chunk_id
stringlengths
3
7
chunk
stringlengths
1
823
source_url
stringclasses
416 values
title
stringclasses
416 values
chunk_idx
int64
0
294
chunk_start_char
int64
0
139k
chunk_end_char
int64
303
139k
29_24
$\lambda (t)\in [1,+\infty )$ is a predefined parameter to impose a larger penalty if the type $t$ is incorrectly predicted as positive. Since the problem of overfitting the weakly annotated labels is more severe for person mentions, we set $\lambda (t)=\lambda _P$ if $t$ is a fine-grained person type, and $\lambda (t...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
24
11,943
12,385
29_25
So that the model can perform well for mentions that cannot be linked to the KB at test time. Method ::: Entity Linking Algorithm In this paper, we use a simple EL algorithm that directly links the mention to the entity with the greatest commonness score. Commonness BIBREF17, BIBREF18 is calculated base on the ancho...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
25
12,385
12,799
29_26
In our FET approach, the commonness score is also used as the confidence on the linking result (i.e., the $\mathbf {g}$ used in the prediction part of Subsection SECREF5). Within a same document, we also use the same heuristic used in BIBREF19 to find coreferences of generic mentions of persons (e.g., “Matt”) to more ...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
26
12,799
13,289
29_27
Experimental results of using the EL system proposed in BIBREF19 is provided in Section SECREF4. Experiments ::: Setup We use two datasets: FIGER (GOLD) BIBREF0 and BBN BIBREF5. The sizes of their tag sets are 113 and 47, respectively. FIGER (GOLD) allows mentions to have multiple type paths, but BBN does not.
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
27
13,289
13,603
29_28
Another commonly used dataset, OntoNotes BIBREF1, is not used since it contains many pronoun and common noun phrase mentions such as “it,” “he,” “a thrift institution,” which are not suitable to directly apply entity linking on. Following BIBREF0, we generate weakly labeled datasets for training with Wikipedia anchor...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
28
13,603
14,132
29_29
We also manually annotated 50 person mentions collected from news articles for tuning the parameter $\lambda _P$. We use the 300 dimensional pretrained GloVe word vectors provided by BIBREF20. The hidden layer sizes of the two layers of BiLSTMs are both set to 250. For the three-layer MLP, the size of the two hidden ...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
29
14,132
14,547
29_30
We also apply batch normalization and dropout to the input of each dense layer in our three-layer MLP during training. We use strict accuracy, Macro F1, and Micro F1 to evaluate fine-grained typing performance BIBREF0. Experiments ::: Compared Methods We compare with the following existing approaches: AFET BIBREF3,...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
30
14,547
14,915
29_31
We use Ours (Full) to represent our full model, and also compare with five variants of our own approach: Ours (DirectTrain) is trained without adding random person types while obtaining the KB type representation, and $\lambda _P$ is set to 1; Ours (NoEL) does not use entity linking, i.e., the KB type representation ...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
31
14,915
15,444
29_32
version of Ours (NonDeep); Ours (LocAttEL) uses the entity linking approach proposed in BIBREF19 instead of our own commonness based approach. Ours (Full), Ours (DirectTrain), and Ours (NonDeep) all use our own commonness based entity linking approach. Experiments ::: Results The experimental results are listed in T...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
32
15,444
15,981
29_33
The performance on both datasets decreases if the entity linking part is removed. Especially on FIGER (GOLD), the strict accuracy drops from 75.5 to 69.8. Using entity linking improves less on BBN.
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
33
15,981
16,179
29_34
We think this is because of three reasons: 1) BBN has a much smaller tag set than FIGER (GOLD); 2) BBN does not allow a mention to be annotated with multiple type paths (e.g., labeling a mention with both /building and /location is not allowed), thus the task is easier; 3) By making the model deep, the performance on ...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
34
16,179
16,569
29_35
The improvement of our full approach over Ours (DirectTrain) on FIGER (GOLD) indicates that the techniques we use to avoid overfitting the weakly labeled data are also effective. Ours (LocAttEL), which uses a more advanced EL system, does not achieve better performance than Ours (Full), which uses our own EL approac...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
35
16,569
17,096
29_36
Sometimes, if it “thinks” that the type information provided by EL is incorrect, it may not use it. 2) The performances of different EL approaches also depends on the dataset and the types of entities used for evaluation. We find that on FIGER (GOLD), the approach in BIBREF19 is better at distinguishing locations and ...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
36
17,096
17,623
29_37
3) For some mentions, although the EL system links it to an incorrect entity, the type of this entity is the same with the correct entity. Conclusions We propose a deep neural model to improve fine-grained entity typing with entity linking. The problem of overfitting the weakly labeled training data is addressed by ...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
37
17,623
18,221
29_38
26206717) from Research Grants Council in Hong Kong and WeChat-HKUST WHAT Lab on Artificial Intelligence Technology. Figure 1: Our approach. The example sentence is “Earlier on Tuesday, Donald Trump pledged to help hard-hit U.S. farmers caught in the middle of the escalating trade war.” Here, the correct label for th...
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
38
18,221
18,721
29_39
The performance of “Ours (DirectTrain)” on BBN is omitted since this dataset does not have fine-grained types for person.
https://arxiv.org/abs/1909.12079
Improving Fine-grained Entity Typing with Entity Linking
39
18,721
18,843
30_0
Common-Knowledge Concept Recognition for SEVA We build a common-knowledge concept recognition system for a Systems Engineer's Virtual Assistant (SEVA) which can be used for downstream tasks such as relation extraction, knowledge graph construction, and question-answering. The problem is formulated as a token classific...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
0
0
680
30_1
In addition, we also create some essential datasets for information such as abbreviations and definitions from the systems engineering domain. Finally, we construct a simple knowledge graph using these extracted concepts along with some hyponym relations. INTRODUCTION The Systems Engineer's Virtual Assistant (SEVA) ...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
1
680
1,239
30_2
In this work, we address a system element by constructing a common-knowledge concept recognition system for improving the performance of SEVA, using the static knowledge collected from the Systems Engineering Handbook BIBREF1 that is widely used in projects across the organization as domain-specific commonsense knowle...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
2
1,239
1,918
30_3
Although constructing an assistant like SEVA system is the overarching objective, a key problem to first address is to extract elementary common-knowledge concepts using the SE handbook and domain experts. We use the term `common-knowledge' as the `commonsense' knowledge of a specific domain. This knowledge can be see...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
3
1,918
2,581
30_4
Implementation with demo and dataset is available at: https://github.com/jitinkrishnan/NASA-SE . BACKGROUND AND MOTIVATION Creating commonsense AI still remains an important and challenging task in AI research today. Some of the inspiring works are the CYC project BIBREF5 that tries to serve as a foundational knowle...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
4
2,581
3,187
30_5
In NASA's context, systems engineering combines several engineering disciplines requiring extreme coordination and is prone to human errors. This, in combination with the lack of efficient knowledge transfer of generic lessons-learned makes most technology-based missions risk-averse. Thus, a comprehensive commonsense ...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
5
3,187
3,727
30_6
A typical NER task seeks to identify entities like name of a person such as `Shakespeare', a geographical location such as `London', or name of an organisation such as `NASA' from unstructured text. A supervised NER dataset consists of the above mentioned entities annotated at the word-token level using labelling sche...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
6
3,727
4,263
30_7
Off-the-shelf named entity extractors do not suffice in the SE common-knowledge scenario because the entities we want to extract are domain-specific concepts such as `system architecture' or `functional requirements' rather than physical entities such as `Shakespeare' or `London'. This requires defining new labels and...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
7
4,263
4,783
30_8
Relations can be as simple as using hand-built patterns or as challenging as using unsupervised methods like Open IE BIBREF10; with bootstrapping, supervised, and semi-supervised methods in between. BIBREF11 and BIBREF12 are some of the high performing models that extract relations from New York Times Corpus BIBREF13 ...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
8
4,783
5,258
30_9
For instance, a well-known work by BIBREF15 uses syntactic patterns such as [Y such as A, B, C], [Y including X], or [Y, including X] to extract hyponyms. Our goal is to extract preliminary hyponym relations from the concepts extracted by the CR and to connect the entities through verb phrases. CONCEPT RECOGNITION S...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
9
5,258
5,687
30_10
For example, the word `system' usually means the same when referring to a `complex system', `system structure', or `management system' in the SE domain. In generic text, the meaning of terms like `evaluation', `requirement', or `analysis' may contextually differ. We would like domain specific phrases such as `system e...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
10
5,687
6,259
30_11
CONCEPT RECOGNITION ::: BIO Labelling Scheme abb: represents abbreviations such as TRL representing Technology Readiness Level. grp: represents a group of people or an individual such as Electrical Engineers, Systems Engineers or a Project Manager. syscon: represents any system concepts such as engineering unit, p...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
11
6,259
6,792
30_12
seterm: represents generic terms that are frequently used in SE text and those that do not fall under syscon or opcon such as project, mission, key performance parameter, audit etc. event: represents event-like information in SE text such as Pre-Phase A, Phase A, Phase B, etc. org: represents an organization such a...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
12
6,792
7,286
30_13
loc: represents location-like entities such as component facilities or centralized facility. mea: represents measures, features, or behaviors such as cost, risk, or feasibility. CONCEPT RECOGNITION ::: Abbreviations Abbreviations are used frequently in SE text. We automatically extract abbreviations using simple p...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
13
7,286
7,887
30_14
`A process to determine a system’s technological maturity based on Technology Readiness Levels (TRLs)' produces the abbreviation TRL which stands for Technology Readiness Levels. `Define one or more initial Concept of Operations (ConOps) scenarios' produces the abbreviation ConOps which stands for Concept of Operation...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
14
7,887
8,545
30_15
We collect these and compile a comprehensive definitions document which is also used for the concept recognition task. An example definition and its description is shown below: Definition: Acceptable Risk Description: The risk that is understood and agreed to by the program/project, governing authority, mission dire...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
15
8,545
8,956
30_16
CONCEPT RECOGNITION ::: CR Dataset Construction and Pre-processing Using python tools such as PyPDF2, NLTK, and RegEx we build a pipeline to convert PDF to raw text along with extensive pre-processing which includes joining sentences that are split, removing URLs, shortening duplicate non-alpha characters, and repla...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
16
8,956
9,498
30_17
Using domain keywords and a domain expert, we annotate roughly 3700 sentences at the word-token level. An example is shown in Figure 2 and the unique tag count is shown in Table 1. CONCEPT RECOGNITION ::: Fine tuning with BERT Any language model can be used for the purpose of customizing an NER problem to CR. We cho...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
17
9,498
9,984
30_18
The idea is to perform multi-class classification using BERT's pre-trained cased language model. We use pytorch transformers and hugging face as per the tutorial by BIBREF17 which uses $BertForTokenClassification$. The text is embedded as tokens and masks with a maximum token length. This embedded tokens are provided ...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
18
9,984
10,499
30_19
Detailed performance of the CR is shown in Table 2 and 3. Additionally, we also implemented CR using spaCy BIBREF18 which also produced similar results. RELATION EXTRACTION In this work, for relation extraction, we focus on hyponyms and verb phrase chunking. Hyponyms are more specific concepts such as earth to plane...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
19
10,499
11,061
30_20
We iteratively construct entities in increasing order of number of words in the definitions with the help of their parts-of-speech tags. This helps in creating subset-of relation between a lower-word entity and a higher-word entity. Each root entity is lemmatized such that entities like processes and process appear on...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
20
11,061
11,710
30_21
`SE' has tag NNP and `functions' has tag NNS. We create a relation called subset-of between `SE functions' and `SE'. RELATION EXTRACTION ::: Relations from Abbreviations Relations from abbreviations are simple direct connections between the abbreviation and its full form described in the abbreviations dataset. Figur...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
21
11,710
12,170
30_22
RELATION EXTRACTION ::: Relation Extraction using Verb Phrase Chunking Finally, we explore creating contextual triples from sentences using all the entities extracted using the CR model and entities from definitions. Only those phrases that connect two entities are selected for verb phrase extraction. Using NLTK's r...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
22
12,170
12,618
30_23
>|<TO>)*} with at least one verb, can extract relation-like phrases from the phrase that links two concepts. An example is shown in Figure FIGREF27. Further investigation of relation extraction from SE handbook is left as future work. CONCLUSION AND FUTURE WORK We presented a common-knowledge concept extractor for t...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
23
12,618
13,110
30_24
We construct a word-level annotated dataset with the help of a domain expert by carefully defining a labelling scheme to train a sequence labelling task to recognize SE concepts. Further, we also construct some essential datasets from the SE domain which can be used for future research. Future directions include const...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
24
13,110
13,637
30_25
Figure 1: Common-knowledge concept recognition and simple relation extraction Table 2: Performance of different labels Table 1: Unique Tag Count from the CR dataset Figure 2: A Snippet of the concept-labelled dataset Table 3: Overall Performance of CR; For fairness, we also provide the accuracy when the most comm...
https://arxiv.org/abs/2003.11687
Common-Knowledge Concept Recognition for SEVA
25
13,637
14,101
31_0
Automatic Argumentative-Zoning Using Word2vec In comparison with document summarization on the articles from social media and newswire, argumentative zoning (AZ) is an important task in scientific paper analysis. Traditional methodology to carry on this task relies on feature engineering from different levels. In this...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
0
0
717
31_1
Using 10-cross-validation scheme, evaluation was conducted on the Argumentative-Zoning (AZ) annotated articles. The results showed that simply averaging the word vectors in a sentence works better than the paragraph to vector algorithm and by integrating specific cuewords into the loss function of the neural network c...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
1
717
1,273
31_2
Introduction One of the crucial tasks for researchers to carry out scientific investigations is to detect existing ideas that are related to their research topics. Research ideas are usually documented in scientific publications. Normally, there is one main idea stated in the abstract, explicitly presenting the aim ...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
2
1,273
1,896
31_3
Therefore, computer assisted technologies such as document summarization are expected to play a role in condensing information and providing readers with more relevant short texts. Unlike document summarization from news circles, where the task is to identify centroid sentences BIBREF0 or to extract the first few sent...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
3
1,896
2,605
31_4
The results of AZ provide readers with general discourse context from which the scientific ideas could be better linked, compared and analyzed. For example, given a specific task, which sentences should be shown to the reader is related to the features of the sentences. For the task of identifying a paper's unique con...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
4
2,605
3,273
31_5
Their scheme resulted seven categories of rhetorical status and the categories are assigned to full sentences. Examples of human annotated sentences with their rhetorical status are shown in Table. TABREF2 . The seven categories are aim, contrast, own, background, other, basis and textual. Analyzing the rhetorical st...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
5
3,273
3,979
31_6
NLP covers the applications from document retrieval, text categorization BIBREF6 , document summarization BIBREF7 to sentiment analysis BIBREF8 , BIBREF9 . Those applications are targeting different types of text resources, such as articles from social media BIBREF10 and scientific publications BIBREF2 . There are sev...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
6
3,979
4,590
31_7
Those problems have been addressed from many angles, one of which is feature extraction and representation. At the early stage of document summarization, features are usually engineered manually. Although the hand-crafted features have shown the ability for document summarization and sentiment analysis BIBREF13 , BIBR...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
7
4,590
5,177
31_8
BIBREF2 have built feature pool of sixteen types of features to classify sentences, such as the position of sentence, sentence length and tense. Widyantoro et. al. used content features, qualifying adjectives and meta-discourse features BIBREF14 to explore AZ task. It took efforts to engineer these features and it is ...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
8
5,177
5,778
31_9
There are plenty of cases where word embedding and sentence representations have been applied to short text classification BIBREF17 and paraphrase detection BIBREF18 . However, the effectiveness of this technique on AZ needs further study. The research question is, is it possible to extract word embeddings as features...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
9
5,778
6,315
31_10
With word2vec tool, word embeddings can be learnt from big amount of text corpus and the semantic relationships between words can be measured by the cosine distances between the vectors. The idea behind word embeddings is to use distributed representation BIBREF19 to map each word into k-dimension vector. How these ve...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
10
6,315
6,928
31_11
BIBREF16 introduced two architectures: Skip-gram model and continuous bag of words (CBOW) model. Each of the model has two different training strategies, such as hierarchical softmax and negative sampling. Both these two models have three layers: input, projection and output layer. The word vectors are obtained once t...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
11
6,928
7,636
31_12
Word2vec has been applied to sentiment analysis BIBREF21 , BIBREF22 , BIBREF23 and text classification BIBREF24 . Sadeghian and Sharafat BIBREF25 explored averaging of the word vectors in a sentiment review statement. Their results indicated that word2vec models significantly outperform the vanilla bag-of-words model...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
12
7,636
8,178
31_13
They also trained sentiment-specific word embeddings to improve the twitter sentiment classification results. This work is aiming at learning word embeddings for the task of AZ. The results were compared from three aspects: the impact of the training corpus, the effectiveness of specific word embeddings and different ...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
13
8,178
8,578
31_14
Le and Mikolov BIBREF26 introduced the concept of word vector representation in a formal way: Given a sequence of training words INLINEFORM0 , the objective of the word2vec model is to maximize the average log probability: INLINEFORM0 INLINEFORM1 INLINEFORM2 p INLINEFORM3 (1) Using softmax technique, the predicti...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
14
8,578
9,111
31_15
learned from large text corpus as features to classify sentences into seven categories in the task of AZ. Three models were explored to obtain the sentence vectors: averaging the vectors of the words in one sentence, paragraph vectors and specific word vectors. The first model, averaging word vectors ( INLINEFORM0 ),...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
15
9,111
9,488
31_16
The main process in this model is to learn the word embedding matrix INLINEFORM2 : INLINEFORM0 INLINEFORM1 INLINEFORM2 INLINEFORM3 (4) where INLINEFORM0 is the word embedding for word INLINEFORM1 , which is learned by the classical word2vec algorithm BIBREF16 . The second model, INLINEFORM0 , is aiming at training...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
16
9,488
9,945
31_17
In comparison with the word2vec framework, the only change in PV-DM is in the equation (3), where INLINEFORM1 is constructed from INLINEFORM2 and INLINEFORM3 , where matrix INLINEFORM4 is the word vector and INLINEFORM5 holds the paragraph vectors in such a way that every paragraph is mapped to a unique vector represe...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
17
9,945
10,517
31_18
In this study, INLINEFORM1 specific word embeddings were trained ( INLINEFORM2 ) inspired by Tang et al. BIBREF21 's model: Sentiment-Specific Word Embedding (unified model: INLINEFORM3 ). After obtaining the word vectors via INLINEFORM4 , the same scheme was used to average the vectors in one sentence as in the model...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
18
10,517
10,998
31_19
Similar to sentiment classification using word embeddings BIBREF21 , where they try to predict each tweet to be either positive or negative, in the task of AZ, the embeddings are used to classify each sentence into one of the seven categories. To evaluate the classification performance, precision, recall and F-measur...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
19
10,998
11,567
31_20
INLINEFORM0 collection contains 6,778 sentences, extracted from the titles and abstracts of publications provided by WEB OF SCIENCE . Test Dataset Argumentative Zoning Corpus ( INLINEFORM0 corpus) consists of 80 AZ INLINEFORM1 annotated conference articles in computational linguistics, originally drawn from the Cm...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
20
11,567
12,140
31_21
The word2vec were also trained using different parameters, such as different dimension of features. To evaluate the impact from different domains, the first model was trained on different corpus. The characteristics of word embeddings based on different model and dataset are listed in Table. TABREF12 . Parameters I...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
21
12,140
12,676
31_22
Strategy of dealing with unbalanced data In imbalanced data sets, some classes are significantly outnumbered by other classes BIBREF27 , which affects the classification results. In this experiment, the test dataset is an imbalanced data set. Table. TABREF16 shows the distribution of rhetorical categories from the I...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
22
12,676
13,260
31_23
10-cross validation scheme was adopted and the results were averaged from 10 iterations. Results of classification for per category Table. TABREF19 and TABREF20 show the classification performance of different methods. The results were examined from the following aspects: When the feature dimension is set to 100 ...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
23
13,260
13,727
31_24
Looking at the F-measure, AVGWVEC performs better than PARAVEC, but PARAVEC gave a better precision results on several categories, such as AIM, CTR, TXT and OWN. The results showed that PARAVEC model is not robust, for example, it performs badly for the category of BAS. For specific category classification, take the B...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
24
13,727
14,263
31_25
From the F-measure, it can be seen that for some categories, 300-dimension features perform better than the 100-dimension ones, for example, CTR and BKG, but they are not as good as 100-dimension features for some categories, such as BAS. When the model is set to AVGWVEC and the feature dimension is 100, the results ...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
25
14,263
14,673
31_26
ACL+AZ outperforms others and brown corpus is better than MixedAbs for most of the categories, but brown corpus is not as good as MixedAbs for the category of OWN. Finally, the results were compared between word embeddings and the methods of cuewords, Teufel 2002 and baseline. To evaluate word embeddings on AZ, the m...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
26
14,673
15,174
31_27
It also outperforms Teufel 2002 for most of the cases, except AIM, BAS and OWN. It won baseline for most of the categories, except OWN. Discussion The classification results showed that the type of word embeddings and the training corpus affect the AZ performance. As the simple model, INLINEFORM0 performs better tha...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
27
15,174
15,774
31_28
Feature dimension doesn't dominate the results. There is no significant difference between the resutls generated by 300-dimension of features and 100 dimensions. Training corpus affects the results. ACL+AZ outperforming others indicates that the topics of the training corpus are important factors in argumentative zo...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
28
15,774
16,295
31_29
But for classifying the categories AIM, BAS and OWN, the manually crafted features proposed by Teufel et al. BIBREF2 gave better results. Conclusion In this paper, different word embedding models on the task of argumentative zoning were compared . The results showed that word embeddings are effective on sentence cla...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
29
16,295
16,926
31_30
The size of the feature pool doesn't matter too much on the results, nor does the vocabulary size. In comparison, the domain of the training corpus affects the classification performance. Table 1. Examples of annotated sentences with their rhetorical status Table 2. Characteristics of word embeddings based on differ...
https://arxiv.org/abs/1703.10152
Automatic Argumentative-Zoning Using Word2vec
30
16,926
17,500
32_0
Multitask Learning for Blackmarket Tweet Detection Online social media platforms have made the world more connected than ever before, thereby making it easier for everyone to spread their content across a wide variety of audiences. Twitter is one such popular platform where people publish tweets to spread their messag...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
0
0
674
32_1
In this paper, we attempt to detect such tweets that have been posted on these blackmarket services in order to gain artificially boosted retweets. We use a multitask learning framework to leverage soft parameter sharing between a classification and a regression based task on separate inputs. This allows us to effecti...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
1
674
1,338
32_2
Twitter provides users with multiple ways of showing their support towards a tweet in the form of Likes, Retweets and Quotes. These content-level appraisals help in spreading the content further and act as a measure of users' agreement on the value of the content. The count of these content-level appraisals therefore ...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
2
1,338
1,720
32_3
This has led to the creation of certain blackmarket services such as FreeFollowers (https://www.freefollowers.io/), Like4Like (https://like4like.org/), YouLikeHits (https://www.youlikehits.com/), JustRetweet (http://justretweet.com), which allow users to post their tweets in order to gain inorganic appraisals in the f...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
3
1,720
2,093
32_4
There has been a lot of research on the detection of fraudulent activities on Twitter such as detection of bots BIBREF2 , fake followers BIBREF3 , collusive retweeters BIBREF0 , BIBREF1 , and social spam BIBREF4 . However, the problem of detecting tweets that are posted to these blackmarket services has not been tack...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
4
2,093
2,687
32_5
However, existing spam tweet detection approaches do not work that well in identifying individual tweets as blackmarket tweets (as shown in Table TABREF29 ). Table TABREF1 shows a sample tweet that was posted on a blackmarket service and another sample tweet that was not. In this paper, we make the first attempt to d...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
5
2,687
3,266
32_6
Instead, we only rely on the features and representations extracted from the content of the tweets. We curate a novel dataset of tweets that have been posted to blackmarket services, and a corresponding set of tweets that haven't. We propose a multitask learning approach to combine properties from the characterizatio...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
6
3,266
3,710
32_7
We train a neural network which takes as input both the traditional feature representation as well as the deep learning based representation generated using the Tweet2Vec model BIBREF5 , and utilizes cross-stitch units BIBREF6 to learn an optimal combination of shared and task-specific knowledge via soft parameter sha...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
7
3,710
4,247
32_8
In short, the contributions of the paper are threefold: a new dataset, characterization of blackmarket tweets, and a novel multitask learning framework to detect tweets posted on blackmarket services. Related Work Several studies have focused on detecting malicious activities such as spam, fake content and blackmark...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
8
4,247
4,881
32_9
BIBREF7 showed that the network structure of spammers and non-spammers is different, and also tracked the life cycle of endogenous Twitter content. Chen et al. BIBREF8 conducted a comprehensive evaluation of several machine learning algorithms for timely detection of spam. Fake tweets, on the other hand, are the tweet...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
9
4,881
5,492
32_10
Blackmarket Services: Blackmarket services have recently received considerable attention due to the increase in the number of users using them. Analysis of such underground services was first documented in BIBREF12 where the authors examined the properties of social networks formed for blackmarket services. Liu et al...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
10
5,492
6,176
32_11
BIBREF0 investigated the customers involved in gaining fake retweets. Chetan et al. BIBREF1 proposed CoReRank, an unsupervised model and CoReRank+, a semi-supervised model which extends CoReRank to detect collusive users involved in retweeting activities. Multitask Learning: Multitask learning is used whenever we hav...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
11
6,176
6,780
32_12
Cross-Stitch units were introduced by Misra et al. BIBREF6 , which can learn an optimal combination of shared and task-specific representations. Gupta et al. BIBREF15 proposed GIRNet, a unified position-sensitive multitask recurrent neural network architecture. Blackmarket Services blackAs studied in BIBREF0 , there...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
12
6,780
7,306
32_13
The unpaid services are available to the users when they contribute to the blackmarket by providing appraisals for other users' content. Here, we mainly concentrate on freemium services.
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
13
7,306
7,493
32_14
The freemium services can be further divided into three categories: (i) social-share services (request customers to spread the content on social media), (ii) credit-based services (customers earn credits by providing appraisals, and can then use the credits earned to gain appraisals for their content), and (iii) auto-...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
14
7,493
8,091
32_15
We crawled two blackmarket sites – YouLikeHits and Like4Like, between the period of February and April 2019. We created dummy accounts (after careful IRB approval) on these sites to participate in the platform and recorded Tweet IDs of the tweets that were posted for gaining retweets. We used Twitter's REST API to col...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
15
8,091
8,623
32_16
Dataset Description In total, we collected INLINEFORM0 tweets posted on blackmarket sites. Out of these, we removed non-English tweets and tweets with a length of less than two characters. Finally, we were left with INLINEFORM1 blackmarket tweets. Then, from the timelines of the authors of these tweets, we randomly ...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
16
8,623
9,119
32_17
Analysis of Blackmarket Tweets To further understand the purpose of the collusive users behind the usage of blackmarket services, we annotated blackmarket tweets in our test set into a few discrete categories. The statistics of the categories are as follows: Promotional - 43.75%, Entertainment - 15.89%, Spam - 13.57...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
17
9,119
9,588
32_18
Most of the tweets in the Others category include personal tweets without any call to action or promotion, but this also can be considered as self-promotion. We further noticed that there were about 5% of normal tweets on concerning issues such as “pray for ...", indicating that blackmarket services are also being use...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
18
9,588
10,219
32_19
Tweet Content Features We use the following features based on the tweet content: INLINEFORM0 : Number of user mentions in the tweet INLINEFORM0 : Number of hashtags in the tweet INLINEFORM0 : Number of URLs in the tweet INLINEFORM0 : Count of media content in the tweet INLINEFORM0 : Is the tweet a reply to...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
19
10,219
10,554
32_20
INLINEFORM0 : Number of special characters (non alpha-numeric) in the tweet INLINEFORM0 : Length of the content (number of characters) in the tweet INLINEFORM0 : Sentiment score of the tweet obtained using SentiWordNet, ranging from -1 (negative) to +1 (positive) INLINEFORM0 : Number of noun words in the tweet...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
20
10,554
11,092
32_21
5 to generate a vector-space representation of each of the tweets. Tweet2Vec is a character-level deep learning based encoder for social media posts trained on the task of predicting the associated hashtags. It considers the assumption that posts with the same hashtags should have similar representation. It uses a bi-d...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
21
11,092
11,709
32_22
BIBREF5 , which is trained on a dataset of 2 million tweets, to get the tweet representation. This gives us a 500-dimensional representation of each tweet, based on its content. Proposed Model The architecture of our model is shown in Figure FIGREF21 . We adopt multitask learning to develop our model. The primary ta...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
22
11,709
12,296
32_23
The model takes a different input feature vector for each of the tasks. Primary Input: The primary task takes as input the tweet content representation generated by the Tweet2Vec model, which is a 500-dimensional vector for each of the tweets, as described above. Secondary Input: The secondary task takes as input t...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
23
12,296
12,857
32_24
The cross-stitch units find the best shared representations using linear combinations, and learn the optimal linear combinations for a given set of tasks. The cross-stitch units essentially allow us to unify two separate networks for two separate tasks into a single network wherein each layer of the network shares the...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
24
12,857
13,434
32_25
The output layer of the second task predicts the numerical values for the number of retweets and likes that a tweet will gain after five days of being posted by using a Mean Squared Error (MSE) loss. Note that the performance of the secondary task is not of importance to us, however, the secondary task helps the prima...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
25
13,434
13,866
32_26
Baseline Methods Since there is no prior work on blackmarket tweet detection, we chose state-of-the-art Twitter spam detection methods as baselines, along with training some state-of-the-art classifiers on the features we generated for our dataset. Spam Detection 1: We use the Twitter spam detection method proposed...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
26
13,866
14,386