguous cases. More concretely, we construct pre-defined textual hints about the gender and number of the speaker and the audience (the interlocutors), which we concatenate to a given input sentence that we would like to translate accordingly. We then show that a black-box NMT system makes the desired morphological decisions according to the given hint, even when no other evidence is available on the source side. While adding those hints results in additional text on the target side, we show that it is simple to remove, leaving only the desired translation.
Our method is appealing as it only requires simple pre-and-post processing of the inputs and outputs, without considering the system internals, or requiring specific annotated data and training procedure as in previous work BIBREF1 . We show that in spite of its simplicity, it is effective in resolving many of the ambiguities and improves the translation quality in up to 2.3 BLEU when given the correct hints, which may be inferred from text metadata or other sources. Finally, we perform a fine-grained syntactic analysis of the translations generated using our method which shows its effectiveness.
## Morphological Ambiguity in Translation
Different languages use different morphological features marking different properties on different elements. For example, English marks for number, case, aspect, tense, person, and degree of comparison. However, English does not mark gender on nouns and verbs. Even when a certain property is marked, languages differ in the form and location of the marking BIBREF0 . For example, marking can occur on the head of a syntactic dependency construction, on its argument, on both (requiring agreement), or on none of them. Translation systems must generate correct target-language morphology as part of the translation process. This requires knowledge of both the source-side and target-side morphology. Current state-of-the-art translation systems do capture many aspects of natural language, including morphology, when a relevant context is available BIBREF2 , BIBREF3 , but resort to “guessing” based on the training-data statistics when it is not. Complications arise when different languages convey different kinds of information in their morphological systems. In such cases, a translation system may be required to remove information available in the source sentence, or to add information not available in it, where the latter can be especially tricky.
## Black-Box Knowledge Injection
Our goal is to supply an NMT system with knowledge regarding the speaker and interlocutor of first-person sentences, in order to produce the desired target-side morphology when the information is not available in the source sentence. The approach we take in the current work is that of black-box injection, in which we attempt to inject knowledge to the input in order to influence the output of a trained NMT system, without having access to its internals or its training procedure as proposed by vanmassenhove-hardmeier-way:2018:EMNLP.
We are motivated by recent work by BIBREF4 who showed that NMT systems learn to track coreference chains when presented with sufficient discourse context. We conjecture that there are enough sentence-internal pronominal coreference chains appearing in the training data of large-scale NMT systems, such that state-of-the-art NMT systems can and do track sentence-internal coreference. We devise a wrapper method to make use of this coreference tracking ability by introducing artificial antecedents that unambiguously convey the desired gender and number properties of the speaker and audience.
More concretely, a sentence such as “I love you” is ambiguous with respect to the gender of the speaker and the gender and number of the audience. However, sentences such as “I love you, she told him” are unambiguous given the coreference groups {I, she} and {you, him} which determine I to be feminine singular and you to be masculine singular. We can thus inject the desired information by prefixing a sentence with short generic sentence fragment such as “She told him:” or “She told them that”, relying on the NMT system's coreference tracking abilities to trigger the correctly marked translation, and then remove the redundant translated prefix from the generated target sentence. We observed that using a parataxis construction (i.e. “she said to him:”) almost exclusively results in target-side parataxis as well (in 99.8% of our examples), making it easy to identify and strip the translated version from the target side. Moreover, because the parataxis construction is grammatically isolated from the rest of the sentence, it can be stripped without requiring additional changes or modification to the rest of the sentence, ensuring grammaticality.
## Experiments & Results
To demonstrate our method in a black-box setting, we focus our experiments on Google's machine translation system (GMT), accessed through its Cloud API. To test the method on real-world sentences, we consider a monologue from the stand-up comedy show “Sarah Silverman: A Speck of Dust”. The monologue consists of 1,244 English sentences, all by a female speaker conveyed to a plural, gender-neutral audience. Our parallel corpora consists of the 1,244 English sentences from the transcript, and their corresponding Hebrew translations based on the Hebrew subtitles. We translate the monologue one sentence at a time through the Google Cloud API. Eyeballing the results suggest that most of the translations use the incorrect, but default, masculine and singular forms for the speaker and the audience, respectively. We expect that by adding the relevant condition of “female speaking to an audience” we will get better translations, affecting both the gender of the speaker and the number of the audience.
To verify this, we experiment with translating the sentences with the following variations: No Prefix—The baseline translation as returned by the GMT system. “He said:”—Signaling a male speaker. We expect to further skew the system towards masculine forms. “She said:”—Signaling a female speaker and unknown audience. As this matches the actual speaker's gender, we expect an improvement in translation of first-person pronouns and verbs with first-person pronouns as subjects. “I said to them:”—Signaling an unknown speaker and plural audience. “He said to them:”—Masculine speaker and plural audience. “She said to them:”—Female speaker and plural audience—the complete, correct condition. We expect the best translation accuracy on this setup. “He/she said to him/her”—Here we set an (incorrect) singular gender-marked audience, to investigate our ability to control the audience morphology.
## Quantitative Results
We compare the different conditions by comparing BLEU BIBREF5 with respect to the reference Hebrew translations. We use the multi-bleu.perl script from the Moses toolkit BIBREF6 . Table shows BLEU scores for the different prefixes. The numbers match our expectations: Generally, providing an incorrect speaker and/or audience information decreases the BLEU scores, while providing the correct information substantially improves it - we see an increase of up to 2.3 BLEU over the baseline. We note the BLEU score improves in all cases, even when given the wrong gender of either the speaker or the audience. We hypothesise this improvement stems from the addition of the word “said” which hints the model to generate a more “spoken” language which matches the tested scenario. Providing correct information for both speaker and audience usually helps more than providing correct information to either one of them individually. The one outlier is providing “She” for the speaker and “her” for the audience. While this is not the correct scenario, we hypothesise it gives an improvement in BLEU as it further reinforces the female gender in the sentence.
## Qualitative Results
The BLEU score is an indication of how close the automated translation is to the reference translation, but does not tell us what exactly changed concerning the gender and number properties we attempt to control. We perform a finer-grained analysis focusing on the relation between the injected speaker and audience information, and the morphological realizations of the corresponding elements. We parse the translations and the references using a Hebrew dependency parser. In addition to the parse structure, the parser also performs morphological analysis and tagging of the individual tokens. We then perform the following analysis.
Speaker's Gender Effects: We search for first-person singular pronouns with subject case (ani, unmarked for gender, corresponding to the English I), and consider the gender of its governing verb (or adjectives in copular constructions such as `I am nice'). The possible genders are `masculine', `feminine' and `both', where the latter indicates a case where the none-diacriticized written form admits both a masculine and a feminine reading. We expect the gender to match the ones requested in the prefix.
Interlocutors' Gender and Number Effects: We search for second-person pronouns and consider their gender and number. For pronouns in subject position, we also consider the gender and number of their governing verbs (or adjectives in copular constructions). For a singular audience, we expect the gender and number to match the requested ones. For a plural audience, we expect the masculine-plural forms.
Results: Speaker. Figure FIGREF3 shows the result for controlling the morphological properties of the speaker ({he, she, I} said). It shows the proportion of gender-inflected verbs for the various conditions and the reference. We see that the baseline system severely under-predicts the feminine form of verbs as compared to the reference. The “He said” conditions further decreases the number of feminine verbs, while the “I said” conditions bring it back to the baseline level. Finally, the “She said” prefixes substantially increase the number of feminine-marked verbs, bringing the proportion much closer to that of the reference (though still under-predicting some of the feminine cases).
Results: Audience. The chart in Figure FIGREF3 shows the results for controlling the number of the audience (...to them vs nothing). It shows the proportion of singular vs. plural second-person pronouns on the various conditions. It shows a similar trend: the baseline system severely under-predicts the plural forms with respect to the reference translation, while adding the “to them” condition brings the proportion much closer to that of the reference.
## Comparison to vanmassenhove-hardmeier-way:2018:EMNLP
Closely related to our work, vanmassenhove-hardmeier-way:2018:EMNLP proposed a method and an English-French test set to evaluate gender-aware translation, based on the Europarl corpus BIBREF7 . We evaluate our method (using Google Translate and the given prefixes) on their test set to see whether it is applicable to another language pair and domain. Table shows the results of our approach vs. their published results and the Google Translate baseline. As may be expected, Google Translate outperforms their system as it is trained on a different corpus and may use more complex machine translation models. Using our method improves the BLEU score even further.
## Other Languages
To test our method’s outputs on multiple languages, we run our pre-and post-processing steps with Google Translate using examples we sourced from native speakers of different languages. For every example we have an English sentence and two translations in the corresponding language, one in masculine and one in feminine form. Not all examples are using the same source English sentence as different languages mark different information. Table shows that for these specific examples our method worked on INLINEFORM0 of the languages we had examples for, while for INLINEFORM1 languages both translations are masculine, and for 1 language both are feminine.
## Related Work
E17-1101 showed that given input with author traits like gender, it is possible to retain those traits in Statistical Machine Translation (SMT) models. W17-4727 showed that incorporating morphological analysis in the decoder improves NMT performance for morphologically rich languages. burlot:hal-01618387 presented a new protocol for evaluating the morphological competence of MT systems, indicating that current translation systems only manage to capture some morphological phenomena correctly. Regarding the application of constraints in NMT, N16-1005 presented a method for controlling the politeness level in the generated output. DBLP:journals/corr/FiclerG17aa showed how to guide a neural text generation system towards style and content parameters like the level of professionalism, subjective/objective, sentiment and others. W17-4811 showed that incorporating more context when translating subtitles can improve the coherence of the generated translations. Most closely to our work, vanmassenhove-hardmeier-way:2018:EMNLP also addressed the missing gender information by training proprietary models with a gender-indicating-prefix. We differ from this work by treating the problem in a black-box manner, and by addressing additional information like the number of the speaker and the gender and number of the audience.
## Conclusions
We highlight the problem of translating between languages with different morphological systems, in which the target translation must contain gender and number information that is not available in the source. We propose a method for injecting such information into a pre-trained NMT model in a black-box setting. We demonstrate the effectiveness of this method by showing an improvement of 2.3 BLEU in an English-to-Hebrew translation setting where the speaker and audience gender can be inferred. We also perform a fine-grained syntactic analysis that shows how our method enables to control the morphological realization of first and second-person pronouns, together with verbs and adjectives related to them. In future work we would like to explore automatic generation of the injected context, or the use of cross-sentence context to infer the injected information.
| 10
|
1903.07398
|
Deep Text-to-Speech System with Seq2Seq Model
|
# Deep Text-to-Speech System with Seq2Seq Model
## Abstract
Recent trends in neural network based text-to-speech/speech synthesis pipelines have employed recurrent Seq2seq architectures that can synthesize realistic sounding speech directly from text characters. These systems however have complex architectures and takes a substantial amount of time to train. We introduce several modifications to these Seq2seq architectures that allow for faster training time, and also allows us to reduce the complexity of the model architecture at the same time. We show that our proposed model can achieve attention alignment much faster than previous architectures and that good audio quality can be achieved with a model that's much smaller in size. Sample audio available at https://soundcloud.com/gary-wang-23/sets/tts-samples-for-cmpt-419.
## Introduction
Traditional text-to-speech (TTS) systems are composed of complex pipelines BIBREF0 , these often include accoustic frontends, duration model, acoustic prediction model and vocoder models. The complexity of the TTS problem coupled with the requirement for deep domain expertise means these systems are often brittle in design and results in un-natural synthesized speech.
The recent push to utilize deep, end-to-end TTS architectures BIBREF1 BIBREF2 that can be trained on <text,audio> pairs shows that deep neural networks can indeed be used to synthesize realistic sounding speech, while at the same time eliminating the need for complex sub-systems that neede to be developed and trained seperately.
The problem of TTS can be summed up as a signal-inversion problem: given a highly compressed source signal (text), we need to invert or "decompress" it into audio. This is a difficult problem as there're multi ways for the same text to be spoken. In addtion, unlike end-to-end translation or speech recognition, TTS ouptuts are continuous, and output sequences are much longer than input squences.
Recent work on neural TTS can be split into two camps, in one camp Seq2Seq models with recurrent architectures are used BIBREF1 BIBREF3 . In the other camp, full convolutional Seq2Seq models are used BIBREF2 . Our model belongs in the first of these classes using recurrent architectures. Specifically we make the following contributions:
## Related Work
Neural text-to-speech systems have garnered large research interest in the past 2 years. The first to fully explore this avenue of research was Google's tacotron BIBREF1 system. Their architecture based off the original Seq2Seq framework. In addition to encoder/decoder RNNs from the original Seq2Seq , they also included a bottleneck prenet module termed CBHG, which is composed of sets of 1-D convolution networks followed by highway residual layers. The attention mechanism follows the original Seq2Seq BIBREF7 mechanism (often termed Bahdanau attention). This is the first work to propose training a Seq2Seq model to convert text to mel spectrogram, which can then be converted to audio wav via iterative algorithms such as Griffin Lim BIBREF8 .
A parrallel work exploring Seq2Seq RNN architecture for text-to-speech was called Char2Wav BIBREF3 . This work utilized a very similar RNN-based Seq2Seq architecture, albeit without any prenet modules. The attention mechanism is guassian mixture model (GMM) attention from Alex Grave's work. Their model mapped text sequence to 80 dimension vectors used for the WORLD Vocoder BIBREF9 , which invert these vectors into audio wave.
More recently, a fully convolutional Seq2Seq architecture was investigated by Baidu Research BIBREF2 BIBREF10 . The deepvoice architecture is composed of causal 1-D convolution layers for both encoder and decoder. They utilized query-key attention similar to that from the transformer architecure BIBREF5 .
Another fully convolutional Seq2Seq architecture known as DCTTS was proposed BIBREF6 . In this architecture they employ modules composed of Causal 1-D convolution layers combined with Highway networks. In addition they introduced methods for help guide attention alignments early. As well as a forced incremental attention mechanism that ensures monotonic increasing of attention read as the model decodes during inference.
## Model Overview
The architecture of our model utilizes RNN-based Seq2Seq model for generating mel spectrogram from text. The architecture is similar to that of Tacotron 2 BIBREF4 . The generated mel spetrogram can either be inverted via iterative algorithms such as Griffin Lim, or through more complicated neural vocoder networks such as a mel spectrogram conditioned Wavenet BIBREF11 .
Figure FIGREF3 below shows the overall architecture of our model.
## Text Encoder
The encoder acts to encoder the input text sequence into a compact hidden representation, which is consumed by the decoder at every decoding step. The encoder is composed of a INLINEFORM0 -dim embedding layer that maps the input sequence into a dense vector. This is followed by a 1-layer bidirectional LSTM/GRU with INLINEFORM1 hidden dim ( INLINEFORM2 hidden dim total for both directions). two linear projections layers project the LSTM/GRU hidden output into two vectors INLINEFORM3 and INLINEFORM4 of the same INLINEFORM5 -dimension, these are the key and value vectors. DISPLAYFORM0
where INLINEFORM0 .
## Query-Key Attention
Query key attention is similar to that from transformers BIBREF5 . Given INLINEFORM0 and INLINEFORM1 from the encoder, the query, INLINEFORM2 , is computed from a linear transform of the concatenation of previous decoder-rnn hidden state, INLINEFORM3 , combined with attention-rnn hidden state, INLINEFORM4 ). DISPLAYFORM0
Given INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , the attention at each decoding step is computed by the scaled dot-product operation as: DISPLAYFORM0
Note that similar to transformers BIBREF5 , we apply a scale the dot-product by INLINEFORM0 to prevent softmax function into regions where it has extremely small gradients.
## Decoder
The decoder is an autoregressive recurrent neural network that predicts mel spectrogram from the encoded input sentence one frame at a time.
The decoder decodes the hidden representation from the encoder, with the guidance of attention. The decoder is composed of two uni-directional LSTM/GRU with INLINEFORM0 hidden dimensions. The first LSTM/GRU, called the AttentionRNN, is for computing attention-mechanism related items such as the attention query INLINEFORM1 . DISPLAYFORM0
The second LSTM/GRU, DecoderRNN, is used to compute the decoder hidden output, INLINEFORM0 . DISPLAYFORM0
A 2-layer dense prenet of dimensions (256,256) projects the previous mel spectrogram output INLINEFORM0 into hidden dimension INLINEFORM1 . Similar to Tacotron 2, the prenet acts as an information bottleneck to help produce useful representation for the downstream attention mechanism. Our model differs from Tacotron 2 in that we jointly project 5 consequetive mel frames at once into our hidden representation, which is faster and unlike Tacotron 2 which project 1 mel frame at at time.
The DecoderRNN's hidden state INLINEFORM0 is also projected to mel spectrogram INLINEFORM1 . A residual post-net composed of 2 dense layer followed by a tanh activation function also projects the same decoder hidden state INLINEFORM2 to mel spectrogram INLINEFORM3 , which is added to the linear projected mel INLINEFORM4 to produce the final mel spectrogram INLINEFORM5 . DISPLAYFORM0
A linear spectrogram INLINEFORM0 is also computed from a linear projection of the decoder hidden state INLINEFORM1 . This acts as an additional condition on the decoder hidden input. DISPLAYFORM0
A single scalar stop token is computed from a linear projection of the decoder hidden state INLINEFORM0 to a scalar, followed by INLINEFORM1 , or sigmoid function. This stop token allows the model to learn when to stop decoding during inference. During inference, if stop token output is INLINEFORM2 , we stop decoding. DISPLAYFORM0
## Training and Loss
Total loss on the model is computed as the sum of 3 component losses: 1. Mean-Squared-Error(MSE) of predicted and ground-truth mel spectrogram 2. MSE of Linear Spectrogram 3. Binary Cross Entropy Loss of our stop token. Adam optimizer is used to optimize the model with learning rate of INLINEFORM0 .
Model is trained via teacher forcing, where the ground-truth mel spectrogram is supplied at every decoding step instead of the model's own predicted mel spectrogram. To ensure the model can learn for long term sequences, teacher forcing ratio is annealed from 1.0 (full teacher forcing) to 0.2 (20 percent teacher forcing) over 300 epochs.
## Proposed Improvements
Our proposed improvements come from the observation that employing generic Seq2seq models for TTS application misses out on further optimization that can be achieved when we consider the specific problem of TTS. Specifically, we notice that in TTS, unlike in applications like machine translation, the Seq2Seq attention mechanism should be mostly monotonic. In other words, when one reads a sequence of text, it is natural to assume that the text position progress nearly linearly in time with the sequence of output mel spectrogram. With this insight, we can make 3 modifications to the model that allows us to train faster while using a a smaller model.
## Changes to Attention Mechanism
In the original Tacotron 2, the attention mechanism used was location sensitive attention BIBREF12 combined the original additive Seq2Seq BIBREF7 Bahdanau attention.
We propose to replace this attention with the simpler query-key attention from transformer model. As mentioned earlier, since for TTS the attention mechanism is an easier problem than say machine translation, we employ query-key attention as it's simple to implement and requires less parameters than the original Bahdanau attention.
## Guided Attention Mask
Following the logic above, we utilize a similar method from BIBREF6 that adds an additional guided attention loss to the overall loss objective, which acts to help the attention mechanism become monotoic as early as possible.
As seen from FIGREF24 , an attention loss mask, INLINEFORM0 , is created applies a loss to force the attention alignment, INLINEFORM1 , to be nearly diagonal. That is: DISPLAYFORM0
Where INLINEFORM0 , INLINEFORM1 is the INLINEFORM2 -th character, INLINEFORM3 is the max character length, INLINEFORM4 is the INLINEFORM5 -th mel frame, INLINEFORM6 is the max mel frame, and INLINEFORM7 is set at 0.2. This modification dramatically speed up the attention alignment and model convergence.
Figure 3 below shows the results visually. The two images are side by side comparison of the model's attention after 10k training steps. The image on the left is trained with the atention mask, and the image on the right is not. We can see that with the attention mask, clear attention alignment is achieved much faster.
## Forced Incremental Attention
During inference, the attention INLINEFORM0 occasionally skips multiple charaters or stall on the same character for multiple output frames. To make generation more robust, we modify INLINEFORM1 during inference to force it to be diagonal.
The Forced incremental attention is implemented as follows:
Given INLINEFORM0 , the position of character read at INLINEFORM1 -th time frame, where INLINEFORM2 , if INLINEFORM3 , the current attention is forcibly set to INLINEFORM4 , so that attention is incremental, i.e INLINEFORM5 .
## Experiment Dataset
The open source LJSpeech Dataset was used to train our TTS model. This dataset contains around 13k <text,audio> pairs of a single female english speaker collect from across 7 different non-fictional books. The total training data time is around 21 hours of audio.
One thing to note that since this is open-source audio recorded in a semi-professional setting, the audio quality is not as good as that of proprietary internal data from Google or Baidu. As most things with deep learning, the better the data, the better the model and results.
## Experiment Procedure
Our model was trained for 300 epochs, with batch size of 32. We used pre-trained opensource implementation of Tactron 2 (https://github.com/NVIDIA/tacotron2) as baseline comparison. Note this open-source version is trained for much longer (around 1000 epochs) however due to our limited compute we only trained our model up to 300 epochs
## Evaluation Metrics
We decide to evaluate our model against previous baselines on two fronts, Mean Opnion Score (MOS) and training speed.
Typical TTS system evaluation is done with mean opinion score (MOS). To compute this score, many samples of a TTS system is given to human evaluators and rated on a score from 1 (Bad) to 5 (Excellent). the MOS is then computed as the arithmetic mean of these score: DISPLAYFORM0
Where INLINEFORM0 are individual ratings for a given sample by N subjects.
For TTS models from google and Baidu, they utilized Amazon mechanical Turk to collect and generate MOS score from larger number of workers. However due to our limited resources, we chose to collect MOS score from friends and families (total 6 people).
For training time comparison, we choose the training time as when attention alignment start to become linear and clear. After digging through the git issues in the Tacotron 2 open-source implementation, we found a few posts where users posted their training curve and attention alignment during training (they also used the default batch size of 32). We used their training steps to roughly estimate the training time of Tacotron 2 when attention roughly aligns. For all other models the training time is not comparable as they either don't apply (e.g parametric model) or are not reported (Tacotron griffin lim, Deepvoice 3).
Direct comparison of model parameters between ours and the open-source tacotron 2, our model contains 4.5 million parameters, whereas the Tacotron 2 contains around 13 million parameters with default setting. By helping our model learn attention alignment faster, we can afford to use a smaller overall model to achieve similar quality speech quality.
## Conclusion
We introduce a new architecture for end-to-end neural text-to-speech system. Our model relies on RNN-based Seq2seq architecture with a query-key attention. We introduce novel guided attention mask to improve model training speed, and at the same time is able to reduce model parameters. This allows our model to achieve attention alignment at least 3 times faster than previous RNN-based Seq2seq models such as Tacotron 2. We also introduce forced incremental attention during synthesis to prevent attention alignment mistakes and allow model to generate coherent speech for very long sentences.
| 15
|
1903.11437
|
Using Monolingual Data in Neural Machine Translation: a Systematic Study
|
# Using Monolingual Data in Neural Machine Translation: a Systematic Study
## Abstract
Neural Machine Translation (MT) has radically changed the way systems are developed. A major difference with the previous generation (Phrase-Based MT) is the way monolingual target data, which often abounds, is used in these two paradigms. While Phrase-Based MT can seamlessly integrate very large language models trained on billions of sentences, the best option for Neural MT developers seems to be the generation of artificial parallel data through \textsl{back-translation} - a technique that fails to fully take advantage of existing datasets. In this paper, we conduct a systematic study of back-translation, comparing alternative uses of monolingual data, as well as multiple data generation procedures. Our findings confirm that back-translation is very effective and give new explanations as to why this is the case. We also introduce new data simulation techniques that are almost as effective, yet much cheaper to implement.
## Introduction
The new generation of Neural Machine Translation (NMT) systems is known to be extremely data hungry BIBREF0 . Yet, most existing NMT training pipelines fail to fully take advantage of the very large volume of monolingual source and/or parallel data that is often available. Making a better use of data is particularly critical in domain adaptation scenarios, where parallel adaptation data is usually assumed to be small in comparison to out-of-domain parallel data, or to in-domain monolingual texts. This situation sharply contrasts with the previous generation of statistical MT engines BIBREF1 , which could seamlessly integrate very large amounts of non-parallel documents, usually with a large positive effect on translation quality.
Such observations have been made repeatedly and have led to many innovative techniques to integrate monolingual data in NMT, that we review shortly. The most successful approach to date is the proposal of BIBREF2 , who use monolingual target texts to generate artificial parallel data via backward translation (BT). This technique has since proven effective in many subsequent studies. It is however very computationally costly, typically requiring to translate large sets of data. Determining the “right” amount (and quality) of BT data is another open issue, but we observe that experiments reported in the literature only use a subset of the available monolingual resources. This suggests that standard recipes for BT might be sub-optimal.
This paper aims to better understand the strengths and weaknesses of BT and to design more principled techniques to improve its effects. More specifically, we seek to answer the following questions: since there are many ways to generate pseudo parallel corpora, how important is the quality of this data for MT performance? Which properties of back-translated sentences actually matter for MT quality? Does BT act as some kind of regularizer BIBREF3 ? Can BT be efficiently simulated? Does BT data play the same role as a target-side language modeling, or are they complementary? BT is often used for domain adaptation: can the effect of having more in-domain data be sorted out from the mere increase of training material BIBREF2 ? For studies related to the impact of varying the size of BT data, we refer the readers to the recent work of BIBREF4 .
To answer these questions, we have reimplemented several strategies to use monolingual data in NMT and have run experiments on two language pairs in a very controlled setting (see § SECREF2 ). Our main results (see § SECREF4 and § SECREF5 ) suggest promising directions for efficient domain adaptation with cheaper techniques than conventional BT.
## In-domain and out-of-domain data
We are mostly interested with the following training scenario: a large out-of-domain parallel corpus, and limited monolingual in-domain data. We focus here on the Europarl domain, for which we have ample data in several languages, and use as in-domain training data the Europarl corpus BIBREF5 for two translation directions: English INLINEFORM0 German and English INLINEFORM1 French. As we study the benefits of monolingual data, most of our experiments only use the target side of this corpus. The rationale for choosing this domain is to (i) to perform large scale comparisons of synthetic and natural parallel corpora; (ii) to study the effect of BT in a well-defined domain-adaptation scenario. For both language pairs, we use the Europarl tests from 2007 and 2008 for evaluation purposes, keeping test 2006 for development. When measuring out-of-domain performance, we will use the WMT newstest 2014.
## NMT setups and performance
Our baseline NMT system implements the attentional encoder-decoder approach BIBREF6 , BIBREF7 as implemented in Nematus BIBREF8 on 4 million out-of-domain parallel sentences. For French we use samples from News-Commentary-11 and Wikipedia from WMT 2014 shared translation task, as well as the Multi-UN BIBREF9 and EU-Bookshop BIBREF10 corpora. For German, we use samples from News-Commentary-11, Rapid, Common-Crawl (WMT 2017) and Multi-UN (see table TABREF5 ). Bilingual BPE units BIBREF11 are learned with 50k merge operations, yielding vocabularies of about respectively 32k and 36k for English INLINEFORM0 French and 32k and 44k for English INLINEFORM1 German.
Both systems use 512-dimensional word embeddings and a single hidden layer with 1024 cells. They are optimized using Adam BIBREF12 and early stopped according to the validation performance. Training lasted for about three weeks on an Nvidia K80 GPU card.
Systems generating back-translated data are trained using the same out-of-domain corpus, where we simply exchange the source and target sides. They are further documented in § SECREF8 . For the sake of comparison, we also train a system that has access to a large batch of in-domain parallel data following the strategy often referred to as “fine-tuning”: upon convergence of the baseline model, we resume training with a 2M sentence in-domain corpus mixed with an equal amount of randomly selected out-of-domain natural sentences, with the same architecture and training parameters, running validation every 2000 updates with a patience of 10. Since BPE units are selected based only on the out-of-domain statistics, fine-tuning is performed on sentences that are slightly longer (ie. they contain more units) than for the initial training. This system defines an upper-bound of the translation performance and is denoted below as natural.
Our baseline and topline results are in Table TABREF6 , where we measure translation performance using BLEU BIBREF13 , BEER BIBREF14 (higher is better) and characTER BIBREF15 (smaller is better). As they are trained from much smaller amounts of data than current systems, these baselines are not quite competitive to today's best system, but still represent serious baselines for these datasets. Given our setups, fine-tuning with in-domain natural data improves BLEU by almost 4 points for both translation directions on in-domain tests; it also improves, albeit by a smaller margin, the BLEU score of the out-of-domain tests.
## Using artificial parallel data in NMT
A simple way to use monolingual data in MT is to turn it into synthetic parallel data and let the training procedure run as usual BIBREF16 . In this section, we explore various ways to implement this strategy. We first reproduce results of BIBREF2 with BT of various qualities, that we then analyze thoroughly.
## The quality of Back-Translation
BT requires the availability of an MT system in the reverse translation direction. We consider here three MT systems of increasing quality:
backtrans-bad: this is a very poor SMT system trained using only 50k parallel sentences from the out-of-domain data, and no additional monolingual data. For this system as for the next one, we use Moses BIBREF17 out-of-the-box, computing alignments with Fastalign BIBREF18 , with a minimal pre-processing (basic tokenization). This setting provides us with a pessimistic estimate of what we could get in low-resource conditions.
backtrans-good: these are much larger SMT systems, which use the same parallel data as the baseline NMTs (see § SECREF4 ) and all the English monolingual data available for the WMT 2017 shared tasks, totalling approximately 174M sentences. These systems are strong, yet relatively cheap to build.
backtrans-nmt: these are the best NMT systems we could train, using settings that replicate the forward translation NMTs.
Note that we do not use any in-domain (Europarl) data to train these systems. Their performance is reported in Table TABREF7 , where we observe a 12 BLEU points gap between the worst and best systems (for both languages).
As noted eg. in BIBREF19 , BIBREF20 , artificial parallel data obtained through forward-translation (FT) can also prove advantageous and we also consider a FT system (fwdtrans-nmt): in this case the target side of the corpus is artificial and is generated using the baseline NMT applied to a natural source.
Our results (see Table TABREF6 ) replicate the findings of BIBREF2 : large gains can be obtained from BT (nearly INLINEFORM0 BLEU in French and German); better artificial data yields better translation systems. Interestingly, our best Moses system is almost as good as the NMT and an order of magnitude faster to train. Improvements obtained with the bad system are much smaller; contrary to the better MTs, this system is even detrimental for the out-of-domain test.
Gains with forward translation are significant, as in BIBREF21 , albeit about half as good as with BT, and result in small improvements for the in-domain and for the out-of-domain tests. Experiments combining forward and backward translation (backfwdtrans-nmt), each using a half of the available artificial data, do not outperform the best BT results.
We finally note the large remaining difference between BT data and natural data, even though they only differ in their source side. This shows that at least in our domain-adaptation settings, BT does not really act as a regularizer, contrarily to the findings of BIBREF4 , BIBREF11 . Figure FIGREF13 displays the learning curves of these two systems. We observe that backtrans-nmt improves quickly in the earliest updates and then stays horizontal, whereas natural continues improving, even after 400k updates. Therefore BT does not help to avoid overfitting, it actually encourages it, which may be due “easier” training examples (cf. § SECREF15 ).
## Properties of back-translated data
Comparing the natural and artificial sources of our parallel data wrt. several linguistic and distributional properties, we observe that (see Fig. FIGREF21 - FIGREF22 ):
artificial sources are on average shorter than natural ones: when using BT, cases where the source is shorter than the target are rarer; cases when they have the same length are more frequent.
automatic word alignments between artificial sources tend to be more monotonic than when using natural sources, as measured by the average Kendall INLINEFORM0 of source-target alignments BIBREF22 : for French-English the respective numbers are 0.048 (natural) and 0.018 (artificial); for German-English 0.068 and 0.053. Using more monotonic sentence pairs turns out to be a facilitating factor for NMT, as also noted by BIBREF20 .
syntactically, artificial sources are simpler than real data; We observe significant differences in the distributions of tree depths.
distributionally, plain word occurrences in artificial sources are more concentrated; this also translates into both a slower increase of the number of types wrt. the number of sentences and a smaller number of rare events.
The intuition is that properties (i) and (ii) should help translation as compared to natural source, while property (iv) should be detrimental. We checked (ii) by building systems with only 10M words from the natural parallel data selecting these data either randomly or based on the regularity of their word alignments. Results in Table TABREF23 show that the latter is much preferable for the overall performance. This might explain that the mostly monotonic BT from Moses are almost as good as the fluid BT from NMT and that both boost the baseline.
## Stupid Back-Translation
We now analyze the effect of using much simpler data generation schemes, which do not require the availability of a backward translation engine.
## Setups
We use the following cheap ways to generate pseudo-source texts:
copy: in this setting, the source side is a mere copy of the target-side data. Since the source vocabulary of the NMT is fixed, copying the target sentences can cause the occurrence of OOVs. To avoid this situation, BIBREF24 decompose the target words into source-side units to make the copy look like source sentences. Each OOV found in the copy is split into smaller units until all the resulting chunks are in the source vocabulary.
copy-marked: another way to integrate copies without having to deal with OOVs is to augment the source vocabulary with a copy of the target vocabulary. In this setup, BIBREF25 ensure that both vocabularies never overlap by marking the target word copies with a special language identifier. Therefore the English word resume cannot be confused with the homographic French word, which is marked @fr@resume.
copy-dummies: instead of using actual copies, we replace each word with “dummy” tokens. We use this unrealistic setup to observe the training over noisy and hardly informative source sentences.
We then use the procedures described in § SECREF4 , except that the pseudo-source embeddings in the copy-marked setup are pretrained for three epochs on the in-domain data, while all remaining parameters are frozen. This prevents random parameters from hurting the already trained model.
## Copy+marking+noise is not so stupid
We observe that the copy setup has only a small impact on the English-French system, for which the baseline is already strong. This is less true for English-German where simple copies yield a significant improvement. Performance drops for both language pairs in the copy-dummies setup.
We achieve our best gains with the copy-marked setup, which is the best way to use a copy of the target (although the performance on the out-of-domain tests is at most the same as the baseline). Such gains may look surprising, since the NMT model does not need to learn to translate but only to copy the source. This is indeed what happens: to confirm this, we built a fake test set having identical source and target side (in French). The average cross-entropy for this test set is 0.33, very close to 0, to be compared with an average cost of 58.52 when we process an actual source (in English). This means that the model has learned to copy words from source to target with no difficulty, even for sentences not seen in training. A follow-up question is whether training a copying task instead of a translation task limits the improvement: would the NMT learn better if the task was harder? To measure this, we introduce noise in the target sentences copied onto the source, following the procedure of BIBREF26 : it deletes random words and performs a small random permutation of the remaining words. Results (+ Source noise) show no difference for the French in-domain test sets, but bring the out-of-domain score to the level of the baseline. Finally, we observe a significant improvement on German in-domain test sets, compared to the baseline (about +1.5 BLEU). This last setup is even almost as good as the backtrans-nmt condition (see § SECREF8 ) for German. This shows that learning to reorder and predict missing words can more effectively serve our purposes than simply learning to copy.
## Towards more natural pseudo-sources
Integrating monolingual data into NMT can be as easy as copying the target into the source, which already gives some improvement; adding noise makes things even better. We now study ways to make pseudo-sources look more like natural data, using the framework of Generative Adversarial Networks (GANs) BIBREF27 , an idea borrowed from BIBREF26 .
## GAN setups
In our setups, we use a marked target copy, viewed as a fake source, which a generator encodes so as to fool a discriminator trained to distinguish a fake from a natural source. Our architecture contains two distinct encoders, one for the natural source and one for the pseudo-source. The latter acts as the generator ( INLINEFORM0 ) in the GAN framework, computing a representation of the pseudo-source that is then input to a discriminator ( INLINEFORM1 ), which has to sort natural from artificial encodings. INLINEFORM2 assigns a probability of a sentence being natural.
During training, the cost of the discriminator is computed over two batches, one with natural (out-of-domain) sentences INLINEFORM0 and one with (in-domain) pseudo-sentences INLINEFORM1 . The discriminator is a bidirectional-Recurrent Neural Network (RNN) of dimension 1024. Averaged states are passed to a single feed-forward layer, to which a sigmoid is applied. It inputs encodings of natural ( INLINEFORM2 ) and pseudo-sentences ( INLINEFORM3 ) and is trained to optimize:
INLINEFORM0
INLINEFORM0 's parameters are updated to maximally fool INLINEFORM1 , thus the loss INLINEFORM2 : INLINEFORM3
Finally, we keep the usual MT objective. ( INLINEFORM0 is a real or pseudo-sentence): INLINEFORM1
We thus need to train three sets of parameters: INLINEFORM0 and INLINEFORM1 (MT parameters), with INLINEFORM2 . The pseudo-source encoder and embeddings are updated wrt. both INLINEFORM3 and INLINEFORM4 . Following BIBREF28 , INLINEFORM5 is updated only when INLINEFORM6 's accuracy exceeds INLINEFORM7 . On the other hand, INLINEFORM8 is not updated when its accuracy exceeds INLINEFORM9 . At each update, two batches are generated for each type of data, which are encoded with the real or pseudo-encoder. The encoder outputs serve to compute INLINEFORM10 and INLINEFORM11 . Finally, the pseudo-source is encoded again (once INLINEFORM12 is updated), both encoders are plugged into the translation model and the MT cost is back-propagated down to the real and pseudo-word embeddings. Pseudo-encoder and discriminator parameters are pre-trained for 10k updates. At test time, the pseudo-encoder is ignored and inference is run as usual.
## GANs can help
Results are in Table TABREF32 , assuming the same fine-tuning procedure as above. On top of the copy-marked setup, our GANs do not provide any improvement in both language pairs, with the exception of a small improvement for English-French on the out-of-domain test, which we understand as a sign that the model is more robust to domain variations, just like when adding pseudo-source noise. When combined with noise, the French model yields the best performance we could obtain with stupid BT on the in-domain tests, at least in terms of BLEU and BEER. On the News domain, we remain close to the baseline level, with slight improvements in German.
A first observation is that this method brings stupid BT models closer to conventional BT, at a greatly reduced computational cost. While French still remains 0.4 to 1.0 BLEU below very good backtranslation, both approaches are in the same ballpark for German - may be because BTs are better for the former system than for the latter.
Finally note that the GAN architecture has two differences with basic copy-marked: (a) a distinct encoder for real and pseudo-sentence; (b) a different training regime for these encoders. To sort out the effects of (a) and (b), we reproduce the GAN setup with BT sentences, instead of copies. Using a separate encoder for the pseudo-source in the backtrans-nmt setup can be detrimental to performance (see Table TABREF32 ): translation degrades in French for all metrics. Adding GANs on top of the pseudo-encoder was not able to make up for the degradation observed in French, but allowed the German system to slightly outperform backtrans-nmt. Even though this setup is unrealistic and overly costly, it shows that GANs are actually helping even good systems.
## Using Target Language Models
In this section, we compare the previous methods with the use of a target side Language Model (LM). Several proposals exist in the literature to integrate LMs in NMT: for instance, BIBREF3 strengthen the decoder by integrating an extra, source independent, RNN layer in a conventional NMT architecture. Training is performed either with parallel, or monolingual data. In the latter case, word prediction only relies on the source independent part of the network.
## LM Setup
We have followed BIBREF29 and reimplemented their deep-fusion technique. It requires to first independently learn a RNN-LM on the in-domain target data with a cross-entropy objective; then to train the optimal combination of the translation and the language models by adding the hidden state of the RNN-LM as an additional input to the softmax layer of the decoder. Our RNN-LMs are trained using dl4mt with the target side of the parallel data and the Europarl corpus (about 6M sentences for both French and German), using a one-layer GRU with the same dimension as the MT decoder (1024).
## LM Results
Results are in Table TABREF33 . They show that deep-fusion hardly improves the Europarl results, while we obtain about +0.6 BLEU over the baseline on newstest-2014 for both languages. deep-fusion differs from stupid BT in that the model is not directly optimized on the in-domain data, but uses the LM trained on Europarl to maximize the likelihood of the out-of-domain training data. Therefore, no specific improvement is to be expected in terms of domain adaptation, and the performance increases in the more general domain. Combining deep-fusion and copy-marked + noise + GANs brings slight improvements on the German in-domain test sets, and performance out of the domain remains near the baseline level.
## Re-analyzing the effects of BT
As a follow up of previous discussions, we analyze the effect of BT on the internals of the network. Arguably, using a copy of the target sentence instead of a natural source should not be helpful for the encoder, but is it also the case with a strong BT? What are the effects on the attention model?
## Parameter freezing protocol
To investigate these questions, we run the same fine-tuning using the copy-marked, backtrans-nmt and backtrans-nmt setups. Note that except for the last one, all training scenarios have access to same target training data. We intend to see whether the overall performance of the NMT system degrades when we selectively freeze certain sets of parameters, meaning that they are not updated during fine-tuning.
## Results
BLEU scores are in Table TABREF39 . The backtrans-nmt setup is hardly impacted by selective updates: updating the only decoder leads to a degradation of at most 0.2 BLEU. For copy-marked, we were not able to freeze the source embeddings, since these are initialized when fine-tuning begins and therefore need to be trained. We observe that freezing the encoder and/or the attention parameters has no impact on the English-German system, whereas it slightly degrades the English-French one. This suggests that using artificial sources, even of the poorest quality, has a positive impact on all the components of the network, which makes another big difference with the LM integration scenario.
The largest degradation is for natural, where the model is prevented from learning from informative source sentences, which leads to a decrease of 0.4 to over 1.0 BLEU. We assume from these experiments that BT impacts most of all the decoder, and learning to encode a pseudo-source, be it a copy or an actual back-translation, only marginally helps to significantly improve the quality. Finally, in the fwdtrans-nmt setup, freezing the decoder does not seem to harm learning with a natural source.
## Related work
The literature devoted to the use of monolingual data is large, and quickly expanding. We already alluded to several possible ways to use such data: using back- or forward-translation or using a target language model. The former approach is mostly documented in BIBREF2 , and recently analyzed in BIBREF19 , which focus on fully artificial settings as well as pivot-based artificial data; and BIBREF4 , which studies the effects of increasing the size of BT data. The studies of BIBREF20 , BIBREF19 also consider forward translation and BIBREF21 expand these results to domain adaptation scenarios. Our results are complementary to these earlier studies.
As shown above, many alternatives to BT exist. The most obvious is to use target LMs BIBREF3 , BIBREF29 , as we have also done here; but attempts to improve the encoder using multi-task learning also exist BIBREF30 .
This investigation is also related to recent attempts to consider supplementary data with a valid target side, such as multi-lingual NMT BIBREF31 , where source texts in several languages are fed in the same encoder-decoder architecture, with partial sharing of the layers. This is another realistic scenario where additional resources can be used to selectively improve parts of the model.
Round trip training is another important source of inspiration, as it can be viewed as a way to use BT to perform semi-unsupervised BIBREF32 or unsupervised BIBREF33 training of NMT. The most convincing attempt to date along these lines has been proposed by BIBREF26 , who propose to use GANs to mitigate the difference between artificial and natural data.
## Conclusion
In this paper, we have analyzed various ways to integrate monolingual data in an NMT framework, focusing on their impact on quality and domain adaptation. While confirming the effectiveness of BT, our study also proposed significantly cheaper ways to improve the baseline performance, using a slightly modified copy of the target, instead of its full BT. When no high quality BT is available, using GANs to make the pseudo-source sentences closer to natural source sentences is an efficient solution for domain adaptation.
To recap our answers to our initial questions: the quality of BT actually matters for NMT (cf. § SECREF8 ) and it seems that, even though artificial source are lexically less diverse and syntactically complex than real sentence, their monotonicity is a facilitating factor. We have studied cheaper alternatives and found out that copies of the target, if properly noised (§ SECREF4 ), and even better, if used with GANs, could be almost as good as low quality BTs (§ SECREF5 ): BT is only worth its cost when good BT can be generated. Finally, BT seems preferable to integrating external LM - at least in our data condition (§ SECREF6 ). Further experiments with larger LMs are needed to confirm this observation, and also to evaluate the complementarity of both strategies. More work is needed to better understand the impact of BT on subparts of the network (§ SECREF7 ).
In future work, we plan to investigate other cheap ways to generate artificial data. The experimental setup we proposed may also benefit from a refining of the data selection strategies to focus on the most useful monolingual sentences.
| 20
|
1904.00648
|
Recognizing Musical Entities in User-generated Content
|
# Recognizing Musical Entities in User-generated Content
## Abstract
Recognizing Musical Entities is important for Music Information Retrieval (MIR) since it can improve the performance of several tasks such as music recommendation, genre classification or artist similarity. However, most entity recognition systems in the music domain have concentrated on formal texts (e.g. artists' biographies, encyclopedic articles, etc.), ignoring rich and noisy user-generated content. In this work, we present a novel method to recognize musical entities in Twitter content generated by users following a classical music radio channel. Our approach takes advantage of both formal radio schedule and users' tweets to improve entity recognition. We instantiate several machine learning algorithms to perform entity recognition combining task-specific and corpus-based features. We also show how to improve recognition results by jointly considering formal and user-generated content
## Introduction
The increasing use of social media and microblogging services has broken new ground in the field of Information Extraction (IE) from user-generated content (UGC). Understanding the information contained in users' content has become one of the main goal for many applications, due to the uniqueness and the variety of this data BIBREF0 . However, the highly informal and noisy status of these sources makes it difficult to apply techniques proposed by the NLP community for dealing with formal and structured content BIBREF1 .
In this work, we analyze a set of tweets related to a specific classical music radio channel, BBC Radio 3, interested in detecting two types of musical named entities, Contributor and Musical Work.
The method proposed makes use of the information extracted from the radio schedule for creating links between users' tweets and tracks broadcasted. Thanks to this linking, we aim to detect when users refer to entities included into the schedule. Apart from that, we consider a series of linguistic features, partly taken from the NLP literature and partly specifically designed for this task, for building statistical models able to recognize the musical entities. To that aim, we perform several experiments with a supervised learning model, Support Vector Machine (SVM), and a recurrent neural network architecture, a bidirectional LSTM with a CRF layer (biLSTM-CRF).
The contributions in this work are summarized as follows:
The paper is structured as follows. In Section 2, we present a review of the previous works related to Named Entity Recognition, focusing on its application on UGC and MIR. Afterwards, in Section 3 it is presented the methodology of this work, describing the dataset and the method proposed. In Section 4, the results obtained are shown. Finally, in Section 5 conclusions are discussed.
## Related Work
Named Entity Recognition (NER), or alternatively Named Entity Recognition and Classification (NERC), is the task of detecting entities in an input text and to assign them to a specific class. It starts to be defined in the early '80, and over the years several approaches have been proposed BIBREF2 . Early systems were based on handcrafted rule-based algorithms, while recently several contributions by Machine Learning scientists have helped in integrating probabilistic models into NER systems.
In particular, new developments in neural architectures have become an important resource for this task. Their main advantages are that they do not need language-specific knowledge resources BIBREF3 , and they are robust to the noisy and short nature of social media messages BIBREF4 . Indeed, according to a performance analysis of several Named Entity Recognition and Linking systems presented in BIBREF5 , it has been found that poor capitalization is one of the main issues when dealing with microblog content. Apart from that, typographic errors and the ubiquitous occurrence of out-of-vocabulary (OOV) words also cause drops in NER recall and precision, together with shortenings and slang, particularly pronounced in tweets.
Music Information Retrieval (MIR) is an interdisciplinary field which borrows tools of several disciplines, such as signal processing, musicology, machine learning, psychology and many others, for extracting knowledge from musical objects (be them audio, texts, etc.) BIBREF6 . In the last decade, several MIR tasks have benefited from NLP, such as sound and music recommendation BIBREF7 , automatic summary of song review BIBREF8 , artist similarity BIBREF9 and genre classification BIBREF10 .
In the field of IE, a first approach for detecting musical named entities from raw text, based on Hidden Markov Models, has been proposed in BIBREF11 . In BIBREF12 , the authors combine state-of-the-art Entity Linking (EL) systems to tackle the problem of detecting musical entities from raw texts. The method proposed relies on the argumentum ad populum intuition, so if two or more different EL systems perform the same prediction in linking a named entity mention, the more likely this prediction is to be correct. In detail, the off-the-shelf systems used are: DBpedia Spotlight BIBREF13 , TagMe BIBREF14 , Babelfy BIBREF15 . Moreover, a first Musical Entity Linking, MEL has been presented in BIBREF16 which combines different state-of-the-art NLP libraries and SimpleBrainz, an RDF knowledge base created from MusicBrainz after a simplification process.
Furthermore, Twitter has also been at the center of many studies done by the MIR community. As example, for building a music recommender system BIBREF17 analyzes tweets containing keywords like nowplaying or listeningto. In BIBREF9 , a similar dataset it is used for discovering cultural listening patterns. Publicly available Twitter corpora built for MIR investigations have been created, among others the Million Musical Tweets dataset BIBREF18 and the #nowplaying dataset BIBREF19 .
## Methodology
We propose a hybrid method which recognizes musical entities in UGC using both contextual and linguistic information. We focus on detecting two types of entities: Contributor: person who is related to a musical work (composer, performer, conductor, etc). Musical Work: musical composition or recording (symphony, concerto, overture, etc).
As case study, we have chosen to analyze tweets extracted from the channel of a classical music radio, BBC Radio 3. The choice to focus on classical music has been mostly motivated by the particular discrepancy between the informal language used in the social platform and the formal nomenclature of contributors and musical works. Indeed, users when referring to a musician or to a classical piece in a tweet, rarely use the full name of the person or of the work, as shown in Table 2.
We extract information from the radio schedule for recreating the musical context to analyze user-generated tweets, detecting when they are referring to a specific work or contributor recently played. We manage to associate to every track broadcasted a list of entities, thanks to the tweets automatically posted by the BBC Radio3 Music Bot, where it is described the track actually on air in the radio. In Table 3, examples of bot-generated tweets are shown.
Afterwards, we detect the entities on the user-generated content by means of two methods: on one side, we use the entities extracted from the radio schedule for generating candidates entities in the user-generated tweets, thanks to a matching algorithm based on time proximity and string similarity. On the other side, we create a statistical model capable of detecting entities directly from the UGC, aimed to model the informal language of the raw texts. In Figure 1, an overview of the system proposed is presented.
## Dataset
In May 2018, we crawled Twitter using the Python library Tweepy, creating two datasets on which Contributor and Musical Work entities have been manually annotated, using IOB tags.
The first set contains user-generated tweets related to the BBC Radio 3 channel. It represents the source of user-generated content on which we aim to predict the named entities. We create it filtering the messages containing hashtags related to BBC Radio 3, such as #BBCRadio3 or #BBCR3. We obtain a set of 2,225 unique user-generated tweets. The second set consists of the messages automatically generated by the BBC Radio 3 Music Bot. This set contains 5,093 automatically generated tweets, thanks to which we have recreated the schedule.
In Table 4, the amount of tokens and relative entities annotated are reported for the two datasets. For evaluation purposes, both sets are split in a training part (80%) and two test sets (10% each one) randomly chosen. Within the user-generated corpora, entities annotated are only about 5% of the whole amount of tokens. In the case of the automatically generated tweets, the percentage is significantly greater and entities represent about the 50%.
## NER system
According to the literature reviewed, state-of-the-art NER systems proposed by the NLP community are not tailored to detect musical entities in user-generated content. Consequently, our first objective has been to understand how to adapt existing systems for achieving significant results in this task.
In the following sections, we describe separately the features, the word embeddings and the models considered. All the resources used are publicy available.
We define a set of features for characterizing the text at the token level. We mix standard linguistic features, such as Part-Of-Speech (POS) and chunk tag, together with several gazetteers specifically built for classical music, and a series of features representing tokens' left and right context. For extracting the POS and the chunk tag we use the Python library twitter_nlp, presented in BIBREF1 .
In total, we define 26 features for describing each token: 1)POS tag; 2)Chunk tag; 3)Position of the token within the text, normalized between 0 and 1; 4)If the token starts with a capital letter; 5)If the token is a digit. Gazetteers: 6)Contributor first names; 7)Contributor last names; 8)Contributor types ("soprano", "violinist", etc.); 9)Classical work types ("symphony", "overture", etc.); 10)Musical instruments; 11)Opus forms ("op", "opus"); 12)Work number forms ("no", "number"); 13)Work keys ("C", "D", "E", "F" , "G" , "A", "B", "flat", "sharp"); 14)Work Modes ("major", "minor", "m"). Finally, we complete the tokens' description including as token's features the surface form, the POS and the chunk tag of the previous and the following two tokens (12 features).
We consider two sets of GloVe word embeddings BIBREF20 for training the neural architecture, one pre-trained with 2B of tweets, publicy downloadable, one trained with a corpora of 300K tweets collected during the 2014-2017 BBC Proms Festivals and disjoint from the data used in our experiments.
The first model considered for this task has been the John Platt's sequential minimal optimization algorithm for training a support vector classifier BIBREF21 , implemented in WEKA BIBREF22 . Indeed, in BIBREF23 results shown that SVM outperforms other machine learning models, such as Decision Trees and Naive Bayes, obtaining the best accuracy when detecting named entities from the user-generated tweets.
However, recent advances in Deep Learning techniques have shown that the NER task can benefit from the use of neural architectures, such as biLSTM-networks BIBREF3 , BIBREF4 . We use the implementation proposed in BIBREF24 for conducting three different experiments. In the first, we train the model using only the word embeddings as feature. In the second, together with the word embeddings we use the POS and chunk tag. In the third, all the features previously defined are included, in addition to the word embeddings. For every experiment, we use both the pre-trained embeddings and the ones that we created with our Twitter corpora. In section 4, results obtained from the several experiments are reported.
## Schedule matching
The bot-generated tweets present a predefined structure and a formal language, which facilitates the entities detection. In this dataset, our goal is to assign to each track played on the radio, represented by a tweet, a list of entities extracted from the tweet raw text. For achieving that, we experiment with the algorithms and features presented previously, obtaining an high level of accuracy, as presented in section 4. The hypothesis considered is that when a radio listener posts a tweet, it is possible that she is referring to a track which has been played a relatively short time before. In this cases, we want to show that knowing the radio schedule can help improving the results when detecting entities.
Once assigned a list of entities to each track, we perform two types of matching. Firstly, within the tracks we identify the ones which have been played in a fixed range of time (t) before and after the generation of the user's tweet. Using the resulting tracks, we create a list of candidates entities on which performing string similarity. The score of the matching based on string similarity is computed as the ratio of the number of tokens in common between an entity and the input tweet, and the total number of token of the entity: DISPLAYFORM0
In order to exclude trivial matches, tokens within a list of stop words are not considered while performing string matching. The final score is a weighted combination of the string matching score and the time proximity of the track, aimed to enhance matches from tracks played closer to the time when the user is posting the tweet.
The performance of the algorithm depends, apart from the time proximity threshold t, also on other two thresholds related to the string matching, one for the Musical Work (w) and one for the Contributor (c) entities. It has been necessary for avoiding to include candidate entities matched against the schedule with a low score, often source of false positives or negatives. Consequently, as last step Contributor and Musical Work candidates entities with respectively a string matching score lower than c and w, are filtered out. In Figure 2, an example of Musical Work entity recognized in an user-generated tweet using the schedule information is presented.
The entities recognized from the schedule matching are joined with the ones obtained directly from the statistical models. In the joined results, the criteria is to give priority to the entities recognized from the machine learning techniques. If they do not return any entities, the entities predicted by the schedule matching are considered. Our strategy is justified by the poorer results obtained by the NER based only on the schedule matching, compared to the other models used in the experiments, to be presented in the next section.
## Results
The performances of the NER experiments are reported separately for three different parts of the system proposed.
Table 6 presents the comparison of the various methods while performing NER on the bot-generated corpora and the user-generated corpora. Results shown that, in the first case, in the training set the F1 score is always greater than 97%, with a maximum of 99.65%. With both test sets performances decrease, varying between 94-97%. In the case of UGC, comparing the F1 score we can observe how performances significantly decrease. It can be considered a natural consequence of the complex nature of the users' informal language in comparison to the structured message created by the bot.
In Table 7, results of the schedule matching are reported. We can observe how the quality of the linking performed by the algorithm is correlated to the choice of the three thresholds. Indeed, the Precision score increase when the time threshold decrease, admitting less candidates as entities during the matching, and when the string similarity thresholds increase, accepting only candidates with an higher degree of similarity. The behaviour of the Recall score is inverted.
Finally, we test the impact of using the schedule matching together with a biLSTM-CRF network. In this experiment, we consider the network trained using all the features proposed, and the embeddings not pre-trained. Table 8 reports the results obtained. We can observe how generally the system benefits from the use of the schedule information. Especially in the testing part, where the neural network recognizes with less accuracy, the explicit information contained in the schedule can be exploited for identifying the entities at which users are referring while listening to the radio and posting the tweets.
## Conclusion
We have presented in this work a novel method for detecting musical entities from user-generated content, modelling linguistic features with statistical models and extracting contextual information from a radio schedule. We analyzed tweets related to a classical music radio station, integrating its schedule to connect users' messages to tracks broadcasted. We focus on the recognition of two kinds of entities related to the music field, Contributor and Musical Work.
According to the results obtained, we have seen a pronounced difference between the system performances when dealing with the Contributor instead of the Musical Work entities. Indeed, the former type of entity has been shown to be more easily detected in comparison to the latter, and we identify several reasons behind this fact. Firstly, Contributor entities are less prone to be shorten or modified, while due to their longness, Musical Work entities often represent only a part of the complete title of a musical piece. Furthermore, Musical Work titles are typically composed by more tokens, including common words which can be easily misclassified. The low performances obtained in the case of Musical Work entities can be a consequences of these observations. On the other hand, when referring to a Contributor users often use only the surname, but in most of the cases it is enough for the system to recognizing the entities.
From the experiments we have seen that generally the biLSTM-CRF architecture outperforms the SVM model. The benefit of using the whole set of features is evident in the training part, but while testing the inclusion of the features not always leads to better results. In addition, some of the features designed in our experiments are tailored to the case of classical music, hence they might not be representative if applied to other fields. We do not exclude that our method can be adapted for detecting other kinds of entity, but it might be needed to redefine the features according to the case considered. Similarly, it has not been found a particular advantage of using the pre-trained embeddings instead of the one trained with our corpora. Furthermore, we verified the statistical significance of our experiment by using Wilcoxon Rank-Sum Test, obtaining that there have been not significant difference between the various model considered while testing.
The information extracted from the schedule also present several limitations. In fact, the hypothesis that a tweet is referring to a track broadcasted is not always verified. Even if it is common that radios listeners do comments about tracks played, or give suggestion to the radio host about what they would like to listen, it is also true that they might refer to a Contributor or Musical Work unrelated to the radio schedule.
| 8
|
1904.04055
|
Evaluating KGR10 Polish word embeddings in the recognition of temporal expressions using BiLSTM-CRF
|
# Evaluating KGR10 Polish word embeddings in the recognition of temporal expressions using BiLSTM-CRF
## Abstract
The article introduces a new set of Polish word embeddings, built using KGR10 corpus, which contains more than 4 billion words. These embeddings are evaluated in the problem of recognition of temporal expressions (timexes) for the Polish language. We described the process of KGR10 corpus creation and a new approach to the recognition problem using Bidirectional Long-Short Term Memory (BiLSTM) network with additional CRF layer, where specific embeddings are essential. We presented experiments and conclusions drawn from them.
## Introduction
Recent studies in information extraction domain (but also in other natural language processing fields) show that deep learning models produce state-of-the-art results BIBREF0 . Deep architectures employ multiple layers to learn hierarchical representations of the input data. In the last few years, neural networks based on dense vector representations provided the best results in various NLP tasks, including named entities recognition BIBREF1 , semantic role labelling BIBREF2 , question answering BIBREF3 and multitask learning BIBREF4 . The core element of most deep learning solutions is the dense distributed semantic representation of words, often called word embeddings. Distributional vectors follow the distributional hypothesis that words with a similar meaning tend to appear in similar contexts. Word embeddings capture the similarity between words and are often used as the first layer in deep learning models. Two of the most common and very efficient methods to produce word embeddings are Continuous Bag-of-Words (CBOW) and Skip-gram (SG), which produce distributed representations of words in a vector space, grouping them by similarity BIBREF5 , BIBREF6 . With the progress of machine learning techniques, it is possible to train such models on much larger data sets, and these often outperform the simple ones. It is possible to use a set of text documents containing even billions of words as training data. Both architectures (CBOW and SG) describe how the neural network learns the vector word representations for each word. In CBOW architecture the task is predicting the word given its context and in SG the task in predicting the context given the word.
Due to a significant increase of quality using deep learning methods together with word embeddings as the input layer for neural networks, many word vector sets have been created, using different corpora. The widest range of available word embeddings is available for English BIBREF7 and there were not so many options for less popular languages, e.g. Polish. There was a definite need within CLARIN-PL project and Sentimenti to increase the quality of NLP methods for Polish which were utilising available Polish word vectors BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 but only FastText modification of Skip-gram BIBREF9 was able to produce vectors for unknown words, based on character n-grams. The observation was that even using a sophisticated deep neural structure, the result strongly depends on the initial distributional representation. There was a need to build a massive corpus of Polish and create high-quality word vectors from that corpus. This work describes how we extended KGR7 1G corpus to become KGR10 with 4 billion words. Next, we present the different variants of word embeddings produced using this corpus. In the article about the recognition of named entities for Polish from the previous year, these embeddings were used in one of the three voting models to obtain the best results and the final system PolDeepNer BIBREF12 took the second place in PolEval2018 Task 2 BIBREF13 . In this article, we evaluated KGR10 FastText word embeddings in recognition of timexes.
## Available word embeddings
At the time we were testing word embeddings for different applications, there were 2 most popular sources of word vectors. The first one, called IPIPAN, is the result of the project Compositional distributional semantic models for identification, discrimination and disambiguation of senses in Polish texts, the process of creating word embeddings is described in article BIBREF10 and corpora used were National Corpus of Polish (NKJP) BIBREF14 and Wikipedia (Wiki). The second one, called FASTTEXT, is original FastText word embeddings set, created for 157 languages (including Polish). Authors used Wikipedia and Common Crawl as the linguistic data source. Table TABREF6 shows the number of tokens in each corpus and the name of the institution which prepared it. There is also information about the public availability of the resource.
Table TABREF7 presents the most commonly used word embeddings in CLARIN-PL before the creation of our embeddings.
## Building a larger corpus
KGR7 corpus (also called plWordNet Corpus 7.0, PLWNC 7.0) BIBREF15 , BIBREF16 was created at the Wroclaw University of Science and Technology by G4.19 Group. Due to the licences of documents in this corpus, this resource is not publicly available. Table TABREF8 contains KGR7 subcorpora and statistics BIBREF17 . One of the subcorpora in KGR7 is KIPI (the IPI PAN Corpus) BIBREF18 . KGR7 covers texts from a wide range of domains like: blogs, science, stenographic recordings, news, journalism, books and parliamentary transcripts. All texts come from the second half of the 20th century and represent the modern Polish language.
## plWordNet Corpus 10.0 (KGR10)
KGR10, also known as plWordNet Corpus 10.0 (PLWNC 10.0), is the result of the work on the toolchain to automatic acquisition and extraction of the website content, called CorpoGrabber BIBREF19 . It is a pipeline of tools to get the most relevant content of the website, including all subsites (up to the user-defined depth). The proposed toolchain can be used to build a big Web corpus of text documents. It requires the list of the root websites as the input. Tools composing CorpoGrabber are adapted to Polish, but most subtasks are language independent. The whole process can be run in parallel on a single machine and includes the following tasks: download of the HTML subpages of each input page URL with HTTrack, extraction of plain text from each subpage by removing boilerplate content (such as navigation links, headers, footers, advertisements from HTML pages) BIBREF20 , deduplication of plain text BIBREF20 , bad quality documents removal utilising Morphological Analysis Converter and Aggregator (MACA) BIBREF21 , documents tagging using Wrocław CRF Tagger (WCRFT) BIBREF22 . Last two steps are available only for Polish.
In order to significantly expand the set of documents in KGR7, we utilised DMOZ (short for directory.mozilla.org) – a multilingual open content directory of World Wide Web links, also known as Open Directory Project (ODP). The website with directory was closed in 2017, but the database still can be found on the web. Polish part of this directory contains more than 30,000 links to Polish websites. We used these links as root URLs for CorpoGrabber, and we downloaded more than 7TB of HTML web pages. After the extraction of text from HTML pages, deduplication of documents (including texts from KGR7) and removing bad quality documents (containing more than 30% of words outside the Morfeusz BIBREF23 dictionary) the result is KGR10 corpus, which contains 4,015,569,051 tokens and 18,084,712 unique words. Due to component licenses, KGR10 corpus is not publicly available.
## KGR10 word embeddings
We created a new Polish word embeddings models using the KGR10 corpus. We built 16 models of word embeddings using the implementation of CBOW and Skip-gram methods in the FastText tool BIBREF9 . These models are available under an open license in the CLARIN-PL project DSpace repository. The internal encoding solution based on embeddings of n-grams composing each word makes it possible to obtain FastText vector representations, also for words which were not processed during the creation of the model. A vector representation is associated with character n-gram and each word is represented as the sum of its n-gram vector representations. Previous solutions ignored the morphology of words and were assigning a distinct vector to each word. This is a limitation for languages with large vocabularies and many rare words, like Turkish, Finnish or Polish BIBREF9 . Authors observed that using word representations trained with subword information outperformed the plain Skip-gram model and the improvement was most significant for morphologically rich Slavic languages such as Czech (8% reduction of perplexity over SG) and Russian (13% reduction) BIBREF9 . We expected that word embeddings created that way for Polish should also provide such improvements. There were also previous attempts to build KGR10 word vectors with other methods (including FastText), and the results are presented in the article BIBREF8 . We selected the best models from that article – with embedding ID prefix EP (embeddings, previous) in Table TABREF13 – to compare with new models, marked as embedding ID prefix EC in Table TABREF13 ).
The word embeddings models used in PolDeepNer for recognition of timexes and named entities were EE1, . It was built on a plain KGR10. The dimension of word embedding is 300, the method of constructing vectors was Skip-gram BIBREF9 , and the number of negative samples for each positive example was 10.
## Temporal expressions
Temporal expressions (henceforth timexes) tell us when something happens, how long something lasts, or how often something occurs. The correct interpretation of a timex often involves knowing the context. Usually, a person is aware of their location in time, i.e., they know what day, month and year it is, and whether it is the beginning or the end of week or month. Therefore, they refer to specific dates, using incomplete expressions such as 12 November, Thursday, the following week, after three days. The temporal context is often necessary to determine to which specific date and time timexes refer. These examples do not exhaust the complexity of the problem of recognising timexes.
TimeML BIBREF24 is a markup language for describing timexes that has been adapted to many languages. One of the best-known methods of recognition of timexes called HeidelTime BIBREF25 , which uses the TIMEX3 annotation standard, currently supports 13 languages (with the use of hand-crafted resources). PLIMEX is a specification for the description of Polish timexes. It is based on TIMEX3 used in TimeML. Classes proposed in TimeML are adapted, namely: date, time, duration, set.
## Recognition of timexes
There are many methods for recognising timexes that are widely used in natural language engineering. For English (but not exclusively), in approaches based on supervised learning, sequence labelling methods are often used, especially Conditional Random Fields BIBREF26 . A review of the methods in the article BIBREF27 about the recognition of timexes for English and Spanish has shown a certain shift within the most popular solutions. As with the normalisation of timexes, the best results are still achieved with rule-based methods, many new solutions have been introduced in the area of recognition. The best systems listed in BIBREF27 , called TIPSem BIBREF28 and ClearTK BIBREF29 , use CRFs for recognition, so initially, we decided to apply the CRF-based approach for this task. The results were described in BIBREF30 , BIBREF31 .
In recent years, solutions based on deep neural networks, using word representation in the form of word embeddings, created with the use of large linguistic corpus, have begun to dominate in the field of recognition of word expressions. The most popular solutions include bidirectional long short-term memory neural networks (henceforth Bi-LSTM), often in combination with conditional random fields, as presented in the paper BIBREF32 dedicated to the recognition of proper names. For the Polish language, deep networks have also recently been used to recognise word expressions. In the issue of recognition of timexes, a bidirectional gated recurrent unit network (GRU) has been used BIBREF33 , BIBREF34 . GRU network is described in detail in the article BIBREF35 . In case of recognition of event descriptions using Bi-LSTM and Bi-GRU, where most of the Liner2 features were included in the input feature vector, better results were obtained BIBREF36 than for the Liner2 method (but without taking into account domain dictionaries). In last year's publication on the issue of named entities recognition using BiLSTM+CRF (together with G4.19 Group members), we received a statistically significant improvement in the quality of recognition compared to a solution using CRF only. The solution has been called PolDeepNer BIBREF12 .
## Experiments and Results
Experiments were carried out by the method proposed in BIBREF27 . The first part is described as Task A, the purpose of which is to identify the boundaries of timexes and assign them to one of the following classes: date, time, duration, set.
We trained the final models using the train set and we evaluated it using the test set, which was the reproduction of analysis performed in articles BIBREF37 , BIBREF38 . The division is presented in Table TABREF16 . We used BiLSTM+CRF classifier as in previous work BIBREF12 . We used precision, recall and F1 metrics from the classic NER task BIBREF12 , where true positive system answer has the same boundaries and type as annotation in gold data set. We evaluated all 17 word embeddings models using these metrics. The results are presented in Tables TABREF17 , TABREF18 and TABREF19 .
We chose the best 3 results from each word embeddings group (EE, EP, EC) from Table TABREF19 presenting F1-scores for all models. Then we evaluated these results using more detailed measures for timexes, presented in BIBREF27 . The following measures were used to evaluate the quality of boundaries and class recognition, so-called strict match: strict precision (Str.P), strict recall (Str.R) and strict F1-score (Str.F1). A relaxed match (Rel.P, Rel.R, Rel.F1) evaluation has also been carried out to determine whether there is an overlap between the system entity and gold entity, e.g. [Sunday] and [Sunday morning] BIBREF27 . If there was an overlap, a relaxed type F1-score (Type.F1) was calculated BIBREF27 . The results are presented in Table TABREF20 .
## Conclusions
The analysis of results from Tables TABREF17 , TABREF18 and TABREF19 show that 12 of 15 best results were obtained using new word embeddings. The evaluation results presented in Table TABREF20 (the chosen best embeddings models from Table TABREF19 ) prove that the best group of word embeddings is EC. The highest type F1-score was obtained for EC1 model, built using binary FastText Skip-gram method utilising subword information, with vector dimension equal to 300 and negative sampling equal to 10. The ability of the model to provide vector representation for the unknown words seems to be the most important. Also, previous models built using KGR10 (EP) are probably less accurate due to an incorrect tokenisation of the corpus. We used WCRFT tagger BIBREF22 , which utilises Toki BIBREF21 to tokenise the input text before the creation of the embeddings model. The comparison of EC1 with previous results obtained using only CRF BIBREF38 show the significant improvement across all the tested metrics: 3.6pp increase in strict F1-score, 1.36pp increase in relaxed precision, 5.61pp increase in relaxed recall and 3.51pp increase in relaxed F1-score.
## Acknowledgements
Work co-financed as part of the investment in the CLARIN-PL research infrastructure funded by the Polish Ministry of Science and Higher Education and in part by the National Centre for Research and Development, Poland, under grant no POIR.01.01.01-00-0472/16.
| 10
|
1904.04358
|
Deep Learning the EEG Manifold for Phonological Categorization from Active Thoughts
|
# Deep Learning the EEG Manifold for Phonological Categorization from Active Thoughts
## Abstract
Speech-related Brain Computer Interfaces (BCI) aim primarily at finding an alternative vocal communication pathway for people with speaking disabilities. As a step towards full decoding of imagined speech from active thoughts, we present a BCI system for subject-independent classification of phonological categories exploiting a novel deep learning based hierarchical feature extraction scheme. To better capture the complex representation of high-dimensional electroencephalography (EEG) data, we compute the joint variability of EEG electrodes into a channel cross-covariance matrix. We then extract the spatio-temporal information encoded within the matrix using a mixed deep neural network strategy. Our model framework is composed of a convolutional neural network (CNN), a long-short term network (LSTM), and a deep autoencoder. We train the individual networks hierarchically, feeding their combined outputs in a final gradient boosting classification step. Our best models achieve an average accuracy of 77.9% across five different binary classification tasks, providing a significant 22.5% improvement over previous methods. As we also show visually, our work demonstrates that the speech imagery EEG possesses significant discriminative information about the intended articulatory movements responsible for natural speech synthesis.
## Introduction
Decoding intended speech or motor activity from brain signals is one of the major research areas in Brain Computer Interface (BCI) systems BIBREF0 , BIBREF1 . In particular, speech-related BCI technologies attempt to provide effective vocal communication strategies for controlling external devices through speech commands interpreted from brain signals BIBREF2 . Not only do they provide neuro-prosthetic help for people with speaking disabilities and neuro-muscular disorders like locked-in-syndrome, nasopharyngeal cancer, and amytotropic lateral sclerosis (ALS), but also equip people with a better medium to communicate and express thoughts, thereby improving the quality of rehabilitation and clinical neurology BIBREF3 , BIBREF4 . Such devices also have applications in entertainment, preventive treatments, personal communication, games, etc. Furthermore, BCI technologies can be utilized in silent communication, as in noisy environments, or situations where any sort of audio-visual communication is infeasible.
Among the various brain activity-monitoring modalities in BCI, electroencephalography (EEG) BIBREF5 , BIBREF6 has demonstrated promising potential to differentiate between various brain activities through measurement of related electric fields. EEG is non-invasive, portable, low cost, and provides satisfactory temporal resolution. This makes EEG suitable to realize BCI systems. EEG data, however, is challenging: these data are high dimensional, have poor SNR, and suffer from low spatial resolution and a multitude of artifacts. For these reasons, it is not particularly obvious how to decode the desired information from raw EEG signals. Although the area of BCI based speech intent recognition has received increasing attention among the research community in the past few years, most research has focused on classification of individual speech categories in terms of discrete vowels, phonemes and words BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 . This includes categorization of imagined EEG signal into binary vowel categories like /a/, /u/ and rest BIBREF7 , BIBREF8 , BIBREF9 ; binary syllable classes like /ba/ and /ku/ BIBREF1 , BIBREF10 , BIBREF11 , BIBREF12 ; a handful of control words like 'up', 'down', 'left', 'right' and 'select' BIBREF15 or others like 'water', 'help', 'thanks', 'food', 'stop' BIBREF13 , Chinese characters BIBREF14 , etc. Such works mostly involve traditional signal processing or manual feature handcrafting along with linear classifiers (e.g., SVMs). In our recent work BIBREF16 , we introduced deep learning models for classification of vowels and words that achieved 23.45% improvement of accuracy over the baseline.
Production of articulatory speech is an extremely complicated process, thereby rendering understanding of the discriminative EEG manifold corresponding to imagined speech highly challenging. As a result, most of the existing approaches failed to achieve satisfactory accuracy on decoding speech tokens from the speech imagery EEG data. Perhaps, for these reasons, very little work has been devoted to relating the brain signals to the underlying articulation. The few exceptions include BIBREF17 , BIBREF18 . In BIBREF17 , Zhao et al. used manually handcrafted features from EEG data, combined with speech audio and facial features to achieve classification of the phonological categories varying based on the articulatory steps. However, the imagined speech classification accuracy based on EEG data alone, as reported in BIBREF17 , BIBREF18 , are not satisfactory in terms of accuracy and reliability. We now turn to describing our proposed models.
## Proposed Framework
Cognitive learning process underlying articulatory speech production involves incorporation of intermediate feedback loops and utilization of past information stored in the form of memory as well as hierarchical combination of several feature extractors. To this end, we develop our mixed neural network architecture composed of three supervised and a single unsupervised learning step, discussed in the next subsections and shown in Fig. FIGREF1 . We formulate the problem of categorizing EEG data based on speech imagery as a non-linear mapping INLINEFORM0 of a multivariate time-series input sequence INLINEFORM1 to fixed output INLINEFORM2 , i.e, mathematically INLINEFORM3 : INLINEFORM4 , where c and t denote the EEG channels and time instants respectively.
## Preprocessing step
We follow similar pre-processing steps on raw EEG data as reported in BIBREF17 (ocular artifact removal using blind source separation, bandpass filtering and subtracting mean value from each channel) except that we do not perform Laplacian filtering step since such high-pass filtering may decrease information content from the signals in the selected bandwidth.
## Joint variability of electrodes
Multichannel EEG data is high dimensional multivariate time series data whose dimensionality depends on the number of electrodes. It is a major hurdle to optimally encode information from these EEG data into lower dimensional space. In fact, our investigation based on a development set (as we explain later) showed that well-known deep neural networks (e.g., fully connected networks such as convolutional neural networks, recurrent neural networks and autoencoders) fail to individually learn such complex feature representations from single-trial EEG data. Besides, we found that instead of using the raw multi-channel high-dimensional EEG requiring large training times and resource requirements, it is advantageous to first reduce its dimensionality by capturing the information transfer among the electrodes. Instead of the conventional approach of selecting a handful of channels as BIBREF17 , BIBREF18 , we address this by computing the channel cross-covariance, resulting in positive, semi-definite matrices encoding the connectivity of the electrodes. We define channel cross-covariance (CCV) between any two electrodes INLINEFORM0 and INLINEFORM1 as: INLINEFORM2 . Next, we reject the channels which have significantly lower cross-covariance than auto-covariance values (where auto-covariance implies CCV on same electrode). We found this measure to be essential as the higher cognitive processes underlying speech planning and synthesis involve frequent information exchange between different parts of the brain. Hence, such matrices often contain more discriminative features and hidden information than mere raw signals. This is essentially different than our previous work BIBREF16 where we extract per-channel 1-D covariance information and feed it to the networks. We present our sample 2-D EEG cross-covariance matrices (of two individuals) in Fig. FIGREF2 .
## CNN & LSTM
In order to decode spatial connections between the electrodes from the channel covariance matrix, we use a CNN BIBREF19 , in particular a four-layered 2D CNN stacking two convolutional and two fully connected hidden layers. The INLINEFORM0 feature map at a given CNN layer with input INLINEFORM1 , weight matrix INLINEFORM2 and bias INLINEFORM3 is obtained as: INLINEFORM4 . At this first level of hierarchy, the network is trained with the corresponding labels as target outputs, optimizing a cross-entropy cost function. In parallel, we apply a four-layered recurrent neural network on the channel covariance matrices to explore the hidden temporal features of the electrodes. Namely, we exploit an LSTM BIBREF20 consisting of two fully connected hidden layers, stacked with two LSTM layers and trained in a similar manner as CNN.
## Deep autoencoder for spatio-temporal information
As we found the individually-trained parallel networks (CNN and LSTM) to be useful (see Table TABREF12 ), we suspected the combination of these two networks could provide a more powerful discriminative spatial and temporal representation of the data than each independent network. As such, we concatenate the last fully-connected layer from the CNN with its counterpart in the LSTM to compose a single feature vector based on these two penultimate layers. Ultimately, this forms a joint spatio-temporal encoding of the cross-covariance matrix.
In order to further reduce the dimensionality of the spatio-temporal encodings and cancel background noise effects BIBREF21 , we train an unsupervised deep autoenoder (DAE) on the fused heterogeneous features produced by the combined CNN and LSTM information. The DAE forms our second level of hierarchy, with 3 encoding and 3 decoding layers, and mean squared error (MSE) as the cost function.
## Classification with Extreme Gradient Boost
At the third level of hierarchy, the discrete latent vector representation of the deep autoencoder is fed into an Extreme Gradient Boost based classification layer BIBREF22 , BIBREF23 motivated by BIBREF21 . It is a regularized gradient boosted decision tree that performs well on structured problems. Since our EEG-phonological pairwise classification has an internal structure involving individual phonemes and words, it seems to be a reasonable choice of classifier. The classifier receives its input from the latent vectors of the deep autoencoder and is trained in a supervised manner to output the final predicted classes corresponding to the speech imagery.
## Dataset
We evaluate our model on a publicly available dataset, KARA ONE BIBREF17 , composed of multimodal data for stimulus-based, imagined and articulated speech state corresponding to 7 phonemic/syllabic ( /iy/, /piy/, /tiy/, /diy/, /uw/, /m/, /n/ ) as well as 4 words(pat, pot, knew and gnaw). The dataset consists of 14 participants, with each prompt presented 11 times to each individual. Since our intention is to classify the phonological categories from human thoughts, we discard the facial and audio information and only consider the EEG data corresponding to imagined speech. It is noteworthy that given the mixed nature of EEG signals, it is reportedly challenging to attain a pairwise EEG-phoneme mapping BIBREF18 . In order to explore the problem space, we thus specifically target five binary classification problems addressed in BIBREF17 , BIBREF18 , i.e presence/absence of consonants, phonemic nasal, bilabial, high-front vowels and high-back vowels.
## Training and hyperparameter selection
We performed two sets of experiments with the single-trial EEG data. In PHASE-ONE, our goals was to identify the best architectures and hyperparameters for our networks with a reasonable number of runs. For PHASE-ONE, we randomly shuffled and divided the data (1913 signals from 14 individuals) into train (80%), development (10%) and test sets (10%). In PHASE-TWO, in order to perform a fair comparison with the previous methods reported on the same dataset, we perform a leave-one-subject out cross-validation experiment using the best settings we learn from PHASE-ONE.
The architectural parameters and hyperparameters listed in Table TABREF6 were selected through an exhaustive grid-search based on the validation set of PHASE-ONE. We conducted a series of empirical studies starting from single hidden-layered networks for each of the blocks and, based on the validation accuracy, we increased the depth of each given network and selected the optimal parametric set from all possible combinations of parameters. For the gradient boosting classification, we fixed the maximum depth at 10, number of estimators at 5000, learning rate at 0.1, regularization coefficient at 0.3, subsample ratio at 0.8, and column-sample/iteration at 0.4. We did not find any notable change of accuracy while varying other hyperparameters while training gradient boost classifier.
## Performance analysis and discussion
To demonstrate the significance of the hierarchical CNN-LSTM-DAE method, we conducted separate experiments with the individual networks in PHASE-ONE of experiments and summarized the results in Table TABREF12 From the average accuracy scores, we observe that the mixed network performs much better than individual blocks which is in agreement with the findings in BIBREF21 . A detailed analysis on repeated runs further shows that in most of the cases, LSTM alone does not perform better than chance. CNN, on the other hand, is heavily biased towards the class label which sees more training data corresponding to it. Though the situation improves with combined CNN-LSTM, our analysis clearly shows the necessity of a better encoding scheme to utilize the combined features rather than mere concatenation of the penultimate features of both networks.
The very fact that our combined network improves the classification accuracy by a mean margin of 14.45% than the CNN-LSTM network indeed reveals that the autoencoder contributes towards filtering out the unrelated and noisy features from the concatenated penultimate feature set. It also proves that the combined supervised and unsupervised neural networks, trained hierarchically, can learn the discriminative manifold better than the individual networks and it is crucial for improving the classification accuracy. In addition to accuracy, we also provide the kappa coefficients BIBREF24 of our method in Fig. FIGREF14 . Here, a higher mean kappa value corresponding to a task implies that the network is able to find better discriminative information from the EEG data beyond random decisions. The maximum above-chance accuracy (75.92%) is recorded for presence/absence of the vowel task and the minimum (49.14%) is recorded for the INLINEFORM0 .
To further investigate the feature representation achieved by our model, we plot T-distributed Stochastic Neighbor Embedding (tSNE) corresponding to INLINEFORM0 and V/C classification tasks in Fig. FIGREF8 . We particularly select these two tasks as our model exhibits respectively minimum and maximum performance for these two. The tSNE visualization reveals that the second set of features are more easily separable than the first one, thereby giving a rationale for our performance.
Next, we provide performance comparison of the proposed approach with the baseline methods for PHASE-TWO of our study (cross-validation experiment) in Table TABREF15 . Since the model encounters the unseen data of a new subject for testing, and given the high inter-subject variability of the EEG data, a reduction in the accuracy was expected. However, our network still managed to achieve an improvement of 18.91, 9.95, 67.15, 2.83 and 13.70 % over BIBREF17 . Besides, our best model shows more reliability compared to previous works: The standard deviation of our model's classification accuracy across all the tasks is reduced from 22.59% BIBREF17 and 17.52% BIBREF18 to a mere 5.41%.
## Conclusion and future direction
In an attempt to move a step towards understanding the speech information encoded in brain signals, we developed a novel mixed deep neural network scheme for a number of binary classification tasks from speech imagery EEG data. Unlike previous approaches which mostly deal with subject-dependent classification of EEG into discrete vowel or word labels, this work investigates a subject-invariant mapping of EEG data with different phonological categories, varying widely in terms of underlying articulator motions (eg: involvement or non-involvement of lips and velum, variation of tongue movements etc). Our model takes an advantage of feature extraction capability of CNN, LSTM as well as the deep learning benefit of deep autoencoders. We took BIBREF17 , BIBREF18 as the baseline works investigating the same problem and compared our performance with theirs. Our proposed method highly outperforms the existing methods across all the five binary classification tasks by a large average margin of 22.51%.
## Acknowledgments
This work was funded by the Natural Sciences and Engineering Research Council (NSERC) of Canada and Canadian Institutes for Health Research (CIHR).
| 12
|
1904.05584
|
Gating Mechanisms for Combining Character and Word-level Word Representations: An Empirical Study
|
# Gating Mechanisms for Combining Character and Word-level Word Representations: An Empirical Study
## Abstract
In this paper we study how different ways of combining character and word-level representations affect the quality of both final word and sentence representations. We provide strong empirical evidence that modeling characters improves the learned representations at the word and sentence levels, and that doing so is particularly useful when representing less frequent words. We further show that a feature-wise sigmoid gating mechanism is a robust method for creating representations that encode semantic similarity, as it performed reasonably well in several word similarity datasets. Finally, our findings suggest that properly capturing semantic similarity at the word level does not consistently yield improved performance in downstream sentence-level tasks. Our code is available at https://github.com/jabalazs/gating
## Introduction
Incorporating sub-word structures like substrings, morphemes and characters to the creation of word representations significantly increases their quality as reflected both by intrinsic metrics and performance in a wide range of downstream tasks BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 .
The reason for this improvement is related to sub-word structures containing information that is usually ignored by standard word-level models. Indeed, when representing words as vectors extracted from a lookup table, semantically related words resulting from inflectional processes such as surf, surfing, and surfed, are treated as being independent from one another. Further, word-level embeddings do not account for derivational processes resulting in syntactically-similar words with different meanings such as break, breakable, and unbreakable. This causes derived words, which are usually less frequent, to have lower-quality (or no) vector representations.
Previous works have successfully combined character-level and word-level word representations, obtaining overall better results than using only word-level representations. For example BIBREF1 achieved state-of-the-art results in a machine translation task by representing unknown words as a composition of their characters. BIBREF4 created word representations by adding the vector representations of the words' surface forms and their morphemes ( INLINEFORM0 ), obtaining significant improvements on intrinsic evaluation tasks, word similarity and machine translation. BIBREF5 concatenated character-level and word-level representations for creating word representations, and then used them as input to their models for obtaining state-of-the-art results in Named Entity Recognition on several languages.
What these works have in common is that the models they describe first learn how to represent subword information, at character BIBREF1 , morpheme BIBREF4 , or substring BIBREF0 levels, and then combine these learned representations at the word level. The incorporation of information at a finer-grained hierarchy results in higher-quality modeling of rare words, morphological processes, and semantics BIBREF6 .
There is no consensus, however, on which combination method works better in which case, or how the choice of a combination method affects downstream performance, either measured intrinsically at the word level, or extrinsically at the sentence level.
In this paper we aim to provide some intuitions about how the choice of mechanism for combining character-level with word-level representations influences the quality of the final word representations, and the subsequent effect these have in the performance of downstream tasks. Our contributions are as follows:
## Background
We are interested in studying different ways of combining word representations, obtained from different hierarchies, into a single word representation. Specifically, we want to study how combining word representations (1) taken directly from a word embedding lookup table, and (2) obtained from a function over the characters composing them, affects the quality of the final word representations.
Let INLINEFORM0 be a set, or vocabulary, of words with INLINEFORM1 elements, and INLINEFORM2 a vocabulary of characters with INLINEFORM3 elements. Further, let INLINEFORM4 be a sequence of words, and INLINEFORM5 be the sequence of characters composing INLINEFORM6 . Each token INLINEFORM7 can be represented as a vector INLINEFORM8 extracted directly from an embedding lookup table INLINEFORM9 , pre-trained or otherwise, and as a vector INLINEFORM10 built from the characters that compose it; in other words, INLINEFORM11 , where INLINEFORM12 is a function that maps a sequence of characters to a vector.
The methods for combining word and character-level representations we study, are of the form INLINEFORM0 where INLINEFORM1 is the final word representation.
## Mapping Characters to Character-level Word Representations
The function INLINEFORM0 is composed of an embedding layer, an optional context function, and an aggregation function.
The embedding layer transforms each character INLINEFORM0 into a vector INLINEFORM1 of dimension INLINEFORM2 , by directly taking it from a trainable embedding lookup table INLINEFORM3 . We define the matrix representation of word INLINEFORM4 as INLINEFORM5 .
The context function takes INLINEFORM0 as input and returns a context-enriched matrix representation INLINEFORM1 , in which each INLINEFORM2 contains a measure of information about its context, and interactions with its neighbors. In particular, we chose to do this by feeding INLINEFORM3 to a BiLSTM BIBREF7 , BIBREF8 .
Informally, we can think of LSTM BIBREF10 as a function INLINEFORM0 that takes a matrix INLINEFORM1 as input and returns a context-enriched matrix representation INLINEFORM2 , where each INLINEFORM3 encodes information about the previous elements INLINEFORM4 .
A BiLSTM is simply composed of 2 LSTM, one that reads the input from left to right (forward), and another that does so from right to left (backward). The output of the forward and backward LSTM are INLINEFORM0 and INLINEFORM1 respectively. In the backward case the LSTM reads INLINEFORM2 first and INLINEFORM3 last, therefore INLINEFORM4 will encode the context from INLINEFORM5 .
The aggregation function takes the context-enriched matrix representation of word INLINEFORM0 for both directions, INLINEFORM1 and INLINEFORM2 , and returns a single vector INLINEFORM3 . To do so we followed BIBREF11 , and defined the character-level representation INLINEFORM4 of word INLINEFORM5 as the linear combination of the forward and backward last hidden states returned by the context function: DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are trainable parameters, and INLINEFORM2 represents the concatenation operation between two vectors.
## Combining Character and Word-level Representations
We tested three different methods for combining INLINEFORM0 with INLINEFORM1 : simple concatenation, a learned scalar gate BIBREF11 , and a learned vector gate (also referred to as feature-wise sigmoidal gate). Additionally, we compared these methods to two baselines: using pre-trained word vectors only, and using character-only features for representing words. See fig:methods for a visual description of the proposed methods.
word-only (w) considers only INLINEFORM0 and ignores INLINEFORM1 : DISPLAYFORM0
char-only (c) considers only INLINEFORM0 and ignores INLINEFORM1 : DISPLAYFORM0
concat (cat) concatenates both word and character-level representations: DISPLAYFORM0
scalar gate (sg) implements the scalar gating mechanism described by BIBREF11 : DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are trainable parameters, INLINEFORM2 , and INLINEFORM3 is the sigmoid function.
vector gate (vg): DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are trainable parameters, INLINEFORM2 , INLINEFORM3 is the element-wise sigmoid function, INLINEFORM4 is the element-wise product for vectors, and INLINEFORM5 is a vector of ones.
The vector gate is inspired by BIBREF11 and BIBREF12 , but is different to the former in that the gating mechanism acts upon each dimension of the word and character-level vectors, and different to the latter in that it does not rely on external sources of information for calculating the gating mechanism.
Finally, note that word only and char only are special cases of both gating mechanisms: INLINEFORM0 (scalar gate) and INLINEFORM1 (vector gate) correspond to word only; INLINEFORM2 and INLINEFORM3 correspond to char only.
## Obtaining Sentence Representations
To enable sentence-level classification we need to obtain a sentence representation from the word vectors INLINEFORM0 . We achieved this by using a BiLSTM with max pooling, which was shown to be a good universal sentence encoding mechanism BIBREF13 .
Let INLINEFORM0 , be an input sentence and INLINEFORM1 its matrix representation, where each INLINEFORM2 was obtained by one of the methods described in subsec:methods. INLINEFORM3 is the context-enriched matrix representation of INLINEFORM4 obtained by feeding INLINEFORM5 to a BiLSTM of output dimension INLINEFORM6 . Lastly, INLINEFORM11 is the final sentence representation of INLINEFORM12 obtained by max-pooling INLINEFORM13 along the sequence dimension.
Finally, we initialized the word representations INLINEFORM0 using GloVe embeddings BIBREF14 , and fine-tuned them during training. Refer to app:hyperparams for details on the other hyperparameters we used.
## Experimental Setup
We trained our models for solving the Natural Language Inference (NLI) task in two datasets, SNLI BIBREF15 and MultiNLI BIBREF16 , and validated them in each corresponding development set (including the matched and mismatched development sets of MultiNLI).
For each dataset-method combination we trained 7 models initialized with different random seeds, and saved each when it reached its best validation accuracy. We then evaluated the quality of each trained model's word representations INLINEFORM0 in 10 word similarity tasks, using the system created by BIBREF17 .
Finally, we fed these obtained word vectors to a BiLSTM with max-pooling and evaluated the final sentence representations in 11 downstream transfer tasks BIBREF13 , BIBREF18 .
## Datasets
Word-level Semantic Similarity A desirable property of vector representations of words is that semantically similar words should have similar vector representations. Assessing whether a set of word representations possesses this quality is referred to as the semantic similarity task. This is the most widely-used evaluation method for evaluating word representations, despite its shortcomings BIBREF20 .
This task consists of comparing the similarity between word vectors measured by a distance metric (usually cosine distance), with a similarity score obtained from human judgements. High correlation between these similarities is an indicator of good performance.
A problem with this formulation though, is that the definition of “similarity” often confounds the meaning of both similarity and relatedness. For example, cup and tea are related but dissimilar words, and this type of distinction is not always clear BIBREF21 , BIBREF22 .
To face the previous problem, we tested our methods in a wide variety of datasets, including some that explicitly model relatedness (WS353R), some that explicitly consider similarity (WS353S, SimLex999, SimVerb3500), and some where the distinction is not clear (MEN, MTurk287, MTurk771, RG, WS353). We also included the RareWords (RW) dataset for evaluating the quality of rare word representations. See appendix:datasets for a more complete description of the datasets we used.
Sentence-level Evaluation Tasks Unlike word-level representations, there is no consensus on the desirable properties sentence representations should have. In response to this, BIBREF13 created SentEval, a sentence representation evaluation benchmark designed for assessing how well sentence representations perform in various downstream tasks BIBREF23 .
Some of the datasets included in SentEval correspond to sentiment classification (CR, MPQA, MR, SST2, and SST5), subjectivity classification (SUBJ), question-type classification (TREC), recognizing textual entailment (SICK E), estimating semantic relatedness (SICK R), and measuring textual semantic similarity (STS16, STSB). The datasets are described by BIBREF13 , and we provide pointers to their original sources in the appendix table:sentence-eval-datasets.
To evaluate these sentence representations SentEval trained a linear model on top of them, and evaluated their performance in the validation sets accompanying each dataset. The only exception was the STS16 task, in which our representations were evaluated directly.
## Word Similarity
table:wordlevelresults shows the quality of word representations in terms of the correlation between word similarity scores obtained by the proposed models and word similarity scores defined by humans.
First, we can see that for each task, character only models had significantly worse performance than every other model trained on the same dataset. The most likely explanation for this is that these models are the only ones that need to learn word representations from scratch, since they have no access to the global semantic knowledge encoded by the GloVe embeddings.
Further, bold results show the overall trend that vector gates outperformed the other methods regardless of training dataset. This implies that learning how to combine character and word-level representations at the dimension level produces word vector representations that capture a notion of word similarity and relatedness that is closer to that of humans.
Additionally, results from the MNLI row in general, and underlined results in particular, show that training on MultiNLI produces word representations better at capturing word similarity. This is probably due to MultiNLI data being richer than that of SNLI. Indeed, MultiNLI data was gathered from various sources (novels, reports, letters, and telephone conversations, among others), rather than the single image captions dataset from which SNLI was created.
Exceptions to the previous rule are models evaluated in MEN and RW. The former case can be explained by the MEN dataset containing only words that appear as image labels in the ESP-Game and MIRFLICKR-1M image datasets BIBREF24 , and therefore having data that is more closely distributed to SNLI than to MultiNLI.
More notably, in the RareWords dataset BIBREF25 , the word only, concat, and scalar gate methods performed equally, despite having been trained in different datasets ( INLINEFORM0 ), and the char only method performed significantly worse when trained in MultiNLI. The vector gate, however, performed significantly better than its counterpart trained in SNLI. These facts provide evidence that this method is capable of capturing linguistic phenomena that the other methods are unable to model.
table:word-similarity-dataset lists the word-similarity datasets and their corresponding reference. As mentioned in subsec:datasets, all the word-similarity datasets contain pairs of words annotated with similarity or relatedness scores, although this difference is not always explicit. Below we provide some details for each.
MEN contains 3000 annotated word pairs with integer scores ranging from 0 to 50. Words correspond to image labels appearing in the ESP-Game and MIRFLICKR-1M image datasets.
MTurk287 contains 287 annotated pairs with scores ranging from 1.0 to 5.0. It was created from words appearing in both DBpedia and in news articles from The New York Times.
MTurk771 contains 771 annotated pairs with scores ranging from 1.0 to 5.0, with words having synonymy, holonymy or meronymy relationships sampled from WordNet BIBREF56 .
RG contains 65 annotated pairs with scores ranging from 0.0 to 4.0 representing “similarity of meaning”.
RW contains 2034 pairs of words annotated with similarity scores in a scale from 0 to 10. The words included in this dataset were obtained from Wikipedia based on their frequency, and later filtered depending on their WordNet synsets, including synonymy, hyperonymy, hyponymy, holonymy and meronymy. This dataset was created with the purpose of testing how well models can represent rare and complex words.
SimLex999 contains 999 word pairs annotated with similarity scores ranging from 0 to 10. In this case the authors explicitly considered similarity and not relatedness, addressing the shortcomings of datasets that do not, such as MEN and WS353. Words include nouns, adjectives and verbs.
SimVerb3500 contains 3500 verb pairs annotated with similarity scores ranging from 0 to 10. Verbs were obtained from the USF free association database BIBREF66 , and VerbNet BIBREF63 . This dataset was created to address the lack of representativity of verbs in SimLex999, and the fact that, at the time of creation, the best performing models had already surpassed inter-annotator agreement in verb similarity evaluation resources. Like SimLex999, this dataset also explicitly considers similarity as opposed to relatedness.
WS353 contains 353 word pairs annotated with similarity scores from 0 to 10.
WS353R is a subset of WS353 containing 252 word pairs annotated with relatedness scores. This dataset was created by asking humans to classify each WS353 word pair into one of the following classes: synonyms, antonyms, identical, hyperonym-hyponym, hyponym-hyperonym, holonym-meronym, meronym-holonym, and none-of-the-above. These annotations were later used to group the pairs into: similar pairs (synonyms, antonyms, identical, hyperonym-hyponym, and hyponym-hyperonym), related pairs (holonym-meronym, meronym-holonym, and none-of-the-above with a human similarity score greater than 5), and unrelated pairs (classified as none-of-the-above with a similarity score less than or equal to 5). This dataset is composed by the union of related and unrelated pairs.
WS353S is another subset of WS353 containing 203 word pairs annotated with similarity scores. This dataset is composed by the union of similar and unrelated pairs, as described previously.
## Word Frequencies and Gating Values
fig:gatingviz shows that for more common words the vector gate mechanism tends to favor only a few dimensions while keeping a low average gating value across dimensions. On the other hand, values are greater and more homogeneous across dimensions in rarer words. Further, fig:freqvsgatevalue shows this mechanism assigns, on average, a greater gating value to less frequent words, confirming the findings by BIBREF11 , and BIBREF12 .
In other words, the less frequent the word, the more this mechanism allows the character-level representation to influence the final word representation, as shown by eq:vg. A possible interpretation of this result is that exploiting character information becomes increasingly necessary as word-level representations' quality decrease.
Another observable trend in both figures is that gating values tend to be low on average. Indeed, it is possible to see in fig:freqvsgatevalue that the average gating values range from INLINEFORM0 to INLINEFORM1 . This result corroborates the findings by BIBREF11 , stating that setting INLINEFORM2 in eq:scalar-gate, was better than setting it to higher values.
In summary, the gating mechanisms learn how to compensate the lack of expressivity of underrepresented words by selectively combining their representations with those of characters.
## Sentence-level Evaluation
table:sentlevelresults shows the impact that different methods for combining character and word-level word representations have in the quality of the sentence representations produced by our models.
We can observe the same trend mentioned in subsec:word-similarity-eval, and highlighted by the difference between bold values, that models trained in MultiNLI performed better than those trained in SNLI at a statistically significant level, confirming the findings of BIBREF13 . In other words, training sentence encoders on MultiNLI yields more general sentence representations than doing so on SNLI.
The two exceptions to the previous trend, SICKE and SICKR, benefited more from models trained on SNLI. We hypothesize this is again due to both SNLI and SICK BIBREF26 having similar data distributions.
Additionally, there was no method that significantly outperformed the word only baseline in classification tasks. This means that the added expressivity offered by explicitly modeling characters, be it through concatenation or gating, was not significantly better than simply fine-tuning the pre-trained GloVe embeddings for this type of task. We hypothesize this is due to the conflation of two effects. First, the fact that morphological processes might not encode important information for solving these tasks; and second, that SNLI and MultiNLI belong to domains that are too dissimilar to the domains in which the sentence representations are being tested.
On the other hand, the vector gate significantly outperformed every other method in the STSB task when trained in both datasets, and in the STS16 task when trained in SNLI. This again hints at this method being capable of modeling phenomena at the word level, resulting in improved semantic representations at the sentence level.
## Relationship Between Word- and Sentence-level Evaluation Tasks
It is clear that the better performance the vector gate had in word similarity tasks did not translate into overall better performance in downstream tasks. This confirms previous findings indicating that intrinsic word evaluation metrics are not good predictors of downstream performance BIBREF29 , BIBREF30 , BIBREF20 , BIBREF31 .
subfig:mnli-correlations shows that the word representations created by the vector gate trained in MultiNLI had positively-correlated results within several word-similarity tasks. This hints at the generality of the word representations created by this method when modeling similarity and relatedness.
However, the same cannot be said about sentence-level evaluation performance; there is no clear correlation between word similarity tasks and sentence-evaluation tasks. This is clearly illustrated by performance in the STSBenchmark, the only in which the vector gate was significantly superior, not being correlated with performance in any word-similarity dataset. This can be interpreted simply as word-level representations capturing word-similarity not being a sufficient condition for good performance in sentence-level tasks.
In general, fig:correlations shows that there are no general correlation effects spanning both training datasets and combination mechanisms. For example, subfig:snli-correlations shows that, for both word-only and concat models trained in SNLI, performance in word similarity tasks correlates positively with performance in most sentence evaluation tasks, however, this does not happen as clearly for the same models trained in MultiNLI (subfig:mnli-correlations).
## Gating Mechanisms for Combining Characters and Word Representations
To the best of our knowledge, there are only two recent works that specifically study how to combine word and subword-level vector representations.
BIBREF11 propose to use a trainable scalar gating mechanism capable of learning a weighting scheme for combining character-level and word-level representations. They compared their proposed method to manually weighting both levels; using characters only; words only; or their concatenation. They found that in some datasets a specific manual weighting scheme performed better, while in others the learned scalar gate did.
BIBREF12 further expand the gating concept by making the mechanism work at a finer-grained level, learning how to weight each vector's dimensions independently, conditioned on external word-level features such as part-of-speech and named-entity tags. Similarly, they compared their proposed mechanism to using words only, characters only, and a concatenation of both, with and without external features. They found that their vector gate performed better than the other methods in all the reported tasks, and beat the state of the art in two reading comprehension tasks.
Both works showed that the gating mechanisms assigned greater importance to character-level representations in rare words, and to word-level representations in common ones, reaffirming the previous findings that subword structures in general, and characters in particular, are beneficial for modeling uncommon words.
## Sentence Representation Learning
The problem of representing sentences as fixed-length vectors has been widely studied.
BIBREF32 suggested a self-adaptive hierarchical model that gradually composes words into intermediate phrase representations, and adaptively selects specific hierarchical levels for specific tasks. BIBREF33 proposed an encoder-decoder model trained by attempting to reconstruct the surrounding sentences of an encoded passage, in a fashion similar to Skip-gram BIBREF34 . BIBREF35 overcame the previous model's need for ordered training sentences by using autoencoders for creating the sentence representations. BIBREF36 implemented a model simpler and faster to train than the previous two, while having competitive performance. Similar to BIBREF33 , BIBREF37 suggested predicting future sentences with a hierarchical CNN-LSTM encoder.
BIBREF13 trained several sentence encoding architectures on a combination of the SNLI and MultiNLI datasets, and showed that a BiLSTM with max-pooling was the best at producing highly transferable sentence representations. More recently, BIBREF18 empirically showed that sentence representations created in a multi-task setting BIBREF38 , performed increasingly better the more tasks they were trained in. BIBREF39 proposed using an autoencoder that relies on multi-head self-attention over the concatenation of the max and mean pooled encoder outputs for producing sentence representations. Finally, BIBREF40 show that modern sentence embedding methods are not vastly superior to random methods.
The works mentioned so far usually evaluate the quality of the produced sentence representations in sentence-level downstream tasks. Common benchmarks grouping these kind of tasks include SentEval BIBREF23 , and GLUE BIBREF41 . Another trend, however, is to probe sentence representations to understand what linguistic phenomena they encode BIBREF42 , BIBREF43 , BIBREF44 , BIBREF45 , BIBREF46 .
## General Feature-wise Transformations
BIBREF47 provide a review on feature-wise transformation methods, of which the mechanisms presented in this paper form a part of. In a few words, the INLINEFORM0 parameter, in both scalar gate and vector gate mechanisms, can be understood as a scaling parameter limited to the INLINEFORM1 range and conditioned on word representations, whereas adding the scaled INLINEFORM2 and INLINEFORM3 representations can be seen as biasing word representations conditioned on character representations.
The previous review extends the work by BIBREF48 , which describes the Feature-wise Linear Modulation (FiLM) framework as a generalization of Conditional Normalization methods, and apply it in visual reasoning tasks. Some of the reported findings are that, in general, scaling has greater impact than biasing, and that in a setting similar to the scalar gate, limiting the scaling parameter to INLINEFORM0 hurt performance. Future decisions involving the design of mechanisms for combining character and word-level representations should be informed by these insights.
## Conclusions
We presented an empirical study showing the effect that different ways of combining character and word representations has in word-level and sentence-level evaluation tasks.
We showed that a vector gate performed consistently better across a variety of word similarity and relatedness tasks. Additionally, despite showing inconsistent results in sentence evaluation tasks, it performed significantly better than the other methods in semantic similarity tasks.
We further showed through this mechanism, that learning character-level representations is always beneficial, and becomes increasingly so with less common words.
In the future it would be interesting to study how the choice of mechanism for combining subword and word representations affects the more recent language-model-based pretraining methods such as ELMo BIBREF49 , GPT BIBREF50 , BIBREF51 and BERT BIBREF52 .
## Acknowledgements
Thanks to Edison Marrese-Taylor and Pablo Loyola for their feedback on early versions of this manuscript. We also gratefully acknowledge the support of the NVIDIA Corporation with the donation of one of the GPUs used for this research. Jorge A. Balazs is partially supported by the Japanese Government MEXT Scholarship.
## Hyperparameters
We only considered words that appear at least twice, for each dataset. Those that appeared only once were considered UNK. We used the Treebank Word Tokenizer as implemented in NLTK for tokenizing the training and development datasets.
In the same fashion as conneau2017supervised, we used a batch size of 64, an SGD optmizer with an initial learning rate of INLINEFORM0 , and at each epoch divided the learning rate by 5 if the validation accuracy decreased. We also used gradient clipping when gradients where INLINEFORM1 .
We defined character vector representations as 50-dimensional vectors randomly initialized by sampling from the uniform distribution in the INLINEFORM0 range.
The output dimension of the character-level BiLSTM was 300 per direction, and remained of such size after combining forward and backward representations as depicted in eq. EQREF9 .
Word vector representations where initialized from the 300-dimensional GloVe vectors BIBREF14 , trained in 840B tokens from the Common Crawl, and finetuned during training. Words not present in the GloVe vocabulary where randomly initialized by sampling from the uniform distribution in the INLINEFORM0 range.
The input size of the word-level LSTM was 300 for every method except concat in which it was 600, and its output was always 2048 per direction, resulting in a 4096-dimensional sentence representation.
## Sentence Evaluation Datasets
table:sentence-eval-datasets lists the sentence-level evaluation datasets used in this paper. The provided URLs correspond to the original sources, and not necessarily to the URLs where SentEval got the data from.
The version of the CR, MPQA, MR, and SUBJ datasets used in this paper were the ones preprocessed by BIBREF75 . Both SST2 and SST5 correspond to preprocessed versions of the SST dataset by BIBREF74 . SST2 corresponds to a subset of SST used by BIBREF54 containing flat representations of sentences annotated with binary sentiment labels, and SST5 to another subset annotated with more fine-grained sentiment labels (very negative, negative, neutral, positive, very positive).
| 18
|
1904.07342
|
Learning Twitter User Sentiments on Climate Change with Limited Labeled Data
|
# Learning Twitter User Sentiments on Climate Change with Limited Labeled Data
## Abstract
While it is well-documented that climate change accepters and deniers have become increasingly polarized in the United States over time, there has been no large-scale examination of whether these individuals are prone to changing their opinions as a result of natural external occurrences. On the sub-population of Twitter users, we examine whether climate change sentiment changes in response to five separate natural disasters occurring in the U.S. in 2018. We begin by showing that relevant tweets can be classified with over 75% accuracy as either accepting or denying climate change when using our methodology to compensate for limited labeled data; results are robust across several machine learning models and yield geographic-level results in line with prior research. We then apply RNNs to conduct a cohort-level analysis showing that the 2018 hurricanes yielded a statistically significant increase in average tweet sentiment affirming climate change. However, this effect does not hold for the 2018 blizzard and wildfires studied, implying that Twitter users' opinions on climate change are fairly ingrained on this subset of natural disasters.
## Background
Much prior work has been done at the intersection of climate change and Twitter, such as tracking climate change sentiment over time BIBREF2 , finding correlations between Twitter climate change sentiment and seasonal effects BIBREF3 , and clustering Twitter users based on climate mentalities using network analysis BIBREF4 . Throughout, Twitter has been accepted as a powerful tool given the magnitude and reach of samples unattainable from standard surveys. However, the aforementioned studies are not scalable with regards to training data, do not use more recent sentiment analysis tools (such as neural nets), and do not consider unbiased comparisons pre- and post- various climate events (which would allow for a more concrete evaluation of shocks to climate change sentiment). This paper aims to address these three concerns as follows.
First, we show that machine learning models formed using our labeling technique can accurately predict tweet sentiment (see Section SECREF2 ). We introduce a novel method to intuit binary sentiments of large numbers of tweets for training purposes. Second, we quantify unbiased outcomes from these predicted sentiments (see Section SECREF4 ). We do this by comparing sentiments within the same cohort of Twitter users tweeting both before and after specific natural disasters; this removes bias from over-weighting Twitter users who are only compelled to compose tweets after a disaster.
## Data
We henceforth refer to a tweet affirming climate change as a “positive" sample (labeled as 1 in the data), and a tweet denying climate change as a “negative" sample (labeled as -1 in the data). All data were downloaded from Twitter in two separate batches using the “twint" scraping tool BIBREF5 to sample historical tweets for several different search terms; queries always included either “climate change" or “global warming", and further included disaster-specific search terms (e.g., “bomb cyclone," “blizzard," “snowstorm," etc.). We refer to the first data batch as “influential" tweets, and the second data batch as “event-related" tweets.
The first data batch consists of tweets relevant to blizzards, hurricanes, and wildfires, under the constraint that they are tweeted by “influential" tweeters, who we define as individuals certain to have a classifiable sentiment regarding the topic at hand. For example, we assume that any tweet composed by Al Gore regarding climate change is a positive sample, whereas any tweet from conspiracy account @ClimateHiJinx is a negative sample. The assumption we make in ensuing methods (confirmed as reasonable in Section SECREF2 ) is that influential tweeters can be used to label tweets in bulk in the absence of manually-labeled tweets. Here, we enforce binary labels for all tweets composed by each of the 133 influential tweeters that we identified on Twitter (87 of whom accept climate change), yielding a total of 16,360 influential tweets.
The second data batch consists of event-related tweets for five natural disasters occurring in the U.S. in 2018. These are: the East Coast Bomb Cyclone (Jan. 2 - 6); the Mendocino, California wildfires (Jul. 27 - Sept. 18); Hurricane Florence (Aug. 31 - Sept. 19); Hurricane Michael (Oct. 7 - 16); and the California Camp Fires (Nov. 8 - 25). For each disaster, we scraped tweets starting from two weeks prior to the beginning of the event, and continuing through two weeks after the end of the event. Summary statistics on the downloaded event-specific tweets are provided in Table TABREF1 . Note that the number of tweets occurring prior to the two 2018 sets of California fires are relatively small. This is because the magnitudes of these wildfires were relatively unpredictable, whereas blizzards and hurricanes are often forecast weeks in advance alongside public warnings. The first (influential tweet data) and second (event-related tweet data) batches are de-duplicated to be mutually exclusive. In Section SECREF2 , we perform geographic analysis on the event-related tweets from which we can scrape self-reported user city from Twitter user profile header cards; overall this includes 840 pre-event and 5,984 post-event tweets.
To create a model for predicting sentiments of event-related tweets, we divide the first data batch of influential tweets into training and validation datasets with a 90%/10% split. The training set contains 49.2% positive samples, and the validation set contains 49.0% positive samples. We form our test set by manually labeling a subset of 500 tweets from the the event-related tweets (randomly chosen across all five natural disasters), of which 50.0% are positive samples.
## Labeling Methodology
Our first goal is to train a sentiment analysis model (on training and validation datasets) in order to perform classification inference on event-based tweets. We experimented with different feature extraction methods and classification models. Feature extractions examined include Tokenizer, Unigram, Bigram, 5-char-gram, and td-idf methods. Models include both neural nets (e.g. RNNs, CNNs) and standard machine learning tools (e.g. Naive Bayes with Laplace Smoothing, k-clustering, SVM with linear kernel). Model accuracies are reported in Table FIGREF3 .
The RNN pre-trained using GloVe word embeddings BIBREF6 achieved the higest test accuracy. We pass tokenized features into the embedding layer, followed by an LSTM BIBREF7 with dropout and ReLU activation, and a dense layer with sigmoid activation. We apply an Adam optimizer on the binary crossentropy loss. Implementing this simple, one-layer LSTM allows us to surpass the other traditional machine learning classification methods. Note the 13-point spread between validation and test accuracies achieved. Ideally, the training, validation, and test datasets have the same underlying distribution of tweet sentiments; the assumption made with our labeling technique is that the influential accounts chosen are representative of all Twitter accounts. Critically, when choosing the influential Twitter users who believe in climate change, we highlighted primarily politicians or news sources (i.e., verifiably affirming or denying climate change); these tweets rarely make spelling errors or use sarcasm. Due to this skew, the model yields a high rate of false negatives. It is likely that we could lessen the gap between validation and test accuracies by finding more “real" Twitter users who are climate change believers, e.g. by using the methodology found in BIBREF4 .
## Outcome Analysis
Our second goal is to compare the mean values of users' binary sentiments both pre- and post- each natural disaster event. Applying our highest-performing RNN to event-related tweets yields the following breakdown of positive tweets: Bomb Cyclone (34.7%), Mendocino Wildfire (80.4%), Hurricane Florence (57.2%), Hurricane Michael (57.6%), and Camp Fire (70.1%). As sanity checks, we examine the predicted sentiments on a subset with geographic user information and compare results to the prior literature.
In Figure FIGREF3 , we map 4-clustering results on three dimensions: predicted sentiments, latitude, and longitude. The clusters correspond to four major regions of the U.S.: the Northeast (green), Southeast (yellow), Midwest (blue), and West Coast (purple); centroids are designated by crosses. Average sentiments within each cluster confirm prior knowledge BIBREF1 : the Southeast and Midwest have lower average sentiments ( INLINEFORM0 and INLINEFORM1 , respectively) than the West Coast and Northeast (0.22 and 0.09, respectively). In Figure FIGREF5 , we plot predicted sentiment averaged by U.S. city of event-related tweeters. The majority of positive tweets emanate from traditionally liberal hubs (e.g. San Francisco, Los Angeles, Austin), while most negative tweets come from the Philadelphia metropolitan area. These regions aside, rural areas tended to see more negative sentiment tweeters post-event, whereas urban regions saw more positive sentiment tweeters; however, overall average climate change sentiment pre- and post-event was relatively stable geographically. This map further confirms findings that coastal cities tend to be more aware of climate change BIBREF8 .
From these mapping exercises, we claim that our “influential tweet" labeling is reasonable. We now discuss our final method on outcomes: comparing average Twitter sentiment pre-event to post-event. In Figure FIGREF8 , we display these metrics in two ways: first, as an overall average of tweet binary sentiment, and second, as a within-cohort average of tweet sentiment for the subset of tweets by users who tweeted both before and after the event (hence minimizing awareness bias). We use Student's t-test to calculate the significance of mean sentiment differences pre- and post-event (see Section SECREF4 ). Note that we perform these mean comparisons on all event-related data, since the low number of geo-tagged samples would produce an underpowered study.
## Results & Discussion
In Figure FIGREF8 , we see that overall sentiment averages rarely show movement post-event: that is, only Hurricane Florence shows a significant difference in average tweet sentiment pre- and post-event at the 1% level, corresponding to a 0.12 point decrease in positive climate change sentiment. However, controlling for the same group of users tells a different story: both Hurricane Florence and Hurricane Michael have significant tweet sentiment average differences pre- and post-event at the 1% level. Within-cohort, Hurricane Florence sees an increase in positive climate change sentiment by 0.21 points, which is contrary to the overall average change (the latter being likely biased since an influx of climate change deniers are likely to tweet about hurricanes only after the event). Hurricane Michael sees an increase in average tweet sentiment of 0.11 points, which reverses the direction of tweets from mostly negative pre-event to mostly positive post-event. Likely due to similar bias reasons, the Mendocino wildfires in California see a 0.06 point decrease in overall sentiment post-event, but a 0.09 point increase in within-cohort sentiment. Methodologically, we assert that overall averages are not robust results to use in sentiment analyses.
We now comment on the two events yielding similar results between overall and within-cohort comparisons. Most tweets regarding the Bomb Cyclone have negative sentiment, though sentiment increases by 0.02 and 0.04 points post-event for overall and within-cohort averages, respectively. Meanwhile, the California Camp Fires yield a 0.11 and 0.27 point sentiment decline in overall and within-cohort averages, respectively. This large difference in sentiment change can be attributed to two factors: first, the number of tweets made regarding wildfires prior to the (usually unexpected) event is quite low, so within-cohort users tend to have more polarized climate change beliefs. Second, the root cause of the Camp Fires was quickly linked to PG&E, bolstering claims that climate change had nothing to do with the rapid spread of fire; hence within-cohort users were less vocally positive regarding climate change post-event.
There are several caveats in our work: first, tweet sentiment is rarely binary (this work could be extended to a multinomial or continuous model). Second, our results are constrained to Twitter users, who are known to be more negative than the general U.S. population BIBREF9 . Third, we do not take into account the aggregate effects of continued natural disasters over time. Going forward, there is clear demand in discovering whether social networks can indicate environmental metrics in a “nowcasting" fashion. As climate change becomes more extreme, it remains to be seen what degree of predictive power exists in our current model regarding climate change sentiments with regards to natural disasters.
| 5
|
1904.11942
|
Contextualized Word Embeddings Enhanced Event Temporal Relation Extraction for Story Understanding
|
# Contextualized Word Embeddings Enhanced Event Temporal Relation Extraction for Story Understanding
## Abstract
Learning causal and temporal relationships between events is an important step towards deeper story and commonsense understanding. Though there are abundant datasets annotated with event relations for story comprehension, many have no empirical results associated with them. In this work, we establish strong baselines for event temporal relation extraction on two under-explored story narrative datasets: Richer Event Description (RED) and Causal and Temporal Relation Scheme (CaTeRS). To the best of our knowledge, these are the first results reported on these two datasets. We demonstrate that neural network-based models can outperform some strong traditional linguistic feature-based models. We also conduct comparative studies to show the contribution of adopting contextualized word embeddings (BERT) for event temporal relation extraction from stories. Detailed analyses are offered to better understand the results.
## Introduction
Event temporal relation understanding is a major component of story/narrative comprehension. It is an important natural language understanding (NLU) task with broad applications to downstream tasks such as story understanding BIBREF0 , BIBREF1 , BIBREF2 , question answering BIBREF3 , BIBREF4 , and text summarization BIBREF5 , BIBREF6 .
The goal of event temporal relation extraction is to build a directed graph where nodes correspond to events, and edges reflect temporal relations between the events. Figure FIGREF1 illustrates an example of such a graph for the text shown above. Different types of edges specify different temporal relations: the event assassination is before slaughtered, slaughtered is included in rampage, and the relation between rampage and war is vague.
Modeling event temporal relations is crucial for story/narrative understanding and storytelling, because a story is typically composed of a sequence of events BIBREF7 . Several story corpora are thus annotated with various event-event relations to understand commonsense event knowledge. CaTeRS BIBREF8 is created by annotating 320 five-sentence stories sampled from ROCStories BIBREF7 dataset. RED BIBREF9 contains annotations of rich relations between event pairs for storyline understanding, including co-reference and partial co-reference relations, temporal; causal, and sub-event relations.
Despite multiple productive research threads on temporal and causal relation modeling among events BIBREF10 , BIBREF11 , BIBREF12 and event relation annotation for story understanding BIBREF8 , the intersection of these two threads seems flimsy. To the best of our knowledge, no event relation extraction results have been reported on CaTeRS and RED.
We apply neural network models that leverage recent advances in contextualized embeddings (BERT BIBREF13 ) to event-event relation extraction tasks for CaTeRS and RED. Our goal in this paper is to increase understanding of how well the state-of-the-art event relation models work for story/narrative comprehension.
In this paper, we report the first results of event temporal relation extraction on two under-explored story comprehension datasets: CaTeRS and RED. We establish strong baselines with neural network models enhanced by recent breakthrough of contextualized embeddings, BERT BIBREF13 . We summarize the contributions of the paper as follows:
## Models
We investigate both neural network-based models and traditional feature-based models. We briefly introduce them in this section.
## Data
is created by annotating 1600 sentences of 320 five-sentence stories sampled from ROCStories BIBREF7 dataset. CaTeRS contains both temporal and causal relations in an effort to understand and predict commonsense relations between events.
As demonstrated in Table TABREF16 , we split all stories into 220 training and 80 test. We do not construct the development set because the dataset is small. Note that some relations have compounded labels such as “CAUSE_BEFORE”, “ENABLE_BEFORE”, etc. We only take the temporal portion of the annotations.
annotates a wide range of relations of event pairs including their coreference and partial coreference relations, and temporal, causal and subevent relationships. We split data according to the standard train, development, test sets, and only focus on the temporal relations.
The common issue of these two datasets is that they are not densely annotated – not every pair of events is annotated with a relation. We provide one way to handle negative (unannotated) pairs in this paper. When constructing negative examples, we take all event pairs that occur within the same or neighboring sentences with no annotations, labeling them as “NONE”. The negative to positive samples ratio is 1.00 and 11.5 for CaTeRS and RED respectively. Note that RED data has much higher negative ratio (as shown in Table TABREF16 ) because it contains longer articles, more complicated sentence structures, and richer entity types than CaTeRS where all stories consist of 5 (mostly short) sentences.
In both the development and test sets, we add all negative pairs as candidates for the relation prediction. During training, the number of negative pairs we add is based on a hyper-parameter that we tune to control the negative-to-positive sample ratio.
To justify our decision of selecting negative pairs within the same or neighboring sentences, we show the distribution of distances across positive sentence pairs in Table TABREF18 . Although CaTeRS data has pair distance more evenly distributed than RED, we observe that the vast majority (85.87% and 93.99% respectively) of positive pairs have sentence distance less than or equal to one.
To handle negative pairs that are more than two sentences away, we automatically predict all out-of-window pairs as “NONE”. This means that some positive pairs will be automatically labeled as negative pairs. Since the percentage of out-of-window positive pairs is small, we believe the impact on performance is small. We can investigate expanding the prediction window in future research, but the trade-off is that we will get more negative pairs that are hard to predict.
## Implementation Details
CAEVO consists of both linguistic-rule-based sieves and feature-based trainable sieves. We train CAEVO sieves with our train set and evaluate them on both dev and test sets. CAEVO is an end-to-end system that automatically annotates both events and relations. In order to resolve label annotation mismatch between CAEVO and our gold data, we create our own final input files to CAEVO system. Default parameter settings are used when running the CAEVO system.
In an effort of building a general model and reducing the number of hand-crafted features, we leverage pre-trained (GloVe 300) embeddings in place of linguistic features. The only linguistic feature we use in our experiment is token distance. We notice in our experiments that hidden layer size, dropout ratio and negative sample ratio impact model performance significantly. We conduct grid search to find the best hyper-parameter combination according to the performance of the development set.
Note that since the CaTeRS data is small and there is no standard train, development, and test splits, we conduct cross-validation on training data to choose the best hyper-parameters and predict on test. For RED data, the standard train, development, test splits are used.
As we mentioned briefly in the introduction, using BERT output as word embeddings could provide an additional performance boost in our NN architecture. We pre-process our raw data by feeding original sentences into a pre-trained BERT model and output the last layer of BERT as token representations. In this experiment, we fix the negative sample ratio according to the result obtained from the previous step and only search for the best hidden layer size and dropout ratio.
## Result and Analysis
Table TABREF25 contains the best hyper-parameters and Table TABREF26 contains micro-average F1 scores for both datasets on dev and test sets. We only consider positive pairs, i.e. correct predictions on NONE pairs are excluded for evaluation. In general, the baseline model CAEVO is outperformed by both NN models, and NN model with BERT embedding achieves the greatest performance. We now provide more detailed analysis and discussion for each dataset.
## Temporal Relation Data
Collecting dense TempRel corpora with event pairs fully annotated has been reported challenging since annotators could easily overlook some pairs BIBREF18 , BIBREF19 , BIBREF10 . TimeBank BIBREF20 is an example with events and their relations annotated sparsely. TB-Dense dataset mitigates this issue by forcing annotators to examine all pairs of events within the same or neighboring sentences. However, densely annotated datasets are relatively small both in terms of number of documents and event pairs, which restricts the complexity of machine learning models used in previous research.
## Feature-based Models
The series of TempEval competitions BIBREF21 , BIBREF22 , BIBREF23 have attracted many research interests in predicting event temporal relations. Early attempts by BIBREF24 , BIBREF21 , BIBREF25 , BIBREF26 only use pair-wise classification models. State-of-the-art local methods, such as ClearTK BIBREF27 , UTTime BIBREF28 , and NavyTime BIBREF29 improve on earlier work by feature engineering with linguistic and syntactic rules. As we mention in the Section 2, CAEVO is the current state-of-the-art system for feature-based temporal event relation extraction BIBREF10 . It's widely used as the baseline for evaluating TB-Dense data. We adopt it as our baseline for evaluating CaTeRS and RED datasets. Additionally, several models BramsenDLB2006, ChambersJ2008, DoLuRo12, NingWuRo18, P18-1212 have successfully incorporated global inference to impose global prediction consistency such as temporal transitivity.
## Neural Network Model
Neural network-based methods have been employed for event temporal relation extraction BIBREF14 , BIBREF15 , BIBREF16 , BIBREF12 which achieved impressive results. However, the dataset they focus on is TB-Dense. We have explored neural network models on CaTeRS and RED, which are more related to story narrative understanding and generation.
In our NN model, we also leverage Bidrectional Encoder Representations from Transformers (BERT) BIBREF30 which has shown significant improvement in many NLP tasks by allowing fine-tuning of pre-trained language representations. Unlike the Generative Pre-trained Transformer (OpenAI GPT) BIBREF31 , BERT uses a biderctional Transformer BIBREF32 instead of a unidirectional (left-to-right) Transformer to incorporate context from both directions. As mentioned earlier, we do not fine-tune BERT in our experiments and simply leverage the last layer as our contextualized word representations.
## Conclusion
We established strong baselines for two story narrative understanding datasets: CaTeRS and RED. We have shown that neural network-based models can outperform feature-based models with wide margins, and we conducted an ablation study to show that contextualized representation learning can boost performance of NN models. Further research can focus on more systematic study or build stronger NN models over the same datasets used in this work. Exploring possibilities to directly apply temporal relation extraction to enhance performance of story generation systems is another promising research direction.
## Acknowledgement
We thank the anonymous reviewers for their constructive comments, as well as the members of the USC PLUS lab for their early feedback. This work is supported by Contract W911NF-15-1-0543 with the US Defense Advanced Research Projects Agency (DARPA).
| 10
|
1905.08949
|
Recent Advances in Neural Question Generation
|
# Recent Advances in Neural Question Generation
## Abstract
Emerging research in Neural Question Generation (NQG) has started to integrate a larger variety of inputs, and generating questions requiring higher levels of cognition. These trends point to NQG as a bellwether for NLP, about how human intelligence embodies the skills of curiosity and integration. We present a comprehensive survey of neural question generation, examining the corpora, methodologies, and evaluation methods. From this, we elaborate on what we see as emerging on NQG's trend: in terms of the learning paradigms, input modalities, and cognitive levels considered by NQG. We end by pointing out the potential directions ahead.
## Introduction
Question Generation (QG) concerns the task of “automatically generating questions from various inputs such as raw text, database, or semantic representation" BIBREF0 . People have the ability to ask rich, creative, and revealing questions BIBREF1 ; e.g., asking Why did Gollum betray his master Frodo Baggins? after reading the fantasy novel The Lord of the Rings. How can machines be endowed with the ability to ask relevant and to-the-point questions, given various inputs? This is a challenging, complementary task to Question Answering (QA). Both QA and QG require an in-depth understanding of the input source and the ability to reason over relevant contexts. But beyond understanding, QG additionally integrates the challenges of Natural Language Generation (NLG), i.e., generating grammatically and semantically correct questions.
QG is of practical importance: in education, forming good questions are crucial for evaluating students’ knowledge and stimulating self-learning. QG can generate assessments for course materials BIBREF2 or be used as a component in adaptive, intelligent tutoring systems BIBREF3 . In dialog systems, fluent QG is an important skill for chatbots, e.g., in initiating conversations or obtaining specific information from human users. QA and reading comprehension also benefit from QG, by reducing the needed human labor for creating large-scale datasets. We can say that traditional QG mainly focused on generating factoid questions from a single sentence or a paragraph, spurred by a series of workshops during 2008–2012 BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 .
Recently, driven by advances in deep learning, QG research has also begun to utilize “neural” techniques, to develop end-to-end neural models to generate deeper questions BIBREF8 and to pursue broader applications BIBREF9 , BIBREF10 .
While there have been considerable advances made in NQG, the area lacks a comprehensive survey. This paper fills this gap by presenting a systematic survey on recent development of NQG, focusing on three emergent trends that deep learning has brought in QG: (1) the change of learning paradigm, (2) the broadening of the input spectrum, and (3) the generation of deep questions.
## Fundamental Aspects of NQG
For the sake of clean exposition, we first provide a broad overview of QG by conceptualizing the problem from the perspective of the three introduced aspects: (1) its learning paradigm, (2) its input modalities, and (3) the cognitive level it involves. This combines past research with recent trends, providing insights on how NQG connects to traditional QG research.
## Learning Paradigm
QG research traditionally considers two fundamental aspects in question asking: “What to ask” and “How to ask”. A typical QG task considers the identification of the important aspects to ask about (“what to ask”), and learning to realize such identified aspects as natural language (“how to ask”). Deciding what to ask is a form of machine understanding: a machine needs to capture important information dependent on the target application, akin to automatic summarization. Learning how to ask, however, focuses on aspects of the language quality such as grammatical correctness, semantically preciseness and language flexibility.
Past research took a reductionist approach, separately considering these two problems of “what” and “how” via content selection and question construction. Given a sentence or a paragraph as input, content selection selects a particular salient topic worthwhile to ask about and determines the question type (What, When, Who, etc.). Approaches either take a syntactic BIBREF11 , BIBREF12 , BIBREF13 or semantic BIBREF14 , BIBREF3 , BIBREF15 , BIBREF16 tack, both starting by applying syntactic or semantic parsing, respectively, to obtain intermediate symbolic representations. Question construction then converts intermediate representations to a natural language question, taking either a tranformation- or template-based approach. The former BIBREF17 , BIBREF18 , BIBREF13 rearranges the surface form of the input sentence to produce the question; the latter BIBREF19 , BIBREF20 , BIBREF21 generates questions from pre-defined question templates. Unfortunately, such QG architectures are limiting, as their representation is confined to the variety of intermediate representations, transformation rules or templates.
In contrast, neural models motivate an end-to-end architectures. Deep learned frameworks contrast with the reductionist approach, admitting approaches that jointly optimize for both the “what” and “how” in an unified framework. The majority of current NQG models follow the sequence-to-sequence (Seq2Seq) framework that use a unified representation and joint learning of content selection (via the encoder) and question construction (via the decoder). In this framework, traditional parsing-based content selection has been replaced by more flexible approaches such as attention BIBREF22 and copying mechanism BIBREF23 . Question construction has become completely data-driven, requiring far less labor compared to transformation rules, enabling better language flexibility compared to question templates.
However, unlike other Seq2Seq learning NLG tasks, such as Machine Translation, Image Captioning, and Abstractive Summarization, which can be loosely regarded as learning a one-to-one mapping, generated questions can differ significantly when the intent of asking differs (e.g., the target answer, the target aspect to ask about, and the question's depth). In Section "Methodology" , we summarize different NQG methodologies based on Seq2Seq framework, investigating how some of these QG-specific factors are integrated with neural models, and discussing what could be further explored. The change of learning paradigm in NQG era is also represented by multi-task learning with other NLP tasks, for which we discuss in Section "Multi-task Learning" .
## Input Modality
Question generation is an NLG task for which the input has a wealth of possibilities depending on applications. While a host of input modalities have been considered in other NLG tasks, such as text summarization BIBREF24 , image captioning BIBREF25 and table-to-text generation BIBREF26 , traditional QG mainly focused on textual inputs, especially declarative sentences, explained by the original application domains of question answering and education, which also typically featured textual inputs.
Recently, with the growth of various QA applications such as Knowledge Base Question Answering (KBQA) BIBREF27 and Visual Question Answering (VQA) BIBREF28 , NQG research has also widened the spectrum of sources to include knowledge bases BIBREF29 and images BIBREF10 . This trend is also spurred by the remarkable success of neural models in feature representation, especially on image features BIBREF30 and knowledge representations BIBREF31 . We discuss adapting NQG models to other input modalities in Section "Wider Input Modalities" .
## Cognitive Levels
Finally, we consider the required cognitive process behind question asking, a distinguishing factor for questions BIBREF32 . A typical framework that attempts to categorize the cognitive levels involved in question asking comes from Bloom's taxonomy BIBREF33 , which has undergone several revisions and currently has six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating and Creating BIBREF32 .
Traditional QG focuses on shallow levels of Bloom's taxonomy: typical QG research is on generating sentence-based factoid questions (e.g., Who, What, Where questions), whose answers are simple constituents in the input sentence BIBREF2 , BIBREF13 . However, a QG system achieving human cognitive level should be able to generate meaningful questions that cater to higher levels of Bloom's taxonomy BIBREF34 , such as Why, What-if, and How questions. Traditionally, those “deep” questions are generated through shallow methods such as handcrafted templates BIBREF20 , BIBREF21 ; however, these methods lack a real understanding and reasoning over the input.
Although asking deep questions is complex, NQG's ability to generalize over voluminous data has enabled recent research to explore the comprehension and reasoning aspects of QG BIBREF35 , BIBREF1 , BIBREF8 , BIBREF34 . We investigate this trend in Section "Generation of Deep Questions" , examining the limitations of current Seq2Seq model in generating deep questions, and the efforts made by existing works, indicating further directions ahead.
The rest of this paper provides a systematic survey of NQG, covering corpus and evaluation metrics before examining specific neural models.
## Corpora
As QG can be regarded as a dual task of QA, in principle any QA dataset can be used for QG as well. However, there are at least two corpus-related factors that affect the difficulty of question generation. The first is the required cognitive level to answer the question, as we discussed in the previous section. Current NQG has achieved promising results on datasets consisting mainly of shallow factoid questions, such as SQuAD BIBREF36 and MS MARCO BIBREF38 . However, the performance drops significantly on deep question datasets, such as LearningQ BIBREF8 , shown in Section "Generation of Deep Questions" . The second factor is the answer type, i.e., the expected form of the answer, typically having four settings: (1) the answer is a text span in the passage, which is usually the case for factoid questions, (2) human-generated, abstractive answer that may not appear in the passage, usually the case for deep questions, (3) multiple choice question where question and its distractors should be jointly generated, and (4) no given answer, which requires the model to automatically learn what is worthy to ask. The design of NQG system differs accordingly.
Table 1 presents a listing of the NQG corpora grouped by their cognitive level and answer type, along with their statistics. Among them, SQuAD was used by most groups as the benchmark to evaluate their NQG models. This provides a fair comparison between different techniques. However, it raises the issue that most NQG models work on factoid questions with answer as text span, leaving other types of QG problems less investigated, such as generating deep multi-choice questions. To overcome this, a wider variety of corpora should be benchmarked against in future NQG research.
## Evaluation Metrics
Although the datasets are commonly shared between QG and QA, it is not the case for evaluation: it is challenging to define a gold standard of proper questions to ask. Meaningful, syntactically correct, semantically sound and natural are all useful criteria, yet they are hard to quantify. Most QG systems involve human evaluation, commonly by randomly sampling a few hundred generated questions, and asking human annotators to rate them on a 5-point Likert scale. The average rank or the percentage of best-ranked questions are reported and used for quality marks.
As human evaluation is time-consuming, common automatic evaluation metrics for NLG, such as BLEU BIBREF41 , METEOR BIBREF42 , and ROUGE BIBREF43 , are also widely used. However, some studies BIBREF44 , BIBREF45 have shown that these metrics do not correlate well with fluency, adequacy, coherence, as they essentially compute the $n$ -gram similarity between the source sentence and the generated question. To overcome this, BIBREF46 proposed a new metric to evaluate the “answerability” of a question by calculating the scores for several question-specific factors, including question type, content words, function words, and named entities. However, as it is newly proposed, it has not been applied to evaluate any NQG system yet.
To accurately measure what makes a good question, especially deep questions, improved evaluation schemes are required to specifically investigate the mechanism of question asking.
## Methodology
Many current NQG models follow the Seq2Seq architecture. Under this framework, given a passage (usually a sentence) $X = (x_1, \cdots , x_n)$ and (possibly) a target answer $A$ (a text span in the passage) as input, an NQG model aims to generate a question $Y = (y_1, \cdots , y_m)$ asking about the target answer $A$ in the passage $X$ , which is defined as finding the best question $\bar{Y}$ that maximizes the conditional likelihood given the passage $X$ and the answer $A$ :
$$\bar{Y} & = \arg \max _Y P(Y \vert X, A) \\
\vspace{-14.22636pt}
& = \arg \max _Y \sum _{t=1}^m P(y_t \vert X, A, y_{< t})$$ (Eq. 5)
BIBREF47 pioneered the first NQG model using an attention Seq2Seq model BIBREF22 , which feeds a sentence into an RNN-based encoder, and generate a question about the sentence through a decoder. The attention mechanism is applied to help decoder pay attention to the most relevant parts of the input sentence while generating a question. Note that this base model does not take the target answer as input. Subsequently, neural models have adopted attention mechanism as a default BIBREF48 , BIBREF49 , BIBREF50 .
Although these NQG models all share the Seq2Seq framework, they differ in the consideration of — (1) QG-specific factors (e.g., answer encoding, question word generation, and paragraph-level contexts), and (2) common NLG techniques (e.g., copying mechanism, linguistic features, and reinforcement learning) — discussed next.
## Encoding Answers
The most commonly considered factor by current NQG systems is the target answer, which is typically taken as an additional input to guide the model in deciding which information to focus on when generating; otherwise, the NQG model tend to generate questions without specific target (e.g., “What is mentioned?"). Models have solved this by either treating the answer's position as an extra input feature BIBREF48 , BIBREF51 , or by encoding the answer with a separate RNN BIBREF49 , BIBREF52 .
The first type of method augments each input word vector with an extra answer indicator feature, indicating whether this word is within the answer span. BIBREF48 implement this feature using the BIO tagging scheme, while BIBREF50 directly use a binary indicator. In addition to the target answer, BIBREF53 argued that the context words closer to the answer also deserve more attention from the model, since they are usually more relevant. To this end, they incorporate trainable position embeddings $(d_{p_1}, d_{p_2}, \cdots , d_{p_n})$ into the computation of attention distribution, where $p_i$ is the relative distance between the $i$ -th word and the answer, and $d_{p_i}$ is the embedding of $p_i$ . This achieved an extra BLEU-4 gain of $0.89$ on SQuAD.
To generate answer-related questions, extra answer indicators explicitly emphasize the importance of answer; however, it also increases the tendency that generated questions include words from the answer, resulting in useless questions, as observed by BIBREF52 . For example, given the input “John Francis O’Hara was elected president of Notre Dame in 1934.", an improperly generated question would be “Who was elected John Francis?", which exposes some words in the answer. To address this, they propose to replace the answer into a special token for passage encoding, and a separate RNN is used to encode the answer. The outputs from two encoders are concatenated as inputs to the decoder. BIBREF54 adopted a similar idea that separately encodes passage and answer, but they instead use the multi-perspective matching between two encodings as an extra input to the decoder.
We forecast treating the passage and the target answer separately as a future trend, as it results in a more flexible model, which generalizes to the abstractive case when the answer is not a text span in the input passage. However, this inevitably increases the model complexity and difficulty in training.
## Question Word Generation
Question words (e.g., “when”, “how”, and “why”) also play a vital role in QG; BIBREF53 observed that the mismatch between generated question words and answer type is common for current NQG systems. For example, a when-question should be triggered for answer “the end of the Mexican War" while a why-question is generated by the model. A few works BIBREF49 , BIBREF53 considered question word generation separately in model design.
BIBREF49 proposed to first generate a question template that contains question word (e.g., “how to #", where # is the placeholder), before generating the rest of the question. To this end, they train two Seq2Seq models; the former learns to generate question templates for a given text , while the latter learns to fill the blank of template to form a complete question. Instead of a two-stage framework, BIBREF53 proposed a more flexible model by introducing an additional decoding mode that generates the question word. When entering this mode, the decoder produces a question word distribution based on a restricted set of vocabulary using the answer embedding, the decoder state, and the context vector. The switch between different modes is controlled by a discrete variable produced by a learnable module of the model in each decoding step.
Determining the appropriate question word harks back to question type identification, which is correlated with the question intention, as different intents may yield different questions, even when presented with the same (passage, answer) input pair. This points to the direction of exploring question pragmatics, where external contextual information (such as intent) can inform and influence how questions should optimally be generated.
## Paragraph-level Contexts
Leveraging rich paragraph-level contexts around the input text is another natural consideration to produce better questions. According to BIBREF47 , around 20% of questions in SQuAD require paragraph-level information to be answered. However, as input texts get longer, Seq2Seq models have a tougher time effectively utilizing relevant contexts, while avoiding irrelevant information.
To address this challenge, BIBREF51 proposed a gated self-attention encoder to refine the encoded context by fusing important information with the context's self-representation properly, which has achieved state-of-the-art results on SQuAD. The long passage consisting of input texts and its context is first embedded via LSTM with answer position as an extra feature. The encoded representation is then fed through a gated self-matching network BIBREF55 to aggregate information from the entire passage and embed intra-passage dependencies. Finally, a feature fusion gate BIBREF56 chooses relevant information between the original and self-matching enhanced representations.
Instead of leveraging the whole context, BIBREF57 performed a pre-filtering by running a coreference resolution system on the context passage to obtain coreference clusters for both the input sentence and the answer. The co-referred sentences are then fed into a gating network, from which the outputs serve as extra features to be concatenated with the original input vectors.
## Answer-unaware QG
The aforementioned models require the target answer as an input, in which the answer essentially serves as the focus of asking. However, in the case that only the input passage is given, a QG system should automatically identify question-worthy parts within the passage. This task is synonymous with content selection in traditional QG. To date, only two works BIBREF58 , BIBREF59 have worked in this setting. They both follow the traditional decomposition of QG into content selection and question construction but implement each task using neural networks. For content selection, BIBREF58 learn a sentence selection task to identify question-worthy sentences from the input paragraph using a neural sequence tagging model. BIBREF59 train a neural keyphrase extractor to predict keyphrases of the passage. For question construction, they both employed the Seq2Seq model, for which the input is either the selected sentence or the input passage with keyphrases as target answer.
However, learning what aspect to ask about is quite challenging when the question requires reasoning over multiple pieces of information within the passage; cf the Gollum question from the introduction. Beyond retrieving question-worthy information, we believe that studying how different reasoning patterns (e.g., inductive, deductive, causal and analogical) affects the generation process will be an aspect for future study.
## Technical Considerations
Common techniques of NLG have also been considered in NQG model, summarized as 3 tactics:
1. Copying Mechanism. Most NQG models BIBREF48 , BIBREF60 , BIBREF61 , BIBREF50 , BIBREF62 employ the copying mechanism of BIBREF23 , which directly copies relevant words from the source sentence to the question during decoding. This idea is widely accepted as it is common to refer back to phrases and entities appearing in the text when formulating factoid questions, and difficult for a RNN decoder to generate such rare words on its own.
2. Linguistic Features. Approaches also seek to leverage additional linguistic features that complements word embeddings, including word case, POS and NER tags BIBREF48 , BIBREF61 as well as coreference BIBREF50 and dependency information BIBREF62 . These categorical features are vectorized and concatenated with word embeddings. The feature vectors can be either one-hot or trainable and serve as input to the encoder.
3. Policy Gradient. Optimizing for just ground-truth log likelihood ignores the many equivalent ways of asking a question. Relevant QG work BIBREF60 , BIBREF63 have adopted policy gradient methods to add task-specific rewards (such as BLEU or ROUGE) to the original objective. This helps to diversify the questions generated, as the model learns to distribute probability mass among equivalent expressions rather than the single ground truth question.
## The State of the Art
In Table 2 , we summarize existing NQG models with their employed techniques and their best-reported performance on SQuAD. These methods achieve comparable results; as of this writing, BIBREF51 is the state-of-the-art.
Two points deserve mention. First, while the copying mechanism has shown marked improvements, there exist shortcomings. BIBREF52 observed many invalid answer-revealing questions attributed to the use of the copying mechanism; cf the John Francis example in Section "Emerging Trends" . They abandoned copying but still achieved a performance rivaling other systems. In parallel application areas such as machine translation, the copy mechanism has been to a large extent replaced with self-attention BIBREF64 or transformer BIBREF65 . The future prospect of the copying mechanism requires further investigation. Second, recent approaches that employ paragraph-level contexts have shown promising results: not only boosting performance, but also constituting a step towards deep question generation, which requires reasoning over rich contexts.
## Emerging Trends
We discuss three trends that we wish to call practitioners' attention to as NQG evolves to take the center stage in QG: Multi-task Learning, Wider Input Modalities and Deep Question Generation.
## Multi-task Learning
As QG has become more mature, work has started to investigate how QG can assist in other NLP tasks, and vice versa. Some NLP tasks benefit from enriching training samples by QG to alleviate the data shortage problem. This idea has been successfully applied to semantic parsing BIBREF66 and QA BIBREF67 . In the semantic parsing task that maps a natural language question to a SQL query, BIBREF66 achieved a 3 $\%$ performance gain with an enlarged training set that contains pseudo-labeled $(SQL, question)$ pairs generated by a Seq2Seq QG model. In QA, BIBREF67 employed the idea of self-training BIBREF68 to jointly learn QA and QG. The QA and QG models are first trained on a labeled corpus. Then, the QG model is used to create more questions from an unlabeled text corpus and the QA model is used to answer these newly-created questions. The newly-generated question–answer pairs form an enlarged dataset to iteratively retrain the two models. The process is repeated while performance of both models improve.
Investigating the core aspect of QG, we say that a well-trained QG system should have the ability to: (1) find the most salient information in the passage to ask questions about, and (2) given this salient information as target answer, to generate an answer related question. BIBREF69 leveraged the first characteristic to improve text summarization by performing multi-task learning of summarization with QG, as both these two tasks require the ability to search for salient information in the passage. BIBREF49 applied the second characteristic to improve QA. For an input question $q$ and a candidate answer $\hat{a}$ , they generate a question $\hat{q}$ for $\hat{a}$ by way of QG system. Since the generated question $\hat{q}$ is closely related to $\hat{a}$ , the similarity between $q$ and $\hat{q}$ helps to evaluate whether $\hat{a}$ is the correct answer.
Other works focus on jointly training to combine QG and QA. BIBREF70 simultaneously train the QG and QA models in the same Seq2Seq model by alternating input data between QA and QG examples. BIBREF71 proposed a training algorithm that generalizes Generative Adversarial Network (GANs) BIBREF72 under the question answering scenario. The model improves QG by incorporating an additional QA-specific loss, and improving QA performance by adding artificially generated training instances from QG. However, while joint training has shown some effectiveness, due to the mixed objectives, its performance on QG are lower than the state-of-the-art results, which leaves room for future exploration.
## Wider Input Modalities
QG work now has incorporated input from knowledge bases (KBQG) and images (VQG).
Inspired by the use of SQuAD as a question benchmark, BIBREF9 created a 30M large-scale dataset of (KB triple, question) pairs to spur KBQG work. They baselined an attention seq2seq model to generate the target factoid question. Due to KB sparsity, many entities and predicates are unseen or rarely seen at training time. BIBREF73 address these few-/zero-shot issues by applying the copying mechanism and incorporating textual contexts to enrich the information for rare entities and relations. Since a single KB triple provides only limited information, KB-generated questions also overgeneralize — a model asks “Who was born in New York?" when given the triple (Donald_Trump, Place_of_birth, New_York). To solve this, BIBREF29 enrich the input with a sequence of keywords collected from its related triples.
Visual Question Generation (VQG) is another emerging topic which aims to ask questions given an image. We categorize VQG into grounded- and open-ended VQG by the level of cognition. Grounded VQG generates visually grounded questions, i.e., all relevant information for the answer can be found in the input image BIBREF74 . A key purpose of grounded VQG is to support the dataset construction for VQA. To ensure the questions are grounded, existing systems rely on image captions to varying degrees. BIBREF75 and BIBREF76 simply convert image captions into questions using rule-based methods with textual patterns. BIBREF74 proposed a neural model that can generate questions with diverse types for a single image, using separate networks to construct dense image captions and to select question types.
In contrast to grounded QG, humans ask higher cognitive level questions about what can be inferred rather than what can be seen from an image. Motivated by this, BIBREF10 proposed open-ended VQG that aims to generate natural and engaging questions about an image. These are deep questions that require high cognition such as analyzing and creation. With significant progress in deep generative models, marked by variational auto-encoders (VAEs) and GANs, such models are also used in open-ended VQG to bring “creativity” into generated questions BIBREF77 , BIBREF78 , showing promising results. This also brings hope to address deep QG from text, as applied in NLG: e.g., SeqGAN BIBREF79 and LeakGAN BIBREF80 .
## Generation of Deep Questions
Endowing a QG system with the ability to ask deep questions will help us build curious machines that can interact with humans in a better manner. However, BIBREF81 pointed out that asking high-quality deep questions is difficult, even for humans. Citing the study from BIBREF82 to show that students in college asked only about 6 deep-reasoning questions per hour in a question–encouraging tutoring session. These deep questions are often about events, evaluation, opinions, syntheses or reasons, corresponding to higher-order cognitive levels.
To verify the effectiveness of existing NQG models in generating deep questions, BIBREF8 conducted an empirical study that applies the attention Seq2Seq model on LearningQ, a deep-question centric dataset containing over 60 $\%$ questions that require reasoning over multiple sentences or external knowledge to answer. However, the results were poor; the model achieved miniscule BLEU-4 scores of $< 4$ and METEOR scores of $< 9$ , compared with $> 12$ (BLEU-4) and $> 16$ (METEOR) on SQuAD. Despite further in-depth analysis are needed to explore the reasons behind, we believe there are two plausible explanations: (1) Seq2Seq models handle long inputs ineffectively, and (2) Seq2Seq models lack the ability to reason over multiple pieces of information.
Despite still having a long way to go, some works have set out a path forward. A few early QG works attempted to solve this through building deep semantic representations of the entire text, using concept maps over keywords BIBREF83 or minimal recursion semantics BIBREF84 to reason over concepts in the text. BIBREF35 proposed a crowdsourcing-based workflow that involves building an intermediate ontology for the input text, soliciting question templates through crowdsourcing, and generating deep questions based on template retrieval and ranking. Although this process is semi-automatic, it provides a practical and efficient way towards deep QG. In a separate line of work, BIBREF1 proposed a framework that simulates how people ask deep questions by treating questions as formal programs that execute on the state of the world, outputting an answer.
Based on our survey, we believe the roadmap towards deep NGQ points towards research that will (1) enhance the NGQ model with the ability to consider relationships among multiple source sentences, (2) explicitly model typical reasoning patterns, and (3) understand and simulate the mechanism behind human question asking.
## Conclusion – What's the Outlook?
We have presented a comprehensive survey of NQG, categorizing current NQG models based on different QG-specific and common technical variations, and summarizing three emerging trends in NQG: multi-task learning, wider input modalities, and deep question generation.
What's next for NGQ? We end with future potential directions by applying past insights to current NQG models; the “unknown unknown", promising directions yet explored.
When to Ask: Besides learning what and how to ask, in many real-world applications that question plays an important role, such as automated tutoring and conversational systems, learning when to ask become an important issue. In contrast to general dialog management BIBREF85 , no research has explored when machine should ask an engaging question in dialog. Modeling question asking as an interactive and dynamic process may become an interesting topic ahead.
Personalized QG: Question asking is quite personalized: people with different characters and knowledge background ask different questions. However, integrating QG with user modeling in dialog management or recommendation system has not yet been explored. Explicitly modeling user state and awareness leads us towards personalized QG, which dovetails deep, end-to-end QG with deep user modeling and pairs the dual of generation–comprehension much in the same vein as in the vision–image generation area.
| 19
|
1905.10810
|
Evaluation of basic modules for isolated spelling error correction in Polish texts
|
# Evaluation of basic modules for isolated spelling error correction in Polish texts
## Abstract
Spelling error correction is an important problem in natural language processing, as a prerequisite for good performance in downstream tasks as well as an important feature in user-facing applications. For texts in Polish language, there exist works on specific error correction solutions, often developed for dealing with specialized corpora, but not evaluations of many different approaches on big resources of errors. We begin to address this problem by testing some basic and promising methods on PlEWi, a corpus of annotated spelling extracted from Polish Wikipedia. These modules may be further combined with appropriate solutions for error detection and context awareness. Following our results, combining edit distance with cosine distance of semantic vectors may be suggested for interpretable systems, while an LSTM, particularly enhanced by ELMo embeddings, seems to offer the best raw performance.
## Introduction
Spelling error correction is a fundamental NLP task. Most language processing applications benefit greatly from being provided clean texts for their best performance. Human users of computers also often expect competent help in making spelling of their texts correct.
Because of the lack of tests of many common spelling correction methods for Polish, it is useful to establish how they perform in a simple scenario. We constrain ourselves to the pure task of isolated correction of non-word errors. They are traditionally separated in error correction literature BIBREF0 . Non-word errors are here incorrect word forms that not only differ from what was intended, but also do not constitute another, existing word themselves. Much of the initial research on error correction focused on this simple task, tackled without means of taking the context of the nearest words into account.
It is true that, especially in the case of neural networks, it is often possible and desirable to combine problems of error detection, correction and context awareness into one task trained with a supervised training procedure. In language correction research for English language also grammatical and regular spelling errors have been treated uniformly with much success BIBREF1 .
However, when more traditional methods are used, because of their predictability and interpretability for example, one can mix and match various approaches to dealing with the subproblems of detection, correction and context handling (often equivalent to employing some kind of a language model). We call it a modular approach to building spelling error correction systems. There is recent research where this paradigm was applied, interestingly, to convolutional networks trained separately for various subtasks BIBREF2 . In similar setups it is more useful to assess abilities of various solutions in isolation. The exact architecture of a spelling correction system should depend on characteristics of texts it will work on.
Similar considerations eliminated from our focus handcrafted solutions for the whole spelling correction pipeline, primarily the LanguageTool BIBREF3 . Its performance in fixing spelling of Polish tweets was already tested BIBREF4 . For our purposes it would be given an unfair advantage, since it is a rule-based system making heavy use of words in context of the error.
## Problems of spelling correction for Polish
Published work on language correction for Polish dates back at least to 1970s, when simplest Levenshtein distance solutions were used for cleaning mainframe inputs BIBREF5 , BIBREF6 . Spelling correction tests described in literature have tended to focus on one approach applied to a specific corpus. Limited examples include works on spellchecking mammography reports and tweets BIBREF7 , BIBREF4 . These works emphasized the importance of tailoring correction systems to specific problems of corpora they are applied to. For example, mammography reports suffer from poor typing, which in this case is a repetitive work done in relative hurry. Tweets, on the other hand, tend to contain emoticons and neologisms that can trick solutions based on rules and dictionaries, such as LanguageTool. The latter is, by itself, fairly well suited for Polish texts, since a number of extensions to the structure of this application was inspired by problems with morphology of Polish language BIBREF3 .
These existing works pointed out more general, potentially useful qualities specific to spelling errors in Polish language texts. It is, primarily, the problem of leaving out diacritical signs, or, more rarely, adding them in wrong places. This phenomenon stems from using a variant of the US keyboard layout, where combinations of AltGr with some alphabetic keys produces characters unique to Polish. When the user forgets or neglects to press the AltGr key, typos such as writing *olowek instead of ołówek appear. In fact, BIBREF4 managed to get substantial performance on Twitter corpus by using this ”diacritical swapping” alone.
## Baseline methods
The methods that we evaluated are baselines are the ones we consider to be basic and with moderate potential of yielding particularly good results. Probably the most straightforward approach to error correction is selecting known words from a dictionary that are within the smallest edit distance from the error. We used the Levenshtein distance metric BIBREF8 implemented in Apache Lucene library BIBREF9 . It is a version of edit distance that treats deletions, insertions and replacements as adding one unit distance, without giving a special treatment to character swaps. The SGJP – Grammatical Dictionary of Polish BIBREF10 was used as the reference vocabulary.
Another simple approach is the aforementioned diacritical swapping, which is a term that we introduce here for referring to a solution inspired by the work of BIBREF4 . Namely, from the incorrect form we try to produce all strings obtainable by either adding or removing diacritical marks from characters. We then exclude options that are not present in SGJP, and select as the correction the one within the smallest edit distance from the error. It is possible for the number of such diacritically-swapped options to become very big. For example, the token Modlin-Zegrze-Pultusk-Różan-Ostrołęka-Łomża-Osowiec (taken from PlEWi corpus of spelling errors, see below) can yield over INLINEFORM0 states with this method, such as Módłiń-Żęgrzę-Pułtuśk-Roźąń-Óśtróleką-Lómzą-Óśówięć. The actual correction here is just fixing the ł in Pułtusk. Hence we only try to correct in this way tokens that are shorter than 17 characters.
## Vector distance
A promising method, adapted from work on correcting texts by English language learners BIBREF11 , expands on the concept of selecting a correction nearest to the spelling error according to some notion of distance. Here, the Levenshtein distance is used in a weighted sum to cosine distance between word vectors. This is based on the observation that trained vectors models of distributional semantics contain also representations of spelling errors, if they were not pruned. Their representations tend to be similar to those of their correct counterparts. For example, the token enginir will appear in similar contexts as engineer, and therefore will be assigned a similar vector embedding.
The distance between two tokens INLINEFORM0 and INLINEFORM1 is thus defined as INLINEFORM2
Here INLINEFORM0 is just Levenshtein distance between strings, and INLINEFORM1 – cosine distance between vectors. INLINEFORM2 denotes the word vector for INLINEFORM3 . Both distance metrics are in our case roughly in the range [0,1] thanks to the scaling of edit distance performed automatically by Apache Lucene. We used a pretrained set of word embeddings of Polish BIBREF12 , obtained with the flavor word2vec procedure using skipgrams and negative sampling BIBREF13 .
## Recurrent neural networks
Another powerful approach, if conceptually simple in linguistic terms, is using a character-based recurrent neural network. Here, we test uni- and bidirectional Long Short-Term Memory networks BIBREF14 that are fed characters of the error as their input and are expected to output its correct form, character after character. This is similar to traditional solutions conceptualizing the spelling error as a chain of characters, which are used as evidence to predict the most likely chain of replacements (original characters). This was done with n-gram methods, Markov chains and other probabilistic models BIBREF15 . Since nowadays neural networks enjoy a large awareness as an element of software infrastructure, with actively maintained packages readily available, their evaluation seems to be the most practically useful. We used the PyTorch BIBREF16 implementation of LSTM in particular.
The bidirectional version BIBREF17 of LSTM reads the character chains forward and backwards at the same time. Predictions from networks running in both directions are averaged.
In order to provide the network an additional, broad picture peek at the whole error form we also evaluated a setup where the internal state of LSTM cells, instead of being initialized randomly, is computed from an ELMo embedding BIBREF18 of the token. The ELMo embedder is capable of integrating linguistic information carried by the whole form (probably often not much in case of errors), as well as the string as a character chain. The latter is processed with a convolutional neural network. How this representation is constructed is informed by the whole corpus on which the embedder was trained. The pretrained ELMo model that we used BIBREF19 was trained on Wikipedia and Common Crawl corpora of Polish.
The ELMo embedding network outputs three layers as matrices, which are supposed to reflect subsequent compositional layers of language, from phonetic phenomena at the bottom to lexical ones at the top. A weighted sum of these layers is computed, with weights trained along with the LSTM error-correcting network. Then we apply a trained linear transformation, followed by INLINEFORM0 non-linearity: INLINEFORM1
(applied cellwise) in order to obtain the initial setting of parameters for the main LSTM. Our ELMo-augmented LSTM is bidirectional.
## Experimental setup
PlEWi BIBREF20 is an early version of WikEd BIBREF21 error corpus, containing error type annotations allowing us to select only non-word errors for evaluation. Specifically, PlEWi supplied 550,755 [error, correction] pairs, from which 298,715 were unique. The corpus contains data extracted from histories of page versions of Polish Wikipedia. An algorithm designed by the corpus author determined where the changes were correcting spelling errors, as opposed to expanding content and disagreements among Wikipedia editors.
The corpus features texts that are descriptive rather than conversational, contain relatively many proper names and are more likely to have been at least skimmed by the authors before submitting for online publication. Error cases provided by PlEWi are, therefore, not a balanced representation of spelling errors in written Polish language. PlEWi does have the advantage of scale in comparison to existing literature, such as BIBREF4 operating on a set of only 740 annotated errors in tweets.
All methods were tested on a test subset of 25% of cases, with 75% left for training (where needed) and 5% for development.
The methods that required training – namely recurrent neural networks – had their loss measured as cross-entropy loss measure between correct character labels and predictions. This value was minimized with Adam algorithm BIBREF22 . The networks were trained for 35 epochs.
## Results
The experimental results are presented in Table TABREF4 . Diacritic swapping showed a remarkably poor performance, despite promising mentions in existing literature. This might be explained by the already mentioned feature of Wikipedia edits, which can be expected to be to some degree self-reviewed before submission. This can very well limit the number of most trivial mistakes.
On the other hand, the vector distance method was able to bring a discernible improvement over pure Levenshtein distance, comparable even with the most basic LSTM. It is possible that assigning more fine-tuned weights to edit distance and semantic distance would make the quality of predictions even higher. The idea of using vector space measurements explicitly can be also expanded if we were to consider the problem of contextualizing corrections. For example, the semantic distance of proposed corrections to the nearest words is likely to carry much information about their appropriateness. Looking from another angle, searching for words that seem semantically off in context may be a good heuristic for detecting errors that are not nonword (that is, they lead to wrong forms appearing in text which are nevertheless in-vocabulary).
The good performance of recurrent network methods is hardly a surprise, given observed effectiveness of neural networks in many NLP tasks in the recent decade. It seems that bidirectional LSTM augmented with ELMo may already hit the limit for correcting Polish spelling errors without contextual information. While it improves accuracy in comparison to LSTM initialized withrandom noise, it makes the test cross-entropy slightly worse, which hints at overfitting. The perplexity measures actually increase sharply for more sophisticated architectures. Perplexity should show how little probability is assigned by the model to true answers. We measure it as INLINEFORM0
where INLINEFORM0 is a sequence of INLINEFORM1 characters, forming the correct version of the word, and INLINEFORM2 is the estimated probability of the INLINEFORM3 th character, given previous predicted characters and the incorrect form. The observed increase of perplexity for increasingly accurate models is most likely due to more refined predicted probability distributions, which go beyond just assigning the bulk of probability to the best answer.
Interesting insights can be gained from weights assigned by optimization to layers of ELMo network, which are taken as the word form embedding (Table TABREF5 ). The first layer, and the one that is nearest to input of the network, is given relatively the least importance, while the middle one dominates both others taken together. This suggests that in error correction, at least for Polish, the middle level of morphemes and other characteristic character chunks is more important than phenomena that are low-level or tied to some specific words. This observation should be taken into account in further research on practical solutions for spelling correction.
## Conclusion
Among the methods tested the bidirectional LSTM, especially initialized by ELMo embeddings, offers the best accuracy and raw performance. Adding ELMo to a straightforward PyTorch implementation of LSTM may be easier now than at the time of performing our tests, as since then the authors of ELMoForManyLangs package BIBREF19 improved their programmatic interface. However, if a more interpretable and explainable output is required, some version of vector distance combined with edit distance may be the best direction. It should be noted that this method produces multiple candidate corrections with their similarity scores, as opposed to only one “best guess“ correction that can be obtained from a character-based LSTM. This is important in applications where it is up to humans to the make the final decision, and they are only to be aided by a machine.
It is desirable for further reasearch to expand the corpus material into a wider and more representative set of texts. Nevertheless, the solution for any practical case has to be tailored to its characteristic error patterns. Works on language correction for English show that available corpora can be ”boosted” BIBREF1 , i.e. expanded by generating new errors consistent with a generative model inferred from the data. This may greatly aid in developing models that are dependent on learning from error corpora.
A deliberate omission in this paper are the elements accompanying most real-word error correction solutions. Some fairly obvious approaches to integrating evidence from context include n-grams and Markov chains, although the possibility of using measurements in spaces of semantic vectors was already mentioned in this article. Similarly, non-word errors can be easily detected with comparing tokens against reference vocabulary, but in practice one should have ways of detecting mistakes masquerading as real words and fixing bad segmentation (tokens that are glued together or improperly separated). Testing how performant are various methods for dealing with these problems in Polish language is left for future research.
| 8
|
1905.11268
|
Combating Adversarial Misspellings with Robust Word Recognition
|
# Combating Adversarial Misspellings with Robust Word Recognition
## Abstract
To combat adversarial spelling mistakes, we propose placing a word recognition model in front of the downstream classifier. Our word recognition models build upon the RNN semicharacter architecture, introducing several new backoff strategies for handling rare and unseen words. Trained to recognize words corrupted by random adds, drops, swaps, and keyboard mistakes, our method achieves 32% relative (and 3.3% absolute) error reduction over the vanilla semi-character model. Notably, our pipeline confers robustness on the downstream classifier, outperforming both adversarial training and off-the-shelf spell checkers. Against a BERT model fine-tuned for sentiment analysis, a single adversarially-chosen character attack lowers accuracy from 90.3% to 45.8%. Our defense restores accuracy to 75%
1 . Surprisingly, better word recognition does not always entail greater robustness. Our analysis reveals that robustness also depends upon a quantity that we denote the sensitivity.
## Introduction
Despite the rapid progress of deep learning techniques on diverse supervised learning tasks, these models remain brittle to subtle shifts in the data distribution. Even when the permissible changes are confined to barely-perceptible perturbations, training robust models remains an open challenge. Following the discovery that imperceptible attacks could cause image recognition models to misclassify examples BIBREF0 , a veritable sub-field has emerged in which authors iteratively propose attacks and countermeasures.
For all the interest in adversarial computer vision, these attacks are rarely encountered outside of academic research. However, adversarial misspellings constitute a longstanding real-world problem. Spammers continually bombard email servers, subtly misspelling words in efforts to evade spam detection while preserving the emails' intended meaning BIBREF1 , BIBREF2 . As another example, programmatic censorship on the Internet has spurred communities to adopt similar methods to communicate surreptitiously BIBREF3 .
In this paper, we focus on adversarially-chosen spelling mistakes in the context of text classification, addressing the following attack types: dropping, adding, and swapping internal characters within words. These perturbations are inspired by psycholinguistic studies BIBREF4 , BIBREF5 which demonstrated that humans can comprehend text altered by jumbling internal characters, provided that the first and last characters of each word remain unperturbed.
First, in experiments addressing both BiLSTM and fine-tuned BERT models, comprising four different input formats: word-only, char-only, word+char, and word-piece BIBREF6 , we demonstrate that an adversary can degrade a classifier's performance to that achieved by random guessing. This requires altering just two characters per sentence. Such modifications might flip words either to a different word in the vocabulary or, more often, to the out-of-vocabulary token UNK. Consequently, adversarial edits can degrade a word-level model by transforming the informative words to UNK. Intuitively, one might suspect that word-piece and character-level models would be less susceptible to spelling attacks as they can make use of the residual word context. However, our experiments demonstrate that character and word-piece models are in fact more vulnerable. We show that this is due to the adversary's effective capacity for finer grained manipulations on these models. While against a word-level model, the adversary is mostly limited to UNK-ing words, against a word-piece or character-level model, each character-level add, drop, or swap produces a distinct input, providing the adversary with a greater set of options.
Second, we evaluate first-line techniques including data augmentation and adversarial training, demonstrating that they offer only marginal benefits here, e.g., a BERT model achieving $90.3$ accuracy on a sentiment classification task, is degraded to $64.1$ by an adversarially-chosen 1-character swap in the sentence, which can only be restored to $69.2$ by adversarial training.
Third (our primary contribution), we propose a task-agnostic defense, attaching a word recognition model that predicts each word in a sentence given a full sequence of (possibly misspelled) inputs. The word recognition model's outputs form the input to a downstream classification model. Our word recognition models build upon the RNN-based semi-character word recognition model due to BIBREF7 . While our word recognizers are trained on domain-specific text from the task at hand, they often predict UNK at test time, owing to the small domain-specific vocabulary. To handle unobserved and rare words, we propose several backoff strategies including falling back on a generic word recognizer trained on a larger corpus. Incorporating our defenses, BERT models subject to 1-character attacks are restored to $88.3$ , $81.1$ , $78.0$ accuracy for swap, drop, add attacks respectively, as compared to $69.2$ , $63.6$ , and $50.0$ for adversarial training
Fourth, we offer a detailed qualitative analysis, demonstrating that a low word error rate alone is insufficient for a word recognizer to confer robustness on the downstream task. Additionally, we find that it is important that the recognition model supply few degrees of freedom to an attacker. We provide a metric to quantify this notion of sensitivity in word recognition models and study its relation to robustness empirically. Models with low sensitivity and word error rate are most robust.
## Related Work
Several papers address adversarial attacks on NLP systems. Changes to text, whether word- or character-level, are all perceptible, raising some questions about what should rightly be considered an adversarial example BIBREF8 , BIBREF9 . BIBREF10 address the reading comprehension task, showing that by appending distractor sentences to the end of stories from the SQuAD dataset BIBREF11 , they could cause models to output incorrect answers. Inspired by this work, BIBREF12 demonstrate an attack that breaks entailment systems by replacing a single word with either a synonym or its hypernym. Recently, BIBREF13 investigated the problem of producing natural-seeming adversarial examples, noting that adversarial examples in NLP are often ungrammatical BIBREF14 .
In related work on character-level attacks, BIBREF8 , BIBREF15 explored gradient-based methods to generate string edits to fool classification and translation systems, respectively. While their focus is on efficient methods for generating adversaries, ours is on improving the worst case adversarial performance. Similarly, BIBREF9 studied how synthetic and natural noise affects character-level machine translation. They considered structure invariant representations and adversarial training as defenses against such noise. Here, we show that an auxiliary word recognition model, which can be trained on unlabeled data, provides a strong defense.
Spelling correction BIBREF16 is often viewed as a sub-task of grammatical error correction BIBREF17 , BIBREF18 . Classic methods rely on a source language model and a noisy channel model to find the most likely correction for a given word BIBREF19 , BIBREF20 . Recently, neural techniques have been applied to the task BIBREF7 , BIBREF21 , which model the context and orthography of the input together. Our work extends the ScRNN model of BIBREF7 .
## Robust Word Recognition
To tackle character-level adversarial attacks, we introduce a simple two-stage solution, placing a word recognition model ( $W$ ) before the downstream classifier ( $C$ ). Under this scheme, all inputs are classified by the composed model $C \circ W$ . This modular approach, with $W$ and $C$ trained separately, offers several benefits: (i) we can deploy the same word recognition model for multiple downstream classification tasks/models; and (ii) we can train the word recognition model with larger unlabeled corpora.
Against adversarial mistakes, two important factors govern the robustness of this combined model: $W$ 's accuracy in recognizing misspelled words and $W$ 's sensitivity to adversarial perturbations on the same input. We discuss these aspects in detail below.
## ScRNN with Backoff
We now describe semi-character RNNs for word recognition, explain their limitations, and suggest techniques to improve them.
Inspired by the psycholinguistic studies BIBREF5 , BIBREF4 , BIBREF7 proposed a semi-character based RNN (ScRNN) that processes a sentence of words with misspelled characters, predicting the correct words at each step. Let $s = \lbrace w_1, w_2, \dots , w_n\rbrace $ denote the input sentence, a sequence of constituent words $w_i$ . Each input word ( $w_i$ ) is represented by concatenating (i) a one hot vector of the first character ( $\mathbf {w_{i1}}$ ); (ii) a one hot representation of the last character ( $\mathbf {w_{il}}$ , where $l$ is the length of word $w_i$ ); and (iii) a bag of characters representation of the internal characters ( $\sum _{j=2}^{l-1}\mathbf {w_{ij}})$ . ScRNN treats the first and the last characters individually, and is agnostic to the ordering of the internal characters. Each word, represented accordingly, is then fed into a BiLSTM cell. At each sequence step, the training target is the correct corresponding word (output dimension equal to vocabulary size), and the model is optimized with cross-entropy loss.
While BIBREF7 demonstrate strong word recognition performance, a drawback of their evaluation setup is that they only attack and evaluate on the subset of words that are a part of their training vocabulary. In such a setting, the word recognition performance is unreasonably dependent on the chosen vocabulary size. In principle, one can design models to predict (correctly) only a few chosen words, and ignore the remaining majority and still reach 100% accuracy. For the adversarial setting, rare and unseen words in the wild are particularly critical, as they provide opportunities for the attackers. A reliable word-recognizer should handle these cases gracefully. Below, we explore different ways to back off when the ScRNN predicts UNK (a frequent outcome for rare and unseen words):
Pass-through: word-recognizer passes on the (possibly misspelled) word as is.
Backoff to neutral word: Alternatively, noting that passing $\colorbox {gray!20}{\texttt {UNK}}$ -predicted words through unchanged exposes the downstream model to potentially corrupted text, we consider backing off to a neutral word like `a', which has a similar distribution across classes.
Backoff to background model: We also consider falling back upon a more generic word recognition model trained upon a larger, less-specialized corpus whenever the foreground word recognition model predicts UNK. Figure 1 depicts this scenario pictorially.
Empirically, we find that the background model (by itself) is less accurate, because of the large number of words it is trained to predict. Thus, it is best to train a precise foreground model on an in-domain corpus and focus on frequent words, and then to resort to a general-purpose background model for rare and unobserved words. Next, we delineate our second consideration for building robust word-recognizers.
## Model Sensitivity
In computer vision, an important factor determining the success of an adversary is the norm constraint on the perturbations allowed to an image ( $|| \bf x - \bf x^{\prime }||_{\infty } < \epsilon $ ). Higher values of $\epsilon $ lead to a higher chance of mis-classification for at least one $\bf x^{\prime }$ . Defense methods such as quantization BIBREF22 and thermometer encoding BIBREF23 try to reduce the space of perturbations available to the adversary by making the model invariant to small changes in the input.
In NLP, we often get such invariance for free, e.g., for a word-level model, most of the perturbations produced by our character-level adversary lead to an UNK at its input. If the model is robust to the presence of these UNK tokens, there is little room for an adversary to manipulate it. Character-level models, on the other hand, despite their superior performance in many tasks, do not enjoy such invariance. This characteristic invariance could be exploited by an attacker. Thus, to limit the number of different inputs to the classifier, we wish to reduce the number of distinct word recognition outputs that an attacker can induce, not just the number of words on which the model is “fooled”. We denote this property of a model as its sensitivity.
We can quantify this notion for a word recognition system $W$ as the expected number of unique outputs it assigns to a set of adversarial perturbations. Given a sentence $s$ from the set of sentences $\mathcal {S}$ , let $A(s) = {s_1}^{\prime } , {s_2}^{\prime }, \dots , {s_n}^{\prime }$ denote the set of $n$ perturbations to it under attack type $A$ , and let $V$ be the function that maps strings to an input representation for the downstream classifier. For a word level model, $V$ would transform sentences to a sequence of word ids, mapping OOV words to the same UNK ID. Whereas, for a char (or word+char, word-piece) model, $V$ would map inputs to a sequence of character IDs. Formally, sensitivity is defined as
$$S_{W,V}^A=\mathbb {E}_{s}\left[\frac{\#_{u}(V \circ W({s_1}^{\prime }), \dots , V \circ W({s_n}^{\prime }))}{n}\right] ,$$ (Eq. 12)
where $V \circ W (s_i)$ returns the input representation (of the downstream classifier) for the output string produced by the word-recognizer $W$ using $s_i$ and $\#_{u}(\cdot )$ counts the number of unique arguments.
Intuitively, we expect a high value of $S_{W, V}^A$ to lead to a lower robustness of the downstream classifier, since the adversary has more degrees of freedom to attack the classifier. Thus, when using word recognition as a defense, it is prudent to design a low sensitivity system with a low error rate. However, as we will demonstrate, there is often a trade-off between sensitivity and error rate.
## Synthesizing Adversarial Attacks
Suppose we are given a classifier $C: \mathcal {S} \rightarrow \mathcal {Y}$ which maps natural language sentences $s \in \mathcal {S}$ to a label from a predefined set $y \in \mathcal {Y}$ . An adversary for this classifier is a function $A$ which maps a sentence $s$ to its perturbed versions $\lbrace s^{\prime }_1, s^{\prime }_2, \ldots , s^{\prime }_{n}\rbrace $ such that each $s^{\prime }_i$ is close to $s$ under some notion of distance between sentences. We define the robustness of classifier $C$ to the adversary $A$ as:
$$R_{C,A} = \mathbb {E}_s \left[\min _{s^{\prime } \in A(s)} \mathbb {1}[C(s^{\prime }) = y]\right],$$ (Eq. 14)
where $y$ represents the ground truth label for $s$ . In practice, a real-world adversary may only be able to query the classifier a few times, hence $R_{C,A}$ represents the worst-case adversarial performance of $C$ . Methods for generating adversarial examples, such as HotFlip BIBREF8 , focus on efficient algorithms for searching the $\min $ above. Improving $R_{C,A}$ would imply better robustness against all these methods.
We explore adversaries which perturb sentences with four types of character-level edits:
(1) Swap: swapping two adjacent internal characters of a word. (2) Drop: removing an internal character of a word. (3) Keyboard: substituting an internal character with adjacent characters of QWERTY keyboard (4) Add: inserting a new character internally in a word. In line with the psycholinguistic studies BIBREF5 , BIBREF4 , to ensure that the perturbations do not affect human ability to comprehend the sentence, we only allow the adversary to edit the internal characters of a word, and not edit stopwords or words shorter than 4 characters.
For 1-character attacks, we try all possible perturbations listed above until we find an adversary that flips the model prediction. For 2-character attacks, we greedily fix the edit which had the least confidence among 1-character attacks, and then try all the allowed perturbations on the remaining words. Higher order attacks can be performed in a similar manner. The greedy strategy reduces the computation required to obtain higher order attacks, but also means that the robustness score is an upper bound on the true robustness of the classifier.
## Experiments and Results
In this section, we first discuss our experiments on the word recognition systems.
## Word Error Correction
Data: We evaluate the spell correctors from § "Robust Word Recognition" on movie reviews from the Stanford Sentiment Treebank (SST) BIBREF24 . The SST dataset consists of 8544 movie reviews, with a vocabulary of over 16K words. As a background corpus, we use the IMDB movie reviews BIBREF25 , which contain 54K movie reviews, and a vocabulary of over 78K words. The two datasets do not share any reviews in common. The spell-correction models are evaluated on their ability to correct misspellings. The test setting consists of reviews where each word (with length $\ge 4$ , barring stopwords) is attacked by one of the attack types (from swap, add, drop and keyboard attacks). In the all attack setting, we mix all attacks by randomly choosing one for each word. This most closely resembles a real world attack setting.
In addition to our word recognition models, we also compare to After The Deadline (ATD), an open-source spell corrector. We found ATD to be the best freely-available corrector. We refer the reader to BIBREF7 for comparisons of ScRNN to other anonymized commercial spell checkers.
For the ScRNN model, we use a single-layer Bi-LSTM with a hidden dimension size of 50. The input representation consists of 198 dimensions, which is thrice the number of unique characters (66) in the vocabulary. We cap the vocabulary size to 10K words, whereas we use the entire vocabulary of 78470 words when we backoff to the background model. For training these networks, we corrupt the movie reviews according to all attack types, i.e., applying one of the 4 attack types to each word, and trying to reconstruct the original words via cross entropy loss.
We calculate the word error rates (WER) of each of the models for different attacks and present our findings in Table 2 . Note that ATD incorrectly predicts $11.2$ words for every 100 words (in the `all' setting), whereas, all of the backoff variations of the ScRNN reconstruct better. The most accurate variant involves backing off to the background model, resulting in a low error rate of $6.9\%$ , leading to the best performance on word recognition. This is a $32\%$ relative error reduction compared to the vanilla ScRNN model with a pass-through backoff strategy. We can attribute the improved performance to the fact that there are $5.25\%$ words in the test corpus that are unseen in the training corpus, and are thus only recoverable by backing off to a larger corpus. Notably, only training on the larger background corpus does worse, at $8.7\%$ , since the distribution of word frequencies is different in the background corpus compared to the foreground corpus.
## Robustness to adversarial attacks
We use sentiment analysis and paraphrase detection as downstream tasks, as for these two tasks, 1-2 character edits do not change the output labels.
For sentiment classification, we systematically study the effect of character-level adversarial attacks on two architectures and four different input formats. The first architecture encodes the input sentence into a sequence of embeddings, which are then sequentially processed by a BiLSTM. The first and last states of the BiLSTM are then used by the softmax layer to predict the sentiment of the input. We consider three input formats for this architecture: (1) Word-only: where the input words are encoded using a lookup table; (2) Char-only: where the input words are encoded using a separate single-layered BiLSTM over their characters; and (3) Word $+$ Char: where the input words are encoded using a concatenation of (1) and (2) .
The second architecture uses the fine-tuned BERT model BIBREF26 , with an input format of word-piece tokenization. This model has recently set a new state-of-the-art on several NLP benchmarks, including the sentiment analysis task we consider here. All models are trained and evaluated on the binary version of the sentence-level Stanford Sentiment Treebank BIBREF24 dataset with only positive and negative reviews.
We also consider the task of paraphrase detection. Here too, we make use of the fine-tuned BERT BIBREF26 , which is trained and evaluated on the Microsoft Research Paraphrase Corpus (MRPC) BIBREF27 .
Two common methods for dealing with adversarial examples include: (1) data augmentation (DA) BIBREF28 ; and (2) adversarial training (Adv) BIBREF29 . In DA, the trained model is fine-tuned after augmenting the training set with an equal number of examples randomly attacked with a 1-character edit. In Adv, the trained model is fine-tuned with additional adversarial examples (selected at random) that produce incorrect predictions from the current-state classifier. The process is repeated iteratively, generating and adding newer adversarial examples from the updated classifier model, until the adversarial accuracy on dev set stops improving.
In Table 3 , we examine the robustness of the sentiment models under each attack and defense method. In the absence of any attack or defense, BERT (a word-piece model) performs the best ( $90.3\%$ ) followed by word+char models ( $80.5\%$ ), word-only models ( $79.2\%$ ) and then char-only models ( $70.3\%$ ). However, even single-character attacks (chosen adversarially) can be catastrophic, resulting in a significantly degraded performance of $46\%$ , $57\%$ , $59\%$ and $33\%$ , respectively under the `all' setting.
Intuitively, one might suppose that word-piece and character-level models would be more robust to such attacks given they can make use of the remaining context. However, we find that they are the more susceptible. To see why, note that the word `beautiful' can only be altered in a few ways for word-only models, either leading to an UNK or an existing vocabulary word, whereas, word-piece and character-only models treat each unique character combination differently. This provides more variations that an attacker can exploit. Following similar reasoning, add and key attacks pose a greater threat than swap and drop attacks. The robustness of different models can be ordered as word-only $>$ word+char $>$ char-only $\sim $ word-piece, and the efficacy of different attacks as add $>$ key $>$ drop $>$ swap.
Next, we scrutinize the effectiveness of defense methods when faced against adversarially chosen attacks. Clearly from table 3 , DA and Adv are not effective in this case. We observed that despite a low training error, these models were not able to generalize to attacks on newer words at test time. ATD spell corrector is the most effective on keyboard attacks, but performs poorly on other attack types, particularly the add attack strategy.
The ScRNN model with pass-through backoff offers better protection, bringing back the adversarial accuracy within $5\%$ range for the swap attack. It is also effective under other attack classes, and can mitigate the adversarial effect in word-piece models by $21\%$ , character-only models by $19\%$ , and in word, and word+char models by over $4.5\%$ . This suggests that the direct training signal of word error correction is more effective than the indirect signal of sentiment classification available to DA and Adv for model robustness.
We observe additional gains by using background models as a backoff alternative, because of its lower word error rate (WER), especially, under the swap and drop attacks. However, these gains do not consistently translate in all other settings, as lower WER is necessary but not sufficient. Besides lower error rate, we find that a solid defense should furnish the attacker the fewest options to attack, i.e. it should have a low sensitivity.
As we shall see in section § "Understanding Model Sensitivity" , the backoff neutral variation has the lowest sensitivity due to mapping UNK predictions to a fixed neutral word. Thus, it results in the highest robustness on most of the attack types for all four model classes.
Table 4 shows the accuracy of BERT on 200 examples from the dev set of the MRPC paraphrase detection task under various attack and defense settings. We re-trained the ScRNN model variants on the MRPC training set for these experiments. Again, we find that simple 1-2 character attacks can bring down the accuracy of BERT significantly ( $89\%$ to $31\%$ ). Word recognition models can provide an effective defense, with both our pass-through and neutral variants recovering most of the accuracy. While the neutral backoff model is effective on 2-char attacks, it hurts performance in the no attack setting, since it incorrectly modifies certain correctly spelled entity names. Since the two variants are already effective, we did not train a background model for this task.
## Understanding Model Sensitivity
To study model sensitivity, for each sentence, we perturb one randomly-chosen word and replace it with all possible perturbations under a given attack type. The resulting set of perturbed sentences is then fed to the word recognizer (whose sensitivity is to be estimated). As described in equation 12 , we count the number of unique predictions from the output sentences. Two corrections are considered unique if they are mapped differently by the downstream classifier.
The neutral backoff variant has the lowest sensitivity (Table 5 ). This is expected, as it returns a fixed neutral word whenever the ScRNN predicts an UNK, therefore reducing the number of unique outputs it predicts. Open vocabulary (i.e. char-only, word+char, word-piece) downstream classifiers consider every unique combination of characters differently, whereas word-only classifiers internally treat all out of vocabulary (OOV) words alike. Hence, for char-only, word+char, and word-piece models, the pass-through version is more sensitive than the background variant, as it passes words as is (and each combination is considered uniquely). However, for word-only models, pass-through is less sensitive as all the OOV character combinations are rendered identical.
Ideally, a preferred defense is one with low sensitivity and word error rate. In practice, however, we see that a low error rate often comes at the cost of sensitivity. We see this trade-off in Figure 2 , where we plot WER and sensitivity on the two axes, and depict the robustness when using different backoff variants. Generally, sensitivity is the more dominant factor out of the two, as the error rates of the considered variants are reasonably low.
We verify if the sentiment (of the reviews) is preserved with char-level attacks. In a human study with 50 attacked (and subsequently misclassified), and 50 unchanged reviews, it was noted that 48 and 49, respectively, preserved the sentiment.
## Conclusion
As character and word-piece inputs become commonplace in modern NLP pipelines, it is worth highlighting the vulnerability they add. We show that minimally-doctored attacks can bring down accuracy of classifiers to random guessing. We recommend word recognition as a safeguard against this and build upon RNN-based semi-character word recognizers. We discover that when used as a defense mechanism, the most accurate word recognition models are not always the most robust against adversarial attacks. Additionally, we highlight the need to control the sensitivity of these models to achieve high robustness.
## Acknowledgements
The authors are grateful to Graham Neubig, Eduard Hovy, Paul Michel, Mansi Gupta, and Antonios Anastasopoulos for suggestions and feedback.
| 12
|
1905.11901
|
Revisiting Low-Resource Neural Machine Translation: A Case Study
|
# Revisiting Low-Resource Neural Machine Translation: A Case Study
## Abstract
It has been shown that the performance of neural machine translation (NMT) drops starkly in low-resource conditions, underperforming phrase-based statistical machine translation (PBSMT) and requiring large amounts of auxiliary data to achieve competitive results. In this paper, we re-assess the validity of these results, arguing that they are the result of lack of system adaptation to low-resource settings. We discuss some pitfalls to be aware of when training low-resource NMT systems, and recent techniques that have shown to be especially helpful in low-resource settings, resulting in a set of best practices for low-resource NMT. In our experiments on German--English with different amounts of IWSLT14 training data, we show that, without the use of any auxiliary monolingual or multilingual data, an optimized NMT system can outperform PBSMT with far less data than previously claimed. We also apply these techniques to a low-resource Korean-English dataset, surpassing previously reported results by 4 BLEU.
## Introduction
While neural machine translation (NMT) has achieved impressive performance in high-resource data conditions, becoming dominant in the field BIBREF0 , BIBREF1 , BIBREF2 , recent research has argued that these models are highly data-inefficient, and underperform phrase-based statistical machine translation (PBSMT) or unsupervised methods in low-data conditions BIBREF3 , BIBREF4 . In this paper, we re-assess the validity of these results, arguing that they are the result of lack of system adaptation to low-resource settings. Our main contributions are as follows:
## Low-Resource Translation Quality Compared Across Systems
Figure FIGREF4 reproduces a plot by BIBREF3 which shows that their NMT system only outperforms their PBSMT system when more than 100 million words (approx. 5 million sentences) of parallel training data are available. Results shown by BIBREF4 are similar, showing that unsupervised NMT outperforms supervised systems if few parallel resources are available. In both papers, NMT systems are trained with hyperparameters that are typical for high-resource settings, and the authors did not tune hyperparameters, or change network architectures, to optimize NMT for low-resource conditions.
## Improving Low-Resource Neural Machine Translation
The bulk of research on low-resource NMT has focused on exploiting monolingual data, or parallel data involving other language pairs. Methods to improve NMT with monolingual data range from the integration of a separately trained language model BIBREF5 to the training of parts of the NMT model with additional objectives, including a language modelling objective BIBREF5 , BIBREF6 , BIBREF7 , an autoencoding objective BIBREF8 , BIBREF9 , or a round-trip objective, where the model is trained to predict monolingual (target-side) training data that has been back-translated into the source language BIBREF6 , BIBREF10 , BIBREF11 . As an extreme case, models that rely exclusively on monolingual data have been shown to work BIBREF12 , BIBREF13 , BIBREF14 , BIBREF4 . Similarly, parallel data from other language pairs can be used to pre-train the network or jointly learn representations BIBREF15 , BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 , BIBREF20 , BIBREF21 .
While semi-supervised and unsupervised approaches have been shown to be very effective for some language pairs, their effectiveness depends on the availability of large amounts of suitable auxiliary data, and other conditions being met. For example, the effectiveness of unsupervised methods is impaired when languages are morphologically different, or when training domains do not match BIBREF22
More broadly, this line of research still accepts the premise that NMT models are data-inefficient and require large amounts of auxiliary data to train. In this work, we want to re-visit this point, and will focus on techniques to make more efficient use of small amounts of parallel training data. Low-resource NMT without auxiliary data has received less attention; work in this direction includes BIBREF23 , BIBREF24 .
## Mainstream Improvements
We consider the hyperparameters used by BIBREF3 to be our baseline. This baseline does not make use of various advances in NMT architectures and training tricks. In contrast to the baseline, we use a BiDeep RNN architecture BIBREF25 , label smoothing BIBREF26 , dropout BIBREF27 , word dropout BIBREF28 , layer normalization BIBREF29 and tied embeddings BIBREF30 .
## Language Representation
Subword representations such as BPE BIBREF31 have become a popular choice to achieve open-vocabulary translation. BPE has one hyperparameter, the number of merge operations, which determines the size of the final vocabulary. For high-resource settings, the effect of vocabulary size on translation quality is relatively small; BIBREF32 report mixed results when comparing vocabularies of 30k and 90k subwords.
In low-resource settings, large vocabularies result in low-frequency (sub)words being represented as atomic units at training time, and the ability to learn good high-dimensional representations of these is doubtful. BIBREF33 propose a minimum frequency threshold for subword units, and splitting any less frequent subword into smaller units or characters. We expect that such a threshold reduces the need to carefully tune the vocabulary size to the dataset, leading to more aggressive segmentation on smaller datasets.
## Hyperparameter Tuning
Due to long training times, hyperparameters are hard to optimize by grid search, and are often re-used across experiments. However, best practices differ between high-resource and low-resource settings. While the trend in high-resource settings is towards using larger and deeper models, BIBREF24 use smaller and fewer layers for smaller datasets. Previous work has argued for larger batch sizes in NMT BIBREF35 , BIBREF36 , but we find that using smaller batches is beneficial in low-resource settings. More aggressive dropout, including dropping whole words at random BIBREF37 , is also likely to be more important. We report results on a narrow hyperparameter search guided by previous work and our own intuition.
## Lexical Model
Finally, we implement and test the lexical model by BIBREF24 , which has been shown to be beneficial in low-data conditions. The core idea is to train a simple feed-forward network, the lexical model, jointly with the original attentional NMT model. The input of the lexical model at time step INLINEFORM0 is the weighted average of source embeddings INLINEFORM1 (the attention weights INLINEFORM2 are shared with the main model). After a feedforward layer (with skip connection), the lexical model's output INLINEFORM3 is combined with the original model's hidden state INLINEFORM4 before softmax computation. INLINEFORM5
Our implementation adds dropout and layer normalization to the lexical model.
## Data and Preprocessing
We use the TED data from the IWSLT 2014 German INLINEFORM0 English shared translation task BIBREF38 . We use the same data cleanup and train/dev split as BIBREF39 , resulting in 159000 parallel sentences of training data, and 7584 for development.
As a second language pair, we evaluate our systems on a Korean–English dataset with around 90000 parallel sentences of training data, 1000 for development, and 2000 for testing.
For both PBSMT and NMT, we apply the same tokenization and truecasing using Moses scripts. For NMT, we also learn BPE subword segmentation with 30000 merge operations, shared between German and English, and independently for Korean INLINEFORM0 English.
To simulate different amounts of training resources, we randomly subsample the IWSLT training corpus 5 times, discarding half of the data at each step. Truecaser and BPE segmentation are learned on the full training corpus; as one of our experiments, we set the frequency threshold for subword units to 10 in each subcorpus (see SECREF7 ). Table TABREF14 shows statistics for each subcorpus, including the subword vocabulary.
Translation outputs are detruecased, detokenized, and compared against the reference with cased BLEU using sacreBLEU BIBREF40 , BIBREF41 . Like BIBREF39 , we report BLEU on the concatenated dev sets for IWSLT 2014 (tst2010, tst2011, tst2012, dev2010, dev2012).
## PBSMT Baseline
We use Moses BIBREF42 to train a PBSMT system. We use MGIZA BIBREF43 to train word alignments, and lmplz BIBREF44 for a 5-gram LM. Feature weights are optimized on the dev set to maximize BLEU with batch MIRA BIBREF45 – we perform multiple runs where indicated. Unlike BIBREF3 , we do not use extra data for the LM. Both PBSMT and NMT can benefit from monolingual data, so the availability of monolingual data is no longer an exclusive advantage of PBSMT (see SECREF5 ).
## NMT Systems
We train neural systems with Nematus BIBREF46 . Our baseline mostly follows the settings in BIBREF3 ; we use adam BIBREF47 and perform early stopping based on dev set BLEU. We express our batch size in number of tokens, and set it to 4000 in the baseline (comparable to a batch size of 80 sentences used in previous work).
We subsequently add the methods described in section SECREF3 , namely the bideep RNN, label smoothing, dropout, tied embeddings, layer normalization, changes to the BPE vocabulary size, batch size, model depth, regularization parameters and learning rate. Detailed hyperparameters are reported in Appendix SECREF7 .
## Results
Table TABREF18 shows the effect of adding different methods to the baseline NMT system, on the ultra-low data condition (100k words of training data) and the full IWSLT 14 training corpus (3.2M words). Our "mainstream improvements" add around 6–7 BLEU in both data conditions.
In the ultra-low data condition, reducing the BPE vocabulary size is very effective (+4.9 BLEU). Reducing the batch size to 1000 token results in a BLEU gain of 0.3, and the lexical model yields an additional +0.6 BLEU. However, aggressive (word) dropout (+3.4 BLEU) and tuning other hyperparameters (+0.7 BLEU) has a stronger effect than the lexical model, and adding the lexical model (9) on top of the optimized configuration (8) does not improve performance. Together, the adaptations to the ultra-low data setting yield 9.4 BLEU (7.2 INLINEFORM2 16.6). The model trained on full IWSLT data is less sensitive to our changes (31.9 INLINEFORM3 32.8 BLEU), and optimal hyperparameters differ depending on the data condition. Subsequently, we still apply the hyperparameters that were optimized to the ultra-low data condition (8) to other data conditions, and Korean INLINEFORM4 English, for simplicity.
For a comparison with PBSMT, and across different data settings, consider Figure FIGREF19 , which shows the result of PBSMT, our NMT baseline, and our optimized NMT system. Our NMT baseline still performs worse than the PBSMT system for 3.2M words of training data, which is consistent with the results by BIBREF3 . However, our optimized NMT system shows strong improvements, and outperforms the PBSMT system across all data settings. Some sample translations are shown in Appendix SECREF8 .
For comparison to previous work, we report lowercased and tokenized results on the full IWSLT 14 training set in Table TABREF20 . Our results far outperform the RNN-based results reported by BIBREF48 , and are on par with the best reported results on this dataset.
Table TABREF21 shows results for Korean INLINEFORM0 English, using the same configurations (1, 2 and 8) as for German–English. Our results confirm that the techniques we apply are successful across datasets, and result in stronger systems than previously reported on this dataset, achieving 10.37 BLEU as compared to 5.97 BLEU reported by gu-EtAl:2018:EMNLP1.
## Conclusions
Our results demonstrate that NMT is in fact a suitable choice in low-data settings, and can outperform PBSMT with far less parallel training data than previously claimed. Recently, the main trend in low-resource MT research has been the better exploitation of monolingual and multilingual resources. Our results show that low-resource NMT is very sensitive to hyperparameters such as BPE vocabulary size, word dropout, and others, and by following a set of best practices, we can train competitive NMT systems without relying on auxiliary resources. This has practical relevance for languages where large amounts of monolingual data, or multilingual data involving related languages, are not available. Even though we focused on only using parallel data, our results are also relevant for work on using auxiliary data to improve low-resource MT. Supervised systems serve as an important baseline to judge the effectiveness of semisupervised or unsupervised approaches, and the quality of supervised systems trained on little data can directly impact semi-supervised workflows, for instance for the back-translation of monolingual data.
## Acknowledgments
Rico Sennrich has received funding from the Swiss National Science Foundation in the project CoNTra (grant number 105212_169888). Biao Zhang acknowledges the support of the Baidu Scholarship.
## Hyperparameters
Table TABREF23 lists hyperparameters used for the different experiments in the ablation study (Table 2). Hyperparameters were kept constant across different data settings, except for the validation interval and subword vocabulary size (see Table 1).
## Sample Translations
Table TABREF24 shows some sample translations that represent typical errors of our PBSMT and NMT systems, trained with ultra-low (100k words) and low (3.2M words) amounts of data. For unknown words such as blutbefleckten (`bloodstained') or Spaniern (`Spaniards', `Spanish'), PBSMT systems default to copying, while NMT systems produce translations on a subword-level, with varying success (blue-flect, bleed; spaniers, Spanians). NMT systems learn some syntactic disambiguation even with very little data, for example the translation of das and die as relative pronouns ('that', 'which', 'who'), while PBSMT produces less grammatical translation. On the flip side, the ultra low-resource NMT system ignores some unknown words in favour of a more-or-less fluent, but semantically inadequate translation: erobert ('conquered') is translated into doing, and richtig aufgezeichnet ('registered correctly', `recorded correctly') into really the first thing.
| 15
|
1905.12801
|
Reducing Gender Bias in Word-Level Language Models with a Gender-Equalizing Loss Function
|
# Reducing Gender Bias in Word-Level Language Models with a Gender-Equalizing Loss Function
## Abstract
Gender bias exists in natural language datasets which neural language models tend to learn, resulting in biased text generation. In this research, we propose a debiasing approach based on the loss function modification. We introduce a new term to the loss function which attempts to equalize the probabilities of male and female words in the output. Using an array of bias evaluation metrics, we provide empirical evidence that our approach successfully mitigates gender bias in language models without increasing perplexity. In comparison to existing debiasing strategies, data augmentation, and word embedding debiasing, our method performs better in several aspects, especially in reducing gender bias in occupation words. Finally, we introduce a combination of data augmentation and our approach, and show that it outperforms existing strategies in all bias evaluation metrics.
## Introduction
Natural Language Processing (NLP) models are shown to capture unwanted biases and stereotypes found in the training data which raise concerns about socioeconomic, ethnic and gender discrimination when these models are deployed for public use BIBREF0 , BIBREF1 .
There are numerous studies that identify algorithmic bias in NLP applications. BIBREF2 showed ethnic bias in Google autocomplete suggestions whereas BIBREF3 found gender bias in advertisement delivery systems. Additionally, BIBREF1 demonstrated that coreference resolution systems exhibit gender bias.
Language modelling is a pivotal task in NLP with important downstream applications such as text generation BIBREF4 . Recent studies by BIBREF0 and BIBREF5 have shown that this task is vulnerable to gender bias in the training corpus. Two prior works focused on reducing bias in language modelling by data preprocessing BIBREF0 and word embedding debiasing BIBREF5 . In this study, we investigate the efficacy of bias reduction during training by introducing a new loss function which encourages the language model to equalize the probabilities of predicting gendered word pairs like he and she. Although we recognize that gender is non-binary, for the purpose of this study, we focus on female and male words.
Our main contributions are summarized as follows: i) to our best knowledge, this study is the first one to investigate bias alleviation in text generation by direct modification of the loss function; ii) our new loss function effectively reduces gender bias in the language models during training by equalizing the probabilities of male and female words in the output; iii) we show that end-to-end debiasing of the language model can achieve word embedding debiasing; iv) we provide an interpretation of our results and draw a comparison to other existing debiasing methods. We show that our method, combined with an existing method, counterfactual data augmentation, achieves the best result and outperforms all existing methods.
## Related Work
Recently, the study of bias in NLP applications has received increasing attention from researchers. Most relevant work in this domain can be broadly divided into two categories: word embedding debiasing and data debiasing by preprocessing.
## Dataset
For the training data, we use Daily Mail news articles released by BIBREF9 . This dataset is composed of 219,506 articles covering a diverse range of topics including business, sports, travel, etc., and is claimed to be biased and sensational BIBREF5 . For manageability, we randomly subsample 5% of the text. The subsample has around 8.25 million tokens in total.
## Language Model
We use a pre-trained 300-dimensional word embedding, GloVe, by BIBREF10 . We apply random search to the hyperparameter tuning of the LSTM language model. The best hyperparameters are as follows: 2 hidden layers each with 300 units, a sequence length of 35, a learning rate of 20 with an annealing schedule of decay starting from 0.25 to 0.95, a dropout rate of 0.25 and a gradient clip of 0.25. We train our models for 150 epochs, use a batch size of 48, and set early stopping with a patience of 5.
## Loss Function
Language models are usually trained using cross-entropy loss. Cross-entropy loss at time step INLINEFORM0 is INLINEFORM1
where INLINEFORM0 is the vocabulary, INLINEFORM1 is the one hot vector of ground truth and INLINEFORM2 indicates the output softmax probability of the model.
We introduce a loss term INLINEFORM0 , which aims to equalize the predicted probabilities of gender pairs such as woman and man. INLINEFORM1
INLINEFORM0 and INLINEFORM1 are a set of corresponding gender pairs, INLINEFORM2 is the size of the gender pairs set, and INLINEFORM3 indicates the output softmax probability. We use gender pairs provided by BIBREF7 . By considering only gender pairs we ensure that only gender information is neutralized and distribution over semantic concepts is not altered. For example, it will try to equalize the probabilities of congressman with congresswoman and actor with actress but distribution of congressman, congresswoman versus actor, actress will not be affected. Overall loss can be written as INLINEFORM4
where INLINEFORM0 is a hyperparameter and INLINEFORM1 is the corpus size. We observe that among the similar minima of the loss function, INLINEFORM2 encourages the model to converge towards a minimum that exhibits the lowest gender bias.
## Model Evaluation
Language models are evaluated using perplexity, which is a standard measure of performance for unseen data. For bias evaluation, we use an array of metrics to provide a holistic diagnosis of the model behavior under debiasing treatment. These metrics are discussed in detail below. In all the evaluation metrics requiring gender pairs, we use gender pairs provided by BIBREF7 . This list contains 223 pairs, all other words are considered gender-neutral.
Co-occurrence bias is computed from the model-generated texts by comparing the occurrences of all gender-neutral words with female and male words. A word is considered to be biased towards a certain gender if it occurs more frequently with words of that gender. This definition was first used by BIBREF7 and later adapted by BIBREF5 . Using the definition of gender bias similar to the one used by BIBREF5 , we define gender bias as INLINEFORM0
where INLINEFORM0 is a set of gender-neutral words, and INLINEFORM1 is the occurrences of a word INLINEFORM2 with words of gender INLINEFORM3 in the same window. This score is designed to capture unequal co-occurrences of neutral words with male and female words. Co-occurrences are computed using a sliding window of size 10 extending equally in both directions. Furthermore, we only consider words that occur more than 20 times with gendered words to exclude random effects.
We also evaluate a normalized version of INLINEFORM0 which we denote by conditional co-occurrence bias, INLINEFORM1 . This is defined as INLINEFORM2
where INLINEFORM0
INLINEFORM0 is less affected by the disparity in the general distribution of male and female words in the text. The disparity between the occurrences of the two genders means that text is more inclined to mention one over the other, so it can also be considered a form of bias. We report the ratio of occurrence of male and female words in the model generated text, INLINEFORM1 , as INLINEFORM2
Another way of quantifying bias in NLP models is based on the idea of causal testing. The model is exposed to paired samples which differ only in one attribute (e.g. gender) and the disparity in the output is interpreted as bias related to that attribute. BIBREF1 and BIBREF0 applied this method to measure bias in coreference resolution and BIBREF0 also used it for evaluating gender bias in language modelling.
Following the approach similar to BIBREF0 , we limit this bias evaluation to a set of gender-neutral occupations. We create a list of sentences based on a set of templates. There are two sets of templates used for evaluating causal occupation bias (Table TABREF7 ). The first set of templates is designed to measure how the probabilities of occupation words depend on the gender information in the seed. Below is an example of the first set of templates: INLINEFORM0
Here, the vertical bar separates the seed sequence that is fed into the language models from the target occupation, for which we observe the output softmax probability. We measure causal occupation bias conditioned on gender as INLINEFORM0
where INLINEFORM0 is a set of gender-neutral occupations and INLINEFORM1 is the size of the gender pairs set. For example, INLINEFORM2 is the softmax probability of the word INLINEFORM3 where the seed sequence is He is a. The second set of templates like below, aims to capture how the probabilities of gendered words depend on the occupation words in the seed. INLINEFORM4
Causal occupation bias conditioned on occupation is represented as INLINEFORM0
where INLINEFORM0 is a set of gender-neutral occupations and INLINEFORM1 is the size of the gender pairs set. For example, INLINEFORM2 is the softmax probability of man where the seed sequence is The doctor is a.
We believe that both INLINEFORM0 and INLINEFORM1 contribute to gender bias in the model-generated texts. We also note that INLINEFORM2 is more easily influenced by the general disparity in male and female word probabilities.
Our debiasing approach does not explicitly address the bias in the embedding layer. Therefore, we use gender-neutral occupations to measure the embedding bias to observe if debiasing the output layer also decreases the bias in the embedding. We define the embedding bias, INLINEFORM0 , as the difference between the Euclidean distance of an occupation word to male words and the distance of the occupation word to the female counterparts. This definition is equivalent to bias by projection described by BIBREF6 . We define INLINEFORM1 as INLINEFORM2
where INLINEFORM0 is a set of gender-neutral occupations, INLINEFORM1 is the size of the gender pairs set and INLINEFORM2 is the word-to-vector dictionary.
## Existing Approaches
We apply CDA where we swap all the gendered words using a bidirectional dictionary of gender pairs described by BIBREF0 . This creates a dataset twice the size of the original data, with exactly the same contextual distributions for both genders and we use it to train the language models.
We also implement the bias regularization method of BIBREF5 which debiases the word embedding during language model training by minimizing the projection of neutral words on the gender axis. We use hyperparameter tuning to find the best regularization coefficient and report results from the model trained with this coefficient. We later refer to this strategy as REG.
## Experiments
Initially, we measure the co-occurrence bias in the training data. After training the baseline model, we implement our loss function and tune for the INLINEFORM0 hyperparameter. We test the existing debiasing approaches, CDA and REG, as well but since BIBREF5 reported that results fluctuate substantially with different REG regularization coefficients, we perform hyperparameter tuning and report the best results in Table TABREF12 . Additionally, we implement a combination of our loss function and CDA and tune for INLINEFORM1 . Finally, bias evaluation is performed for all the trained models. Causal occupation bias is measured directly from the models using template datasets discussed above and co-occurrence bias is measured from the model-generated texts, which consist of 10,000 documents of 500 words each.
## Results
Results for the experiments are listed in Table TABREF12 . It is interesting to observe that the baseline model amplifies the bias in the training data set as measured by INLINEFORM0 and INLINEFORM1 . From measurements using the described bias metrics, our method effectively mitigates bias in language modelling without a significant increase in perplexity. At INLINEFORM2 value of 1, it reduces INLINEFORM3 by 58.95%, INLINEFORM4 by 45.74%, INLINEFORM5 by 100%, INLINEFORM6 by 98.52% and INLINEFORM7 by 98.98%. Compared to the results of CDA and REG, it achieves the best results in both occupation biases, INLINEFORM8 and INLINEFORM9 , and INLINEFORM10 . We notice that all methods result in INLINEFORM11 around 1, indicating that there are near equal amounts of female and male words in the generated texts. In our experiments we note that with increasing INLINEFORM12 , the bias steadily decreases and perplexity tends to slightly increase. This indicates that there is a trade-off between bias and perplexity.
REG is not very effective in mitigating bias when compared to other methods, and fails to achieve the best result in any of the bias metrics that we used. But REG results in the best perplexity and even does better than the baseline model in this respect. This indicates that REG has a slight regularization effect. Additionally, it is interesting to note that our loss function outperforms REG in INLINEFORM0 even though REG explicitly aims to reduce gender bias in the embeddings. Although our method does not explicitly attempt geometric debiasing of the word embedding, the results show that it results in the most debiased embedding as compared to other methods. Furthermore, BIBREF8 emphasizes that geometric gender bias in word embeddings is not completely understood and existing word embedding debiasing strategies are insufficient. Our approach provides an appealing end-to-end solution for model debiasing without relying on any measure of bias in the word embedding. We believe this concept is generalizable to other NLP applications.
Our method outperforms CDA in INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 . While CDA achieves slightly better results for co-occurrence biases, INLINEFORM3 and INLINEFORM4 , and results in a better perplexity. With a marginal differences, our results are comparable to those of CDA and both models seem to have similar bias mitigation effects. However, our method does not require a data augmentation step and allows training of an unbiased model directly from biased datasets. For this reason, it also requires less time to train than CDA since its training data has a smaller size without data augmentation. Furthermore, CDA fails to effectively mitigate occupation bias when compared to our approach. Although the training data for CDA does not contain gender bias, the model still exhibits some gender bias when measured with our causal occupation bias metrics. This reinforces the concept that some model-level constraints are essential to debiasing a model and dataset debiasing alone cannot be trusted.
Finally, we note that the combination of CDA and our loss function outperforms all the methods in all measures of biases without compromising perplexity. Therefore, it can be argued that a cascade of these approaches can be used to optimally debias the language models.
## Conclusion and Discussion
In this research, we propose a new approach for mitigating gender bias in neural language models and empirically show its effectiveness in reducing bias as measured with different evaluation metrics. Our research also highlights the fact that debiasing the model with bias penalties in the loss function is an effective method. We emphasize that loss function based debiasing is powerful and generalizable to other downstream NLP applications. The research also reinforces the idea that geometric debiasing of the word embedding is not a complete solution for debiasing the downstream applications but encourages end-to-end approaches to debiasing.
All the debiasing techniques experimented in this paper rely on a predefined set of gender pairs in some way. CDA used gender pairs for flipping, REG uses it for gender space definition and our technique uses them for computing loss. This reliance on pre-defined set of gender pairs can be considered a limitation of these methods. It also results in another concern. There are gender associated words which do not have pairs, like pregnant. These words are not treated properly by techniques relying on gender pairs.
Future work includes designing a context-aware version of our loss function which can distinguish between the unbiased and biased mentions of the gendered words and only penalize the biased version. Another interesting direction is exploring the application of this method in mitigating racial bias which brings more challenges.
## Acknowledgment
We are grateful to Sam Bowman for helpful advice, Shikha Bordia, Cuiying Yang, Gang Qian, Xiyu Miao, Qianyi Fan, Tian Liu, and Stanislav Sobolevsky for discussions, and reviewers for detailed feedback.
| 11
|
1906.01840
|
Improving Textual Network Embedding with Global Attention via Optimal Transport
|
# Improving Textual Network Embedding with Global Attention via Optimal Transport
## Abstract
Constituting highly informative network embeddings is an important tool for network analysis. It encodes network topology, along with other useful side information, into low-dimensional node-based feature representations that can be exploited by statistical modeling. This work focuses on learning context-aware network embeddings augmented with text data. We reformulate the network-embedding problem, and present two novel strategies to improve over traditional attention mechanisms: ($i$) a content-aware sparse attention module based on optimal transport, and ($ii$) a high-level attention parsing module. Our approach yields naturally sparse and self-normalized relational inference. It can capture long-term interactions between sequences, thus addressing the challenges faced by existing textual network embedding schemes. Extensive experiments are conducted to demonstrate our model can consistently outperform alternative state-of-the-art methods.
## Introduction
When performing network embedding, one maps network nodes into vector representations that reside in a low-dimensional latent space. Such techniques seek to encode topological information of the network into the embedding, such as affinity BIBREF0 , local interactions (e.g, local neighborhoods) BIBREF1 , and high-level properties such as community structure BIBREF2 . Relative to classical network-representation learning schemes BIBREF3 , network embeddings provide a more fine-grained representation that can be easily repurposed for other downstream applications (e.g., node classification, link prediction, content recommendation and anomaly detection).
For real-world networks, one naturally may have access to rich side information about each node. Of particular interest are textual networks, where the side information comes in the form of natural language sequences BIBREF4 . For example, user profiles or their online posts on social networks (e.g., Facebook, Twitter), and documents in citation networks (e.g., Cora, arXiv). The integration of text information promises to significantly improve embeddings derived solely from the noisy, sparse edge representations BIBREF5 .
Recent work has started to explore the joint embedding of network nodes and the associated text for abstracting more informative representations. BIBREF5 reformulated DeepWalk embedding as a matrix factorization problem, and fused text-embedding into the solution, while BIBREF6 augmented the network with documents as auxiliary nodes. Apart from direct embedding of the text content, one can first model the topics of the associated text BIBREF7 and then supply the predicted labels to facilitate embedding BIBREF8 .
Many important downstream applications of network embeddings are context-dependent, since a static vector representation of the nodes adapts to the changing context less effectively BIBREF9 . For example, the interactions between social network users are context-dependent (e.g., family, work, interests), and contextualized user profiling can promote the specificity of recommendation systems. This motivates context-aware embedding techniques, such as CANE BIBREF9 , where the vector embedding dynamically depends on the context. For textual networks, the associated texts are natural candidates for context. CANE introduced a simple mutual attention weighting mechanism to derive context-aware dynamic embeddings for link prediction. Following the CANE setup, WANE BIBREF10 further improved the contextualized embedding, by considering fine-grained text alignment.
Despite the promising results reported thus far, we identify three major limitations of existing context-aware network embedding solutions. First, mutual (or cross) attentions are computed from pairwise similarities between local text embeddings (word/phrase matching), whereas global sequence-level modeling is known to be more favorable across a wide range of NLP tasks BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 . Second, related to the above point, low-level affinity scores are directly used as mutual attention without considering any high-level parsing. Such an over-simplified operation denies desirable features, such as noise suppression and relational inference BIBREF15 , thereby compromising model performance. Third, mutual attention based on common similarity measures (e.g., cosine similarity) typically yields dense attention matrices, while psychological and computational evidence suggests a sparse attention mechanism functions more effectively BIBREF16 , BIBREF17 . Thus such naive similarity-based approaches can be suboptimal, since they are more likely to incorporate irrelevant word/phrase matching.
This work represents an attempt to improve context-aware textual network embedding, by addressing the above issues. Our contributions include: ( INLINEFORM0 ) We present a principled and more-general formulation of the network embedding problem, under reproducing kernel Hilbert spaces (RKHS) learning; this formulation clarifies aspects of the existing literature and provides a flexible framework for future extensions. ( INLINEFORM1 ) A novel global sequence-level matching scheme is proposed, based on optimal transport, which matches key concepts between text sequences in a sparse attentive manner. ( INLINEFORM2 ) We develop a high-level attention-parsing mechanism that operates on top of low-level attention, which is capable of capturing long-term interactions and allows relational inference for better contextualization. We term our model Global Attention Network Embedding (GANE). To validate the effectiveness of GANE, we benchmarked our models against state-of-the-art counterparts on multiple datasets. Our models consistently outperform competing methods.
## Problem setup
We introduce basic notation and definitions used in this work.
## Proposed Method
## Model framework overview
To capture both the topological information (network structure INLINEFORM0 ) and the semantic information (text content INLINEFORM1 ) in the textual network embedding, we explicitly model two types of embeddings for each node INLINEFORM2 : ( INLINEFORM3 ) the topological embedding INLINEFORM4 , and ( INLINEFORM5 ) the semantic embedding INLINEFORM6 . The final embedding is constructed by concatenating the topological and semantic embeddings, i.e., INLINEFORM7 . We consider the topological embedding INLINEFORM8 as a static property of the node, fixed regardless of the context. On the other hand, the semantic embedding INLINEFORM9 dynamically depends on the context, which is the focus of this study.
Motivated by the work of BIBREF9 , we consider the following probabilistic objective to train the network embeddings:
DISPLAYFORM0
where INLINEFORM0 represents sampled edges from the network and INLINEFORM1 is the collection of model parameters. The edge loss INLINEFORM2 is given by the cross entropy DISPLAYFORM0
where INLINEFORM0 denotes the conditional likelihood of observing a (weighted) link between nodes INLINEFORM1 and INLINEFORM2 , with the latter serving as the context. More specifically,
DISPLAYFORM0
where INLINEFORM0 is the normalizing constant and INLINEFORM1 is an inner product operation, to be defined momentarily. Note here we have suppressed the dependency on INLINEFORM2 to simplify notation.
To capture both the topological and semantic information, along with their interactions, we propose to use the following decomposition for our inner product term:
DISPLAYFORM0
Here we use INLINEFORM0 to denote the inner product evaluation between the two feature embeddings INLINEFORM1 and INLINEFORM2 , which can be defined by a semi-positive-definite kernel function INLINEFORM3 BIBREF26 , e.g., Euclidean kernel, Gaussian RBF, IMQ kernel, etc. Note that for INLINEFORM4 , INLINEFORM5 and INLINEFORM6 do not reside on the same feature space. As such, embeddings are first mapped to the same feature space for inner product evaluation. In this study, we use the Euclidean kernel
INLINEFORM0
for inner product evaluation with INLINEFORM0 , and linear mapping
INLINEFORM0
for feature space realignment with INLINEFORM0 . Here INLINEFORM1 is a trainable parameter, and throughout this paper we omit the bias terms in linear maps to avoid notational clutter.
Note that our solution differs from existing network-embedding models in that: ( INLINEFORM0 ) our objective is a principled likelihood loss, while prior works heuristically combine the losses of four different models BIBREF9 , which may fail to capture the non-trivial interactions between the fixed and dynamic embeddings; and ( INLINEFORM1 ) we present a formal derivation of network embedding in a reproducing kernel Hilbert space.
Direct optimization of ( EQREF9 ) requires summing over all nodes in the network, which can be computationally infeasible for large-scale networks. To alleviate this issue, we consider other more computationally efficient surrogate objectives. In particular, we adopt the negative sampling approach BIBREF27 , which replaces the bottleneck Softmax with a more tractable approximation given by
DISPLAYFORM0
where INLINEFORM0 is the sigmoid function, and INLINEFORM1 is a noise distribution over the nodes. Negative sampling can be considered as a special variant of noise contrastive estimation BIBREF28 , which seeks to recover the ground-truth likelihood by contrasting data samples with noise samples, thereby bypassing the need to compute the normalizing constant. As the number of noise samples INLINEFORM2 goes to infinity, this approximation becomes exact BIBREF29 . Following the practice of BIBREF27 , we set our noise distribution to INLINEFORM3 , where INLINEFORM4 denotes the out-degree of node INLINEFORM5 .
We argue that a key to the context-aware network embedding is the design of an effective attention mechanism, which cross-matches the relevant content between the node's associated text and the context. Over-simplified dot-product attention limits the potential of existing textual network embedding schemes. In the following sections, we present two novel, efficient attention designs that fulfill the desiderata listed in our Introduction. Our discussion follows the setup used in CANE BIBREF9 and WANE BIBREF10 , where the text from the interacting node is used as the context. Generalization to other forms of context is straightforward.
## Optimal-transport-based matching
We first consider reformulating content matching as an optimal transport problem, and then re-purpose the transport plan as our attention score to aggregate context-dependent information. More specifically, we see a node's text and context as two (discrete) distributions over the content space. Related content will be matched in the sense that they yield a higher weight in the optimal transport plan INLINEFORM0 . The following two properties make the optimal transport plan more appealing for use as attention score. ( INLINEFORM1 ) Sparsity: when solved exactly, INLINEFORM2 is a sparse matrix with at most INLINEFORM3 non-zero elements, where INLINEFORM4 is the number of contents ( BIBREF30 , INLINEFORM5 ); ( INLINEFORM6 ) Self-normalized: row-sum and column-sum equal the respective marginal distributions.
Implementation-wise, we first feed embedded text sequence INLINEFORM0 and context sequence INLINEFORM1 into our OT solver to compute the OT plan,
DISPLAYFORM0
Note that here we treat pre-embedded sequence INLINEFORM0 as INLINEFORM1 point masses in the feature space, each with weight INLINEFORM2 , and similarly for INLINEFORM3 . Next we “transport” the semantic content from context INLINEFORM4 according to the estimated OT plan with matrix multiplication
DISPLAYFORM0
where we have treated INLINEFORM0 as a INLINEFORM1 matrix. Intuitively, this operation aligns the context with the target text sequence via averaging the context semantic embeddings with respect to the OT plan for each content element in INLINEFORM2 . To finalize the contextualized embedding, we aggregate the information from both INLINEFORM3 and the aligned INLINEFORM4 with an operator INLINEFORM5 ,
DISPLAYFORM0
In this case, we practice the following simple aggregation strategy: first concatenate INLINEFORM0 and the aligned INLINEFORM1 along the feature dimension, and then take max-pooling along the temporal dimension to reduce the feature vector into a INLINEFORM2 vector, followed by a linear mapping to project the embedding vector to the desired dimensionality.
## Attention parsing
Direct application of attention scores based on a low-level similarity-based matching criteria (e.g., dot-product attention) can be problematic in a number of ways: ( INLINEFORM0 ) low-level attention scores can be noisy (i.e., spurious matchings), and ( INLINEFORM1 ) similarity-matching does not allow relational inference. To better understand these points, consider the following cases. For ( INLINEFORM2 ), if the sequence embeddings used do not explicitly address the syntactic structure of the text, a relatively dense attention score matrix can be expected. For ( INLINEFORM3 ), consider the case when the context is a query, and the matching appears as a cue in the node's text data; then the information needed is actually in the vicinity rather than the exact matching location (e.g., shifted a few steps ahead). Inspired by the work of BIBREF31 , we propose a new mechanism called attention parsing to address the aforementioned issues.
As the name suggests, attention parsing re-calibrates the raw low-level attention scores to better integrate the information. To this end, we conceptually treat the raw attention matrix INLINEFORM0 as a two-dimensional image and apply convolutional filters to it:
DISPLAYFORM0
where INLINEFORM0 denotes the filter banks with INLINEFORM1 and INLINEFORM2 respectively as window sizes and channel number. We can stack more convolutional layers, break sequence embedding dimensions to allow multi-group (channel) low-level attention as input, or introduce more-sophisticated model architectures (e.g., ResNet BIBREF32 , Transformer BIBREF18 , etc.) to enhance our model. For now, we focus on the simplest model described above, for the sake of demonstration.
With INLINEFORM0 as the high-level representation of attention, our next step is to reduce it to a weight vector to align information from the context INLINEFORM1 . We apply a max-pooling operation with respect to the context dimension, followed by a linear map to get the logits INLINEFORM2 of the weights DISPLAYFORM0
where INLINEFORM0 is the projection matrix. Then the parsed attention weight INLINEFORM1 is obtained by DISPLAYFORM0
which is used to compute the aligned context embedding
DISPLAYFORM0
Note that here we compute a globally aligned context embedding vector INLINEFORM0 , rather than one for each location in INLINEFORM1 as described in the last section ( INLINEFORM2 ). In the subsequent aggregation operation, INLINEFORM3 is broadcasted to all the locations in INLINEFORM4 . We call this global alignment, to distinguish it from the local alignment strategy described in the last section. Both alignment strategies have their respective merits, and in practice they can be directly combined to produce the final context-aware embedding.
## Related Work
## Experiments
## Experimental setup
We consider three benchmark datasets: ( INLINEFORM0 ) Cora, a paper citation network with text information, built by BIBREF44 . We prune the dataset so that it only has papers on the topic of machine learning. ( INLINEFORM1 ) Hepth, a paper citation network from Arxiv on high energy physics theory, with paper abstracts as text information. ( INLINEFORM2 ) Zhihu, a Q&A network dataset constructed by BIBREF9 , which has 10,000 active users with text descriptions and their collaboration links. Summary statistics of these three datasets are summarized in Table . Pre-processing protocols from prior studies are used for data preparation BIBREF10 , BIBREF34 , BIBREF9 .
For quantitative evaluation, we tested our model on the following tasks: ( INLINEFORM0 ) Link prediction, where we deliberately mask out a portion of the edges to see if the embedding learned from the remaining edges can be used to accurately predict the missing edges. ( INLINEFORM1 ) Multi-label node classification, where we use the learned embedding to predict the labels associated with each node. Note that the label information is not used in our embedding. We also carried out ablation study to identify the gains. In addition to the quantitative results, we also visualized the embedding and the attention matrices to qualitatively verify our hypotheses.
For the link prediction task, we adopt the area under the curve (AUC) score to evaluate the performance, AUC is employed to measure the probability that vertices in existing edges are more similar than those in the nonexistent edge. For each training ratio, the experiment is executed 10 times and the mean AUC scores are reported, where higher AUC indicates better performance. For multi-label classification, we evaluate the performance with Macro-F1 scores. The experiment for each training ratio is also executed 10 times and the average Macro-F1 scores are reported, where a higher value indicates better performance.
To demonstrate the effectiveness of the proposed solutions, we evaluated our model along with the following strong baselines. ( INLINEFORM0 ) Topology only embeddings: MMB BIBREF45 , DeepWalk BIBREF1 , LINE BIBREF33 , Node2vec BIBREF46 . ( INLINEFORM1 ) Joint embedding of topology & text: Naive combination, TADW BIBREF5 , CENE BIBREF6 , CANE BIBREF9 , WANE BIBREF10 , DMTE BIBREF34 . A brief summary of these competing models is provided in the Supplementary Material (SM).
## Results
We consider two variants of our model, denoted as GANE-OT and GANE-AP. GANE-OT employs the most basic OT-based attention model, specifically, global word-by-word alignment model; while GANE-AP additionally uses a one-layer convolutional neural network for the attention parsing. Detailed experimental setups are described in the SM.
Tables and summarize the results from the link-prediction experiments on all three datasets, where a different ratio of edges are used for training. Results from models other than GANE are collected from BIBREF9 , BIBREF10 and BIBREF34 . We have also repeated these experiments on our own, and the results are consistent with the ones reported. Note that BIBREF34 did not report results on DMTE. Both GANE variants consistently outperform competing solutions. In the low-training-sample regime our solutions lead by a large margin, and the performance gap closes as the number of training samples increases. This indicates that our OT-based mutual attention framework can yield more informative textual representations than other methods. Note that GANE-AP delivers better results compared with GANE-OT, suggesting the attention parsing mechanism can further improve the low-level mutual attention matrix. More results on Cora and Hepth are provided in the SM.
To further evaluate the effectiveness of our model, we consider multi-label vertex classification. Following the setup described in BIBREF9 , we first computed all context-aware embeddings. Then we averaged over each node's context-aware embeddings with all other connected nodes, to obtain a global embedding for each node, i.e., INLINEFORM0 , where INLINEFORM1 denotes the degree of node INLINEFORM2 . A linear SVM is employed, instead of a sophisticated deep classifier, to predict the label attribute of a node. We randomly sample a portion of labeled vertices with embeddings ( INLINEFORM3 ) to train the classifier, using the rest of the nodes to evaluate prediction accuracy. We compare our results with those from other state-of-the-art models in Table . The GANE models delivered better results compared with their counterparts, lending strong evidence that the OT attention and attention parsing mechanism promise to capture more meaningful representations.
We further explore the effect of INLINEFORM0 -gram length in our model (i.e., the filter size for the covolutional layers used by the attention parsing module). In Figure FIGREF39 we plot the AUC scores for link prediction on the Cora dataset against varying INLINEFORM1 -gram length. The performance peaked around length 20, then starts to drop, indicating a moderate attention span is more preferable. Similar results are observed on other datasets (results not shown). Experimental details on the ablation study can be found in the SM.
## Qualitative Analysis
We employed t-SNE BIBREF47 to project the network embeddings for the Cora dataset in a two-dimensional space using GANE-OT, with each node color coded according to its label. As shown in Figure FIGREF40 , papers clustered together belong to the same category, with the clusters well-separated from each other in the network embedding space. Note that our network embeddings are trained without any label information. Together with the label classification results, this implies our model is capable of extracting meaningful information from both context and network topological.
To verify that our OT-based attention mechanism indeed produces sparse attention scores, we visualized the OT attention matrices and compared them with those simarlity-based attention matrices (e.g., WANE). Figure FIGREF44 plots one typical example. Our OT solver returns a sparse attention matrix, while dot-product-based WANE attention is effectively dense. This underscores the effectiveness of OT-based attention in terms of noise suppression.
## Conclusion
We have proposed a novel and principled mutual-attention framework based on optimal transport (OT). Compared with existing solutions, the attention mechanisms employed by our GANE model enjoys the following benefits: (i) it is naturally sparse and self-normalized, (ii) it is a global sequence matching scheme, and (iii) it can capture long-term interactions between two sentences. These claims are supported by experimental evidence from link prediction and multi-label vertex classification. Looking forward, our attention mechanism can also be applied to tasks such as relational networks BIBREF15 , natural language inference BIBREF11 , and QA systems BIBREF48 .
## Acknowledgments
This research was supported in part by DARPA, DOE, NIH, ONR and NSF. Appendix Competing models Topology only embeddings Mixed Membership Stochastic Blockmodel (MMB) BIBREF45 : a graphical model for relational data, each node randomly select a different "topic" when forming an edge. DeepWalk BIBREF1 : executes truncated random walks on the graph, and by treating nodes as tokens and random walks as natural language sequences, the node embedding are obtained using the SkipGram model BIBREF27 . Node2vec BIBREF46 : a variant of DeepWalk by executing biased random walks to explore the neighborhood (e.g., Breadth-first or Depth-first sampling). Large-scale Information Network Embedding (LINE) BIBREF33 : scalable network embedding scheme via maximizing the joint and conditional likelihoods. Joint embedding of topology & text Naive combination BIBREF9 : direct combination of the structure embedding and text embedding that best predicts edges. Text-Associated DeepWalk (TADW) BIBREF5 : reformulating embedding as a matrix factorization problem, and fused text-embedding into the solution. Content-Enhanced Network Embedding (CENE) BIBREF6 : treats texts as a special kind of nodes. Context-Aware Network Embedding (CANE) BIBREF9 : decompose the embedding into context-free and context-dependent part, use mutual attention to address the context-dependent embedding. Word-Alignment-based Network Embedding (WANE) BIBREF10 : Using fine-grained alignment to improve context-aware embedding. Diffusion Maps for Textual network Embedding (DMTE) BIBREF34 : using truncated diffusion maps to improve the context-free part embedding in CANE. Complete Link prediction results on Cora and Hepth The complete results for Cora and Hepth are listed in Tables and . Results from models other than GANE are collected from BIBREF9 , BIBREF10 , BIBREF34 . We have also repeated these experiments on our own, the results are consistent with the ones reported. Note that DMTE did not report results on Hepth BIBREF34 . Negative sampling approximation In this section we provide a quick justification for the negative sampling approximation. To this end, we first briefly review noise contrastive estimation (NCE) and how it connects to maximal likelihood estimation, then we establish the link to negative sampling. Interested readers are referred to BIBREF50 for a more thorough discussion on this topic. Noise contrastive estimation. NCE seeks to learn the parameters of a likelihood model INLINEFORM0 by optimizing the following discriminative objective: J() = uipd[p(y=1|ui,v) - K Eu'pn [p(y=0|u,v)]], where INLINEFORM1 is the label of whether INLINEFORM2 comes from the data distribution INLINEFORM3 or the tractable noise distribution INLINEFORM4 , and INLINEFORM5 is the context. Using the Monte Carlo estimator for the second term gives us J() = uipd[p(y=1|ui,v) - k=1K [p(y=0|uk,v)]], uk iid pn. Since the goal of INLINEFORM6 is to predict the label of a sample from a mixture distribution with INLINEFORM7 from INLINEFORM8 and INLINEFORM9 from INLINEFORM10 , plugging the model likelihood and noise likelihood into the label likelihood gives us p(y=1;u,v) = p(u|v)p(u|v) + K pn(u|v), p(y=0;u,v) = K pn(u|v)p(u|v) + K pn(u|v). Recall INLINEFORM11 takes the following softmax form DISPLAYFORM0 NCE treats INLINEFORM12 as an learnable parameter and optimized along with INLINEFORM13 . One key observation is that, in practice, one can safely clamp INLINEFORM14 to 1, and the NCE learned model ( INLINEFORM15 ) will self-normalize in the sense that INLINEFORM16 . As such, one can simply plug INLINEFORM17 into the above objective. Another key result is that, as INLINEFORM18 , the gradient of NCE objective recovers the gradient of softmax objective INLINEFORM19 BIBREF49 . Negative sampling as NCE. If we set INLINEFORM20 and let INLINEFORM21 be the uniform distribution on INLINEFORM22 , we have DISPLAYFORM1 where INLINEFORM23 is the sigmoid function. Plugging this back to the INLINEFORM24 covers the negative sampling objective Eqn (6) used in the paper. Combined with the discussion above, we know Eqn (6) provides a valid approximation to the INLINEFORM25 -likelihood in terms of the gradient directions, when INLINEFORM26 is sufficiently large. In this study, we use INLINEFORM27 negative sample for computational efficiency. Using more samples did not significantly improve our results (data not shown). Experiment Setup We use the same codebase from CANE BIBREF9 . The implementation is based on TensorFlow, all experiments are exectuted on a single NVIDIA TITAN X GPU. We set embedding dimension to INLINEFORM28 for all our experiments. To conduct a fair comparison with the baseline models, we follow the experiment setup from BIBREF10 . For all experiments, we set word embedding dimension as 100 trained from scratch. We train the model with Adam optimizer and set learning rate INLINEFORM29 . For GANE-AP model, we use best filte size INLINEFORM30 for convolution from our abalation study. Ablation study setup To test how the INLINEFORM31 -gram length affect our GANE-AP model performance, we re-run our model with different choices of INLINEFORM32 -gram length, namely, the window size in convolutional layer. Each experiment is repeated for 10 times and we report the averaged results to eliminate statistical fluctuations.
| 13
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.