emnlp emnlp2011 emnlp2011-142 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Lanjun Zhou ; Binyang Li ; Wei Gao ; Zhongyu Wei ; Kam-Fai Wong
Abstract: Polarity classification of opinionated sentences with both positive and negative sentiments1 is a key challenge in sentiment analysis. This paper presents a novel unsupervised method for discovering intra-sentence level discourse relations for eliminating polarity ambiguities. Firstly, a discourse scheme with discourse constraints on polarity was defined empirically based on Rhetorical Structure Theory (RST). Then, a small set of cuephrase-based patterns were utilized to collect a large number of discourse instances which were later converted to semantic sequential representations (SSRs). Finally, an unsupervised method was adopted to generate, weigh and filter new SSRs without cue phrases for recognizing discourse relations. Experimental results showed that the proposed methods not only effectively recognized the defined discourse relations but also achieved significant improvement by integrating discourse information in sentence-level polarity classification.
Reference: text
sentIndex sentText sentNum sentScore
1 This paper presents a novel unsupervised method for discovering intra-sentence level discourse relations for eliminating polarity ambiguities. [sent-5, score-1.048]
2 Firstly, a discourse scheme with discourse constraints on polarity was defined empirically based on Rhetorical Structure Theory (RST). [sent-6, score-1.535]
3 Then, a small set of cuephrase-based patterns were utilized to collect a large number of discourse instances which were later converted to semantic sequential representations (SSRs). [sent-7, score-0.719]
4 Finally, an unsupervised method was adopted to generate, weigh and filter new SSRs without cue phrases for recognizing discourse relations. [sent-8, score-0.78]
5 Experimental results showed that the proposed methods not only effectively recognized the defined discourse relations but also achieved significant improvement by integrating discourse information in sentence-level polarity classification. [sent-9, score-1.577]
6 1 Introduction As an important task of sentiment analysis, polarity classification is critically affected by discourse structure (Polanyi and Zaenen, 2006). [sent-10, score-1.055]
7 , 2008) and proved that the utilization of discourse relations could improve the performance of polarity classification on dialogues (Somasundaran et al. [sent-13, score-1.01]
8 However, cur1Defined as ambiguous sentences in this paper 162 rent state-of-the-art methods for sentence-level polarity classification are facing difficulties in ascertaining the polarity of some sentences. [sent-15, score-0.716]
9 Example (a) is difficult for existing polarity classification methods for two reasons: (1) the number of positive expressions is less than negative expressions; (2) the importance of each sentiment expression is unknown. [sent-19, score-0.444]
10 Existing sentence-level polarity classification methods ignoring discourse structure often give wrong results for these sentences. [sent-26, score-0.935]
11 (n and s denote nucleus and satellite segment, respectively) art method (Xu and Kit, 2010) in NTCIR-8 Chinese MOAT as the baseline polarity classifier (BPC) in this paper. [sent-33, score-0.633]
12 In this paper, we focused on the automation of recognizing intra-sentence level discourse relations for polarity classification. [sent-35, score-1.005]
13 Based on the previous work of Rhetorical Structure Theory (RST) (Mann and Thompson, 1988), a discourse scheme with discourse constraints on polarity was defined empirically (see Section 3). [sent-36, score-1.535]
14 From a raw corpus, a small set of cuephrase-based patterns were used to collect discourse instances. [sent-38, score-0.599]
15 Experimental results showed that the proposed methods could effectively recognize the defined discourse relations and achieve significant improvement in sentence-level polarity classification comparing to BPC. [sent-41, score-1.01]
16 Section 3 presents the discourse scheme with discourse constraints on polarity. [sent-44, score-1.2]
17 2 Related Work Research on polarity classification were generally conducted on 4 levels: document-level (Pang et al. [sent-47, score-0.362]
18 There was little research focusing on the automatic recognition of intra-sentence level discourse 163 relations for sentiment analysis in the literature. [sent-52, score-0.73]
19 Polanyi and Zaenen (2006) argued that valence calculation is critically affected by discourse structure. [sent-53, score-0.639]
20 Nevertheless, they did not propose a computational model for their discourse scheme. [sent-56, score-0.573]
21 Nonetheless, contrastive relations were only one type of discourse relations which may help polarity classification. [sent-58, score-1.058]
22 (2008) modeled polarity reversal using HCRFs integrated with inter-sentence discourse structures. [sent-60, score-0.908]
23 However, our work is on intra- sentence level and our purpose is not to find polarity reversals but trying to adapt general discourse schemes (e. [sent-61, score-0.908]
24 , RST) to help determine the overall polarity of ambiguous sentences. [sent-63, score-0.354]
25 , 2009), which proposed opinion frames as a representation ofdiscourse-level associations on dialogue and modeled the scheme to improve opinion polarity classification. [sent-66, score-0.482]
26 Our work differs from their approaches in two key aspects: (1) we distinguished nucleus and satellite in discourse but opinion frames did not; (2) our method for discourse discovery was unsupervisedwhile their method needed annotated data. [sent-68, score-1.51]
27 Most research works about discourse classification were not related to sentiment analysis. [sent-69, score-0.682]
28 Supervised discourse classification methods (Soricut and Marcu, 2003; Duverle and Prendinger, 2009) needed manually annotated data. [sent-70, score-0.6]
29 Marcu and Echihabi (2002) presented an unsupervised method to recognize discourse relations held between arbitrary spans of text. [sent-71, score-0.675]
30 They showed that lexical pairs extracted from massive amount of data can have a major impact on discourse classification. [sent-72, score-0.573]
31 Thus, in additional to lexical features, we incorporated sequential and semantic information in proposed method for discourse relation classification. [sent-76, score-0.644]
32 3 Discourse Scheme for Eliminating Polarity Ambiguities Since not all of the discourse relations in RST would help eliminate polarity ambiguities, the discourse scheme defined in this paper was on a much coarser level. [sent-78, score-1.59]
33 In order to ascertain which relations should be included in our scheme, 500 ambiguous sentences were randomly chosen from NTCIR MOAT Chinese corpus and the most common discourse relations for connecting independent clauses in compound sentences were annotated. [sent-79, score-0.819]
34 We found that 13 relations from RST occupied about 70% of the annotated discourse relations which may help eliminate polarity ambiguities. [sent-80, score-1.058]
35 Inspired by Marcu and Echihabi (2002), to construct relatively lownoise discourse instances for unsupervised methods using cue phrases, we grouped the 13 relations into the following 5 relations: Contrast is a union of Antithesis, Concession, Otherwise and Contrast from RST. [sent-81, score-0.819]
36 Cue-phrase-based patterns could find only limited number of discourse instances with high precision (Marcu and Echihabi, 2002). [sent-88, score-0.643]
37 Thus, we proposed a language independent unsupervised method to identify discourse relations without cue phrases while maintaining relatively high precision. [sent-92, score-0.795]
38 For each discourse relation, we started with several cuephrase-based patterns and collected a large number of discourse instances from raw corpus. [sent-93, score-1.216]
39 Then, discourse instances were converted to semantic sequential representations (SSRs). [sent-94, score-0.661]
40 1 Gathering and representing discourse instances A discourse instance, denoted by Di, consists of two successive segments (Di[1] , Di[2] ) within a sentence. [sent-98, score-1.296]
41 Table 1 listed some examples of cue phrases for each discourse relation. [sent-127, score-0.673]
42 "CUE1" indicated satellite segments and "CUE2" indicated nucleus segments. [sent-133, score-0.417]
43 Note that we did not distinguish satellite from nucleus for Continuation in this paper because the polarity could be determined by either segment. [sent-134, score-0.663]
44 To simplify the problem of discourse segmentation, we split compound sentences into discourse segments using commas and semicolons. [sent-136, score-1.239]
45 Although we collected discourse instances from compound sentences only, the number of instances for each discourse relation was large enough for the proposed unsupervised method. [sent-137, score-1.327]
46 In order to incorporate lexical and semantic information in our method, we represented each word in a discourse instance using a part-of-speech tag, a semantic label and a sentiment tag. [sent-139, score-0.655]
47 The next problem became how to start from current SSRs and generate new SSRs for recognizing discourse relations without cue phrases. [sent-151, score-0.77]
48 that di and dj could generate a common SSR if and only if the orders of nucleus segment and satellite segment were the same. [sent-179, score-0.626]
49 As a result, for each discourse relation rn, a corresponding common SSR set Sn could be obtained by adopting match(di, dj) where i j for all discourse instances. [sent-185, score-1.23]
50 An advantage eof i match(d1 , d2) was eth ina-t the generated common SSRs preserved the sequential structure of original discourse instances. [sent-186, score-0.655]
51 And common SSRs allows us to build high precision discourse classifiers (See Section 5). [sent-187, score-0.611]
52 For example, the adverb "very" in "very brilliant" of D1 was not important for discourse recognition. [sent-189, score-0.573]
53 Nouns (except for named entities) and verbs were most representative words in discourse recognition (Marcu and Echihabi, 2002). [sent-196, score-0.573]
54 In addition, adjectives and adverbs appearing in sentiment lexicons were important for polarity classification. [sent-197, score-0.417]
55 The common SSR set Sn for each discourse relation rn could be directly used in SSR-based unsupervised classifiers or be employed as effective features in supervised methods. [sent-211, score-0.684]
56 7n2)%c e Table 4: Distribution of discourse relations on NTC-7. [sent-214, score-0.648]
57 Others represents discourse relations not included in our discourse scheme. [sent-215, score-1.221]
58 1 Annotation work and Data We extracted all compound sentences which may contain the defined discourse relations from opinionated sentences (neutral ones were dropped) of NT- CIR7 MOAT simplified Chinese training data. [sent-217, score-0.73]
59 1,225 discourse instances were extracted and two annotators were trained to annotate discourse relations according to the discourse scheme defined in Section 3. [sent-218, score-1.872]
60 Note that we annotate both explicit and implicit discourse relations. [sent-219, score-0.573]
61 Table 4 showed the distribution of annotated discourse relations based on the inter-annotator agreement. [sent-223, score-0.648]
62 The proportion of occurrences of each discourse relations varied greatly. [sent-224, score-0.648]
63 The experiments of this paper were performed using the following data sets: NTC-7 contained manually annotated discourse instances (shown in Table 4). [sent-226, score-0.636]
64 The experiments of discourse identification were performed on this data set. [sent-227, score-0.573]
65 The experiments of polarity ambiguity elimination using the identified discourse relations were performed on this data set. [sent-229, score-1.002]
66 Given a discourse instance Di, the probabilities: P(rk | (Di[1] , Di[2] )) for each relation rk were estimated|( on all text from XINHUA. [sent-238, score-0.639]
67 Then, the most likely discourse relation was determined by taking the maximum over argmaxk{P(rk |(Di[1] , Di[2] )}. [sent-239, score-0.649]
68 cSSR used both cue-phrase-based patterns together with common SSRs for recognizing discourse relations. [sent-240, score-0.659]
69 Common SSRs were mined from discourse instances extracted from XINHUA using cuephrase-based patterns. [sent-241, score-0.667]
70 SVM was trained utilizing cue phrases, probabilities from M&E;, topic similarity, structure overlap, polarity of segments and mined common SSRs (Optional). [sent-243, score-0.596]
71 The nucleus and satellite information is Figure 2: Influences of different values of minconf to the performance of cSSR acquired by cSSR if a segment pair could match a cSSR. [sent-250, score-0.507]
72 (2) Apply discourse constraints on polarity to ascertain the polarity for each discourse instance. [sent-252, score-1.855]
73 There may be conflicts between polarities acquired by BPC and discourse constraints on polarity (e. [sent-253, score-0.966]
74 , Two segments with the same polarity holding a Contrast relation). [sent-255, score-0.486]
75 To handle this problem, we chose the segment with higher polarity confidence and adjusted the polarity of the other segment using discourse constraints on polarity. [sent-256, score-1.439]
76 (3) If there was more than one discourse instance in a single sentence, the overall polarity of the sentence was determined by voting of polarities from each discourse instance under the majority rule. [sent-257, score-1.549]
77 Table 5 presented the experimental results for discourse relation classification. [sent-274, score-0.619]
78 8 VS10M R9231s Table 6: Performance of integrating discourse classifiers and constraints to polarity classification. [sent-280, score-0.949]
79 On the other side, M&E; which only considered word pairs between two segments of discourse instances got a higher recall with a large drop of precision. [sent-285, score-0.713]
80 The drop of precision may be caused by the neglect ofstructural and semantic information of discourse instances. [sent-286, score-0.596]
81 Actually, cSSR outperformed Baseline in all discourse relations except for Contrast. [sent-291, score-0.648]
82 Table 6 presented the performance of integrating discourse classifiers to polarity classification. [sent-299, score-0.929]
83 For Baseline and cSSR, the information of nucleus and satellite could be obtained directly from cue- Table 5: Performance of recognizing discourse relations. [sent-300, score-0.893]
84 For SVM+cSSR, the nucleus and satellite information was acquired by cSSR if a segment pair could match a cSSR. [sent-302, score-0.396]
85 It's clear that the performance of polarity classification was enhanced with the improvement of discourse relation recogni- tion. [sent-304, score-0.981]
86 M&E; was not included in this experiment because the performance of polarity classification was decreased by the mis-classified discourse relations. [sent-305, score-0.935]
87 It's straightforward that these words were insignificant in discourse relation classification purpose. [sent-325, score-0.646]
88 0), the proposed method successfully edo wwonr weighed ,th teh e w porrodpso swehdic mh were Figure 3: Improvement from individual discourse relations. [sent-337, score-0.573]
89 Contribution of different discourse relations We also analyzed the contribution of different discourse relations in eliminating polarity ambiguities. [sent-343, score-1.669]
90 Referto Figure 3, the improvement ofpolarity classification mainly came from three discourse relations: Contrast, Continuation and Cause. [sent-344, score-0.6]
91 It was straightforward that Contrast relation could eliminate polarity ambiguities because it held between two segments with opposite polarities. [sent-345, score-0.483]
92 However, recall Table 4, although Cause occurred more often than Contrast, only a part of discourse instances holding Cause relation contained two segments with the opposite polarities. [sent-347, score-0.833]
93 Consequently, the polarity of the second segment should be negative. [sent-355, score-0.405]
94 6 Conclusions and Future work This paper focused on unsupervised discovery of intra-sentence discourse relations for sentence level polarity classification. [sent-356, score-1.01]
95 We firstly presented a discourse scheme based on empirical observations. [sent-357, score-0.607]
96 Then, an unsupervised method was proposed starting from a small set of cue-phrase-based patterns to mine high quality common SSRs for each discourse relation. [sent-358, score-0.664]
97 Experimental results showed that our methods not only effectively recognized discourse relations but also achieved significant improvement (p<0. [sent-360, score-0.648]
98 A novel discourse parser based on support vector machine classification. [sent-407, score-0.573]
99 Supervised and unsupervised methods in employing discourse relations for improving opinion polarity classification. [sent-495, score-1.057]
100 Sentence level discourse parsing using syntactic and lexical information. [sent-502, score-0.573]
wordName wordTfidf (topN-words)
[('discourse', 0.573), ('ssrs', 0.5), ('polarity', 0.335), ('cssr', 0.222), ('nucleus', 0.153), ('satellite', 0.145), ('ssr', 0.133), ('minconf', 0.111), ('continuation', 0.104), ('cue', 0.1), ('mismatches', 0.088), ('sentiment', 0.082), ('holding', 0.078), ('moat', 0.078), ('ssri', 0.078), ('di', 0.075), ('dj', 0.075), ('relations', 0.075), ('segments', 0.073), ('segment', 0.07), ('bpc', 0.067), ('conf', 0.057), ('weighing', 0.052), ('mined', 0.05), ('rst', 0.048), ('opinion', 0.047), ('relation', 0.046), ('somasundaran', 0.045), ('instances', 0.044), ('sn', 0.042), ('rhetorical', 0.04), ('opinionated', 0.04), ('common', 0.038), ('eliminating', 0.038), ('polarities', 0.038), ('echihabi', 0.036), ('confidence', 0.036), ('chinese', 0.035), ('prp', 0.035), ('scheme', 0.034), ('marcu', 0.034), ('asher', 0.033), ('inproceedings', 0.033), ('denoted', 0.033), ('utilized', 0.032), ('cause', 0.031), ('svm', 0.031), ('weigh', 0.03), ('determined', 0.03), ('ambiguities', 0.029), ('um', 0.029), ('adopted', 0.028), ('match', 0.028), ('calculation', 0.028), ('unsupervised', 0.027), ('classification', 0.027), ('polanyi', 0.026), ('patterns', 0.026), ('sequential', 0.025), ('indicated', 0.023), ('drop', 0.023), ('bos', 0.023), ('wiebe', 0.022), ('simplified', 0.022), ('banned', 0.022), ('boris', 0.022), ('hated', 0.022), ('loved', 0.022), ('ofaverage', 0.022), ('ofdiscourse', 0.022), ('sadamitsu', 0.022), ('ufm', 0.022), ('recognizing', 0.022), ('jj', 0.022), ('hong', 0.021), ('filtering', 0.021), ('integrating', 0.021), ('dropped', 0.02), ('compound', 0.02), ('rk', 0.02), ('maintaining', 0.02), ('constraints', 0.02), ('contrast', 0.02), ('che', 0.019), ('preserved', 0.019), ('frames', 0.019), ('converted', 0.019), ('affected', 0.019), ('critically', 0.019), ('elimination', 0.019), ('duverle', 0.019), ('miltsakaki', 0.019), ('ascertain', 0.019), ('corrupted', 0.019), ('brilliant', 0.019), ('ntcir', 0.019), ('contained', 0.019), ('kong', 0.019), ('nn', 0.019), ('ambiguous', 0.019)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000011 142 emnlp-2011-Unsupervised Discovery of Discourse Relations for Eliminating Intra-sentence Polarity Ambiguities
Author: Lanjun Zhou ; Binyang Li ; Wei Gao ; Zhongyu Wei ; Kam-Fai Wong
Abstract: Polarity classification of opinionated sentences with both positive and negative sentiments1 is a key challenge in sentiment analysis. This paper presents a novel unsupervised method for discovering intra-sentence level discourse relations for eliminating polarity ambiguities. Firstly, a discourse scheme with discourse constraints on polarity was defined empirically based on Rhetorical Structure Theory (RST). Then, a small set of cuephrase-based patterns were utilized to collect a large number of discourse instances which were later converted to semantic sequential representations (SSRs). Finally, an unsupervised method was adopted to generate, weigh and filter new SSRs without cue phrases for recognizing discourse relations. Experimental results showed that the proposed methods not only effectively recognized the defined discourse relations but also achieved significant improvement by integrating discourse information in sentence-level polarity classification.
2 0.258028 94 emnlp-2011-Modelling Discourse Relations for Arabic
Author: Amal Al-Saif ; Katja Markert
Abstract: We present the first algorithms to automatically identify explicit discourse connectives and the relations they signal for Arabic text. First we show that, for Arabic news, most adjacent sentences are connected via explicit connectives in contrast to English, making the treatment of explicit discourse connectives for Arabic highly important. We also show that explicit Arabic discourse connectives are far more ambiguous than English ones, making their treatment challenging. In the second part of the paper, we present supervised algorithms to address automatic discourse connective identification and discourse relation recognition. Our connective identifier based on gold standard syntactic features achieves almost human performance. In addition, an identifier based solely on simple lexical and automatically derived morphological and POS features performs with high reliability, essential for languages that do not have high-quality parsers yet. Our algorithm for recognizing discourse relations performs significantly better than a baseline based on the connective surface string alone and therefore reduces the ambiguity in explicit connective interpretation.
3 0.20342599 92 emnlp-2011-Minimally Supervised Event Causality Identification
Author: Quang Do ; Yee Seng Chan ; Dan Roth
Abstract: This paper develops a minimally supervised approach, based on focused distributional similarity methods and discourse connectives, for identifying of causality relations between events in context. While it has been shown that distributional similarity can help identifying causality, we observe that discourse connectives and the particular discourse relation they evoke in context provide additional information towards determining causality between events. We show that combining discourse relation predictions and distributional similarity methods in a global inference procedure provides additional improvements towards determining event causality.
4 0.12486804 8 emnlp-2011-A Model of Discourse Predictions in Human Sentence Processing
Author: Amit Dubey ; Frank Keller ; Patrick Sturt
Abstract: This paper introduces a psycholinguistic model of sentence processing which combines a Hidden Markov Model noun phrase chunker with a co-reference classifier. Both models are fully incremental and generative, giving probabilities of lexical elements conditional upon linguistic structure. This allows us to compute the information theoretic measure of surprisal, which is known to correlate with human processing effort. We evaluate our surprisal predictions on the Dundee corpus of eye-movement data show that our model achieve a better fit with human reading times than a syntax-only model which does not have access to co-reference information.
5 0.11438113 30 emnlp-2011-Compositional Matrix-Space Models for Sentiment Analysis
Author: Ainur Yessenalina ; Claire Cardie
Abstract: We present a general learning-based approach for phrase-level sentiment analysis that adopts an ordinal sentiment scale and is explicitly compositional in nature. Thus, we can model the compositional effects required for accurate assignment of phrase-level sentiment. For example, combining an adverb (e.g., “very”) with a positive polar adjective (e.g., “good”) produces a phrase (“very good”) with increased polarity over the adjective alone. Inspired by recent work on distributional approaches to compositionality, we model each word as a matrix and combine words using iterated matrix multiplication, which allows for the modeling of both additive and multiplicative semantic effects. Although the multiplication-based matrix-space framework has been shown to be a theoretically elegant way to model composition (Rudolph and Giesbrecht, 2010), training such models has to be done carefully: the optimization is nonconvex and requires a good initial starting point. This paper presents the first such algorithm for learning a matrix-space model for semantic composition. In the context of the phrase-level sentiment analysis task, our experimental results show statistically significant improvements in performance over a bagof-words model.
6 0.094828919 105 emnlp-2011-Predicting Thread Discourse Structure over Technical Web Forums
7 0.090810485 120 emnlp-2011-Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions
8 0.087198325 81 emnlp-2011-Learning General Connotation of Words using Graph-based Algorithms
9 0.084741041 126 emnlp-2011-Structural Opinion Mining for Graph-based Sentiment Representation
10 0.080246054 63 emnlp-2011-Harnessing WordNet Senses for Supervised Sentiment Classification
11 0.076977462 84 emnlp-2011-Learning the Information Status of Noun Phrases in Spoken Dialogues
12 0.074631304 33 emnlp-2011-Cooooooooooooooollllllllllllll!!!!!!!!!!!!!! Using Word Lengthening to Detect Sentiment in Microblogs
13 0.051367719 12 emnlp-2011-A Weakly-supervised Approach to Argumentative Zoning of Scientific Documents
14 0.048627079 62 emnlp-2011-Generating Subsequent Reference in Shared Visual Scenes: Computation vs Re-Use
15 0.046954475 128 emnlp-2011-Structured Relation Discovery using Generative Models
16 0.046383858 38 emnlp-2011-Data-Driven Response Generation in Social Media
17 0.042383093 14 emnlp-2011-A generative model for unsupervised discovery of relations and argument classes from clinical texts
18 0.041203734 113 emnlp-2011-Relation Acquisition using Word Classes and Partial Patterns
19 0.040664166 114 emnlp-2011-Relation Extraction with Relation Topics
20 0.040252328 17 emnlp-2011-Active Learning with Amazon Mechanical Turk
topicId topicWeight
[(0, 0.132), (1, -0.14), (2, -0.046), (3, 0.075), (4, 0.189), (5, -0.184), (6, 0.137), (7, 0.144), (8, -0.116), (9, -0.431), (10, -0.149), (11, 0.002), (12, 0.047), (13, 0.131), (14, 0.044), (15, -0.132), (16, 0.085), (17, 0.062), (18, 0.057), (19, -0.003), (20, 0.015), (21, 0.008), (22, 0.06), (23, 0.079), (24, 0.01), (25, -0.051), (26, -0.054), (27, -0.011), (28, -0.036), (29, 0.027), (30, 0.032), (31, -0.076), (32, 0.056), (33, 0.024), (34, 0.03), (35, -0.054), (36, 0.0), (37, -0.034), (38, -0.008), (39, 0.044), (40, 0.026), (41, -0.016), (42, -0.01), (43, -0.05), (44, 0.012), (45, -0.021), (46, 0.061), (47, 0.001), (48, -0.047), (49, -0.003)]
simIndex simValue paperId paperTitle
same-paper 1 0.97292829 142 emnlp-2011-Unsupervised Discovery of Discourse Relations for Eliminating Intra-sentence Polarity Ambiguities
Author: Lanjun Zhou ; Binyang Li ; Wei Gao ; Zhongyu Wei ; Kam-Fai Wong
Abstract: Polarity classification of opinionated sentences with both positive and negative sentiments1 is a key challenge in sentiment analysis. This paper presents a novel unsupervised method for discovering intra-sentence level discourse relations for eliminating polarity ambiguities. Firstly, a discourse scheme with discourse constraints on polarity was defined empirically based on Rhetorical Structure Theory (RST). Then, a small set of cuephrase-based patterns were utilized to collect a large number of discourse instances which were later converted to semantic sequential representations (SSRs). Finally, an unsupervised method was adopted to generate, weigh and filter new SSRs without cue phrases for recognizing discourse relations. Experimental results showed that the proposed methods not only effectively recognized the defined discourse relations but also achieved significant improvement by integrating discourse information in sentence-level polarity classification.
2 0.89845365 94 emnlp-2011-Modelling Discourse Relations for Arabic
Author: Amal Al-Saif ; Katja Markert
Abstract: We present the first algorithms to automatically identify explicit discourse connectives and the relations they signal for Arabic text. First we show that, for Arabic news, most adjacent sentences are connected via explicit connectives in contrast to English, making the treatment of explicit discourse connectives for Arabic highly important. We also show that explicit Arabic discourse connectives are far more ambiguous than English ones, making their treatment challenging. In the second part of the paper, we present supervised algorithms to address automatic discourse connective identification and discourse relation recognition. Our connective identifier based on gold standard syntactic features achieves almost human performance. In addition, an identifier based solely on simple lexical and automatically derived morphological and POS features performs with high reliability, essential for languages that do not have high-quality parsers yet. Our algorithm for recognizing discourse relations performs significantly better than a baseline based on the connective surface string alone and therefore reduces the ambiguity in explicit connective interpretation.
3 0.77283889 92 emnlp-2011-Minimally Supervised Event Causality Identification
Author: Quang Do ; Yee Seng Chan ; Dan Roth
Abstract: This paper develops a minimally supervised approach, based on focused distributional similarity methods and discourse connectives, for identifying of causality relations between events in context. While it has been shown that distributional similarity can help identifying causality, we observe that discourse connectives and the particular discourse relation they evoke in context provide additional information towards determining causality between events. We show that combining discourse relation predictions and distributional similarity methods in a global inference procedure provides additional improvements towards determining event causality.
4 0.46747202 8 emnlp-2011-A Model of Discourse Predictions in Human Sentence Processing
Author: Amit Dubey ; Frank Keller ; Patrick Sturt
Abstract: This paper introduces a psycholinguistic model of sentence processing which combines a Hidden Markov Model noun phrase chunker with a co-reference classifier. Both models are fully incremental and generative, giving probabilities of lexical elements conditional upon linguistic structure. This allows us to compute the information theoretic measure of surprisal, which is known to correlate with human processing effort. We evaluate our surprisal predictions on the Dundee corpus of eye-movement data show that our model achieve a better fit with human reading times than a syntax-only model which does not have access to co-reference information.
5 0.32694134 84 emnlp-2011-Learning the Information Status of Noun Phrases in Spoken Dialogues
Author: Altaf Rahman ; Vincent Ng
Abstract: An entity in a dialogue may be old, new, or mediated/inferrable with respect to the hearer’s beliefs. Knowing the information status of the entities participating in a dialogue can therefore facilitate its interpretation. We address the under-investigated problem of automatically determining the information status of discourse entities. Specifically, we extend Nissim’s (2006) machine learning approach to information-status determination with lexical and structured features, and exploit learned knowledge of the information status of each discourse entity for coreference resolution. Experimental results on a set of Switchboard dialogues reveal that (1) incorporating our proposed features into Nissim’s feature set enables our system to achieve stateof-the-art performance on information-status classification, and (2) the resulting information can be used to improve the performance of learning-based coreference resolvers.
6 0.32313457 81 emnlp-2011-Learning General Connotation of Words using Graph-based Algorithms
7 0.3219921 105 emnlp-2011-Predicting Thread Discourse Structure over Technical Web Forums
8 0.29425466 30 emnlp-2011-Compositional Matrix-Space Models for Sentiment Analysis
9 0.27666649 33 emnlp-2011-Cooooooooooooooollllllllllllll!!!!!!!!!!!!!! Using Word Lengthening to Detect Sentiment in Microblogs
10 0.27540237 126 emnlp-2011-Structural Opinion Mining for Graph-based Sentiment Representation
11 0.27183521 120 emnlp-2011-Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions
12 0.25471863 63 emnlp-2011-Harnessing WordNet Senses for Supervised Sentiment Classification
13 0.21218632 12 emnlp-2011-A Weakly-supervised Approach to Argumentative Zoning of Scientific Documents
14 0.15778579 62 emnlp-2011-Generating Subsequent Reference in Shared Visual Scenes: Computation vs Re-Use
15 0.15738042 113 emnlp-2011-Relation Acquisition using Word Classes and Partial Patterns
16 0.1525328 38 emnlp-2011-Data-Driven Response Generation in Social Media
17 0.14449856 17 emnlp-2011-Active Learning with Amazon Mechanical Turk
18 0.13350722 128 emnlp-2011-Structured Relation Discovery using Generative Models
19 0.12913549 114 emnlp-2011-Relation Extraction with Relation Topics
20 0.1290894 70 emnlp-2011-Identifying Relations for Open Information Extraction
topicId topicWeight
[(15, 0.019), (23, 0.1), (36, 0.013), (37, 0.031), (45, 0.049), (48, 0.382), (53, 0.019), (54, 0.017), (57, 0.017), (62, 0.027), (64, 0.017), (66, 0.04), (69, 0.023), (79, 0.034), (82, 0.011), (87, 0.012), (96, 0.047), (97, 0.016), (98, 0.026)]
simIndex simValue paperId paperTitle
same-paper 1 0.75320053 142 emnlp-2011-Unsupervised Discovery of Discourse Relations for Eliminating Intra-sentence Polarity Ambiguities
Author: Lanjun Zhou ; Binyang Li ; Wei Gao ; Zhongyu Wei ; Kam-Fai Wong
Abstract: Polarity classification of opinionated sentences with both positive and negative sentiments1 is a key challenge in sentiment analysis. This paper presents a novel unsupervised method for discovering intra-sentence level discourse relations for eliminating polarity ambiguities. Firstly, a discourse scheme with discourse constraints on polarity was defined empirically based on Rhetorical Structure Theory (RST). Then, a small set of cuephrase-based patterns were utilized to collect a large number of discourse instances which were later converted to semantic sequential representations (SSRs). Finally, an unsupervised method was adopted to generate, weigh and filter new SSRs without cue phrases for recognizing discourse relations. Experimental results showed that the proposed methods not only effectively recognized the defined discourse relations but also achieved significant improvement by integrating discourse information in sentence-level polarity classification.
2 0.59693396 63 emnlp-2011-Harnessing WordNet Senses for Supervised Sentiment Classification
Author: Balamurali AR ; Aditya Joshi ; Pushpak Bhattacharyya
Abstract: Traditional approaches to sentiment classification rely on lexical features, syntax-based features or a combination of the two. We propose semantic features using word senses for a supervised document-level sentiment classifier. To highlight the benefit of sense-based features, we compare word-based representation of documents with a sense-based representation where WordNet senses of the words are used as features. In addition, we highlight the benefit of senses by presenting a part-ofspeech-wise effect on sentiment classification. Finally, we show that even if a WSD engine disambiguates between a limited set of words in a document, a sentiment classifier still performs better than what it does in absence of sense annotation. Since word senses used as features show promise, we also examine the possibility of using similarity metrics defined on WordNet to address the problem of not finding a sense in the training corpus. We per- form experiments using three popular similarity metrics to mitigate the effect of unknown synsets in a test corpus by replacing them with similar synsets from the training corpus. The results show promising improvement with respect to the baseline.
3 0.53482068 8 emnlp-2011-A Model of Discourse Predictions in Human Sentence Processing
Author: Amit Dubey ; Frank Keller ; Patrick Sturt
Abstract: This paper introduces a psycholinguistic model of sentence processing which combines a Hidden Markov Model noun phrase chunker with a co-reference classifier. Both models are fully incremental and generative, giving probabilities of lexical elements conditional upon linguistic structure. This allows us to compute the information theoretic measure of surprisal, which is known to correlate with human processing effort. We evaluate our surprisal predictions on the Dundee corpus of eye-movement data show that our model achieve a better fit with human reading times than a syntax-only model which does not have access to co-reference information.
4 0.36741674 94 emnlp-2011-Modelling Discourse Relations for Arabic
Author: Amal Al-Saif ; Katja Markert
Abstract: We present the first algorithms to automatically identify explicit discourse connectives and the relations they signal for Arabic text. First we show that, for Arabic news, most adjacent sentences are connected via explicit connectives in contrast to English, making the treatment of explicit discourse connectives for Arabic highly important. We also show that explicit Arabic discourse connectives are far more ambiguous than English ones, making their treatment challenging. In the second part of the paper, we present supervised algorithms to address automatic discourse connective identification and discourse relation recognition. Our connective identifier based on gold standard syntactic features achieves almost human performance. In addition, an identifier based solely on simple lexical and automatically derived morphological and POS features performs with high reliability, essential for languages that do not have high-quality parsers yet. Our algorithm for recognizing discourse relations performs significantly better than a baseline based on the connective surface string alone and therefore reduces the ambiguity in explicit connective interpretation.
5 0.34677115 108 emnlp-2011-Quasi-Synchronous Phrase Dependency Grammars for Machine Translation
Author: Kevin Gimpel ; Noah A. Smith
Abstract: We present a quasi-synchronous dependency grammar (Smith and Eisner, 2006) for machine translation in which the leaves of the tree are phrases rather than words as in previous work (Gimpel and Smith, 2009). This formulation allows us to combine structural components of phrase-based and syntax-based MT in a single model. We describe a method of extracting phrase dependencies from parallel text using a target-side dependency parser. For decoding, we describe a coarse-to-fine approach based on lattice dependency parsing of phrase lattices. We demonstrate performance improvements for Chinese-English and UrduEnglish translation over a phrase-based baseline. We also investigate the use of unsupervised dependency parsers, reporting encouraging preliminary results.
6 0.34168518 1 emnlp-2011-A Bayesian Mixture Model for PoS Induction Using Multiple Features
7 0.33690605 128 emnlp-2011-Structured Relation Discovery using Generative Models
8 0.33579785 59 emnlp-2011-Fast and Robust Joint Models for Biomedical Event Extraction
9 0.33578691 136 emnlp-2011-Training a Parser for Machine Translation Reordering
10 0.33576691 28 emnlp-2011-Closing the Loop: Fast, Interactive Semi-Supervised Annotation With Queries on Features and Instances
11 0.33511263 123 emnlp-2011-Soft Dependency Constraints for Reordering in Hierarchical Phrase-Based Translation
12 0.33320943 98 emnlp-2011-Named Entity Recognition in Tweets: An Experimental Study
13 0.33245996 35 emnlp-2011-Correcting Semantic Collocation Errors with L1-induced Paraphrases
14 0.33188692 137 emnlp-2011-Training dependency parsers by jointly optimizing multiple objectives
15 0.33161664 120 emnlp-2011-Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions
16 0.33155096 79 emnlp-2011-Lateen EM: Unsupervised Training with Multiple Objectives, Applied to Dependency Grammar Induction
17 0.33131319 53 emnlp-2011-Experimental Support for a Categorical Compositional Distributional Model of Meaning
18 0.33090472 126 emnlp-2011-Structural Opinion Mining for Graph-based Sentiment Representation
19 0.32935253 81 emnlp-2011-Learning General Connotation of Words using Graph-based Algorithms
20 0.32935229 30 emnlp-2011-Compositional Matrix-Space Models for Sentiment Analysis