id stringclasses 100 values | document stringclasses 100 values | gold_summary stringlengths 76 223 | aspect_selected stringclasses 3 values | prompt_type stringclasses 2 values | objective stringclasses 2 values | prompt stringlengths 2.5k 16.7k | source_dataset stringclasses 1 value | prompt_variant_id int64 0 2 | section stringclasses 1 value |
|---|---|---|---|---|---|---|---|---|---|
2021.acl-long.67 | Bilingual lexicons map words in one language to their translations in another , and are typically induced by learning linear projections to align monolingual word embedding spaces . In this paper , we show it is possible to produce much higher quality lexicons with methods that combine ( 1 ) unsupervised bitext mining and ( 2 ) unsupervised word alignment . Directly applying a pipeline that uses recent algorithms for both subproblems significantly improves induced lexicon quality and further gains are possible by learning to filter the resulting lexical entries , with both unsupervised and semisupervised schemes . Our final model outperforms the state of the art on the BUCC 2020 shared task by 14 F 1 points averaged over 12 language pairs , while also providing a more interpretable approach that allows for rich reasoning of word meaning in context . Further analysis of our output and the standard reference lexicons suggests they are of comparable quality , and new benchmarks may be needed to measure further progress on this task . 1 Bilingual lexicons map words in one language to their translations in another , and can be automatically induced by learning linear projections to align monolingual word embedding spaces ( Artetxe et al . , 2016 ; Smith et al . , 2017 ; Lample et al . , 2018 , inter alia ) . Although very successful in practice , the linear nature of these methods encodes unrealistic simplifying assumptions ( e.g. all translations of a word have similar embeddings ) . In this paper , we show it is possible to produce much higher quality lexicons without these restrictions by introducing new methods that combine ( 1 ) unsupervised bitext mining and ( 2 ) unsupervised word alignment . We show that simply pipelining recent algorithms for unsupervised bitext mining ( Tran et al . , 2020 ) and unsupervised word alignment ( Sabet et al . , 2020 ) significantly improves bilingual lexicon induction ( BLI ) quality , and that further gains are possible by learning to filter the resulting lexical entries . Improving on a recent method for doing BLI via unsupervised machine translation ( Artetxe et al . , 2019 ) , we show that unsupervised mining produces better bitext for lexicon induction than translation , especially for less frequent words . These core contributions are established by systematic experiments in the class of bitext construction and alignment methods ( Figure 1 ) . Our full induction algorithm filters the lexicon found via the initial unsupervised pipeline . The filtering can be either fully unsupervised or weakly-supervised : for the former , we filter using simple heuristics and global statistics ; for the latter , we train a multi-layer perceptron ( MLP ) to predict the probability of a word pair being in the lexicon , where the features are global statistics of word alignments . In addition to BLI , our method can also be directly adapted to improve word alignment and reach competitive or better alignment accuracy than the state of the art on all investigated language pairs . We find that improved alignment in sentence representations ( Tran et al . , 2020 ) leads to better contextual word alignments using local similarity ( Sabet et al . , 2020 ) . Our final BLI approach outperforms the previous state of the art on the BUCC 2020 shared task ( Rapp et al . , 2020 ) by 14 F 1 points averaged over 12 language pairs . Manual analysis shows that most of our false positives are due to the incompleteness of the reference and that our lexicon is comparable to the reference lexicon and the output of a supervised system . Because both of our key building blocks make use of the pretrainined contextual representations from mBART ( Liu et al . , Word Alignment Statistical Feature Extraction cooccurrence(good , guten ) = 2 one-to-one align(good , guten ) = 2 many-to-one align(good , guten ) = 0 cosine_similarity(good , guten ) = 0.8 inner_product(good , guten ) = 1.8 count(good ) = 2 count(guten ) = 2 We present a direct and effective framework for BLI with unsupervised bitext mining and word alignment , which sets a new state of the art on the task . From the perspective of pretrained multilingual models ( Conneau et al . , 2019 ; Liu et al . , 2020 ; Tran et al . , 2020 , inter alia ) , our work shows that they have successfully captured information about word translation that can be extracted using similarity based alignment and refinement . Although BLI is only about word types , it strongly benefits from contextualized reasoning at the token level . umich.edu/ ˜mihalcea / wpt ( en-fr and ro-en ) ; https : //web.eecs.umich.edu/ ˜mihalcea / wpt05 ( enhi ) | The proposed method achieves the state-of-the-art in the bilingual lexical induction task while keeping the interpretability of their pipeline. | outcome | low_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: outcome.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
Describe the conclusions briefly.
---
Document:
Bilingual lexicons map words in one language to their translations in another , and are typically induced by learning linear projections to align monolingual word embedding spaces . In this paper , we show it is possible to produce much higher quality lexicons with methods that combine ( 1 ) unsupervised bitext mining and ( 2 ) unsupervised word alignment . Directly applying a pipeline that uses recent algorithms for both subproblems significantly improves induced lexicon quality and further gains are possible by learning to filter the resulting lexical entries , with both unsupervised and semisupervised schemes . Our final model outperforms the state of the art on the BUCC 2020 shared task by 14 F 1 points averaged over 12 language pairs , while also providing a more interpretable approach that allows for rich reasoning of word meaning in context . Further analysis of our output and the standard reference lexicons suggests they are of comparable quality , and new benchmarks may be needed to measure further progress on this task . 1 Bilingual lexicons map words in one language to their translations in another , and can be automatically induced by learning linear projections to align monolingual word embedding spaces ( Artetxe et al . , 2016 ; Smith et al . , 2017 ; Lample et al . , 2018 , inter alia ) . Although very successful in practice , the linear nature of these methods encodes unrealistic simplifying assumptions ( e.g. all translations of a word have similar embeddings ) . In this paper , we show it is possible to produce much higher quality lexicons without these restrictions by introducing new methods that combine ( 1 ) unsupervised bitext mining and ( 2 ) unsupervised word alignment . We show that simply pipelining recent algorithms for unsupervised bitext mining ( Tran et al . , 2020 ) and unsupervised word alignment ( Sabet et al . , 2020 ) significantly improves bilingual lexicon induction ( BLI ) quality , and that further gains are possible by learning to filter the resulting lexical entries . Improving on a recent method for doing BLI via unsupervised machine translation ( Artetxe et al . , 2019 ) , we show that unsupervised mining produces better bitext for lexicon induction than translation , especially for less frequent words . These core contributions are established by systematic experiments in the class of bitext construction and alignment methods ( Figure 1 ) . Our full induction algorithm filters the lexicon found via the initial unsupervised pipeline . The filtering can be either fully unsupervised or weakly-supervised : for the former , we filter using simple heuristics and global statistics ; for the latter , we train a multi-layer perceptron ( MLP ) to predict the probability of a word pair being in the lexicon , where the features are global statistics of word alignments . In addition to BLI , our method can also be directly adapted to improve word alignment and reach competitive or better alignment accuracy than the state of the art on all investigated language pairs . We find that improved alignment in sentence representations ( Tran et al . , 2020 ) leads to better contextual word alignments using local similarity ( Sabet et al . , 2020 ) . Our final BLI approach outperforms the previous state of the art on the BUCC 2020 shared task ( Rapp et al . , 2020 ) by 14 F 1 points averaged over 12 language pairs . Manual analysis shows that most of our false positives are due to the incompleteness of the reference and that our lexicon is comparable to the reference lexicon and the output of a supervised system . Because both of our key building blocks make use of the pretrainined contextual representations from mBART ( Liu et al . , Word Alignment Statistical Feature Extraction cooccurrence(good , guten ) = 2 one-to-one align(good , guten ) = 2 many-to-one align(good , guten ) = 0 cosine_similarity(good , guten ) = 0.8 inner_product(good , guten ) = 1.8 count(good ) = 2 count(guten ) = 2 We present a direct and effective framework for BLI with unsupervised bitext mining and word alignment , which sets a new state of the art on the task . From the perspective of pretrained multilingual models ( Conneau et al . , 2019 ; Liu et al . , 2020 ; Tran et al . , 2020 , inter alia ) , our work shows that they have successfully captured information about word translation that can be extracted using similarity based alignment and refinement . Although BLI is only about word types , it strongly benefits from contextualized reasoning at the token level . umich.edu/ ˜mihalcea / wpt ( en-fr and ro-en ) ; https : //web.eecs.umich.edu/ ˜mihalcea / wpt05 ( enhi )
| sobamchan/aclsum | 2 | full_paper |
2020.acl-main.75 | Humor plays an important role in human languages and it is essential to model humor when building intelligence systems . Among different forms of humor , puns perform wordplay for humorous effects by employing words with double entendre and high phonetic similarity . However , identifying and modeling puns are challenging as puns usually involved implicit semantic or phonological tricks . In this paper , we propose Pronunciation-attentive Contextualized Pun Recognition ( PCPR ) to perceive human humor , detect if a sentence contains puns and locate them in the sentence . PCPR derives contextualized representation for each word in a sentence by capturing the association between the surrounding context and its corresponding phonetic symbols . Extensive experiments are conducted on two benchmark datasets . Results demonstrate that the proposed approach significantly outperforms the state-of-the-art methods in pun detection and location tasks . In-depth analyses verify the effectiveness and robustness of PCPR . During the last decades , social media has promoted the creation of a vast amount of humorous web contents ( Nijholt et al . , 2017 ) . Automatic recognition of humor has become an important task in the area of figurative language processing , which can benefit various downstream NLP applications such as dialogue systems , sentiment analysis , and machine translation ( Melby and Warner , 1995 ; Augello et al . , 2008 ; Ghosh et al . , 2015 ; Bertero and Fung , 2016 ; Blinov et al . , 2019 ) . However , humor is one of the most complicated behaviors in natural language semantics and sometimes it is even difficult for humans to interpret . In most cases , understanding humor requires adequate background knowledge and a rich context . Puns are a form of humorous approaches using the different meanings of identical words or words with similar pronunciations to explain texts or utterances . There are two main types of puns . Homographic puns rely on multiple interpretations of the same word . As shown in Table 1 , the phrase all right means good condition or opposite to left ; the word reaction means chemical change or action . The two meanings of the same expression are consistent with its context , which creates a humorous pun in both sentences when there is a clear contrast between two meanings . On the other hand , heterographic puns take advantage of phonologically same or similar words . For example , the word pairs sale and sail , weak and week in Table 1 have the same or similar pronunciations . The sentences are funny because both words fit the same context . Understanding puns is a big fish to fry for deep comprehension of complex semantics . These two forms of puns have been studied in literature from different angles . To recognize puns in a sentence , word sense disambiguation techniques ( WSD ) ( Navigli , 2009 ) have been employed to identify the equitable intention of words in utterances ( Pedersen , 2017 ) . External knowledge bases such as WordNet ( Miller , 1998b ) have been applied in determining word senses of pun words ( Oele and Evang , 2017 ) . However , these methods can not tackle heterographic puns with distinct word spellings and knowledge bases that only contain a limited vocabulary . To resolve the issues of sparseness and heterographics , the word embedding techniques ( Mikolov et al . , 2013 ; Pennington et al . , 2014 ) provide flexible representations to model puns ( Hurtado et al . , 2017 ; Indurthi and Oota , 2017 ; Cai et al . , 2018 ) . However , a word may have different meanings regarding its contexts . Especially , an infrequent meaning of the word might be utilized for creating a pun . Therefore , static word embeddings are insufficient to represent words . In addition , some puns are created by replacing a word with another word with the same or similar pronunciation as examples shown in Table 1 . Therefore , to recognize puns , it is essential to model the association between words in the sentence and the pronunciation of words . Despite existing approaches attempt to leverage phonological structures to understand puns ( Doogan et al . , 2017 ; Jaech et al . , 2016 ) , there is a lack of a general framework to model these two types of signals in a whole . In this paper , we propose Pronunciation-attentive Contextualized Pun Recognition ( PCPR ) to jointly model the contextualized word embeddings and phonological word representations for pun recognition . To capture the phonological structures of words , we break each word into a sequence of phonemes as its pronunciation so that homophones can have similar phoneme sets . For instance , the phonemes of the word pun are { P , AH , N } . In PCPR , we construct a pronunciation attentive module to identify important phonemes of each word , which can be applied in other tasks related to phonology . We jointly encode the contextual and phonological features into a self-attentive embedding to tackle both pun detection and location tasks . We summarize our contributions as following . • To the best of our knowledge , PCPR is the first work to jointly model contextualized word embeddings and pronunciation embeddings to recognize puns . Both contexts and phonological properties are beneficial to pun recognition . In this paper , we propose a novel approach , PCPR , for pun detection and location by leveraging a contextualized word encoder and modeling phonemes as word pronunciations . Moreover , we would love to apply the proposed model to other problems , such as general humor recognition , irony discovery , and sarcasm detection , as the future work . | Puns involve implicit semantic or phonological tricks however there is no general framework to model these two types of signals as a whole. | challenge | high_quality | adherence_first | Objective: ADHERENCE-FIRST.
Task: Write a one-sentence summary focused ONLY on the aspect: challenge.
Be strict: exclude any information not directly tied to the aspect.
If uncertain, omit rather than speculate.
Stay faithful to the document; do not add any unsupported claims.
Extract a short summary of the core issue that the paper targets. Return only the summary in one sentence.
---
Document:
Humor plays an important role in human languages and it is essential to model humor when building intelligence systems . Among different forms of humor , puns perform wordplay for humorous effects by employing words with double entendre and high phonetic similarity . However , identifying and modeling puns are challenging as puns usually involved implicit semantic or phonological tricks . In this paper , we propose Pronunciation-attentive Contextualized Pun Recognition ( PCPR ) to perceive human humor , detect if a sentence contains puns and locate them in the sentence . PCPR derives contextualized representation for each word in a sentence by capturing the association between the surrounding context and its corresponding phonetic symbols . Extensive experiments are conducted on two benchmark datasets . Results demonstrate that the proposed approach significantly outperforms the state-of-the-art methods in pun detection and location tasks . In-depth analyses verify the effectiveness and robustness of PCPR . During the last decades , social media has promoted the creation of a vast amount of humorous web contents ( Nijholt et al . , 2017 ) . Automatic recognition of humor has become an important task in the area of figurative language processing , which can benefit various downstream NLP applications such as dialogue systems , sentiment analysis , and machine translation ( Melby and Warner , 1995 ; Augello et al . , 2008 ; Ghosh et al . , 2015 ; Bertero and Fung , 2016 ; Blinov et al . , 2019 ) . However , humor is one of the most complicated behaviors in natural language semantics and sometimes it is even difficult for humans to interpret . In most cases , understanding humor requires adequate background knowledge and a rich context . Puns are a form of humorous approaches using the different meanings of identical words or words with similar pronunciations to explain texts or utterances . There are two main types of puns . Homographic puns rely on multiple interpretations of the same word . As shown in Table 1 , the phrase all right means good condition or opposite to left ; the word reaction means chemical change or action . The two meanings of the same expression are consistent with its context , which creates a humorous pun in both sentences when there is a clear contrast between two meanings . On the other hand , heterographic puns take advantage of phonologically same or similar words . For example , the word pairs sale and sail , weak and week in Table 1 have the same or similar pronunciations . The sentences are funny because both words fit the same context . Understanding puns is a big fish to fry for deep comprehension of complex semantics . These two forms of puns have been studied in literature from different angles . To recognize puns in a sentence , word sense disambiguation techniques ( WSD ) ( Navigli , 2009 ) have been employed to identify the equitable intention of words in utterances ( Pedersen , 2017 ) . External knowledge bases such as WordNet ( Miller , 1998b ) have been applied in determining word senses of pun words ( Oele and Evang , 2017 ) . However , these methods can not tackle heterographic puns with distinct word spellings and knowledge bases that only contain a limited vocabulary . To resolve the issues of sparseness and heterographics , the word embedding techniques ( Mikolov et al . , 2013 ; Pennington et al . , 2014 ) provide flexible representations to model puns ( Hurtado et al . , 2017 ; Indurthi and Oota , 2017 ; Cai et al . , 2018 ) . However , a word may have different meanings regarding its contexts . Especially , an infrequent meaning of the word might be utilized for creating a pun . Therefore , static word embeddings are insufficient to represent words . In addition , some puns are created by replacing a word with another word with the same or similar pronunciation as examples shown in Table 1 . Therefore , to recognize puns , it is essential to model the association between words in the sentence and the pronunciation of words . Despite existing approaches attempt to leverage phonological structures to understand puns ( Doogan et al . , 2017 ; Jaech et al . , 2016 ) , there is a lack of a general framework to model these two types of signals in a whole . In this paper , we propose Pronunciation-attentive Contextualized Pun Recognition ( PCPR ) to jointly model the contextualized word embeddings and phonological word representations for pun recognition . To capture the phonological structures of words , we break each word into a sequence of phonemes as its pronunciation so that homophones can have similar phoneme sets . For instance , the phonemes of the word pun are { P , AH , N } . In PCPR , we construct a pronunciation attentive module to identify important phonemes of each word , which can be applied in other tasks related to phonology . We jointly encode the contextual and phonological features into a self-attentive embedding to tackle both pun detection and location tasks . We summarize our contributions as following . • To the best of our knowledge , PCPR is the first work to jointly model contextualized word embeddings and pronunciation embeddings to recognize puns . Both contexts and phonological properties are beneficial to pun recognition . In this paper , we propose a novel approach , PCPR , for pun detection and location by leveraging a contextualized word encoder and modeling phonemes as word pronunciations . Moreover , we would love to apply the proposed model to other problems , such as general humor recognition , irony discovery , and sarcasm detection , as the future work .
| sobamchan/aclsum | 2 | full_paper |
P10-1139 | There is a growing research interest in opinion retrieval as on-line users ' opinions are becoming more and more popular in business , social networks , etc . Practically speaking , the goal of opinion retrieval is to retrieve documents , which entail opinions or comments , relevant to a target subject specified by the user 's query . A fundamental challenge in opinion retrieval is information representation . Existing research focuses on document-based approaches and documents are represented by bag-of-word . However , due to loss of contextual information , this representation fails to capture the associative information between an opinion and its corresponding target . It can not distinguish different degrees of a sentiment word when associated with different targets . This in turn seriously affects opinion retrieval performance . In this paper , we propose a sentence-based approach based on a new information representation , namely topic-sentiment word pair , to capture intra-sentence contextual information between an opinion and its target . Additionally , we consider inter-sentence information to capture the relationships among the opinions on the same topic . Finally , the two types of information are combined in a unified graph-based model , which can effectively rank the documents . Compared with existing approaches , experimental results on the COAE08 dataset showed that our graph-based model achieved significant improvement . In recent years , there is a growing interest in sharing personal opinions on the Web , such as product reviews , economic analysis , political polls , etc . These opinions can not only help independent users make decisions , but also obtain valuable feedbacks ( Pang et al . , 2008 ) . Opinion oriented research , including sentiment classifica-tion , opinion extraction , opinion question answering , and opinion summarization , etc . are receiving growing attention ( Wilson , et al . , 2005 ; Liu et al . , 2005 ; Oard et al . , 2006 ) . However , most existing works concentrate on analyzing opinions expressed in the documents , and none on how to represent the information needs required to retrieve opinionated documents . In this paper , we focus on opinion retrieval , whose goal is to find a set of documents containing not only the query keyword(s ) but also the relevant opinions . This requirement brings about the challenge on how to represent information needs for effective opinion retrieval . In order to solve the above problem , previous work adopts a 2-stage approach . In the first stage , relevant documents are determined and ranked by a score , i.e. tf-idf value . In the second stage , an opinion score is generated for each relevant document ( Macdonald and Ounis , 2007 ; Oard et al . , 2006 ) . The opinion score can be acquired by either machine learning-based sentiment classifiers , such as SVM ( Zhang and Yu , 2007 ) , or a sentiment lexicons with weighted scores from training documents ( Amati et al . , 2007 ; Hannah et al . , 2007 ; Na et al . , 2009 ) . Finally , an overall score combining the two is computed by using a score function , e.g. linear combination , to re-rank the retrieved documents . Retrieval in the 2-stage approach is based on document and document is represented by bag-of-word . This representation , however , can only ensure that there is at least one opinion in each relevant document , but it can not determine the relevance pairing of individual opinion to its target . In general , by simply representing a document in bag-of-word , contextual information i.e. the corresponding target of an opinion , is neglected . This may result in possible mismatch between an opinion and a target and in turn affects opinion retrieval performance . By the same token , the effect to documents consisting of mul-tiple topics , which is common in blogs and on-line reviews , is also significant . In this setting , even if a document is regarded opinionated , it can not ensure that all opinions in the document are indeed relevant to the target concerned . Therefore , we argue that existing information representation i.e. bag-of-word , can not satisfy the information needs for opinion retrieval . In this paper , we propose to handle opinion retrieval in the granularity of sentence . It is observed that a complete opinion is always expressed in one sentence , and the relevant target of the opinion is mostly the one found in it . Therefore , it is crucial to maintain the associative information between an opinion and its target within a sentence . We define the notion of a topic-sentiment word pair , which is composed of a topic term ( i.e. the target ) and a sentiment word ( i.e. opinion ) of a sentence . Word pairs can maintain intra-sentence contextual information to express the potential relevant opinions . In addition , inter-sentence contextual information is also captured by word pairs to represent the relationship among opinions on the same topic . In practice , the inter-sentence information reflects the degree of a word pair . Finally , we combine both intra-sentence and inter-sentence contextual information to construct a unified undirected graph to achieve effective opinion retrieval . The rest of the paper is organized as follows . In Section 2 , we describe the motivation of our approach . Section 3 presents a novel unified graph-based model for opinion retrieval . We evaluated our model and the results are presented in Section 4 . We review related works on opinion retrieval in Section 5 . Finally , in Section 6 , the paper is concluded and future work is suggested . In this work we focus on the problem of opinion retrieval . Different from existing approaches , which regard document relevance as the key indicator of opinion relevance , we propose to explore the relevance of individual opinion . To do that , opinion retrieval is performed in the granularity of sentence . We define the notion of word pair , which can not only maintain the association between the opinion and the corresponding target in the sentence , but it can also build up the relationship among sentences through the same word pair . Furthermore , we convert the relationships between word pairs and sentences into a unified graph , and use the HITS algorithm to achieve document ranking for opinion retrieval . Finally , we compare our approach with existing methods . Experimental results show that our proposed model performs well on COAE08 dataset . The novelty of our work lies in using word pairs to represent the information needs for opinion retrieval . On the one hand , word pairs can identify the relevant opinion according to intra-sentence contextual information . On the other hand , word pairs can measure the degree of a relevant opinion by taking inter-sentence contextual information into consideration . With the help of word pairs , the information needs for opinion retrieval can be represented appropriately . In the future , more research is required in the following directions : ( 1 ) Since word pairs can indicate relevant opinions effectively , it is worth further study on how they could be applied to other opinion oriented applications , e.g. opinion summarization , opinion prediction , etc . ( 2 ) The characteristics of blogs will be taken into consideration , i.e. , the post time , which could be helpful to create a more time sensitivity graph to filter out fake opinions . ( 3 ) Opinion holder is another important role of an opinion , and the identification of opinion holder is a main task in NTCIR . It would be interesting to study opinion holders , e.g. its seniority , for opinion retrieval . | Existing approaches to the opinion retrieval task represent documents using bag-of-words disregarding contextual information between an opinion and its corresponding text. | challenge | high_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: challenge.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
From the following paper excerpt, summarize the key problem or challenge the authors aim to address. Return only the summary in one sentence.
---
Document:
There is a growing research interest in opinion retrieval as on-line users ' opinions are becoming more and more popular in business , social networks , etc . Practically speaking , the goal of opinion retrieval is to retrieve documents , which entail opinions or comments , relevant to a target subject specified by the user 's query . A fundamental challenge in opinion retrieval is information representation . Existing research focuses on document-based approaches and documents are represented by bag-of-word . However , due to loss of contextual information , this representation fails to capture the associative information between an opinion and its corresponding target . It can not distinguish different degrees of a sentiment word when associated with different targets . This in turn seriously affects opinion retrieval performance . In this paper , we propose a sentence-based approach based on a new information representation , namely topic-sentiment word pair , to capture intra-sentence contextual information between an opinion and its target . Additionally , we consider inter-sentence information to capture the relationships among the opinions on the same topic . Finally , the two types of information are combined in a unified graph-based model , which can effectively rank the documents . Compared with existing approaches , experimental results on the COAE08 dataset showed that our graph-based model achieved significant improvement . In recent years , there is a growing interest in sharing personal opinions on the Web , such as product reviews , economic analysis , political polls , etc . These opinions can not only help independent users make decisions , but also obtain valuable feedbacks ( Pang et al . , 2008 ) . Opinion oriented research , including sentiment classifica-tion , opinion extraction , opinion question answering , and opinion summarization , etc . are receiving growing attention ( Wilson , et al . , 2005 ; Liu et al . , 2005 ; Oard et al . , 2006 ) . However , most existing works concentrate on analyzing opinions expressed in the documents , and none on how to represent the information needs required to retrieve opinionated documents . In this paper , we focus on opinion retrieval , whose goal is to find a set of documents containing not only the query keyword(s ) but also the relevant opinions . This requirement brings about the challenge on how to represent information needs for effective opinion retrieval . In order to solve the above problem , previous work adopts a 2-stage approach . In the first stage , relevant documents are determined and ranked by a score , i.e. tf-idf value . In the second stage , an opinion score is generated for each relevant document ( Macdonald and Ounis , 2007 ; Oard et al . , 2006 ) . The opinion score can be acquired by either machine learning-based sentiment classifiers , such as SVM ( Zhang and Yu , 2007 ) , or a sentiment lexicons with weighted scores from training documents ( Amati et al . , 2007 ; Hannah et al . , 2007 ; Na et al . , 2009 ) . Finally , an overall score combining the two is computed by using a score function , e.g. linear combination , to re-rank the retrieved documents . Retrieval in the 2-stage approach is based on document and document is represented by bag-of-word . This representation , however , can only ensure that there is at least one opinion in each relevant document , but it can not determine the relevance pairing of individual opinion to its target . In general , by simply representing a document in bag-of-word , contextual information i.e. the corresponding target of an opinion , is neglected . This may result in possible mismatch between an opinion and a target and in turn affects opinion retrieval performance . By the same token , the effect to documents consisting of mul-tiple topics , which is common in blogs and on-line reviews , is also significant . In this setting , even if a document is regarded opinionated , it can not ensure that all opinions in the document are indeed relevant to the target concerned . Therefore , we argue that existing information representation i.e. bag-of-word , can not satisfy the information needs for opinion retrieval . In this paper , we propose to handle opinion retrieval in the granularity of sentence . It is observed that a complete opinion is always expressed in one sentence , and the relevant target of the opinion is mostly the one found in it . Therefore , it is crucial to maintain the associative information between an opinion and its target within a sentence . We define the notion of a topic-sentiment word pair , which is composed of a topic term ( i.e. the target ) and a sentiment word ( i.e. opinion ) of a sentence . Word pairs can maintain intra-sentence contextual information to express the potential relevant opinions . In addition , inter-sentence contextual information is also captured by word pairs to represent the relationship among opinions on the same topic . In practice , the inter-sentence information reflects the degree of a word pair . Finally , we combine both intra-sentence and inter-sentence contextual information to construct a unified undirected graph to achieve effective opinion retrieval . The rest of the paper is organized as follows . In Section 2 , we describe the motivation of our approach . Section 3 presents a novel unified graph-based model for opinion retrieval . We evaluated our model and the results are presented in Section 4 . We review related works on opinion retrieval in Section 5 . Finally , in Section 6 , the paper is concluded and future work is suggested . In this work we focus on the problem of opinion retrieval . Different from existing approaches , which regard document relevance as the key indicator of opinion relevance , we propose to explore the relevance of individual opinion . To do that , opinion retrieval is performed in the granularity of sentence . We define the notion of word pair , which can not only maintain the association between the opinion and the corresponding target in the sentence , but it can also build up the relationship among sentences through the same word pair . Furthermore , we convert the relationships between word pairs and sentences into a unified graph , and use the HITS algorithm to achieve document ranking for opinion retrieval . Finally , we compare our approach with existing methods . Experimental results show that our proposed model performs well on COAE08 dataset . The novelty of our work lies in using word pairs to represent the information needs for opinion retrieval . On the one hand , word pairs can identify the relevant opinion according to intra-sentence contextual information . On the other hand , word pairs can measure the degree of a relevant opinion by taking inter-sentence contextual information into consideration . With the help of word pairs , the information needs for opinion retrieval can be represented appropriately . In the future , more research is required in the following directions : ( 1 ) Since word pairs can indicate relevant opinions effectively , it is worth further study on how they could be applied to other opinion oriented applications , e.g. opinion summarization , opinion prediction , etc . ( 2 ) The characteristics of blogs will be taken into consideration , i.e. , the post time , which could be helpful to create a more time sensitivity graph to filter out fake opinions . ( 3 ) Opinion holder is another important role of an opinion , and the identification of opinion holder is a main task in NTCIR . It would be interesting to study opinion holders , e.g. its seniority , for opinion retrieval .
| sobamchan/aclsum | 0 | full_paper |
P16-1067 | This paper proposes an unsupervised approach for segmenting a multiauthor document into authorial components . The key novelty is that we utilize the sequential patterns hidden among document elements when determining their authorships . For this purpose , we adopt Hidden Markov Model ( HMM ) and construct a sequential probabilistic model to capture the dependencies of sequential sentences and their authorships . An unsupervised learning method is developed to initialize the HMM parameters . Experimental results on benchmark datasets have demonstrated the significant benefit of our idea and our approach has outperformed the state-of-the-arts on all tests . As an example of its applications , the proposed approach is applied for attributing authorship of a document and has also shown promising results . Authorship analysis is a process of inspecting documents in order to extract authorial information about these documents . It is considered as a general concept that embraces several types of authorship subjects , including authorship verif ication , plagiarism detection and author attribution . Authorship verification ( Brocardo et al . , 2013 ; Potha and Stamatatos , 2014 ) decides whether a given document is written by a specific author . Plagiarism detection ( Stein et al . , 2011 ; Kestemont et al . , 2011 ) seeks to expose the similarity between two texts . However , it is un-able to determine if they are written by the same author . In author attribution ( Juola , 2006 ; Savoy , 2015 ) , a real author of an anonymous document is predicted using labeled documents of a set of candidate authors . Another significant subject in authorship analysis , which has received comparatively less attention from research community , is authorship-based document decomposition ( ABDD ) . This subject is to group the sentences of a multi-author document to different classes , of which each contains the sentences written by only one author . Many applications can take advantage of such a subject , especially those in forensic investigation , which aim to determine the authorship of sentences in a multi-author document . Furthermore , this kind of subject is beneficial for detecting plagiarism in a document and defining contributions of authors in a multi-author document for commercial purpose . ABDD can also be applied to identify which source ( regarded as an ' author ' in this paper ) a part of a document is copied from when the document is formed by taking contents from various sources . In despite of the benefits of ABDD , there has been little research reported on this subject . Koppel et al . ( 2011 ) are the first researchers who implemented an unsupervised approach for ABDD . However , their approach is restricted to Hebrew documents only . The authors of Akiva and Koppel ( 2013 ) addressed the drawbacks of the above approach by proposing a generic unsupervised approach for ABDD . Their approach utilized distance measurements to increase the precision and accuracy of clustering and classification phases , respectively . The accuracy of their approach is highly dependent on the number of au-thors . When the number of authors increases , the accuracy of the approach is significantly dropped . Giannella ( 2015 ) presented an improved approach for ABDD when the number of authors of the document is known or unknown . In his approach , a Bayesian segmentation algorithm is applied , which is followed by a segment clustering algorithm . However , the author tested his approach by using only documents with a few transitions among authors . Furthermore , the accuracy of the approach is very sensitive to the setting of its parameters . In Aldebei et al . ( 2015 ) , the authors presented an unsupervised approach ABDD by exploiting the differences in the posterior probabilities of a Naive-Bayesian model in order to increase the precision and the classification accuracy , and to be less dependent on the number of authors in comparing with the approach in Akiva and Koppel ( 2013 ) . Their work was tested on documents with up to 400 transitions among authors and the accuracy of their approach was not sensitive to the setting of parameters , in contrast with the approach in Giannella ( 2015 ) . However , the performance of their approach greatly depends on a threshold , of which the optimal value for an individual document is not easy to find . Some other works have focused on segmenting a document into components according to their topics . For applications where the topics of documents are unavailable , these topicbased solutions will fail . In this paper , the ABDD approach is independent of documents ' topics . All of the existing works have assumed that the observations ( i.e. , sentences ) are independent and identically distributed ( i.i.d . ) . No consideration has been given to the contextual information between the observations . However , in some cases , the i.i.d . assumption is deemed as a poor one ( Rogovschi et al . , 2010 ) . In this paper , we will relax this assumption and consider sentences of a document as a sequence of observations . We make use of the contextual information hidden between sentences in order to identify the authorship of each sentence in a document . In other words , the authorships of the " previous " and " subsequent " sentences have relationships with the authorship of the current sentence . There-fore , in this paper , a well-known sequential model , Hidden Markov Model ( HMM ) , is used for modelling the sequential patterns of the document in order to describe the authorship relationships . The contributions of this article are summarized as follows . 1 . We capture the dependencies between consecutive elements in a document to identify different authorial components and construct an HMM for classification . It is for the first time the sequential patterns hidden among document elements is considered for such a problem . 2 . To build and learn the HMM model , an unsupervised learning method is first proposed to estimate its initial parameters , and it does not require any information of authors or document 's context other than how many authors have contributed to write the document . 3 . Different from the approach in Aldebei et al . ( 2015 ) , the proposed unsupervised approach no longer relies on any predetermined threshold for ABDD . 4 . Comprehensive experiments are conducted to demonstrate the superior performance of our ideas on both widely-used artificial benchmark datasets and an authentic scientific document . As an example of its applications , the proposed approach is also applied for attributing authorship on a popular dataset . The proposed approach can not only correctly determine the author of a disputed document but also provide a way for measuring the confidence level of the authorship decision for the first time . The rest of this article is organised as follows . Section 2 reviews the HMM . Section 3 presents the details of our proposed approach , including the processes for initialization and learning of HMM parameters , and the Viterbi decoding process for classification . Experiments are conducted in Section 4 , followed by the conclusion in Section 5 . We have developed an unsupervised approach for decomposing a multi-author document based on authorship . Different from the stateof-the-art approaches , we have innovatively made use of the sequential information hidden among document elements . For this purpose , we have used HMM and constructed a sequential probabilistic model , which is used to find the best sequence of authors that represents the sentences of the document . An unsupervised learning method has also been developed to estimate the initial parameter values of HMM . Comparative experiments conducted on benchmark datasets have demonstrated the effectiveness of our ideas with superior perfor-mance achieved on both artificial and authentic documents . An application of the proposed approach on authorship attribution has also achieved perfect results of 100 % accuracies together with confidence measurement for the first time . | Experiments with artificial and authentic scientific document datasets show that the proposed model outperforms existing methods and also be able to provide confidence scores. | outcome | high_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: outcome.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
Extract a short summary of the paper’s results and conclusions. Return only the summary in one sentence.
---
Document:
This paper proposes an unsupervised approach for segmenting a multiauthor document into authorial components . The key novelty is that we utilize the sequential patterns hidden among document elements when determining their authorships . For this purpose , we adopt Hidden Markov Model ( HMM ) and construct a sequential probabilistic model to capture the dependencies of sequential sentences and their authorships . An unsupervised learning method is developed to initialize the HMM parameters . Experimental results on benchmark datasets have demonstrated the significant benefit of our idea and our approach has outperformed the state-of-the-arts on all tests . As an example of its applications , the proposed approach is applied for attributing authorship of a document and has also shown promising results . Authorship analysis is a process of inspecting documents in order to extract authorial information about these documents . It is considered as a general concept that embraces several types of authorship subjects , including authorship verif ication , plagiarism detection and author attribution . Authorship verification ( Brocardo et al . , 2013 ; Potha and Stamatatos , 2014 ) decides whether a given document is written by a specific author . Plagiarism detection ( Stein et al . , 2011 ; Kestemont et al . , 2011 ) seeks to expose the similarity between two texts . However , it is un-able to determine if they are written by the same author . In author attribution ( Juola , 2006 ; Savoy , 2015 ) , a real author of an anonymous document is predicted using labeled documents of a set of candidate authors . Another significant subject in authorship analysis , which has received comparatively less attention from research community , is authorship-based document decomposition ( ABDD ) . This subject is to group the sentences of a multi-author document to different classes , of which each contains the sentences written by only one author . Many applications can take advantage of such a subject , especially those in forensic investigation , which aim to determine the authorship of sentences in a multi-author document . Furthermore , this kind of subject is beneficial for detecting plagiarism in a document and defining contributions of authors in a multi-author document for commercial purpose . ABDD can also be applied to identify which source ( regarded as an ' author ' in this paper ) a part of a document is copied from when the document is formed by taking contents from various sources . In despite of the benefits of ABDD , there has been little research reported on this subject . Koppel et al . ( 2011 ) are the first researchers who implemented an unsupervised approach for ABDD . However , their approach is restricted to Hebrew documents only . The authors of Akiva and Koppel ( 2013 ) addressed the drawbacks of the above approach by proposing a generic unsupervised approach for ABDD . Their approach utilized distance measurements to increase the precision and accuracy of clustering and classification phases , respectively . The accuracy of their approach is highly dependent on the number of au-thors . When the number of authors increases , the accuracy of the approach is significantly dropped . Giannella ( 2015 ) presented an improved approach for ABDD when the number of authors of the document is known or unknown . In his approach , a Bayesian segmentation algorithm is applied , which is followed by a segment clustering algorithm . However , the author tested his approach by using only documents with a few transitions among authors . Furthermore , the accuracy of the approach is very sensitive to the setting of its parameters . In Aldebei et al . ( 2015 ) , the authors presented an unsupervised approach ABDD by exploiting the differences in the posterior probabilities of a Naive-Bayesian model in order to increase the precision and the classification accuracy , and to be less dependent on the number of authors in comparing with the approach in Akiva and Koppel ( 2013 ) . Their work was tested on documents with up to 400 transitions among authors and the accuracy of their approach was not sensitive to the setting of parameters , in contrast with the approach in Giannella ( 2015 ) . However , the performance of their approach greatly depends on a threshold , of which the optimal value for an individual document is not easy to find . Some other works have focused on segmenting a document into components according to their topics . For applications where the topics of documents are unavailable , these topicbased solutions will fail . In this paper , the ABDD approach is independent of documents ' topics . All of the existing works have assumed that the observations ( i.e. , sentences ) are independent and identically distributed ( i.i.d . ) . No consideration has been given to the contextual information between the observations . However , in some cases , the i.i.d . assumption is deemed as a poor one ( Rogovschi et al . , 2010 ) . In this paper , we will relax this assumption and consider sentences of a document as a sequence of observations . We make use of the contextual information hidden between sentences in order to identify the authorship of each sentence in a document . In other words , the authorships of the " previous " and " subsequent " sentences have relationships with the authorship of the current sentence . There-fore , in this paper , a well-known sequential model , Hidden Markov Model ( HMM ) , is used for modelling the sequential patterns of the document in order to describe the authorship relationships . The contributions of this article are summarized as follows . 1 . We capture the dependencies between consecutive elements in a document to identify different authorial components and construct an HMM for classification . It is for the first time the sequential patterns hidden among document elements is considered for such a problem . 2 . To build and learn the HMM model , an unsupervised learning method is first proposed to estimate its initial parameters , and it does not require any information of authors or document 's context other than how many authors have contributed to write the document . 3 . Different from the approach in Aldebei et al . ( 2015 ) , the proposed unsupervised approach no longer relies on any predetermined threshold for ABDD . 4 . Comprehensive experiments are conducted to demonstrate the superior performance of our ideas on both widely-used artificial benchmark datasets and an authentic scientific document . As an example of its applications , the proposed approach is also applied for attributing authorship on a popular dataset . The proposed approach can not only correctly determine the author of a disputed document but also provide a way for measuring the confidence level of the authorship decision for the first time . The rest of this article is organised as follows . Section 2 reviews the HMM . Section 3 presents the details of our proposed approach , including the processes for initialization and learning of HMM parameters , and the Viterbi decoding process for classification . Experiments are conducted in Section 4 , followed by the conclusion in Section 5 . We have developed an unsupervised approach for decomposing a multi-author document based on authorship . Different from the stateof-the-art approaches , we have innovatively made use of the sequential information hidden among document elements . For this purpose , we have used HMM and constructed a sequential probabilistic model , which is used to find the best sequence of authors that represents the sentences of the document . An unsupervised learning method has also been developed to estimate the initial parameter values of HMM . Comparative experiments conducted on benchmark datasets have demonstrated the effectiveness of our ideas with superior perfor-mance achieved on both artificial and authentic documents . An application of the proposed approach on authorship attribution has also achieved perfect results of 100 % accuracies together with confidence measurement for the first time .
| sobamchan/aclsum | 2 | full_paper |
P07-1026 | Convolution tree kernel has shown promising results in semantic role classification . However , it only carries out hard matching , which may lead to over-fitting and less accurate similarity measure . To remove the constraint , this paper proposes a grammardriven convolution tree kernel for semantic role classification by introducing more linguistic knowledge into the standard tree kernel . The proposed grammar-driven tree kernel displays two advantages over the previous one : 1 ) grammar-driven approximate substructure matching and 2 ) grammardriven approximate tree node matching . The two improvements enable the grammardriven tree kernel explore more linguistically motivated structure features than the previous one . Experiments on the CoNLL-2005 SRL shared task show that the grammardriven tree kernel significantly outperforms the previous non-grammar-driven one in SRL . Moreover , we present a composite kernel to integrate feature-based and tree kernel-based methods . Experimental results show that the composite kernel outperforms the previously best-reported methods . Given a sentence , the task of Semantic Role Labeling ( SRL ) consists of analyzing the logical forms expressed by some target verbs or nouns and some constituents of the sentence . In particular , for each predicate ( target verb or noun ) all the constituents in the sentence which fill semantic arguments ( roles ) of the predicate have to be recognized . Typical semantic roles include Agent , Patient , Instrument , etc . and also adjuncts such as Locative , Temporal , Manner , and Cause , etc . Generally , semantic role identification and classification are regarded as two key steps in semantic role labeling . Semantic role identification involves classifying each syntactic element in a sentence into either a semantic argument or a non-argument while semantic role classification involves classifying each semantic argument identified into a specific semantic role . This paper focuses on semantic role classification task with the assumption that the semantic arguments have been identified correctly . Both feature-based and kernel-based learning methods have been studied for semantic role classification ( Carreras and Màrquez , 2004 ; Carreras and Màrquez , 2005 ) . In feature-based methods , a flat feature vector is used to represent a predicateargument structure while , in kernel-based methods , a kernel function is used to measure directly the similarity between two predicate-argument structures . As we know , kernel methods are more effective in capturing structured features . Moschitti ( 2004 ) and Che et al . ( 2006 ) used a convolution tree kernel ( Collins and Duffy , 2001 ) for semantic role classification . The convolution tree kernel takes sub-tree as its feature and counts the number of common sub-trees as the similarity between two predicate-arguments . This kernel has shown very promising results in SRL . However , as a general learning algorithm , the tree kernel only carries out hard matching between any two sub-trees without considering any linguistic knowledge in kernel design . This makes the kernel fail to handle similar phrase structures ( e.g. , " buy a car " vs. " buy a red car " ) and near-synonymic grammar tags ( e.g. , the POS variations between " high / JJ degree / NN " 1 and " higher / JJR degree / NN " ) 2 . To some degree , it may lead to over-fitting and compromise performance . This paper reports our preliminary study in addressing the above issue by introducing more linguistic knowledge into the convolution tree kernel . To our knowledge , this is the first attempt in this research direction . In detail , we propose a grammar-driven convolution tree kernel for semantic role classification that can carry out more linguistically motivated substructure matching . Experimental results show that the proposed method significantly outperforms the standard convolution tree kernel on the data set of the CoNLL-2005 SRL shared task . The remainder of the paper is organized as follows : Section 2 reviews the previous work and Section 3 discusses our grammar-driven convolution tree kernel . Section 4 shows the experimental results . We conclude our work in Section 5 . In this paper , we propose a novel grammar-driven convolution tree kernel for semantic role classification . More linguistic knowledge is considered in the new kernel design . The experimental results verify that the grammar-driven kernel is more effective in capturing syntactic structure features than the previous convolution tree kernel because it allows grammar-driven approximate matching of substructures and node features . We also discuss the criteria to determine the optional nodes in a CFG rule in defining our grammar-driven convolution tree kernel . The extension of our work is to improve the performance of the entire semantic role labeling system using the grammar-driven tree kernel , including all four stages : pruning , semantic role identification , classification and post inference . In addition , a more interesting research topic is to study how to integrate linguistic knowledge and tree kernel methods to do feature selection for tree kernelbased NLP applications ( Suzuki et al . , 2004 ) . In detail , a linguistics and statistics-based theory that can suggest the effectiveness of different substructure features and whether they should be generated or not by the tree kernels would be worked out . | They propose to integrate a linguistically motivated grammar-baed convolution tree kernel into a standard tree kernel to achieve better substructure matching and tree node matching. | approach | high_quality | adherence_first | Objective: ADHERENCE-FIRST.
Task: Write a one-sentence summary focused ONLY on the aspect: approach.
Be strict: exclude any information not directly tied to the aspect.
If uncertain, omit rather than speculate.
Stay faithful to the document; do not add any unsupported claims.
Summarize the method or approach the authors propose to tackle the problem. Return only the summary in one sentence.
---
Document:
Convolution tree kernel has shown promising results in semantic role classification . However , it only carries out hard matching , which may lead to over-fitting and less accurate similarity measure . To remove the constraint , this paper proposes a grammardriven convolution tree kernel for semantic role classification by introducing more linguistic knowledge into the standard tree kernel . The proposed grammar-driven tree kernel displays two advantages over the previous one : 1 ) grammar-driven approximate substructure matching and 2 ) grammardriven approximate tree node matching . The two improvements enable the grammardriven tree kernel explore more linguistically motivated structure features than the previous one . Experiments on the CoNLL-2005 SRL shared task show that the grammardriven tree kernel significantly outperforms the previous non-grammar-driven one in SRL . Moreover , we present a composite kernel to integrate feature-based and tree kernel-based methods . Experimental results show that the composite kernel outperforms the previously best-reported methods . Given a sentence , the task of Semantic Role Labeling ( SRL ) consists of analyzing the logical forms expressed by some target verbs or nouns and some constituents of the sentence . In particular , for each predicate ( target verb or noun ) all the constituents in the sentence which fill semantic arguments ( roles ) of the predicate have to be recognized . Typical semantic roles include Agent , Patient , Instrument , etc . and also adjuncts such as Locative , Temporal , Manner , and Cause , etc . Generally , semantic role identification and classification are regarded as two key steps in semantic role labeling . Semantic role identification involves classifying each syntactic element in a sentence into either a semantic argument or a non-argument while semantic role classification involves classifying each semantic argument identified into a specific semantic role . This paper focuses on semantic role classification task with the assumption that the semantic arguments have been identified correctly . Both feature-based and kernel-based learning methods have been studied for semantic role classification ( Carreras and Màrquez , 2004 ; Carreras and Màrquez , 2005 ) . In feature-based methods , a flat feature vector is used to represent a predicateargument structure while , in kernel-based methods , a kernel function is used to measure directly the similarity between two predicate-argument structures . As we know , kernel methods are more effective in capturing structured features . Moschitti ( 2004 ) and Che et al . ( 2006 ) used a convolution tree kernel ( Collins and Duffy , 2001 ) for semantic role classification . The convolution tree kernel takes sub-tree as its feature and counts the number of common sub-trees as the similarity between two predicate-arguments . This kernel has shown very promising results in SRL . However , as a general learning algorithm , the tree kernel only carries out hard matching between any two sub-trees without considering any linguistic knowledge in kernel design . This makes the kernel fail to handle similar phrase structures ( e.g. , " buy a car " vs. " buy a red car " ) and near-synonymic grammar tags ( e.g. , the POS variations between " high / JJ degree / NN " 1 and " higher / JJR degree / NN " ) 2 . To some degree , it may lead to over-fitting and compromise performance . This paper reports our preliminary study in addressing the above issue by introducing more linguistic knowledge into the convolution tree kernel . To our knowledge , this is the first attempt in this research direction . In detail , we propose a grammar-driven convolution tree kernel for semantic role classification that can carry out more linguistically motivated substructure matching . Experimental results show that the proposed method significantly outperforms the standard convolution tree kernel on the data set of the CoNLL-2005 SRL shared task . The remainder of the paper is organized as follows : Section 2 reviews the previous work and Section 3 discusses our grammar-driven convolution tree kernel . Section 4 shows the experimental results . We conclude our work in Section 5 . In this paper , we propose a novel grammar-driven convolution tree kernel for semantic role classification . More linguistic knowledge is considered in the new kernel design . The experimental results verify that the grammar-driven kernel is more effective in capturing syntactic structure features than the previous convolution tree kernel because it allows grammar-driven approximate matching of substructures and node features . We also discuss the criteria to determine the optional nodes in a CFG rule in defining our grammar-driven convolution tree kernel . The extension of our work is to improve the performance of the entire semantic role labeling system using the grammar-driven tree kernel , including all four stages : pruning , semantic role identification , classification and post inference . In addition , a more interesting research topic is to study how to integrate linguistic knowledge and tree kernel methods to do feature selection for tree kernelbased NLP applications ( Suzuki et al . , 2004 ) . In detail , a linguistics and statistics-based theory that can suggest the effectiveness of different substructure features and whether they should be generated or not by the tree kernels would be worked out .
| sobamchan/aclsum | 0 | full_paper |
N18-1108 | Recurrent neural networks ( RNNs ) have achieved impressive results in a variety of linguistic processing tasks , suggesting that they can induce non-trivial properties of language . We investigate here to what extent RNNs learn to track abstract hierarchical syntactic structure . We test whether RNNs trained with a generic language modeling objective in four languages ( Italian , English , Hebrew , Russian ) can predict long-distance number agreement in various constructions . We include in our evaluation nonsensical sentences where RNNs can not rely on semantic or lexical cues ( " The colorless green ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas I ate with the chair sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep furiously " ) , and , for Italian , we compare model performance to human intuitions . Our language-model-trained RNNs make reliable predictions about long-distance agreement , and do not lag much behind human performance . We thus bring support to the hypothesis that RNNs are not just shallowpattern extractors , but they also acquire deeper grammatical competence . Recurrent neural networks ( RNNs ; Elman , 1990 ) are general sequence processing devices that do not explicitly encode the hierarchical structure that is thought to be essential to natural language ( Everaert et al . , 2015 ) . Early work using artificial languages showed that they may nevertheless be able to approximate context-free languages ( Elman , 1991 ) . More recently , RNNs have achieved impressive results in large-scale tasks such as language modeling for speech recognition and machine translation , and are by now standard tools for sequential natural language tasks ( e.g. , Mikolov et al . , 2010 ; Graves , 2012 ; Wu et al . , 2016 ) . This suggests that RNNs may learn to track grammatical structure even when trained on noisier natural data . The conjecture is supported by the success of RNNs as feature extractors for syntactic parsing ( e.g. , Cross and Huang , 2016 ; Kiperwasser and Goldberg , 2016 ; Zhang et al . , 2017 ) . Linzen et al . ( 2016 ) directly evaluated the extent to which RNNs can approximate hierarchical structure in corpus-extracted natural language data . They tested whether RNNs can learn to predict English subject-verb agreement , a task thought to require hierarchical structure in the general case ( " the girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl the boys like . Their experiments confirmed that RNNs can , in principle , handle such constructions . However , in their study RNNs could only succeed when provided with explicit supervision on the target task . Linzen and colleagues argued that the unsupervised language modeling objective is not sufficient for RNNs to induce the syntactic knowledge necessary to cope with long-distance agreement . . . is is is is is is is is is is is is is is is is is The current paper reevaluates these conclusions . We strengthen the evaluation paradigm of Linzen and colleagues in several ways . Most importantly , their analysis did not rule out the possibility that RNNs might be relying on semantic or collocational / frequency-based information , rather than purely on syntactic structure . In " dogs dogs dogs dogs dogs about what typically barks ( dogs , not neighbourhoods ) , without relying on more abstract structural cues . In a follow-up study to Linzen and colleagues ' , Bernardy and Lappin ( 2017 ) observed that RNNs are better at long-distance agreement when they construct rich lexical representations of words , which suggests effects of this sort might indeed be at play . We introduce a method to probe the syntactic abilities of RNNs that abstracts away from potential lexical , semantic and frequency-based confounds . Inspired by Chomsky 's ( 1957 ) We extend the previous work in three additional ways . First , alongside English , which has few morphological cues to agreement , we examine Italian , Hebrew and Russian , which have richer morphological systems . Second , we go beyond subject-verb agreement and develop an automated method to harvest a variety of long-distance number agreement constructions from treebanks . Finally , for Italian , we collect human judgments for the tested sentences , providing an important comparison point for RNN performance . 1We focus on the more interesting unsupervised setup , where RNNs are trained to perform generic , large-scale language modeling ( LM ): they are not given explicit evidence , at training time , that they must focus on long-distance agreement , but they are rather required to track a multitude of cues that might help with word prediction in general . Our results are encouraging . RNNs trained with a LM objective solve the long-distance agreement problem well , even on nonce sentences . The pattern is consistent across languages , and , crucially , not far from human performance in Italian . Moreover , RNN performance on language modeling ( measured in terms of perplexity ) is a good predictor of long-distance agreement accuracy . This suggests that the ability to capture structural generalizations is an important aspect of what makes the best RNN architectures so good at language modeling . Since our positive results contradict , to some extent , those of Linzen et al . ( 2016 ) , we also replicate their relevant experiment using our best RNN ( an LSTM ) . We outperform their models , suggesting that a careful architecture / hyperparameter search is crucial to obtain RNNs that are not only good at language modeling , but able to extract syntactic generalizations . 2 Constructing a long-distance agreement benchmark Overview . We construct our number agreement test sets as follows . Original sentences are automatically extracted from a dependency treebank . They are then converted into nonce sentences by substituting all content words with random words with the same morphology , resulting in grammatical but nonsensical sequences . An LM is evaluated on its predictions for the target ( second ) word in the dependency , in both the original and nonce sentences . Long-distance agreement constructions . Agreement relations , such as subject-verb agreement in English , are an ideal test bed for the syntactic abilities of LMs , because the form of the second item ( the target ) is predictable from the first item ( the cue ) . Crucially , the cue and the target are linked by a structural relation , where linear order in the word sequence does not matter ( Everaert et al . , 2015 ) In all these cases , the number of the main verb " thinks " is determined by its subject ( " girl " ) , and this relation depends on the syntactic structure of the sentence , not on the linear sequence of words . As the last sentence shows , the word directly preceding the verb can even be a noun with the opposite number ( " friends " ) , but this does not influence the structurally-determined form of the verb . When the cue and the target are adjacent ( " the girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks . . . " ) , an LM can predict the target without access to syntactic structure : it can simply extract the relevant morphosyntactic features of words ( e.g. , number ) and record the co-occurrence frequencies of patterns such as N P lur V P lur ( Mikolov et al . , 2013 ) . Thus , we focus here on long-distance agreement , where an arbitrary num- ber of words can occur between the elements of the agreement relation . We limit ourselves to number agreement ( plural or singular ) , as it is the only overt agreement feature shared by all of the languages we study . Identifying candidate constructions . We started by collecting pairs of part-of-speech ( POS ) tags connected by a dependency arc . Independently of which element is the head of the relation , we refer to the first item as the cue and to the second as the target . We additionally refer to the POS sequence characterizing the entire pattern as a construction , and to the elements in the middle as context . For each candidate construction , we collected all of the contexts in the corpus that intervene between the cue and the target ( we define contexts as the sequence of POS tags of the top-level nodes in the dependency subtrees ) . For example , for the English subject-verb agreement construction shown in Fig . 1a , the context is defined by VERB ( head of the relative clause ) and ADV ( adverbial modifier of the target verb ) , which together dominate the sequence " the boys like often " . For the Russian adjective-noun agreement construction in Fig . 1b , the context is NOUN , because in the dependency grammar we use the noun " moment " is the head of the prepositional phrase " at that moment " , which modifies the adjective " deep " . The candidate agreement pair and the context form a construction , which is characterized by a sequence of POS tags , e.g. , NOUN VERB ADV VERB or VERB NOUN CCONJ VERB ( Fig . 1c ) . is is is is is is is is is is is is is is is is is " and " girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls who stayed at home were were were were were were were were were were were were were were were were were " . Conversely , standard syntactic structures might be split between different constructions , e.g. , relative clause contexts occur in both NOUN VERB VERB and NOUN VERB ADV VERB constructions ( the latter is illustrated by the English example in Fig . 1a ) . Construction contexts can contain a variable numbers of words . Since we are interested in challenging cases , we only considered cases in which at least three tokens intervened between the cue and the target . Excluding non-agreement constructions . In the next step , we excluded constructions in which the candidate cue and target did not agree in number in all of the instances of the construction in the treebank ( if both the cue and the target were morphologically annotated for number ) . This step retained English subject-verb constructions , for example , but excluded verb-object constructions , since any form of a verb can appear both with singular and plural objects . To focus on robust agreement patterns , we only kept constructions with at least 10 instances of both plural and singular agreement . When applied to the treebanks we used ( see Section 3 ) , this step resulted in between two ( English ) and 21 ( Russian ) constructions per lan-guage . English has the poorest morphology and consequently the lowest number of patterns with identifiable morphological agreement . Only the VP-conjunction construction ( Fig . 1c ) was identified in all four languages . Subject-verb agreement constructions were extracted in all languages but Russian ; Russian has relatively flexible word order and a noun dependent preceding a head verb is not necessarily its subject . The full list of extracted constructions in English and Italian is given in Tables 2 and 3 , respectively . For the other languages , see the Supplementary Material ( SM ) . 2 Original sentence test set . Our " original " sentence test set included all sentences from each construction where all words from the cue and up to and including the target occurred in the LM vocabulary ( Section 3 ) , and where the singular / plural counterpart of the target occurred in the treebank and in the language model vocabulary ( this is required by the evaluation procedure outlined below ) . The total counts of constructions and original sentences in our test sets are provided in Table 1 . The average number of context words separating the cue and the target ranged from 3.6 ( Hebrew ) to 4.5 ( Italian ) . Generating nonce sentences . We generated nine nonce variants of each original sentence as follows . Each content word ( noun , verb , adjective , proper noun , numeral , adverb ) in the sentence was substituted by another random content word from the treebank with matching POS and morphological features . To avoid forms that are ambiguous between several POS , which are particularly frequent in English ( e.g. , plural noun and singular verb forms ) , we excluded the forms that appeared with a different POS more than 10 % of the time in the treebank . Function words ( determiners , pronouns , adpositions , particles ) and punctuation were left intact . For example , we generated the nonce ( 1b ) from the original sentence ( 1a ): ( ( e.g. , " it stays the shuttle " in ( 1b ) ) . Evaluation procedure . For each sentence in our test set , we retrieved from our treebank the form that is identical to the agreement target in all morphological features except number ( e.g. , " finds " instead of " find " in ( 1b ) ) . Given a sentence with prefix p up to and excluding the target , we then compute the probabilities P ( t 1 |p ) and P ( t 2 |p ) for the singular and plural variants of the target , t 1 and t 2 , based on the language model . Following Linzen et al . ( 2016 ) , we say that the model identified the correct target if it assigned a higher probability to the form with the correct number . In ( 1b ) , for example , the model should assign a higher probability to " finds " than " find".3 3 Experimental setup Treebanks . We extracted our test sets from the Italian , English , Hebrew and Russian Universal Dependency treebanks ( UD , v2.0 , Nivre et al . , 2016 ) . The English and Hebrew treebanks were post-processed to obtain a richer morphological annotation at the word level ( see SM for details ) . We ran an extensive analysis of the abilities of RNNs trained on a generic language-modeling task to predict long-distance number agreement . Results were consistent across four languages and a number of constructions . They were above strong baselines even in the challenging case of nonsense sentences , and not far from human performance . We are not aware of other collections of human long-distance agreement judgments on nonsensical sentences , and we thus consider our publicly available data set an important contribution of our work , of interest to students of human language processing in general . The constructions we considered are quite infrequent ( according to a rough estimate based on the treebanks , the language in which they are most common is Hebrew , and even there they occur with average 0.8 % sentence frequency ) . Moreover , they vary in the contexts that separate the cue and the target . So , RNNs are not simply memorizing frequent morphosyntactic sequences ( which would already be impressive , for systems learning from raw text ) . We tentatively conclude that LMtrained RNNs can construct abstract grammatical representations of their input . This , in turn , suggests that the input itself contains enough information to trigger some form of syntactic learning in a system , such as an RNN , that does not contain an explicit prior bias in favour of syntactic structures . In future work , we would like to better understand what kind of syntactic information RNNs are encoding , and how . On the one hand , we plan to adapt methods to inspect information flow across RNN states ( e.g. , Hupkes et al . , 2017 ) . On the other , we would like to expand our empirical investigation by focusing on other long-distance phenomena , such as overt case assignment ( Blake , 2001 ) or parasitic gap licensing ( Culicover and Postal , 2001 ) . While it is more challenging to extract reliable examples of such phenomena from corpora , their study would probe more sophisticated syntactic capabilities , possibly even shedding light on the theoretical analysis of the underlying linguistic structures . Finally , it may be useful to complement the corpus-driven approach used in the current paper with constructed evaluation sentences that isolate particular syntactic phenomena , independent of their frequency in a natural corpus , as is common in psycholinguistics ( Enguehard et al . , 2017 ) . | The RNNs trained on an LM objective can solve long-distance agreement problems well even on nonsensical sentences consistently across languages indicating their deeper grammatical competence. | outcome | low_quality | adherence_first | Objective: ADHERENCE-FIRST.
Task: Write a one-sentence summary focused ONLY on the aspect: outcome.
Be strict: exclude any information not directly tied to the aspect.
If uncertain, omit rather than speculate.
Stay faithful to the document; do not add any unsupported claims.
Give the outcomes in a sentence.
---
Document:
Recurrent neural networks ( RNNs ) have achieved impressive results in a variety of linguistic processing tasks , suggesting that they can induce non-trivial properties of language . We investigate here to what extent RNNs learn to track abstract hierarchical syntactic structure . We test whether RNNs trained with a generic language modeling objective in four languages ( Italian , English , Hebrew , Russian ) can predict long-distance number agreement in various constructions . We include in our evaluation nonsensical sentences where RNNs can not rely on semantic or lexical cues ( " The colorless green ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas I ate with the chair sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep furiously " ) , and , for Italian , we compare model performance to human intuitions . Our language-model-trained RNNs make reliable predictions about long-distance agreement , and do not lag much behind human performance . We thus bring support to the hypothesis that RNNs are not just shallowpattern extractors , but they also acquire deeper grammatical competence . Recurrent neural networks ( RNNs ; Elman , 1990 ) are general sequence processing devices that do not explicitly encode the hierarchical structure that is thought to be essential to natural language ( Everaert et al . , 2015 ) . Early work using artificial languages showed that they may nevertheless be able to approximate context-free languages ( Elman , 1991 ) . More recently , RNNs have achieved impressive results in large-scale tasks such as language modeling for speech recognition and machine translation , and are by now standard tools for sequential natural language tasks ( e.g. , Mikolov et al . , 2010 ; Graves , 2012 ; Wu et al . , 2016 ) . This suggests that RNNs may learn to track grammatical structure even when trained on noisier natural data . The conjecture is supported by the success of RNNs as feature extractors for syntactic parsing ( e.g. , Cross and Huang , 2016 ; Kiperwasser and Goldberg , 2016 ; Zhang et al . , 2017 ) . Linzen et al . ( 2016 ) directly evaluated the extent to which RNNs can approximate hierarchical structure in corpus-extracted natural language data . They tested whether RNNs can learn to predict English subject-verb agreement , a task thought to require hierarchical structure in the general case ( " the girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl the boys like . Their experiments confirmed that RNNs can , in principle , handle such constructions . However , in their study RNNs could only succeed when provided with explicit supervision on the target task . Linzen and colleagues argued that the unsupervised language modeling objective is not sufficient for RNNs to induce the syntactic knowledge necessary to cope with long-distance agreement . . . is is is is is is is is is is is is is is is is is The current paper reevaluates these conclusions . We strengthen the evaluation paradigm of Linzen and colleagues in several ways . Most importantly , their analysis did not rule out the possibility that RNNs might be relying on semantic or collocational / frequency-based information , rather than purely on syntactic structure . In " dogs dogs dogs dogs dogs about what typically barks ( dogs , not neighbourhoods ) , without relying on more abstract structural cues . In a follow-up study to Linzen and colleagues ' , Bernardy and Lappin ( 2017 ) observed that RNNs are better at long-distance agreement when they construct rich lexical representations of words , which suggests effects of this sort might indeed be at play . We introduce a method to probe the syntactic abilities of RNNs that abstracts away from potential lexical , semantic and frequency-based confounds . Inspired by Chomsky 's ( 1957 ) We extend the previous work in three additional ways . First , alongside English , which has few morphological cues to agreement , we examine Italian , Hebrew and Russian , which have richer morphological systems . Second , we go beyond subject-verb agreement and develop an automated method to harvest a variety of long-distance number agreement constructions from treebanks . Finally , for Italian , we collect human judgments for the tested sentences , providing an important comparison point for RNN performance . 1We focus on the more interesting unsupervised setup , where RNNs are trained to perform generic , large-scale language modeling ( LM ): they are not given explicit evidence , at training time , that they must focus on long-distance agreement , but they are rather required to track a multitude of cues that might help with word prediction in general . Our results are encouraging . RNNs trained with a LM objective solve the long-distance agreement problem well , even on nonce sentences . The pattern is consistent across languages , and , crucially , not far from human performance in Italian . Moreover , RNN performance on language modeling ( measured in terms of perplexity ) is a good predictor of long-distance agreement accuracy . This suggests that the ability to capture structural generalizations is an important aspect of what makes the best RNN architectures so good at language modeling . Since our positive results contradict , to some extent , those of Linzen et al . ( 2016 ) , we also replicate their relevant experiment using our best RNN ( an LSTM ) . We outperform their models , suggesting that a careful architecture / hyperparameter search is crucial to obtain RNNs that are not only good at language modeling , but able to extract syntactic generalizations . 2 Constructing a long-distance agreement benchmark Overview . We construct our number agreement test sets as follows . Original sentences are automatically extracted from a dependency treebank . They are then converted into nonce sentences by substituting all content words with random words with the same morphology , resulting in grammatical but nonsensical sequences . An LM is evaluated on its predictions for the target ( second ) word in the dependency , in both the original and nonce sentences . Long-distance agreement constructions . Agreement relations , such as subject-verb agreement in English , are an ideal test bed for the syntactic abilities of LMs , because the form of the second item ( the target ) is predictable from the first item ( the cue ) . Crucially , the cue and the target are linked by a structural relation , where linear order in the word sequence does not matter ( Everaert et al . , 2015 ) In all these cases , the number of the main verb " thinks " is determined by its subject ( " girl " ) , and this relation depends on the syntactic structure of the sentence , not on the linear sequence of words . As the last sentence shows , the word directly preceding the verb can even be a noun with the opposite number ( " friends " ) , but this does not influence the structurally-determined form of the verb . When the cue and the target are adjacent ( " the girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks . . . " ) , an LM can predict the target without access to syntactic structure : it can simply extract the relevant morphosyntactic features of words ( e.g. , number ) and record the co-occurrence frequencies of patterns such as N P lur V P lur ( Mikolov et al . , 2013 ) . Thus , we focus here on long-distance agreement , where an arbitrary num- ber of words can occur between the elements of the agreement relation . We limit ourselves to number agreement ( plural or singular ) , as it is the only overt agreement feature shared by all of the languages we study . Identifying candidate constructions . We started by collecting pairs of part-of-speech ( POS ) tags connected by a dependency arc . Independently of which element is the head of the relation , we refer to the first item as the cue and to the second as the target . We additionally refer to the POS sequence characterizing the entire pattern as a construction , and to the elements in the middle as context . For each candidate construction , we collected all of the contexts in the corpus that intervene between the cue and the target ( we define contexts as the sequence of POS tags of the top-level nodes in the dependency subtrees ) . For example , for the English subject-verb agreement construction shown in Fig . 1a , the context is defined by VERB ( head of the relative clause ) and ADV ( adverbial modifier of the target verb ) , which together dominate the sequence " the boys like often " . For the Russian adjective-noun agreement construction in Fig . 1b , the context is NOUN , because in the dependency grammar we use the noun " moment " is the head of the prepositional phrase " at that moment " , which modifies the adjective " deep " . The candidate agreement pair and the context form a construction , which is characterized by a sequence of POS tags , e.g. , NOUN VERB ADV VERB or VERB NOUN CCONJ VERB ( Fig . 1c ) . is is is is is is is is is is is is is is is is is " and " girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls who stayed at home were were were were were were were were were were were were were were were were were " . Conversely , standard syntactic structures might be split between different constructions , e.g. , relative clause contexts occur in both NOUN VERB VERB and NOUN VERB ADV VERB constructions ( the latter is illustrated by the English example in Fig . 1a ) . Construction contexts can contain a variable numbers of words . Since we are interested in challenging cases , we only considered cases in which at least three tokens intervened between the cue and the target . Excluding non-agreement constructions . In the next step , we excluded constructions in which the candidate cue and target did not agree in number in all of the instances of the construction in the treebank ( if both the cue and the target were morphologically annotated for number ) . This step retained English subject-verb constructions , for example , but excluded verb-object constructions , since any form of a verb can appear both with singular and plural objects . To focus on robust agreement patterns , we only kept constructions with at least 10 instances of both plural and singular agreement . When applied to the treebanks we used ( see Section 3 ) , this step resulted in between two ( English ) and 21 ( Russian ) constructions per lan-guage . English has the poorest morphology and consequently the lowest number of patterns with identifiable morphological agreement . Only the VP-conjunction construction ( Fig . 1c ) was identified in all four languages . Subject-verb agreement constructions were extracted in all languages but Russian ; Russian has relatively flexible word order and a noun dependent preceding a head verb is not necessarily its subject . The full list of extracted constructions in English and Italian is given in Tables 2 and 3 , respectively . For the other languages , see the Supplementary Material ( SM ) . 2 Original sentence test set . Our " original " sentence test set included all sentences from each construction where all words from the cue and up to and including the target occurred in the LM vocabulary ( Section 3 ) , and where the singular / plural counterpart of the target occurred in the treebank and in the language model vocabulary ( this is required by the evaluation procedure outlined below ) . The total counts of constructions and original sentences in our test sets are provided in Table 1 . The average number of context words separating the cue and the target ranged from 3.6 ( Hebrew ) to 4.5 ( Italian ) . Generating nonce sentences . We generated nine nonce variants of each original sentence as follows . Each content word ( noun , verb , adjective , proper noun , numeral , adverb ) in the sentence was substituted by another random content word from the treebank with matching POS and morphological features . To avoid forms that are ambiguous between several POS , which are particularly frequent in English ( e.g. , plural noun and singular verb forms ) , we excluded the forms that appeared with a different POS more than 10 % of the time in the treebank . Function words ( determiners , pronouns , adpositions , particles ) and punctuation were left intact . For example , we generated the nonce ( 1b ) from the original sentence ( 1a ): ( ( e.g. , " it stays the shuttle " in ( 1b ) ) . Evaluation procedure . For each sentence in our test set , we retrieved from our treebank the form that is identical to the agreement target in all morphological features except number ( e.g. , " finds " instead of " find " in ( 1b ) ) . Given a sentence with prefix p up to and excluding the target , we then compute the probabilities P ( t 1 |p ) and P ( t 2 |p ) for the singular and plural variants of the target , t 1 and t 2 , based on the language model . Following Linzen et al . ( 2016 ) , we say that the model identified the correct target if it assigned a higher probability to the form with the correct number . In ( 1b ) , for example , the model should assign a higher probability to " finds " than " find".3 3 Experimental setup Treebanks . We extracted our test sets from the Italian , English , Hebrew and Russian Universal Dependency treebanks ( UD , v2.0 , Nivre et al . , 2016 ) . The English and Hebrew treebanks were post-processed to obtain a richer morphological annotation at the word level ( see SM for details ) . We ran an extensive analysis of the abilities of RNNs trained on a generic language-modeling task to predict long-distance number agreement . Results were consistent across four languages and a number of constructions . They were above strong baselines even in the challenging case of nonsense sentences , and not far from human performance . We are not aware of other collections of human long-distance agreement judgments on nonsensical sentences , and we thus consider our publicly available data set an important contribution of our work , of interest to students of human language processing in general . The constructions we considered are quite infrequent ( according to a rough estimate based on the treebanks , the language in which they are most common is Hebrew , and even there they occur with average 0.8 % sentence frequency ) . Moreover , they vary in the contexts that separate the cue and the target . So , RNNs are not simply memorizing frequent morphosyntactic sequences ( which would already be impressive , for systems learning from raw text ) . We tentatively conclude that LMtrained RNNs can construct abstract grammatical representations of their input . This , in turn , suggests that the input itself contains enough information to trigger some form of syntactic learning in a system , such as an RNN , that does not contain an explicit prior bias in favour of syntactic structures . In future work , we would like to better understand what kind of syntactic information RNNs are encoding , and how . On the one hand , we plan to adapt methods to inspect information flow across RNN states ( e.g. , Hupkes et al . , 2017 ) . On the other , we would like to expand our empirical investigation by focusing on other long-distance phenomena , such as overt case assignment ( Blake , 2001 ) or parasitic gap licensing ( Culicover and Postal , 2001 ) . While it is more challenging to extract reliable examples of such phenomena from corpora , their study would probe more sophisticated syntactic capabilities , possibly even shedding light on the theoretical analysis of the underlying linguistic structures . Finally , it may be useful to complement the corpus-driven approach used in the current paper with constructed evaluation sentences that isolate particular syntactic phenomena , independent of their frequency in a natural corpus , as is common in psycholinguistics ( Enguehard et al . , 2017 ) .
| sobamchan/aclsum | 1 | full_paper |
N03-1024 | We describe a syntax-based algorithm that automatically builds Finite State Automata ( word lattices ) from semantically equivalent translation sets . These FSAs are good representations of paraphrases . They can be used to extract lexical and syntactic paraphrase pairs and to generate new , unseen sentences that express the same meaning as the sentences in the input sets . Our FSAs can also predict the correctness of alternative semantic renderings , which may be used to evaluate the quality of translations . In the past , paraphrases have come under the scrutiny of many research communities . Information retrieval researchers have used paraphrasing techniques for query reformulation in order to increase the recall of information retrieval engines ( Sparck Jones and Tait , 1984 ) . Natural language generation researchers have used paraphrasing to increase the expressive power of generation systems ( Iordanskaja et al . , 1991 ; Lenke , 1994 ; Stede , 1999 ) . And researchers in multi-document text summarization ( Barzilay et al . , 1999 ) , information extraction ( Shinyama et al . , 2002 ) , and question answering ( Lin and Pantel , 2001 ; Hermjakob et al . , 2002 ) have focused on identifying and exploiting paraphrases in the context of recognizing redundancies , alternative formulations of the same meaning , and improving the performance of question answering systems . In previous work ( Barzilay and McKeown , 2001 ; Lin and Pantel , 2001 ; Shinyama et al . , 2002 ) , paraphrases are represented as sets or pairs of semantically equivalent words , phrases , and patterns . Although this is adequate in the context of some applications , it is clearly too weak from a generative perspective . Assume , for example , that we know that text pairs ( stock market rose , stock market gained ) and ( stock market rose , stock prices rose ) have the same meaning . If we memorized only these two pairs , it would be impossible to infer that , in fact , consistent with our intuition , any of the following sets of phrases are also semantically equivalent : { stock market rose , stock market gained , stock prices rose , stock prices gained } and { stock market , stock prices } in the context of rose or gained ; { market rose } , { market gained } , { prices rose } and { prices gained } in the context of stock ; and so on . In this paper , we propose solutions for two problems : the problem of paraphrase representation and the problem of paraphrase induction . We propose a new , finite-statebased representation of paraphrases that enables one to encode compactly large numbers of paraphrases . We also propose algorithms that automatically derive such representations from inputs that are now routinely released in conjunction with large scale machine translation evaluations ( DARPA , 2002 ) : multiple English translations of many foreign language texts . For instance , when given as input the 11 semantically equivalent English translations in Figure 1 , our algorithm automatically induces the FSA in Figure 2 , which represents compactly 49 distinct renderings of the same semantic meaning . Our FSAs capture both lexical paraphrases , such as { fighting , bat-tle } , { died , were killed } and structural paraphrases such as { last week 's fighting , the battle of last week } . The contexts in which these are correct paraphrases are also conveniently captured in the representation . In previous work , Langkilde and Knight ( 1998 ) used word lattices for language generation , but their method involved hand-crafted rules . Bangalore et al . ( 2001 ) and Barzilay and Lee ( 2002 ) both applied the technique of multi-sequence alignment ( MSA ) to align parallel corpora and produced similar FSAs . For their purposes , they mainly need to ensure the correctness of consensus among different translations , so that different constituent orderings in input sentences do not pose a serious prob- lem . In contrast , we want to ensure the correctness of all paths represented by the FSAs , and direct application of MSA in the presence of different constituent orderings can be problematic . For example , when given as input the same sentences in Figure 1 , one instantiation of the MSA algorithm produces the FSA in Figure 3 , which contains many " bad " paths such as the battle of last week 's fighting took at least 12 people lost their people died in the fighting last week 's fighting ( See Section 4.2.2 for a more quantitative analysis . ) . It 's still possible to use MSA if , for example , the input is pre-clustered to have the same constituent ordering ( Barzilay and Lee ( 2003 ) ) . But we chose to approach this problem from another direction . As a result , we propose a new syntax-based algorithm to produce FSAs . In this paper , we first introduce the multiple translation corpus that we use in our experiments ( see Section 2 ) . We then present the algorithms that we developed to induce finite-state paraphrase representations from such data ( see Section 3 ) . An important part of the paper is dedicated to evaluating the quality of the finite-state representations that we derive ( see Section 4 ) . Since our representations encode thousands and sometimes millions of equivalent verbalizations of the same meaning , we use both manual and automatic evaluation techniques . Some of the automatic evaluations we perform are novel as well . In this paper , we presented a new syntax-based algorithm that learns paraphrases from a newly available dataset . The multiple translation corpus that we use in this paper is the first instance in a series of similar corpora that are built and made publicly available by LDC in the context of a series of DARPA-sponsored MT evaluations . The algorithm we proposed constructs finite state representations of paraphrases that are useful in many contexts : to induce large lists of lexical and structural paraphrases ; to generate semantically equivalent renderings of a given meaning ; and to estimate the quality of machine translation systems . More experiments need to be carried out in order to assess extrinsically whether the FSAs we produce can be used to yield higher agreement scores between human and automatic assessments of translation quality . In our future work , we wish to experiment with more flexible merging algorithms and to integrate better the top-down and bottom-up processes that are used to in-duce FSAs . We also wish to extract more abstract paraphrase patterns from the current representation . Such patterns are more likely to get reused -which would help us get reliable statistics for them in the extraction phase , and also have a better chance of being applicable to unseen data . | They propose a syntax-based algorithm that builds Finite State Automata from translation sets which are good representations of paraphrases. | approach | high_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: approach.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
Summarize the method or approach the authors propose to tackle the problem. Return only the summary in one sentence.
---
Document:
We describe a syntax-based algorithm that automatically builds Finite State Automata ( word lattices ) from semantically equivalent translation sets . These FSAs are good representations of paraphrases . They can be used to extract lexical and syntactic paraphrase pairs and to generate new , unseen sentences that express the same meaning as the sentences in the input sets . Our FSAs can also predict the correctness of alternative semantic renderings , which may be used to evaluate the quality of translations . In the past , paraphrases have come under the scrutiny of many research communities . Information retrieval researchers have used paraphrasing techniques for query reformulation in order to increase the recall of information retrieval engines ( Sparck Jones and Tait , 1984 ) . Natural language generation researchers have used paraphrasing to increase the expressive power of generation systems ( Iordanskaja et al . , 1991 ; Lenke , 1994 ; Stede , 1999 ) . And researchers in multi-document text summarization ( Barzilay et al . , 1999 ) , information extraction ( Shinyama et al . , 2002 ) , and question answering ( Lin and Pantel , 2001 ; Hermjakob et al . , 2002 ) have focused on identifying and exploiting paraphrases in the context of recognizing redundancies , alternative formulations of the same meaning , and improving the performance of question answering systems . In previous work ( Barzilay and McKeown , 2001 ; Lin and Pantel , 2001 ; Shinyama et al . , 2002 ) , paraphrases are represented as sets or pairs of semantically equivalent words , phrases , and patterns . Although this is adequate in the context of some applications , it is clearly too weak from a generative perspective . Assume , for example , that we know that text pairs ( stock market rose , stock market gained ) and ( stock market rose , stock prices rose ) have the same meaning . If we memorized only these two pairs , it would be impossible to infer that , in fact , consistent with our intuition , any of the following sets of phrases are also semantically equivalent : { stock market rose , stock market gained , stock prices rose , stock prices gained } and { stock market , stock prices } in the context of rose or gained ; { market rose } , { market gained } , { prices rose } and { prices gained } in the context of stock ; and so on . In this paper , we propose solutions for two problems : the problem of paraphrase representation and the problem of paraphrase induction . We propose a new , finite-statebased representation of paraphrases that enables one to encode compactly large numbers of paraphrases . We also propose algorithms that automatically derive such representations from inputs that are now routinely released in conjunction with large scale machine translation evaluations ( DARPA , 2002 ) : multiple English translations of many foreign language texts . For instance , when given as input the 11 semantically equivalent English translations in Figure 1 , our algorithm automatically induces the FSA in Figure 2 , which represents compactly 49 distinct renderings of the same semantic meaning . Our FSAs capture both lexical paraphrases , such as { fighting , bat-tle } , { died , were killed } and structural paraphrases such as { last week 's fighting , the battle of last week } . The contexts in which these are correct paraphrases are also conveniently captured in the representation . In previous work , Langkilde and Knight ( 1998 ) used word lattices for language generation , but their method involved hand-crafted rules . Bangalore et al . ( 2001 ) and Barzilay and Lee ( 2002 ) both applied the technique of multi-sequence alignment ( MSA ) to align parallel corpora and produced similar FSAs . For their purposes , they mainly need to ensure the correctness of consensus among different translations , so that different constituent orderings in input sentences do not pose a serious prob- lem . In contrast , we want to ensure the correctness of all paths represented by the FSAs , and direct application of MSA in the presence of different constituent orderings can be problematic . For example , when given as input the same sentences in Figure 1 , one instantiation of the MSA algorithm produces the FSA in Figure 3 , which contains many " bad " paths such as the battle of last week 's fighting took at least 12 people lost their people died in the fighting last week 's fighting ( See Section 4.2.2 for a more quantitative analysis . ) . It 's still possible to use MSA if , for example , the input is pre-clustered to have the same constituent ordering ( Barzilay and Lee ( 2003 ) ) . But we chose to approach this problem from another direction . As a result , we propose a new syntax-based algorithm to produce FSAs . In this paper , we first introduce the multiple translation corpus that we use in our experiments ( see Section 2 ) . We then present the algorithms that we developed to induce finite-state paraphrase representations from such data ( see Section 3 ) . An important part of the paper is dedicated to evaluating the quality of the finite-state representations that we derive ( see Section 4 ) . Since our representations encode thousands and sometimes millions of equivalent verbalizations of the same meaning , we use both manual and automatic evaluation techniques . Some of the automatic evaluations we perform are novel as well . In this paper , we presented a new syntax-based algorithm that learns paraphrases from a newly available dataset . The multiple translation corpus that we use in this paper is the first instance in a series of similar corpora that are built and made publicly available by LDC in the context of a series of DARPA-sponsored MT evaluations . The algorithm we proposed constructs finite state representations of paraphrases that are useful in many contexts : to induce large lists of lexical and structural paraphrases ; to generate semantically equivalent renderings of a given meaning ; and to estimate the quality of machine translation systems . More experiments need to be carried out in order to assess extrinsically whether the FSAs we produce can be used to yield higher agreement scores between human and automatic assessments of translation quality . In our future work , we wish to experiment with more flexible merging algorithms and to integrate better the top-down and bottom-up processes that are used to in-duce FSAs . We also wish to extract more abstract paraphrase patterns from the current representation . Such patterns are more likely to get reused -which would help us get reliable statistics for them in the extraction phase , and also have a better chance of being applicable to unseen data .
| sobamchan/aclsum | 0 | full_paper |
P16-1067 | This paper proposes an unsupervised approach for segmenting a multiauthor document into authorial components . The key novelty is that we utilize the sequential patterns hidden among document elements when determining their authorships . For this purpose , we adopt Hidden Markov Model ( HMM ) and construct a sequential probabilistic model to capture the dependencies of sequential sentences and their authorships . An unsupervised learning method is developed to initialize the HMM parameters . Experimental results on benchmark datasets have demonstrated the significant benefit of our idea and our approach has outperformed the state-of-the-arts on all tests . As an example of its applications , the proposed approach is applied for attributing authorship of a document and has also shown promising results . Authorship analysis is a process of inspecting documents in order to extract authorial information about these documents . It is considered as a general concept that embraces several types of authorship subjects , including authorship verif ication , plagiarism detection and author attribution . Authorship verification ( Brocardo et al . , 2013 ; Potha and Stamatatos , 2014 ) decides whether a given document is written by a specific author . Plagiarism detection ( Stein et al . , 2011 ; Kestemont et al . , 2011 ) seeks to expose the similarity between two texts . However , it is un-able to determine if they are written by the same author . In author attribution ( Juola , 2006 ; Savoy , 2015 ) , a real author of an anonymous document is predicted using labeled documents of a set of candidate authors . Another significant subject in authorship analysis , which has received comparatively less attention from research community , is authorship-based document decomposition ( ABDD ) . This subject is to group the sentences of a multi-author document to different classes , of which each contains the sentences written by only one author . Many applications can take advantage of such a subject , especially those in forensic investigation , which aim to determine the authorship of sentences in a multi-author document . Furthermore , this kind of subject is beneficial for detecting plagiarism in a document and defining contributions of authors in a multi-author document for commercial purpose . ABDD can also be applied to identify which source ( regarded as an ' author ' in this paper ) a part of a document is copied from when the document is formed by taking contents from various sources . In despite of the benefits of ABDD , there has been little research reported on this subject . Koppel et al . ( 2011 ) are the first researchers who implemented an unsupervised approach for ABDD . However , their approach is restricted to Hebrew documents only . The authors of Akiva and Koppel ( 2013 ) addressed the drawbacks of the above approach by proposing a generic unsupervised approach for ABDD . Their approach utilized distance measurements to increase the precision and accuracy of clustering and classification phases , respectively . The accuracy of their approach is highly dependent on the number of au-thors . When the number of authors increases , the accuracy of the approach is significantly dropped . Giannella ( 2015 ) presented an improved approach for ABDD when the number of authors of the document is known or unknown . In his approach , a Bayesian segmentation algorithm is applied , which is followed by a segment clustering algorithm . However , the author tested his approach by using only documents with a few transitions among authors . Furthermore , the accuracy of the approach is very sensitive to the setting of its parameters . In Aldebei et al . ( 2015 ) , the authors presented an unsupervised approach ABDD by exploiting the differences in the posterior probabilities of a Naive-Bayesian model in order to increase the precision and the classification accuracy , and to be less dependent on the number of authors in comparing with the approach in Akiva and Koppel ( 2013 ) . Their work was tested on documents with up to 400 transitions among authors and the accuracy of their approach was not sensitive to the setting of parameters , in contrast with the approach in Giannella ( 2015 ) . However , the performance of their approach greatly depends on a threshold , of which the optimal value for an individual document is not easy to find . Some other works have focused on segmenting a document into components according to their topics . For applications where the topics of documents are unavailable , these topicbased solutions will fail . In this paper , the ABDD approach is independent of documents ' topics . All of the existing works have assumed that the observations ( i.e. , sentences ) are independent and identically distributed ( i.i.d . ) . No consideration has been given to the contextual information between the observations . However , in some cases , the i.i.d . assumption is deemed as a poor one ( Rogovschi et al . , 2010 ) . In this paper , we will relax this assumption and consider sentences of a document as a sequence of observations . We make use of the contextual information hidden between sentences in order to identify the authorship of each sentence in a document . In other words , the authorships of the " previous " and " subsequent " sentences have relationships with the authorship of the current sentence . There-fore , in this paper , a well-known sequential model , Hidden Markov Model ( HMM ) , is used for modelling the sequential patterns of the document in order to describe the authorship relationships . The contributions of this article are summarized as follows . 1 . We capture the dependencies between consecutive elements in a document to identify different authorial components and construct an HMM for classification . It is for the first time the sequential patterns hidden among document elements is considered for such a problem . 2 . To build and learn the HMM model , an unsupervised learning method is first proposed to estimate its initial parameters , and it does not require any information of authors or document 's context other than how many authors have contributed to write the document . 3 . Different from the approach in Aldebei et al . ( 2015 ) , the proposed unsupervised approach no longer relies on any predetermined threshold for ABDD . 4 . Comprehensive experiments are conducted to demonstrate the superior performance of our ideas on both widely-used artificial benchmark datasets and an authentic scientific document . As an example of its applications , the proposed approach is also applied for attributing authorship on a popular dataset . The proposed approach can not only correctly determine the author of a disputed document but also provide a way for measuring the confidence level of the authorship decision for the first time . The rest of this article is organised as follows . Section 2 reviews the HMM . Section 3 presents the details of our proposed approach , including the processes for initialization and learning of HMM parameters , and the Viterbi decoding process for classification . Experiments are conducted in Section 4 , followed by the conclusion in Section 5 . We have developed an unsupervised approach for decomposing a multi-author document based on authorship . Different from the stateof-the-art approaches , we have innovatively made use of the sequential information hidden among document elements . For this purpose , we have used HMM and constructed a sequential probabilistic model , which is used to find the best sequence of authors that represents the sentences of the document . An unsupervised learning method has also been developed to estimate the initial parameter values of HMM . Comparative experiments conducted on benchmark datasets have demonstrated the effectiveness of our ideas with superior perfor-mance achieved on both artificial and authentic documents . An application of the proposed approach on authorship attribution has also achieved perfect results of 100 % accuracies together with confidence measurement for the first time . | There is no method for multiauthor segmentation of a document into author components which can be applied to authorship verification, plagiarism detection and author attribution. | challenge | low_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: challenge.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
Give the problem in short form.
---
Document:
This paper proposes an unsupervised approach for segmenting a multiauthor document into authorial components . The key novelty is that we utilize the sequential patterns hidden among document elements when determining their authorships . For this purpose , we adopt Hidden Markov Model ( HMM ) and construct a sequential probabilistic model to capture the dependencies of sequential sentences and their authorships . An unsupervised learning method is developed to initialize the HMM parameters . Experimental results on benchmark datasets have demonstrated the significant benefit of our idea and our approach has outperformed the state-of-the-arts on all tests . As an example of its applications , the proposed approach is applied for attributing authorship of a document and has also shown promising results . Authorship analysis is a process of inspecting documents in order to extract authorial information about these documents . It is considered as a general concept that embraces several types of authorship subjects , including authorship verif ication , plagiarism detection and author attribution . Authorship verification ( Brocardo et al . , 2013 ; Potha and Stamatatos , 2014 ) decides whether a given document is written by a specific author . Plagiarism detection ( Stein et al . , 2011 ; Kestemont et al . , 2011 ) seeks to expose the similarity between two texts . However , it is un-able to determine if they are written by the same author . In author attribution ( Juola , 2006 ; Savoy , 2015 ) , a real author of an anonymous document is predicted using labeled documents of a set of candidate authors . Another significant subject in authorship analysis , which has received comparatively less attention from research community , is authorship-based document decomposition ( ABDD ) . This subject is to group the sentences of a multi-author document to different classes , of which each contains the sentences written by only one author . Many applications can take advantage of such a subject , especially those in forensic investigation , which aim to determine the authorship of sentences in a multi-author document . Furthermore , this kind of subject is beneficial for detecting plagiarism in a document and defining contributions of authors in a multi-author document for commercial purpose . ABDD can also be applied to identify which source ( regarded as an ' author ' in this paper ) a part of a document is copied from when the document is formed by taking contents from various sources . In despite of the benefits of ABDD , there has been little research reported on this subject . Koppel et al . ( 2011 ) are the first researchers who implemented an unsupervised approach for ABDD . However , their approach is restricted to Hebrew documents only . The authors of Akiva and Koppel ( 2013 ) addressed the drawbacks of the above approach by proposing a generic unsupervised approach for ABDD . Their approach utilized distance measurements to increase the precision and accuracy of clustering and classification phases , respectively . The accuracy of their approach is highly dependent on the number of au-thors . When the number of authors increases , the accuracy of the approach is significantly dropped . Giannella ( 2015 ) presented an improved approach for ABDD when the number of authors of the document is known or unknown . In his approach , a Bayesian segmentation algorithm is applied , which is followed by a segment clustering algorithm . However , the author tested his approach by using only documents with a few transitions among authors . Furthermore , the accuracy of the approach is very sensitive to the setting of its parameters . In Aldebei et al . ( 2015 ) , the authors presented an unsupervised approach ABDD by exploiting the differences in the posterior probabilities of a Naive-Bayesian model in order to increase the precision and the classification accuracy , and to be less dependent on the number of authors in comparing with the approach in Akiva and Koppel ( 2013 ) . Their work was tested on documents with up to 400 transitions among authors and the accuracy of their approach was not sensitive to the setting of parameters , in contrast with the approach in Giannella ( 2015 ) . However , the performance of their approach greatly depends on a threshold , of which the optimal value for an individual document is not easy to find . Some other works have focused on segmenting a document into components according to their topics . For applications where the topics of documents are unavailable , these topicbased solutions will fail . In this paper , the ABDD approach is independent of documents ' topics . All of the existing works have assumed that the observations ( i.e. , sentences ) are independent and identically distributed ( i.i.d . ) . No consideration has been given to the contextual information between the observations . However , in some cases , the i.i.d . assumption is deemed as a poor one ( Rogovschi et al . , 2010 ) . In this paper , we will relax this assumption and consider sentences of a document as a sequence of observations . We make use of the contextual information hidden between sentences in order to identify the authorship of each sentence in a document . In other words , the authorships of the " previous " and " subsequent " sentences have relationships with the authorship of the current sentence . There-fore , in this paper , a well-known sequential model , Hidden Markov Model ( HMM ) , is used for modelling the sequential patterns of the document in order to describe the authorship relationships . The contributions of this article are summarized as follows . 1 . We capture the dependencies between consecutive elements in a document to identify different authorial components and construct an HMM for classification . It is for the first time the sequential patterns hidden among document elements is considered for such a problem . 2 . To build and learn the HMM model , an unsupervised learning method is first proposed to estimate its initial parameters , and it does not require any information of authors or document 's context other than how many authors have contributed to write the document . 3 . Different from the approach in Aldebei et al . ( 2015 ) , the proposed unsupervised approach no longer relies on any predetermined threshold for ABDD . 4 . Comprehensive experiments are conducted to demonstrate the superior performance of our ideas on both widely-used artificial benchmark datasets and an authentic scientific document . As an example of its applications , the proposed approach is also applied for attributing authorship on a popular dataset . The proposed approach can not only correctly determine the author of a disputed document but also provide a way for measuring the confidence level of the authorship decision for the first time . The rest of this article is organised as follows . Section 2 reviews the HMM . Section 3 presents the details of our proposed approach , including the processes for initialization and learning of HMM parameters , and the Viterbi decoding process for classification . Experiments are conducted in Section 4 , followed by the conclusion in Section 5 . We have developed an unsupervised approach for decomposing a multi-author document based on authorship . Different from the stateof-the-art approaches , we have innovatively made use of the sequential information hidden among document elements . For this purpose , we have used HMM and constructed a sequential probabilistic model , which is used to find the best sequence of authors that represents the sentences of the document . An unsupervised learning method has also been developed to estimate the initial parameter values of HMM . Comparative experiments conducted on benchmark datasets have demonstrated the effectiveness of our ideas with superior perfor-mance achieved on both artificial and authentic documents . An application of the proposed approach on authorship attribution has also achieved perfect results of 100 % accuracies together with confidence measurement for the first time .
| sobamchan/aclsum | 1 | full_paper |
E06-1014 | Probabilistic Latent Semantic Analysis ( PLSA ) models have been shown to provide a better model for capturing polysemy and synonymy than Latent Semantic Analysis ( LSA ) . However , the parameters of a PLSA model are trained using the Expectation Maximization ( EM ) algorithm , and as a result , the trained model is dependent on the initialization values so that performance can be highly variable . In this paper we present a method for using LSA analysis to initialize a PLSA model . We also investigated the performance of our method for the tasks of text segmentation and retrieval on personal-size corpora , and present results demonstrating the efficacy of our proposed approach . In modeling a collection of documents for information access applications , the documents are often represented as a " bag of words " , i.e. , as term vectors composed of the terms and corresponding counts for each document . The term vectors for a document collection can be organized into a term by document co-occurrence matrix . When directly using these representations , synonyms and polysemous terms , that is , terms with multiple senses or meanings , are not handled well . Methods for smoothing the term distributions through the use of latent classes have been shown to improve the performance of a number of information access tasks , including retrieval over smaller collections ( Deerwester et al . , 1990 ) , text segmentation ( Brants et al . , 2002 ) , and text classification ( Wu and Gunopulos , 2002 ) . The Probabilistic Latent Semantic Analysis model ( PLSA ) ( Hofmann , 1999 ) provides a probabilistic framework that attempts to capture polysemy and synonymy in text for applications such as retrieval and segmentation . It uses a mixture decomposition to model the co-occurrence data , and the probabilities of words and documents are obtained by a convex combination of the aspects . The mixture approximation has a well defined probability distribution and the factors have a clear probabilistic meaning in terms of the mixture component distributions . The PLSA model computes the relevant probability distributions by selecting the model parameter values that maximize the probability of the observed data , i.e. , the likelihood function . The standard method for maximum likelihood estimation is the Expectation Maximization ( EM ) algorithm . For a given initialization , the likelihood function increases with EM iterations until a local maximum is reached , rather than a global maximum , so that the quality of the solution depends on the initialization of the model . Additionally , the likelihood values across different initializations are not comparable , as we will show . Thus , the likelihood function computed over the training data can not be used as a predictor of model performance across different models . Rather than trying to predict the best performing model from a set of models , in this paper we focus on finding a good way to initialize the PLSA model . We will present a framework for using Latent Semantic Analysis ( LSA ) ( Deerwester et al . , 1990 ) to better initialize the parameters of a corresponding PLSA model . The EM algorithm is then used to further refine the initial estimate . This combination of LSA and PLSA leverages the advantages of both . This paper is organized as follows : in section 2 , we review related work in the area . In section 3 , we summarize related work on LSA and its probabilistic interpretation . In section 4 we review the PLSA model and in section 5 we present our method for initializing a PLSA model using LSA model parameters . In section 6 , we evaluate the performance of our framework on a text segmentation task and several smaller information retrieval tasks . And in section 7 , we summarize our results and give directions for future work . We have presented LSA-PLSA , an approach for improving the performance of PLSA by leveraging the best features of PLSA and LSA . Our approach uses LSA to initialize a PLSA model , allowing for arbitrary weighting schemes to be incorporated into a PLSA model while leveraging the optimization used to improve the estimate of the PLSA parameters . We have evaluated the proposed framework on two tasks : personalsize information retrieval and text segmentation . The LSA-PLSA model outperformed PLSA on all tasks . And in all cases , combining PLSA-based models outperformed a single model . The best performance was obtained with combined models when one of the models was the LSA-PLSA model . When combining multiple PLSA models , the use of LSA-PLSA in combination with either two PLSA models or one PLSA and one LSA model improved performance while reducing the running time over the combination of four or more PLSA models as used by others . Future areas of investigation include quantifying the expected performance of the LSAinitialized PLSA model by comparing performance to that of the empirically best performing model and examining whether tempered EM could further improve performance . | They propose to use Latent Semantic Analysis to initialize probabilistic latent semantic analysis models, EM algorithm is further used to refine the initial estimate. | approach | high_quality | adherence_first | Objective: ADHERENCE-FIRST.
Task: Write a one-sentence summary focused ONLY on the aspect: approach.
Be strict: exclude any information not directly tied to the aspect.
If uncertain, omit rather than speculate.
Stay faithful to the document; do not add any unsupported claims.
How do the authors address the problem? Provide a brief summary of their approach. Return only the summary in one sentence.
---
Document:
Probabilistic Latent Semantic Analysis ( PLSA ) models have been shown to provide a better model for capturing polysemy and synonymy than Latent Semantic Analysis ( LSA ) . However , the parameters of a PLSA model are trained using the Expectation Maximization ( EM ) algorithm , and as a result , the trained model is dependent on the initialization values so that performance can be highly variable . In this paper we present a method for using LSA analysis to initialize a PLSA model . We also investigated the performance of our method for the tasks of text segmentation and retrieval on personal-size corpora , and present results demonstrating the efficacy of our proposed approach . In modeling a collection of documents for information access applications , the documents are often represented as a " bag of words " , i.e. , as term vectors composed of the terms and corresponding counts for each document . The term vectors for a document collection can be organized into a term by document co-occurrence matrix . When directly using these representations , synonyms and polysemous terms , that is , terms with multiple senses or meanings , are not handled well . Methods for smoothing the term distributions through the use of latent classes have been shown to improve the performance of a number of information access tasks , including retrieval over smaller collections ( Deerwester et al . , 1990 ) , text segmentation ( Brants et al . , 2002 ) , and text classification ( Wu and Gunopulos , 2002 ) . The Probabilistic Latent Semantic Analysis model ( PLSA ) ( Hofmann , 1999 ) provides a probabilistic framework that attempts to capture polysemy and synonymy in text for applications such as retrieval and segmentation . It uses a mixture decomposition to model the co-occurrence data , and the probabilities of words and documents are obtained by a convex combination of the aspects . The mixture approximation has a well defined probability distribution and the factors have a clear probabilistic meaning in terms of the mixture component distributions . The PLSA model computes the relevant probability distributions by selecting the model parameter values that maximize the probability of the observed data , i.e. , the likelihood function . The standard method for maximum likelihood estimation is the Expectation Maximization ( EM ) algorithm . For a given initialization , the likelihood function increases with EM iterations until a local maximum is reached , rather than a global maximum , so that the quality of the solution depends on the initialization of the model . Additionally , the likelihood values across different initializations are not comparable , as we will show . Thus , the likelihood function computed over the training data can not be used as a predictor of model performance across different models . Rather than trying to predict the best performing model from a set of models , in this paper we focus on finding a good way to initialize the PLSA model . We will present a framework for using Latent Semantic Analysis ( LSA ) ( Deerwester et al . , 1990 ) to better initialize the parameters of a corresponding PLSA model . The EM algorithm is then used to further refine the initial estimate . This combination of LSA and PLSA leverages the advantages of both . This paper is organized as follows : in section 2 , we review related work in the area . In section 3 , we summarize related work on LSA and its probabilistic interpretation . In section 4 we review the PLSA model and in section 5 we present our method for initializing a PLSA model using LSA model parameters . In section 6 , we evaluate the performance of our framework on a text segmentation task and several smaller information retrieval tasks . And in section 7 , we summarize our results and give directions for future work . We have presented LSA-PLSA , an approach for improving the performance of PLSA by leveraging the best features of PLSA and LSA . Our approach uses LSA to initialize a PLSA model , allowing for arbitrary weighting schemes to be incorporated into a PLSA model while leveraging the optimization used to improve the estimate of the PLSA parameters . We have evaluated the proposed framework on two tasks : personalsize information retrieval and text segmentation . The LSA-PLSA model outperformed PLSA on all tasks . And in all cases , combining PLSA-based models outperformed a single model . The best performance was obtained with combined models when one of the models was the LSA-PLSA model . When combining multiple PLSA models , the use of LSA-PLSA in combination with either two PLSA models or one PLSA and one LSA model improved performance while reducing the running time over the combination of four or more PLSA models as used by others . Future areas of investigation include quantifying the expected performance of the LSAinitialized PLSA model by comparing performance to that of the empirically best performing model and examining whether tempered EM could further improve performance .
| sobamchan/aclsum | 1 | full_paper |
2020.emnlp-main.505 | News headline generation aims to produce a short sentence to attract readers to read the news . One news article often contains multiple keyphrases that are of interest to different users , which can naturally have multiple reasonable headlines . However , most existing methods focus on the single headline generation . In this paper , we propose generating multiple headlines with keyphrases of user interests , whose main idea is to generate multiple keyphrases of interest to users for the news first , and then generate multiple keyphrase-relevant headlines . We propose a multi-source Transformer decoder , which takes three sources as inputs : ( a ) keyphrase , ( b ) keyphrase-filtered article , and ( c ) original article to generate keyphrase-relevant , highquality , and diverse headlines . Furthermore , we propose a simple and effective method to mine the keyphrases of interest in the news article and build a first large-scale keyphraseaware news headline corpus , which contains over 180 K aligned triples of news article , headline , keyphrase . Extensive experimental comparisons on the real-world dataset show that the proposed method achieves state-of-theart results in terms of quality and diversity 1 . News Headline Generation is an under-explored subtask of text summarization ( See et al . , 2017 ; Gehrmann et al . , 2018 ; Zhong et al . , 2019 ) . Unlike text summaries that contain multiple contextrelated sentences to cover the main ideas of a document , news headlines often contain a single short sentence to encourage users to read the news . Since one news article typically contains multiple keyphrases or topics of interest to different users , it is useful to generate multiple headlines covering different keyphrases for the news article . Multiheadline generation aims to generate multiple independent headlines , which allows us to recommend news with different news headlines based on the interests of users . Besides , multi-headline generation can provide multiple hints for human news editors to assist them in writing news headlines . However , most existing methods ( Takase et al . , 2016 ; Ayana et al . , 2016 ; Murao et al . , 2019 ; Colmenares et al . , 2019 ; Zhang et al . , 2018 ) focus on single-headline generation . The headline generation process is treated as an one-to-one mapping ( the input is an article and the output is a headline ) , which trains and tests the models without any additional guiding information or constraints . We argue that this may lead to two problems . Firstly , since it is reasonable to generate multiple headlines for the news , training to generate the single ground-truth might result in a lack of more detailed guidance . Even worse , a single ground-truth without any constraint or guidance is often not enough to measure the quality of the generated headline for model testing . For example , even if a generated headline is considered reasonable by humans , it can get a low score in ROUGE ( Lin , 2004 ) , because it might focus on the keyphrases or aspects that are not consistent with the ground-truth . In this paper , we incorporate the keyphrase information into the headline generation as additional guidance . Unlike one-to-one mapping employed in previous works , we treat the headline generation process as a two-to-one mapping , where the inputs are news articles and keyphrases , and the output is a headline . We propose a keyphrase-aware news multi-headline generation method , which contains two modules : ( a ) Keyphrase Generation Model , which aims to generate multiple keyphrases of interest to users for the news article . ( b ) Keyphrase-Aware Multi-Headline Generation Model , which takes the news article and a keyphrase as input and generates a keyphrase-relevant news headline . For training models , we build a first large-scale news keyphrase-aware headline corpus that contains 180 K aligned triples of news article , headline , keyphrase . As in years past , a lot of the food trends of the year were based on creating perfectly photogenic dishes . An aesthetically pleasing dish , however , does n't mean it will stand the test of time . In fact , it 's not uncommon for food trends to be all the hype one year and die out the next . From broccoli coffee to " bowl food , " here are 10 food trends that you likely wo n't see in 2019 . ... [ 15 sentences with 307 words are abbreviated from here . ] In 2018 , restaurants all over the US decided it was a good idea to place gold foil on everything from ice cream to chicken wings to pizza resulting in an expensive food trend . For example , the Ainsworth in New York City sells $ 1,000 worth of gold covered chicken wings . It seems everyone can agree that this is a food trend that might soon disappear . In this paper , we demonstrate how to enable news headline generation systems to be aware of keyphrases such that the model can generate diverse news headlines in a controlled manner . We also build a first large-scale keyphrase-aware news headline corpus , which is based on mining the keyphrases of users ' interests in news articles with user queries . Moreover , we propose a keyphraseaware news multi-headline generation model that contains a multi-source Transformer decoder with three variants of attention-based fusing mechanisms . Extensive experiments on the real-world dataset show that our approach can generate highquality , keyphrase-relevant , and diverse news headlines , which outperforms many strong baselines . | They propose a multi-source transformer decoder and train it using a new large-scale keyphrase-aware news headline corpus built from a search engine. | approach | low_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: approach.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
Summarize the method.
---
Document:
News headline generation aims to produce a short sentence to attract readers to read the news . One news article often contains multiple keyphrases that are of interest to different users , which can naturally have multiple reasonable headlines . However , most existing methods focus on the single headline generation . In this paper , we propose generating multiple headlines with keyphrases of user interests , whose main idea is to generate multiple keyphrases of interest to users for the news first , and then generate multiple keyphrase-relevant headlines . We propose a multi-source Transformer decoder , which takes three sources as inputs : ( a ) keyphrase , ( b ) keyphrase-filtered article , and ( c ) original article to generate keyphrase-relevant , highquality , and diverse headlines . Furthermore , we propose a simple and effective method to mine the keyphrases of interest in the news article and build a first large-scale keyphraseaware news headline corpus , which contains over 180 K aligned triples of news article , headline , keyphrase . Extensive experimental comparisons on the real-world dataset show that the proposed method achieves state-of-theart results in terms of quality and diversity 1 . News Headline Generation is an under-explored subtask of text summarization ( See et al . , 2017 ; Gehrmann et al . , 2018 ; Zhong et al . , 2019 ) . Unlike text summaries that contain multiple contextrelated sentences to cover the main ideas of a document , news headlines often contain a single short sentence to encourage users to read the news . Since one news article typically contains multiple keyphrases or topics of interest to different users , it is useful to generate multiple headlines covering different keyphrases for the news article . Multiheadline generation aims to generate multiple independent headlines , which allows us to recommend news with different news headlines based on the interests of users . Besides , multi-headline generation can provide multiple hints for human news editors to assist them in writing news headlines . However , most existing methods ( Takase et al . , 2016 ; Ayana et al . , 2016 ; Murao et al . , 2019 ; Colmenares et al . , 2019 ; Zhang et al . , 2018 ) focus on single-headline generation . The headline generation process is treated as an one-to-one mapping ( the input is an article and the output is a headline ) , which trains and tests the models without any additional guiding information or constraints . We argue that this may lead to two problems . Firstly , since it is reasonable to generate multiple headlines for the news , training to generate the single ground-truth might result in a lack of more detailed guidance . Even worse , a single ground-truth without any constraint or guidance is often not enough to measure the quality of the generated headline for model testing . For example , even if a generated headline is considered reasonable by humans , it can get a low score in ROUGE ( Lin , 2004 ) , because it might focus on the keyphrases or aspects that are not consistent with the ground-truth . In this paper , we incorporate the keyphrase information into the headline generation as additional guidance . Unlike one-to-one mapping employed in previous works , we treat the headline generation process as a two-to-one mapping , where the inputs are news articles and keyphrases , and the output is a headline . We propose a keyphrase-aware news multi-headline generation method , which contains two modules : ( a ) Keyphrase Generation Model , which aims to generate multiple keyphrases of interest to users for the news article . ( b ) Keyphrase-Aware Multi-Headline Generation Model , which takes the news article and a keyphrase as input and generates a keyphrase-relevant news headline . For training models , we build a first large-scale news keyphrase-aware headline corpus that contains 180 K aligned triples of news article , headline , keyphrase . As in years past , a lot of the food trends of the year were based on creating perfectly photogenic dishes . An aesthetically pleasing dish , however , does n't mean it will stand the test of time . In fact , it 's not uncommon for food trends to be all the hype one year and die out the next . From broccoli coffee to " bowl food , " here are 10 food trends that you likely wo n't see in 2019 . ... [ 15 sentences with 307 words are abbreviated from here . ] In 2018 , restaurants all over the US decided it was a good idea to place gold foil on everything from ice cream to chicken wings to pizza resulting in an expensive food trend . For example , the Ainsworth in New York City sells $ 1,000 worth of gold covered chicken wings . It seems everyone can agree that this is a food trend that might soon disappear . In this paper , we demonstrate how to enable news headline generation systems to be aware of keyphrases such that the model can generate diverse news headlines in a controlled manner . We also build a first large-scale keyphrase-aware news headline corpus , which is based on mining the keyphrases of users ' interests in news articles with user queries . Moreover , we propose a keyphraseaware news multi-headline generation model that contains a multi-source Transformer decoder with three variants of attention-based fusing mechanisms . Extensive experiments on the real-world dataset show that our approach can generate highquality , keyphrase-relevant , and diverse news headlines , which outperforms many strong baselines .
| sobamchan/aclsum | 0 | full_paper |
D18-1065 | In this paper we show that a simple beam approximation of the joint distribution between attention and output is an easy , accurate , and efficient attention mechanism for sequence to sequence learning . The method combines the advantage of sharp focus in hard attention and the implementation ease of soft attention . On five translation and two morphological inflection tasks we show effortless and consistent gains in BLEU compared to existing attention mechanisms . In structured input-output models as used in tasks like translation and image captioning , the attention variable decides which part of the input aligns to the current output . Many attention mechanisms have been proposed ( Xu et al . , 2015 ; Bahdanau et al . , 2014 ; Luong et al . , 2015 ; Martins and Astudillo , 2016 ) but the de facto standard is a soft attention mechanism that first assigns attention weights to input encoder states , then computes an attention weighted ' soft ' aligned input state , which finally derives the output distribution . This method is end to end differentiable and easy to implement . Another less popular variant is hard attention that aligns each output to exactly one input state but requires intricate training to teach the network to choose that state . When successfully trained , hard attention is often found to be more accurate ( Xu et al . , 2015 ; Zaremba and Sutskever , 2015 ) . In NLP , a recent success has been in a monotonic hard attention setting in morphological inflection tasks ( Yu et al . , 2016 ; Aharoni and Goldberg , 2017 ) . For general seq2seq learning , methods like Sparse-Max ( Martins and Astudillo , 2016 ) and local attention ( Luong et al . , 2015 ) were proposed to bridge the gap between soft and hard attention . * Both authors contributed equally to this work In this paper we propose a surprisingly simpler alternative based on the original joint distribution between output and attention , of which existing soft and hard attention mechanisms are approximations . The joint model couples input states individually to the output like in hard attention , but it combines the advantage of end-to-end trainability of soft attention . When the number of input states is large , we propose to use a simple approximation of the full joint distribution called Beam-joint . This approximation is also easily trainable and does not suffer from the high variance of Monte-Carlo sampling gradients of hard attention . We evaluated our model on five translation tasks and increased BLEU by 0.8 to 1.7 over soft attention , which in turn was better than hard and the recent Sparsemax ( Martins and Astudillo , 2016 ) attention . More importantly , the training process was as easy as soft attention . For further support , we also evaluate on two morphological inflection tasks and got gains over soft and hard attention . In this paper we showed a simple yet effective approximation of the joint attention-output distribution in sequence to sequence learning . Our joint model consistently provides higher accuracy without significant running time overheads in five translation and two morphological inflection tasks . An interesting direction for future work is to extend beam-joint to multi-head attention architectures as in ( Vaswani et al . , 2017 ; Xu Chen , 2018 ) . | The proposed approach outperforms soft attention models and recent hard attention and Sparsemax models on five translation tasks and also on morphological inflection tasks. | outcome | high_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: outcome.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
Summarize the main results or outcomes reported by the authors. Return only the summary in one sentence.
---
Document:
In this paper we show that a simple beam approximation of the joint distribution between attention and output is an easy , accurate , and efficient attention mechanism for sequence to sequence learning . The method combines the advantage of sharp focus in hard attention and the implementation ease of soft attention . On five translation and two morphological inflection tasks we show effortless and consistent gains in BLEU compared to existing attention mechanisms . In structured input-output models as used in tasks like translation and image captioning , the attention variable decides which part of the input aligns to the current output . Many attention mechanisms have been proposed ( Xu et al . , 2015 ; Bahdanau et al . , 2014 ; Luong et al . , 2015 ; Martins and Astudillo , 2016 ) but the de facto standard is a soft attention mechanism that first assigns attention weights to input encoder states , then computes an attention weighted ' soft ' aligned input state , which finally derives the output distribution . This method is end to end differentiable and easy to implement . Another less popular variant is hard attention that aligns each output to exactly one input state but requires intricate training to teach the network to choose that state . When successfully trained , hard attention is often found to be more accurate ( Xu et al . , 2015 ; Zaremba and Sutskever , 2015 ) . In NLP , a recent success has been in a monotonic hard attention setting in morphological inflection tasks ( Yu et al . , 2016 ; Aharoni and Goldberg , 2017 ) . For general seq2seq learning , methods like Sparse-Max ( Martins and Astudillo , 2016 ) and local attention ( Luong et al . , 2015 ) were proposed to bridge the gap between soft and hard attention . * Both authors contributed equally to this work In this paper we propose a surprisingly simpler alternative based on the original joint distribution between output and attention , of which existing soft and hard attention mechanisms are approximations . The joint model couples input states individually to the output like in hard attention , but it combines the advantage of end-to-end trainability of soft attention . When the number of input states is large , we propose to use a simple approximation of the full joint distribution called Beam-joint . This approximation is also easily trainable and does not suffer from the high variance of Monte-Carlo sampling gradients of hard attention . We evaluated our model on five translation tasks and increased BLEU by 0.8 to 1.7 over soft attention , which in turn was better than hard and the recent Sparsemax ( Martins and Astudillo , 2016 ) attention . More importantly , the training process was as easy as soft attention . For further support , we also evaluate on two morphological inflection tasks and got gains over soft and hard attention . In this paper we showed a simple yet effective approximation of the joint attention-output distribution in sequence to sequence learning . Our joint model consistently provides higher accuracy without significant running time overheads in five translation and two morphological inflection tasks . An interesting direction for future work is to extend beam-joint to multi-head attention architectures as in ( Vaswani et al . , 2017 ; Xu Chen , 2018 ) .
| sobamchan/aclsum | 0 | full_paper |
P19-1352 | Word embedding is central to neural machine translation ( NMT ) , which has attracted intensive research interest in recent years . In NMT , the source embedding plays the role of the entrance while the target embedding acts as the terminal . These layers occupy most of the model parameters for representation learning . Furthermore , they indirectly interface via a soft-attention mechanism , which makes them comparatively isolated . In this paper , we propose shared-private bilingual word embeddings , which give a closer relationship between the source and target embeddings , and which also reduce the number of model parameters . For similar source and target words , their embeddings tend to share a part of the features and they cooperatively learn these common representation units . Experiments on 5 language pairs belonging to 6 different language families and written in 5 different alphabets demonstrate that the proposed model provides a significant performance boost over the strong baselines with dramatically fewer model parameters . With the introduction of ever more powerful architectures , neural machine translation ( NMT ) has become the most promising machine translation method ( Kalchbrenner and Blunsom , 2013 ; Sutskever et al . , 2014 ; Bahdanau et al . , 2015 ) . For word representation , different architecturesincluding , but not limited to , recurrence-based ( Chen et al . , 2018 ) , convolution-based ( Gehring et al . , 2017 ) and transformation-based ( Vaswani et al . , 2017 ) NMT models-have been taking advantage of the distributed word embeddings to capture the syntactic and semantic properties of words ( Turian et al . , 2010 ) . Figure 1 : Comparison between ( a ) standard word embeddings and ( b ) shared-private word embeddings . In ( a ) , the English word " Long " and the German word " Lange " , which have similar lexical meanings , are represented by two private d-dimension vectors . While in ( b ) , the two word embeddings are made up of two parts , indicating the shared ( lined nodes ) and the private ( unlined nodes ) features . This enables the two words to make use of common representation units , leading to a closer relationship between them . NMT usually utilizes three matrices to represent source embeddings , target input embeddings , and target output embeddings ( also known as pre-softmax weight ) , respectively . These embeddings occupy most of the model parameters , which constrains the improvements of NMT because the recent methods become increasingly memory-hungry ( Vaswani et al . , 2017 ; Chen et al . , 2018 ) . 1 Even though converting words into subword units ( Sennrich et al . , 2016b ) , nearly 55 % of model parameters are used for word representation in the Transformer model ( Vaswani et al . , 2017 ) . To overcome this difficulty , several methods are proposed to reduce the parameters used for word representation of NMT . Press and Wolf ( 2017 ) propose two weight tying ( WT ) methods , called decoder WT and three-way WT , to substantially reduce the parameters of the word embeddings . Decoder WT ties the target input embedding and target output embedding , which has become the new de facto standard of practical NMT ( Sen- Figure 2 : Shared-private bilingual word embeddings perform between the source and target words or sub-words ( a ) with similar lexical meaning , ( b ) with same word form , and ( c ) without any relationship . Different sharing mechanisms are adapted into different relationship categories . This strikes the right balance between capturing monolingual and bilingual characteristics . The closeness of relationship decides the portion of features to be used for sharing . Words with similar lexical meaning tend to share more features , followed by the words with the same word form , and then the unrelated words , as illustrated by the lined nodes . nrich et al . , 2017 ) . Three-way WT uses only one matrix to represent the three word embeddings , where the source and target words that have the same word form tend to share a word vector . This method can also be adapted to sub-word NMT with a shared source-target sub-word vocabulary and it performs well in language pairs with many of the same characters , such as English-German and English-French ( Vaswani et al . , 2017 ) . Unfortunately , this method is not applicable to languages that are written in different alphabets , such as Chinese-English ( Hassan et al . , 2018 ) . Another challenge facing the source and target word embeddings of NMT is the lack of interactions . This degrades the attention performance , leading to some unaligned translations that hurt the translation quality . Hence , Kuang et al . ( 2018 ) propose to bridge the source and target embeddings , which brings better attention to the related source and target words . Their method is applicable to any language pairs , providing a tight interaction between the source and target word pairs . However , their method requires additional components and model parameters . In this work , we aim to enhance the word representations and the interactions between the source and target words , while using even fewer parameters . To this end , we present a languageindependent method , which is called sharedprivate bilingual word embeddings , to share a part of the embeddings of a pair of source and target words that have some common characteristics ( i.e. similar words should have similar vectors ) . Figure 1 illustrates the difference between the standard word embeddings and shared-private word embeddings of NMT . In the proposed method , each source ( or target ) word is represented by a word embedding that consists of the shared features and the private features . The shared features can also be regarded as the prior alignments connecting the source and target words . The private features allow the words to better learn the monolingual characteristics . Meanwhile , the features shared by the source and target embeddings result in a significant reduction of the number of parameters used for word representations . The experimental results on 6 translation datasets of different scales show that our model with fewer parameters yields consistent improvements over the strong Transformer baselines . In this work , we propose a novel sharing technique to improve the learning of word embeddings for NMT . Each word embedding is composed of shared and private features . The shared features act as a prior alignment guidance for the attention model to improve the quality of attention . Meanwhile , the private features enable the words to better capture the monolingual characteristics , result in an improvement of the overall translation quality . According to the degree of relevance between a parallel word pair , the word pairs are categorized into three different groups and the number of shared features is different . Our experimental results show that the proposed method outperforms the strong Transformer baselines while using fewer model parameters . | They propose a language independet method where a model shares embeddings between source and target only when words have some common characteristics. | approach | high_quality | adherence_first | Objective: ADHERENCE-FIRST.
Task: Write a one-sentence summary focused ONLY on the aspect: approach.
Be strict: exclude any information not directly tied to the aspect.
If uncertain, omit rather than speculate.
Stay faithful to the document; do not add any unsupported claims.
How do the authors address the problem? Provide a brief summary of their approach. Return only the summary in one sentence.
---
Document:
Word embedding is central to neural machine translation ( NMT ) , which has attracted intensive research interest in recent years . In NMT , the source embedding plays the role of the entrance while the target embedding acts as the terminal . These layers occupy most of the model parameters for representation learning . Furthermore , they indirectly interface via a soft-attention mechanism , which makes them comparatively isolated . In this paper , we propose shared-private bilingual word embeddings , which give a closer relationship between the source and target embeddings , and which also reduce the number of model parameters . For similar source and target words , their embeddings tend to share a part of the features and they cooperatively learn these common representation units . Experiments on 5 language pairs belonging to 6 different language families and written in 5 different alphabets demonstrate that the proposed model provides a significant performance boost over the strong baselines with dramatically fewer model parameters . With the introduction of ever more powerful architectures , neural machine translation ( NMT ) has become the most promising machine translation method ( Kalchbrenner and Blunsom , 2013 ; Sutskever et al . , 2014 ; Bahdanau et al . , 2015 ) . For word representation , different architecturesincluding , but not limited to , recurrence-based ( Chen et al . , 2018 ) , convolution-based ( Gehring et al . , 2017 ) and transformation-based ( Vaswani et al . , 2017 ) NMT models-have been taking advantage of the distributed word embeddings to capture the syntactic and semantic properties of words ( Turian et al . , 2010 ) . Figure 1 : Comparison between ( a ) standard word embeddings and ( b ) shared-private word embeddings . In ( a ) , the English word " Long " and the German word " Lange " , which have similar lexical meanings , are represented by two private d-dimension vectors . While in ( b ) , the two word embeddings are made up of two parts , indicating the shared ( lined nodes ) and the private ( unlined nodes ) features . This enables the two words to make use of common representation units , leading to a closer relationship between them . NMT usually utilizes three matrices to represent source embeddings , target input embeddings , and target output embeddings ( also known as pre-softmax weight ) , respectively . These embeddings occupy most of the model parameters , which constrains the improvements of NMT because the recent methods become increasingly memory-hungry ( Vaswani et al . , 2017 ; Chen et al . , 2018 ) . 1 Even though converting words into subword units ( Sennrich et al . , 2016b ) , nearly 55 % of model parameters are used for word representation in the Transformer model ( Vaswani et al . , 2017 ) . To overcome this difficulty , several methods are proposed to reduce the parameters used for word representation of NMT . Press and Wolf ( 2017 ) propose two weight tying ( WT ) methods , called decoder WT and three-way WT , to substantially reduce the parameters of the word embeddings . Decoder WT ties the target input embedding and target output embedding , which has become the new de facto standard of practical NMT ( Sen- Figure 2 : Shared-private bilingual word embeddings perform between the source and target words or sub-words ( a ) with similar lexical meaning , ( b ) with same word form , and ( c ) without any relationship . Different sharing mechanisms are adapted into different relationship categories . This strikes the right balance between capturing monolingual and bilingual characteristics . The closeness of relationship decides the portion of features to be used for sharing . Words with similar lexical meaning tend to share more features , followed by the words with the same word form , and then the unrelated words , as illustrated by the lined nodes . nrich et al . , 2017 ) . Three-way WT uses only one matrix to represent the three word embeddings , where the source and target words that have the same word form tend to share a word vector . This method can also be adapted to sub-word NMT with a shared source-target sub-word vocabulary and it performs well in language pairs with many of the same characters , such as English-German and English-French ( Vaswani et al . , 2017 ) . Unfortunately , this method is not applicable to languages that are written in different alphabets , such as Chinese-English ( Hassan et al . , 2018 ) . Another challenge facing the source and target word embeddings of NMT is the lack of interactions . This degrades the attention performance , leading to some unaligned translations that hurt the translation quality . Hence , Kuang et al . ( 2018 ) propose to bridge the source and target embeddings , which brings better attention to the related source and target words . Their method is applicable to any language pairs , providing a tight interaction between the source and target word pairs . However , their method requires additional components and model parameters . In this work , we aim to enhance the word representations and the interactions between the source and target words , while using even fewer parameters . To this end , we present a languageindependent method , which is called sharedprivate bilingual word embeddings , to share a part of the embeddings of a pair of source and target words that have some common characteristics ( i.e. similar words should have similar vectors ) . Figure 1 illustrates the difference between the standard word embeddings and shared-private word embeddings of NMT . In the proposed method , each source ( or target ) word is represented by a word embedding that consists of the shared features and the private features . The shared features can also be regarded as the prior alignments connecting the source and target words . The private features allow the words to better learn the monolingual characteristics . Meanwhile , the features shared by the source and target embeddings result in a significant reduction of the number of parameters used for word representations . The experimental results on 6 translation datasets of different scales show that our model with fewer parameters yields consistent improvements over the strong Transformer baselines . In this work , we propose a novel sharing technique to improve the learning of word embeddings for NMT . Each word embedding is composed of shared and private features . The shared features act as a prior alignment guidance for the attention model to improve the quality of attention . Meanwhile , the private features enable the words to better capture the monolingual characteristics , result in an improvement of the overall translation quality . According to the degree of relevance between a parallel word pair , the word pairs are categorized into three different groups and the number of shared features is different . Our experimental results show that the proposed method outperforms the strong Transformer baselines while using fewer model parameters .
| sobamchan/aclsum | 1 | full_paper |
2020.emnlp-main.384 | Word embedding models are typically able to capture the semantics of words via the distributional hypothesis , but fail to capture the numerical properties of numbers that appear in a text . This leads to problems with numerical reasoning involving tasks such as question answering . We propose a new methodology to assign and learn embeddings for numbers . Our approach creates Deterministic , Independentof-Corpus Embeddings ( referred to as DICE ) for numbers , such that their cosine similarity reflects the actual distance on the number line . DICE outperforms a wide range of pre-trained word embedding models across multiple examples of two tasks : ( i ) evaluating the ability to capture numeration and magnitude ; and ( ii ) to perform list maximum , decoding , and addition . We further explore the utility of these embeddings in downstream applications by initializing numbers with our approach for the task of magnitude prediction . We also introduce a regularization approach to learn model-based embeddings of numbers in a contextual setting . Word embeddings capture semantic relationships between words by operationalizing the distributional hypothesis ( Harris , 1954 ; Firth , 1957 ) . They can be learned either non-contextually ( Mikolov et al . , 2013b ; Pennington et al . , 2014 ; Bojanowski et al . , 2017 ) or contextually ( Devlin et al . , 2018 ; Peters et al . , 2018 ) . Non-contextual embeddings have worked well on various language understanding and semantic tasks ( Rumelhart et al . , 1988 ; Mikolov et al . , 2013a , b ) . More recently , they have also been used as pre-trained word embeddings to aid more sophisticated contextual models for solving rigorous natural language processing ( NLP ) problems , including translation , paraphrasing , and sentence-similarity tasks ( Kiros et al . , 2015 ; Wieting et al . , 2015 ) . While word embeddings effectively capture semantic relationships between words , they are less effective at capturing numeric properties associated with numbers . Though numbers represent a significant percentage of tokens in a corpus , they are often overlooked . In non-contextual word embedding models , they are treated like any other word , which leads to misinterpretation . For instance , they exhibit unintuitive similarities with other words and do not contain strong prior information about the magnitude of the number they encode . In sentence similarity and reasoning tasks , failure to handle numbers causes as much as 29 % of contradictions ( De Marneffe et al . , 2008 ) . In other data-intensive tasks where numbers are abundant , like neural machine translation , they are masked to hide the translation models inefficiency in dealing with them ( Mitchell and Lapata , 2009 ) . There are a variety of tests proposed to measure the efficiency of number embeddings . For instance , Naik et al . ( 2019 ) shows that GloVe ( Pennington et al . , 2014 ) , word2vec ( Mikolov et al . , 2013b ) , and fastText ( Joulin et al . , 2016 ; Bojanowski et al . , 2017 ) fail to capture numeration and magnitude properties of a number . Numeration is the property of associating numbers with their corresponding word representations ( " 3 " and " three " ) while magnitude represents a number 's actual value ( 3 < 4 ) . Further , Wallace et al . ( 2019 ) proposes several tests for analyzing numerical reasoning of number embeddings that include list maximum , decoding , and addition . In this paper , we experimentally demonstrate that if the cosine similarity between word embeddings of two numbers reflects their actual distance on the number line , the resultant word embeddings are useful in downstream tasks . We first demonstrate how Deterministic , Independent-of-Corpus Embeddings ( DICE ) can be constructed such that they almost perfectly capture properties of numera-tion and magnitude . These non-contextual embeddings also perform well on related tests for numeracy ( Wallace et al . , 2019 ) . To demonstrate the efficacy of DICE for downstream tasks , we explore its utility in two experiments . First , we design a DICE embedding initialized Bi-LSTM network to classify the magnitude of masked numbers in the 600 K dataset ( Chen et al . , 2019 ) . Second , given the popularity of modern contextual model-based embeddings , we devise a regularization procedure that emulates the hypothesis proposed by DICE and can be employed in any task-based fine-tuning process . We demonstrate that adding such regularization helps the model internalize notions of numeracy while learning task-based contextual embeddings for the numbers present in the text . We find promising results in a numerical reasoning task that involves numerical question answering based on a sub-split of the popular SQuAD dataset ( Rajpurkar et al . , 2016 ) . Our contribution can be summarized as follows : • We propose a deterministic technique to learn numerical embeddings . DICE embeddings are learned independently of corpus and effectively capture properties of numeracy . • We prove experimentally that the resultant embeddings learned using the above methods improve a model 's ability to reason about numbers in a variety of tasks , including numeration , magnitude , list maximum , decoding , and addition . • We also demonstrate that properties of DICE can be adapted to contextual models , like BERT ( Devlin et al . , 2018 ) , through a novel regularization technique for solving tasks involving numerical reasoning . In this work , we methodologically assign and learn embeddings for numbers to reflect their numerical properties . We validate our proposed approach with several experiments that test number embeddings . The tests that evaluate the numeral embeddings are fundamentally applicable to all real numbers . Finally , we introduced an approach to jointly learn embeddings of numbers and words that preserve numerical properties and evaluated them on a contextual word embedding based model . In our future work , we would like to extend this idea to unseen numbers in vocabulary as a function of seen ones . | They propose a deterministic technique to learn numerical embeddings where cosine similarity reflects the actual distance and a regularization approach for a contextual setting. | approach | high_quality | coverage_first | Objective: COVERAGE-FIRST.
Task: Write a one-sentence summary focused on the aspect: approach.
Prioritize including the most salient, high-utility information relevant to the aspect.
You may include multiple key points if they fit naturally in one sentence.
Stay faithful to the document; do not add any unsupported claims.
Outline the methodology used by the authors to solve the stated challenge. Return only the summary in one sentence.
---
Document:
Word embedding models are typically able to capture the semantics of words via the distributional hypothesis , but fail to capture the numerical properties of numbers that appear in a text . This leads to problems with numerical reasoning involving tasks such as question answering . We propose a new methodology to assign and learn embeddings for numbers . Our approach creates Deterministic , Independentof-Corpus Embeddings ( referred to as DICE ) for numbers , such that their cosine similarity reflects the actual distance on the number line . DICE outperforms a wide range of pre-trained word embedding models across multiple examples of two tasks : ( i ) evaluating the ability to capture numeration and magnitude ; and ( ii ) to perform list maximum , decoding , and addition . We further explore the utility of these embeddings in downstream applications by initializing numbers with our approach for the task of magnitude prediction . We also introduce a regularization approach to learn model-based embeddings of numbers in a contextual setting . Word embeddings capture semantic relationships between words by operationalizing the distributional hypothesis ( Harris , 1954 ; Firth , 1957 ) . They can be learned either non-contextually ( Mikolov et al . , 2013b ; Pennington et al . , 2014 ; Bojanowski et al . , 2017 ) or contextually ( Devlin et al . , 2018 ; Peters et al . , 2018 ) . Non-contextual embeddings have worked well on various language understanding and semantic tasks ( Rumelhart et al . , 1988 ; Mikolov et al . , 2013a , b ) . More recently , they have also been used as pre-trained word embeddings to aid more sophisticated contextual models for solving rigorous natural language processing ( NLP ) problems , including translation , paraphrasing , and sentence-similarity tasks ( Kiros et al . , 2015 ; Wieting et al . , 2015 ) . While word embeddings effectively capture semantic relationships between words , they are less effective at capturing numeric properties associated with numbers . Though numbers represent a significant percentage of tokens in a corpus , they are often overlooked . In non-contextual word embedding models , they are treated like any other word , which leads to misinterpretation . For instance , they exhibit unintuitive similarities with other words and do not contain strong prior information about the magnitude of the number they encode . In sentence similarity and reasoning tasks , failure to handle numbers causes as much as 29 % of contradictions ( De Marneffe et al . , 2008 ) . In other data-intensive tasks where numbers are abundant , like neural machine translation , they are masked to hide the translation models inefficiency in dealing with them ( Mitchell and Lapata , 2009 ) . There are a variety of tests proposed to measure the efficiency of number embeddings . For instance , Naik et al . ( 2019 ) shows that GloVe ( Pennington et al . , 2014 ) , word2vec ( Mikolov et al . , 2013b ) , and fastText ( Joulin et al . , 2016 ; Bojanowski et al . , 2017 ) fail to capture numeration and magnitude properties of a number . Numeration is the property of associating numbers with their corresponding word representations ( " 3 " and " three " ) while magnitude represents a number 's actual value ( 3 < 4 ) . Further , Wallace et al . ( 2019 ) proposes several tests for analyzing numerical reasoning of number embeddings that include list maximum , decoding , and addition . In this paper , we experimentally demonstrate that if the cosine similarity between word embeddings of two numbers reflects their actual distance on the number line , the resultant word embeddings are useful in downstream tasks . We first demonstrate how Deterministic , Independent-of-Corpus Embeddings ( DICE ) can be constructed such that they almost perfectly capture properties of numera-tion and magnitude . These non-contextual embeddings also perform well on related tests for numeracy ( Wallace et al . , 2019 ) . To demonstrate the efficacy of DICE for downstream tasks , we explore its utility in two experiments . First , we design a DICE embedding initialized Bi-LSTM network to classify the magnitude of masked numbers in the 600 K dataset ( Chen et al . , 2019 ) . Second , given the popularity of modern contextual model-based embeddings , we devise a regularization procedure that emulates the hypothesis proposed by DICE and can be employed in any task-based fine-tuning process . We demonstrate that adding such regularization helps the model internalize notions of numeracy while learning task-based contextual embeddings for the numbers present in the text . We find promising results in a numerical reasoning task that involves numerical question answering based on a sub-split of the popular SQuAD dataset ( Rajpurkar et al . , 2016 ) . Our contribution can be summarized as follows : • We propose a deterministic technique to learn numerical embeddings . DICE embeddings are learned independently of corpus and effectively capture properties of numeracy . • We prove experimentally that the resultant embeddings learned using the above methods improve a model 's ability to reason about numbers in a variety of tasks , including numeration , magnitude , list maximum , decoding , and addition . • We also demonstrate that properties of DICE can be adapted to contextual models , like BERT ( Devlin et al . , 2018 ) , through a novel regularization technique for solving tasks involving numerical reasoning . In this work , we methodologically assign and learn embeddings for numbers to reflect their numerical properties . We validate our proposed approach with several experiments that test number embeddings . The tests that evaluate the numeral embeddings are fundamentally applicable to all real numbers . Finally , we introduced an approach to jointly learn embeddings of numbers and words that preserve numerical properties and evaluated them on a contextual word embedding based model . In our future work , we would like to extend this idea to unseen numbers in vocabulary as a function of seen ones .
| sobamchan/aclsum | 2 | full_paper |
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 5