emnlp emnlp2010 emnlp2010-95 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Xiaohua Liu ; Bo Han ; Kuan Li ; Stephan Hyeonjun Stiller ; Ming Zhou
Abstract: In this paper we develop an approach to tackle the problem of verb selection for learners of English as a second language (ESL) by using features from the output of Semantic Role Labeling (SRL). Unlike existing approaches to verb selection that use local features such as n-grams, our approach exploits semantic features which explicitly model the usage context of the verb. The verb choice highly depends on its usage context which is not consistently captured by local features. We then combine these semantic features with other local features under the generalized perceptron learning framework. Experiments on both indomain and out-of-domain corpora show that our approach outperforms the baseline and achieves state-of-the-art performance. 1
Reference: text
sentIndex sentText sentNum sentScore
1 edu Abstract In this paper we develop an approach to tackle the problem of verb selection for learners of English as a second language (ESL) by using features from the output of Semantic Role Labeling (SRL). [sent-7, score-0.541]
2 Unlike existing approaches to verb selection that use local features such as n-grams, our approach exploits semantic features which explicitly model the usage context of the verb. [sent-8, score-0.904]
3 The verb choice highly depends on its usage context which is not consistently captured by local features. [sent-9, score-0.693]
4 We then combine these semantic features with other local features under the generalized perceptron learning framework. [sent-10, score-0.444]
5 ”, the verb “gained” may indicate that the “scholarship” was competitive and required the agent’s efforts; in contrast, “got” sounds neutral and less descriptive. [sent-16, score-0.352]
6 was visiting Mi- 1068 Since verbs carry multiple important functions, misusing them can be misleading, e. [sent-18, score-0.135]
7 , 2008), more than 30% of the errors in the Chinese Learner English Corpus (CLEC) are verb choice errors. [sent-22, score-0.387]
8 Hence, it is useful to develop an approach to automatically detect and correct verb selection errors made by ESL learners. [sent-23, score-0.496]
9 However, verb selection is a challenging task because verbs often exhibit a variety of usages and each usage depends on a particular context, which can hardly be adequately described by conventional n-gram features. [sent-24, score-0.851]
10 Some researchers (Tetreault and Chodorow, 2008) exploited syntactic information and n-gram features to represent verb usage context. [sent-27, score-0.623]
11 (2008) introduced an unsupervised web-based proofing method for correcting verb-noun collocation errors. [sent-29, score-0.251]
12 (2006) employed phrasal Statistical Machine Translation (SMT) techniques to correct countability errors. [sent-31, score-0.156]
13 c od2s01 in0 N Aastsuorcaialt Lioan g foura Cgeom Prpoucteastisoin ga,l p Laignegsui 1s0ti6c8s–1076, Unlike the other papers, we derive features from the output of an SRL (Màrquez, 2009) system to explicitly model verb usage context. [sent-35, score-0.623]
14 SRL is generally understood as the task of identifying the argu- ments of a given verb and assigning them semantic labels describing the roles they play. [sent-36, score-0.488]
15 For example, given a sentence “I want to watch TV tonight” and the target predicate “watch”, the output of SRL will be something like “I [A0] want to watch [target predicate] TV [A1] tonight [AM-TMP]. [sent-37, score-0.784]
16 ”, meaning that the action “watch” is conducted by the agent “I”, on the patient “TV”, and the action happens “tonight”. [sent-38, score-0.22]
17 For example, in “I want to watch a match tonight. [sent-40, score-0.321]
18 (ii) Predicate-argument structures abstract away syntactic differences in sentences with similar meanings, and therefore can potentially filter out lots of noise from the usage context. [sent-42, score-0.272]
19 For example, consider “I want to watch a football match on TV tonight”: if “match” is successfully identified as the agent of “watch”, “watch football”, which is unrelated to the us… … age of “watch” in this case, can be easily excluded from the usage context. [sent-43, score-0.633]
20 Experimental results show that SRL-derived features are effective in verb selection, but we also observe that noise in SRL output adversely increases feature space dimensions and the number of false suggestions. [sent-46, score-0.46]
21 We propose to exploit SRL-derived features to explicitly model verb usage context. [sent-51, score-0.623]
22 We propose to use the generalized perceptron framework to integrate SRL-derived (and other) features and achieve state-ofthe-art performance on both in-domain and out-of-domain test sets. [sent-53, score-0.276]
23 Ye and Baldwin (2006) apply semantic role–related information to verb sense disambiguation. [sent-61, score-0.43]
24 Narayanan and Harabagiu (2004) use semantic role structures for question answering. [sent-62, score-0.158]
25 However, in the context of ESL error detection and correction, little study has been carried out on clearly exploiting semantic information. [sent-65, score-0.115]
26 The SMT approach on the artificial data set achieves encouraging results for correcting countability errors. [sent-69, score-0.148]
27 (2008) use web frequency counts to identify and correct determiner and verb-noun collocation errors. [sent-71, score-0.136]
28 Compared with these methods, our approach explicitly models verb usage context by leveraging the SRL output. [sent-72, score-0.616]
29 The SRL-based semantic features are integrated, along with the local features, into the generalized perceptron model. [sent-73, score-0.4]
30 Given as input an English sentence written by ESL learners, the system first checks every verb and generates correction candidates by replacing each verb with its confusion set. [sent-75, score-1.104]
31 Then a feature vector that represents verb usage context is derived from the outputs of an SRL sys- tem and then multiplied with the feature weight vector trained by the generalized perceptron. [sent-76, score-0.776]
32 For example, “sees” is a checkpoint in “Jane sees TV every day. [sent-82, score-0.16]
33 Correction candidates are generated by replacing each checkpoint with its confusions. [sent-84, score-0.195]
34 Table 1 shows a sentence with one checkpoint and the corresponding correction candidates. [sent-85, score-0.268]
35 One state-of-the-art SRL system (Riedel and Meza-Ruiz, 2008) is then utilized to extract predicate-argument structures for each verb in the input, as illustrated in Table 2. [sent-90, score-0.397]
36 Semantic features are generated by combining the predicate with each of its arguments; e. [sent-91, score-0.112]
37 , “watches A0 Jane”, “sees A0 Jane”, “watch_A _J _A _J es A1 TV” and “sees A1 TV” are semantic fea_A _T _A _T tures derived from the semantic roles listed in Table 2. [sent-93, score-0.214]
38 Each correction candidate s is represented as a feature vector (s ) , where d is the total number of features. [sent-97, score-0.246]
39 c28o re In the above framework, the basic idea is to generate correction candidates with the help of predefined confusion sets and apply the global linear model to each candidate to compute the degree of its fitness to the usage context that is represented as features derived from SRL results. [sent-102, score-0.712]
40 To make our idea practical, we need to solve the following three subtasks: (i) generating the confusion set that includes possible replacements for a given verb; (ii) representing the context with semantic features and other complementary features; and (iii) training the feature weight. [sent-103, score-0.288]
41 2 Generation of Verb Confusion Sets Verb confusion sets are used to generate correction candidates. [sent-109, score-0.301]
42 Due to the great number of verbs and their diversified usages, manually collecting all verb confusions in all scenarios is prohibitively time-consuming. [sent-110, score-0.528]
43 For every selected verb we manually compile a confusion set using the following data sources: 1. [sent-112, score-0.481]
44 We extract all the synonyms of verbs from the Microsoft Encarta Dictionary, and this forms the major source for our confusion sets. [sent-114, score-0.264]
45 Therefore English verbs in the dictionary sharing more than two Chinese meanings are collected. [sent-120, score-0.179]
46 For example, “see” and “read” are in a confusion set because they share the meanings of both “看” (“to see”, “to read”) and “领 领会” (“to grasp”) in Chinese. [sent-121, score-0.173]
47 We extract paraphrasing verb expressions from a phrasal SMT translation table learnt from parallel corpora (Och and Ney, 2004). [sent-124, score-0.403]
48 This may help us use the implicit semantics of verbs that SMT can capture but a dictionary cannot, such as the fact that the verb Note that verbs in any confusion set that we are not interested in are dropped, and that the verb itself is included in its own confusion set. [sent-125, score-1.232]
49 We leave it to our future work to automatically construct verb confusions. [sent-126, score-0.352]
50 3 Verb Usage Context Features The verb usage refers to its surrounding text, which influences the way one understands the expression. [sent-128, score-0.579]
51 Intuitively, verb usage context can take the form of a collocation, e. [sent-129, score-0.616]
52 We use features derived from the SRL output to represent verb usage context. [sent-135, score-0.623]
53 The SRL system ac- context1 … … … cepts a sentence as input and outputs all arguments and the semantic roles they play for every verb in the sentence. [sent-136, score-0.572]
54 ” and the predicate “opened”, the output of SRL is listed in Table 4, where A0 and A1 are two core roles, representing the agent and patient of an action, respectively, and other roles starting with “AM-”are adjunct roles, e. [sent-138, score-0.27]
55 Predicate-argument structures keep the key participants of a given verb while dropping other unrelated words from its usage context. [sent-141, score-0.624]
56 So “say _ Chinese”, which is irrelevant to the usage of said, is not extracted as a feature. [sent-144, score-0.227]
57 The SRL system, however, may output erroneous predicate-argument structures, which negatively affect the performance of verb selection. [sent-145, score-0.352]
58 For instance, for the sentence “He hasn ’t done anything but take [make] a lot of money”, “lot” is incorrectly identified as the patient of “take”, making it hard to select “make” as the proper verb even though “make money” forms a sound collocation. [sent-146, score-0.448]
59 Back-off features are generated by replacing the word with its POS tag to alleviate data sparseness. [sent-148, score-0.116]
60 4 Perceptron Learning We choose the generalized perceptron algorithm as our training method because of its easy implementation and its capability of incorporating various features. [sent-152, score-0.232]
61 However, there are still two concerns about this perceptron learning approach: its ineffectiveness in dealing with inseparable samples and its ignorance of weight normalization that potentially limits its ability to generalize. [sent-153, score-0.173]
62 4 we show that the training error rate drops significantly to a very low level after several rounds of training, suggesting that the correct candidates can almost be separated from others. [sent-155, score-0.15]
63 We leave it to our future work to replace perceptron learning with other models like Support Vector Machines (Vapnik, 1995). [sent-157, score-0.173]
64 GEN(si ) is the function that outputs all the possible corrections for the input sentence si with each checkpoint substituted by one of its confusions, as described in Section 3. [sent-160, score-0.171]
65 We observe that the generated candidates sometimes contain reasonable outputs for the verb selection task, which 1072 should be removed. [sent-162, score-0.526]
66 For instance, in “… reporters could not take [make] notes or tape the conversa- tion”, both “take” and “make” are suitable verbs in this context. [sent-163, score-0.176]
67 A vector field is filled with 1 if the corresponding feature exists, or 0 otherwise; w is the feature weight vector, where positive elements suggest that the corresponding features support the hypothesis that the candidate is correct. [sent-170, score-0.118]
68 For example, when o is “I want to look TV” and si is “I want to watch TV”, w will be updated. [sent-172, score-0.391]
69 The averaged perceptron weight vector is defined as T1Ni1. [sent-174, score-0.205]
70 Furthermore, we study the contribution of predicate-argument-related features, and the performances on verbs with varying distance to their arguments. [sent-185, score-0.135]
71 1 Experiment Preparation The training corpus for perceptron learning was taken from LDC2005T12. [sent-187, score-0.173]
72 We randomly selected newswires containing target verbs from the New York Times as the training data. [sent-188, score-0.135]
73 We assume that the newswire data is of high quality and free of linguistic errors, and finally we gathered 20000 sentences that contain any of the target verbs we were focusing on. [sent-191, score-0.135]
74 To build the out-of-domain test dataset, we gathered 186 samples that contained errors related to the verbs we were interested in from English blogs written by Chinese and from the CLEC corpus, which were then corrected by an English native speaker. [sent-196, score-0.203]
75 Furthermore, for every error involving the verbs in our target list, both the verb and the word that determines the error are marked by the English native speaker. [sent-197, score-0.52]
76 , 2008): revised precision (RP), recall of the correction (RC) and false alarm (FA). [sent-207, score-0.366]
77 It measures the system’s coverage of verb selection errors. [sent-212, score-0.428]
78 FA # of Incorrect Modified Checkpoints (6) # of All Checkpoints FA is related to the cases where a correct verb is mistakenly replaced by an inappropriate one. [sent-213, score-0.422]
79 4 Results and Analysis The training error curves of perceptron learning with different feature sets are shown in Figure 2. [sent-216, score-0.173]
80 They drop to a low error rate and then stabilize after a few number of training rounds, indicating that most of the cases are linearly separable and that perceptron learning is applicable to the verb selection task. [sent-217, score-0.601]
81 We conducted feature selection by dropping features that occur less than N times. [sent-218, score-0.12]
82 We observe that, after feature selection, some useful features such as “watch_A1_TV” and “see_A1_TV” were kept, but some noisy features like “Jane_A0_sees ” and “Jane_A0_watches ” were removed, suggesting the effectiveness of this feature selection approach. [sent-220, score-0.164]
83 In the in-domain test, the SMT-based approach has the highest false alarm rate, though its output with word insertions or deletions is not considered wrong if the substituted verb is correct. [sent-225, score-0.512]
84 Our approach, regardless of what feature sets are used, outperforms the SMT-based approach in terms of all metrics, showing the effectiveness of perceptron learning for the verb selection task. [sent-226, score-0.601]
85 Under the perceptron learning framework, we can see that the system using only SRL-related features has higher revised precision and recall of correction, but also a slightly higher false alarm rate than the system based on only local features. [sent-227, score-0.457]
86 When local features and SRL-derived features are integrated together, the state-of-the-art performance is achieved with a 5% increase in recall, and minor changes in precision and false alarm. [sent-228, score-0.198]
87 In the out-of-domain test, the SMT-based approach performs much better than in the in-domain test, especially in terms of false alarm rate, indicating the SMT-based approach may favor short sentences. [sent-229, score-0.16]
88 We observe similar performance differences between the systems with different feature sets under the same perceptron learning framework, reaffirming the usefulness of the SRL-based features for verb selection. [sent-231, score-0.569]
89 Furthermore, we studied the performance of our system on verbs with varying distance to their arguments on the out-of-domain test corpus. [sent-236, score-0.182]
90 Table 9 shows that the system with only SRLderived features performs significantly better than the system with only local features on the verb whose usage depends on a distant argument, i. [sent-238, score-0.744]
91 In this case, the longdistance feature devised from AM-MNR helps select the suitable verb, while the trigram features cannot because they cannot represent the long distance verb usage context. [sent-261, score-0.623]
92 However, the SRL system regards “is”, the predicate of the clause, as the patient, resulting in features like “doubt_A1_is” and “suspect_A1_is”, which capture nothing about verb usage context. [sent-266, score-0.691]
93 Besides its incapability of handling verb selection errors involving clauses, the SRL-derived features fail to work when verb selection depends on deep meanings that cannot be captured by current shallow predicate-argument structures. [sent-269, score-1.01]
94 5 Conclusions and Future Work Verb selection is challenging because verb usage highly depends on the usage context, which is hard to capture and represent. [sent-274, score-0.913]
95 In this paper, we propose to utilize the output of an SRL system to explicitly model verb usage context. [sent-275, score-0.579]
96 We also propose to use the generalized perceptron learning framework to integrate SRL-derived features with other features. [sent-276, score-0.276]
97 Experimental results show that our method outperforms the SMT-based system and achieves state- 1075 of-the-art performance when SRL-related features and other local features are integrated. [sent-277, score-0.134]
98 We also show that, for cases where the particular verb usage mainly depends on its distant arguments, a system with only SRL-derived features performs much better than the system with only local features. [sent-278, score-0.7]
99 In the future, we plan to automatically construct confusion sets, expand our approach to more verbs and test our approach on a larger size of real data. [sent-279, score-0.264]
100 Discriminative training methods for hidden Markov models: theory and experiments with perceptron algorithms. [sent-297, score-0.173]
wordName wordTfidf (topN-words)
[('srl', 0.449), ('verb', 0.352), ('watch', 0.289), ('usage', 0.227), ('esl', 0.223), ('perceptron', 0.173), ('correction', 0.172), ('tv', 0.144), ('verbs', 0.135), ('confusion', 0.129), ('collocation', 0.103), ('alarm', 0.096), ('checkpoint', 0.096), ('patient', 0.096), ('smt', 0.092), ('brockett', 0.08), ('semantic', 0.078), ('correcting', 0.076), ('selection', 0.076), ('tonight', 0.074), ('countability', 0.072), ('killing', 0.072), ('proofing', 0.072), ('wear', 0.072), ('learners', 0.069), ('predicate', 0.068), ('sees', 0.064), ('false', 0.064), ('clec', 0.062), ('jane', 0.062), ('candidates', 0.061), ('generalized', 0.059), ('roles', 0.058), ('rounds', 0.056), ('chinese', 0.054), ('phrasal', 0.051), ('agent', 0.048), ('boss', 0.048), ('checkpoints', 0.048), ('habit', 0.048), ('opened', 0.048), ('raincoat', 0.048), ('scholarship', 0.048), ('spending', 0.048), ('srlderived', 0.048), ('tlh', 0.048), ('wandering', 0.048), ('rc', 0.048), ('surdeanu', 0.048), ('riedel', 0.048), ('arguments', 0.047), ('local', 0.046), ('telephone', 0.045), ('structures', 0.045), ('meanings', 0.044), ('features', 0.044), ('jianfeng', 0.043), ('candidate', 0.042), ('narayanan', 0.041), ('rquez', 0.041), ('confusions', 0.041), ('gui', 0.041), ('tetreault', 0.041), ('encarta', 0.041), ('manipulations', 0.041), ('gen', 0.041), ('notes', 0.041), ('yi', 0.041), ('experimentally', 0.039), ('si', 0.038), ('action', 0.038), ('fa', 0.038), ('replacing', 0.038), ('rf', 0.037), ('indomain', 0.037), ('football', 0.037), ('changning', 0.037), ('mistakenly', 0.037), ('cap', 0.037), ('outputs', 0.037), ('context', 0.037), ('rp', 0.036), ('errors', 0.035), ('role', 0.035), ('mihai', 0.034), ('shanghai', 0.034), ('revised', 0.034), ('dolan', 0.034), ('alleviate', 0.034), ('gao', 0.034), ('correct', 0.033), ('native', 0.033), ('money', 0.032), ('harabagiu', 0.032), ('sanda', 0.032), ('vector', 0.032), ('want', 0.032), ('depends', 0.031), ('said', 0.031), ('usages', 0.03)]
simIndex simValue paperId paperTitle
same-paper 1 0.99999934 95 emnlp-2010-SRL-Based Verb Selection for ESL
Author: Xiaohua Liu ; Bo Han ; Kuan Li ; Stephan Hyeonjun Stiller ; Ming Zhou
Abstract: In this paper we develop an approach to tackle the problem of verb selection for learners of English as a second language (ESL) by using features from the output of Semantic Role Labeling (SRL). Unlike existing approaches to verb selection that use local features such as n-grams, our approach exploits semantic features which explicitly model the usage context of the verb. The verb choice highly depends on its usage context which is not consistently captured by local features. We then combine these semantic features with other local features under the generalized perceptron learning framework. Experiments on both indomain and out-of-domain corpora show that our approach outperforms the baseline and achieves state-of-the-art performance. 1
2 0.34761459 68 emnlp-2010-Joint Inference for Bilingual Semantic Role Labeling
Author: Tao Zhuang ; Chengqing Zong
Abstract: We show that jointly performing semantic role labeling (SRL) on bitext can improve SRL results on both sides. In our approach, we use monolingual SRL systems to produce argument candidates for predicates in bitext at first. Then, we simultaneously generate SRL results for two sides of bitext using our joint inference model. Our model prefers the bilingual SRL result that is not only reasonable on each side of bitext, but also has more consistent argument structures between two sides. To evaluate the consistency between two argument structures, we also formulate a log-linear model to compute the probability of aligning two arguments. We have experimented with our model on Chinese-English parallel PropBank data. Using our joint inference model, F1 scores of SRL results on Chinese and English text achieve 79.53% and 77.87% respectively, which are 1.52 and 1.74 points higher than the results of baseline monolingual SRL combination systems respectively.
3 0.20836684 54 emnlp-2010-Generating Confusion Sets for Context-Sensitive Error Correction
Author: Alla Rozovskaya ; Dan Roth
Abstract: In this paper, we consider the problem of generating candidate corrections for the task of correcting errors in text. We focus on the task of correcting errors in preposition usage made by non-native English speakers, using discriminative classifiers. The standard approach to the problem assumes that the set of candidate corrections for a preposition consists of all preposition choices participating in the task. We determine likely preposition confusions using an annotated corpus of nonnative text and use this knowledge to produce smaller sets of candidates. We propose several methods of restricting candidate sets. These methods exclude candidate prepositions that are not observed as valid corrections in the annotated corpus and take into account the likelihood of each preposition confusion in the non-native text. We find that restricting candidates to those that are ob- served in the non-native data improves both the precision and the recall compared to the approach that views all prepositions as possible candidates. Furthermore, the approach that takes into account the likelihood of each preposition confusion is shown to be the most effective.
4 0.20419674 121 emnlp-2010-What a Parser Can Learn from a Semantic Role Labeler and Vice Versa
Author: Stephen Boxwell ; Dennis Mehay ; Chris Brew
Abstract: In many NLP systems, there is a unidirectional flow of information in which a parser supplies input to a semantic role labeler. In this paper, we build a system that allows information to flow in both directions. We make use of semantic role predictions in choosing a single-best parse. This process relies on an averaged perceptron model to distinguish likely semantic roles from erroneous ones. Our system penalizes parses that give rise to low-scoring semantic roles. To explore the consequences of this we perform two experiments. First, we use a baseline generative model to produce n-best parses, which are then re-ordered by our semantic model. Second, we use a modified version of our semantic role labeler to predict semantic roles at parse time. The performance of this modified labeler is weaker than that of our best full SRL, because it is restricted to features that can be computed directly from the parser’s packed chart. For both experiments, the resulting semantic predictions are then used to select parses. Finally, we feed the selected parses produced by each experiment to the full version of our semantic role labeler. We find that SRL performance can be improved over this baseline by selecting parses with likely semantic roles.
5 0.13140328 92 emnlp-2010-Predicting the Semantic Compositionality of Prefix Verbs
Author: Shane Bergsma ; Aditya Bhargava ; Hua He ; Grzegorz Kondrak
Abstract: In many applications, replacing a complex word form by its stem can reduce sparsity, revealing connections in the data that would not otherwise be apparent. In this paper, we focus on prefix verbs: verbs formed by adding a prefix to an existing verb stem. A prefix verb is considered compositional if it can be decomposed into a semantically equivalent expression involving its stem. We develop a classifier to predict compositionality via a range of lexical and distributional features, including novel features derived from web-scale Ngram data. Results on a new annotated corpus show that prefix verb compositionality can be predicted with high accuracy. Our system also performs well when trained and tested on conventional morphological segmentations of prefix verbs.
6 0.084232509 24 emnlp-2010-Automatically Producing Plot Unit Representations for Narrative Text
7 0.08268445 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation
8 0.072850697 56 emnlp-2010-Hashing-Based Approaches to Spelling Correction of Personal Names
9 0.066896349 117 emnlp-2010-Using Unknown Word Techniques to Learn Known Words
10 0.062020201 107 emnlp-2010-Towards Conversation Entailment: An Empirical Investigation
11 0.061275344 69 emnlp-2010-Joint Training and Decoding Using Virtual Nodes for Cascaded Segmentation and Tagging Tasks
12 0.058430672 21 emnlp-2010-Automatic Discovery of Manner Relations and its Applications
13 0.054581497 33 emnlp-2010-Cross Language Text Classification by Model Translation and Semi-Supervised Learning
14 0.054368921 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
15 0.053797271 39 emnlp-2010-EMNLP 044
16 0.053509083 67 emnlp-2010-It Depends on the Translation: Unsupervised Dependency Parsing via Word Alignment
17 0.053302493 47 emnlp-2010-Example-Based Paraphrasing for Improved Phrase-Based Statistical Machine Translation
18 0.052169275 63 emnlp-2010-Improving Translation via Targeted Paraphrasing
19 0.0514239 7 emnlp-2010-A Mixture Model with Sharing for Lexical Semantics
20 0.050862391 20 emnlp-2010-Automatic Detection and Classification of Social Events
topicId topicWeight
[(0, 0.218), (1, 0.033), (2, 0.066), (3, 0.24), (4, -0.065), (5, -0.001), (6, 0.033), (7, -0.483), (8, 0.182), (9, 0.162), (10, -0.147), (11, 0.301), (12, -0.169), (13, -0.059), (14, 0.049), (15, 0.06), (16, -0.047), (17, -0.005), (18, 0.115), (19, 0.13), (20, -0.014), (21, -0.092), (22, 0.045), (23, -0.007), (24, 0.036), (25, 0.064), (26, 0.11), (27, 0.072), (28, -0.063), (29, -0.045), (30, -0.045), (31, 0.036), (32, -0.025), (33, 0.005), (34, 0.061), (35, -0.07), (36, 0.019), (37, 0.057), (38, -0.036), (39, -0.014), (40, -0.051), (41, -0.01), (42, -0.037), (43, 0.049), (44, -0.037), (45, 0.047), (46, 0.024), (47, 0.029), (48, -0.004), (49, 0.045)]
simIndex simValue paperId paperTitle
same-paper 1 0.9559685 95 emnlp-2010-SRL-Based Verb Selection for ESL
Author: Xiaohua Liu ; Bo Han ; Kuan Li ; Stephan Hyeonjun Stiller ; Ming Zhou
Abstract: In this paper we develop an approach to tackle the problem of verb selection for learners of English as a second language (ESL) by using features from the output of Semantic Role Labeling (SRL). Unlike existing approaches to verb selection that use local features such as n-grams, our approach exploits semantic features which explicitly model the usage context of the verb. The verb choice highly depends on its usage context which is not consistently captured by local features. We then combine these semantic features with other local features under the generalized perceptron learning framework. Experiments on both indomain and out-of-domain corpora show that our approach outperforms the baseline and achieves state-of-the-art performance. 1
2 0.77642614 68 emnlp-2010-Joint Inference for Bilingual Semantic Role Labeling
Author: Tao Zhuang ; Chengqing Zong
Abstract: We show that jointly performing semantic role labeling (SRL) on bitext can improve SRL results on both sides. In our approach, we use monolingual SRL systems to produce argument candidates for predicates in bitext at first. Then, we simultaneously generate SRL results for two sides of bitext using our joint inference model. Our model prefers the bilingual SRL result that is not only reasonable on each side of bitext, but also has more consistent argument structures between two sides. To evaluate the consistency between two argument structures, we also formulate a log-linear model to compute the probability of aligning two arguments. We have experimented with our model on Chinese-English parallel PropBank data. Using our joint inference model, F1 scores of SRL results on Chinese and English text achieve 79.53% and 77.87% respectively, which are 1.52 and 1.74 points higher than the results of baseline monolingual SRL combination systems respectively.
3 0.53348821 54 emnlp-2010-Generating Confusion Sets for Context-Sensitive Error Correction
Author: Alla Rozovskaya ; Dan Roth
Abstract: In this paper, we consider the problem of generating candidate corrections for the task of correcting errors in text. We focus on the task of correcting errors in preposition usage made by non-native English speakers, using discriminative classifiers. The standard approach to the problem assumes that the set of candidate corrections for a preposition consists of all preposition choices participating in the task. We determine likely preposition confusions using an annotated corpus of nonnative text and use this knowledge to produce smaller sets of candidates. We propose several methods of restricting candidate sets. These methods exclude candidate prepositions that are not observed as valid corrections in the annotated corpus and take into account the likelihood of each preposition confusion in the non-native text. We find that restricting candidates to those that are ob- served in the non-native data improves both the precision and the recall compared to the approach that views all prepositions as possible candidates. Furthermore, the approach that takes into account the likelihood of each preposition confusion is shown to be the most effective.
4 0.49336559 121 emnlp-2010-What a Parser Can Learn from a Semantic Role Labeler and Vice Versa
Author: Stephen Boxwell ; Dennis Mehay ; Chris Brew
Abstract: In many NLP systems, there is a unidirectional flow of information in which a parser supplies input to a semantic role labeler. In this paper, we build a system that allows information to flow in both directions. We make use of semantic role predictions in choosing a single-best parse. This process relies on an averaged perceptron model to distinguish likely semantic roles from erroneous ones. Our system penalizes parses that give rise to low-scoring semantic roles. To explore the consequences of this we perform two experiments. First, we use a baseline generative model to produce n-best parses, which are then re-ordered by our semantic model. Second, we use a modified version of our semantic role labeler to predict semantic roles at parse time. The performance of this modified labeler is weaker than that of our best full SRL, because it is restricted to features that can be computed directly from the parser’s packed chart. For both experiments, the resulting semantic predictions are then used to select parses. Finally, we feed the selected parses produced by each experiment to the full version of our semantic role labeler. We find that SRL performance can be improved over this baseline by selecting parses with likely semantic roles.
5 0.42183292 92 emnlp-2010-Predicting the Semantic Compositionality of Prefix Verbs
Author: Shane Bergsma ; Aditya Bhargava ; Hua He ; Grzegorz Kondrak
Abstract: In many applications, replacing a complex word form by its stem can reduce sparsity, revealing connections in the data that would not otherwise be apparent. In this paper, we focus on prefix verbs: verbs formed by adding a prefix to an existing verb stem. A prefix verb is considered compositional if it can be decomposed into a semantically equivalent expression involving its stem. We develop a classifier to predict compositionality via a range of lexical and distributional features, including novel features derived from web-scale Ngram data. Results on a new annotated corpus show that prefix verb compositionality can be predicted with high accuracy. Our system also performs well when trained and tested on conventional morphological segmentations of prefix verbs.
6 0.24799719 21 emnlp-2010-Automatic Discovery of Manner Relations and its Applications
7 0.24246684 117 emnlp-2010-Using Unknown Word Techniques to Learn Known Words
8 0.2390223 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation
9 0.230122 24 emnlp-2010-Automatically Producing Plot Unit Representations for Narrative Text
10 0.22128071 7 emnlp-2010-A Mixture Model with Sharing for Lexical Semantics
11 0.21465106 56 emnlp-2010-Hashing-Based Approaches to Spelling Correction of Personal Names
12 0.19539326 107 emnlp-2010-Towards Conversation Entailment: An Empirical Investigation
13 0.19049385 29 emnlp-2010-Combining Unsupervised and Supervised Alignments for MT: An Empirical Study
14 0.18348019 103 emnlp-2010-Tense Sense Disambiguation: A New Syntactic Polysemy Task
15 0.17959645 77 emnlp-2010-Measuring Distributional Similarity in Context
16 0.17398541 2 emnlp-2010-A Fast Decoder for Joint Word Segmentation and POS-Tagging Using a Single Discriminative Model
17 0.16829941 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
18 0.16639197 45 emnlp-2010-Evaluating Models of Latent Document Semantics in the Presence of OCR Errors
19 0.16158955 99 emnlp-2010-Statistical Machine Translation with a Factorized Grammar
20 0.16097328 15 emnlp-2010-A Unified Framework for Scope Learning via Simplified Shallow Semantic Parsing
topicId topicWeight
[(5, 0.017), (10, 0.016), (12, 0.524), (29, 0.078), (30, 0.021), (32, 0.012), (52, 0.021), (56, 0.053), (66, 0.079), (72, 0.037), (76, 0.032), (87, 0.011)]
simIndex simValue paperId paperTitle
1 0.92810982 79 emnlp-2010-Mining Name Translations from Entity Graph Mapping
Author: Gae-won You ; Seung-won Hwang ; Young-In Song ; Long Jiang ; Zaiqing Nie
Abstract: This paper studies the problem of mining entity translation, specifically, mining English and Chinese name pairs. Existing efforts can be categorized into (a) a transliterationbased approach leveraging phonetic similarity and (b) a corpus-based approach exploiting bilingual co-occurrences, each of which suffers from inaccuracy and scarcity respectively. In clear contrast, we use unleveraged resources of monolingual entity co-occurrences, crawled from entity search engines, represented as two entity-relationship graphs extracted from two language corpora respectively. Our problem is then abstracted as finding correct mappings across two graphs. To achieve this goal, we propose a holistic approach, of exploiting both transliteration similarity and monolingual co-occurrences. This approach, building upon monolingual corpora, complements existing corpus-based work, requiring scarce resources of parallel or compa- rable corpus, while significantly boosting the accuracy of transliteration-based work. We validate our proposed system using real-life datasets.
2 0.90822679 101 emnlp-2010-Storing the Web in Memory: Space Efficient Language Models with Constant Time Retrieval
Author: David Guthrie ; Mark Hepple
Abstract: We present three novel methods of compactly storing very large n-gram language models. These methods use substantially less space than all known approaches and allow n-gram probabilities or counts to be retrieved in constant time, at speeds comparable to modern language modeling toolkits. Our basic approach generates an explicit minimal perfect hash function, that maps all n-grams in a model to distinct integers to enable storage of associated values. Extensions of this approach exploit distributional characteristics of n-gram data to reduce storage costs, including variable length coding of values and the use of tiered structures that partition the data for more efficient storage. We apply our approach to storing the full Google Web1T n-gram set and all 1-to-5 grams of the Gigaword newswire corpus. For the 1.5 billion n-grams of Gigaword, for example, we can store full count information at a cost of 1.66 bytes per n-gram (around 30% of the cost when using the current stateof-the-art approach), or quantized counts for 1.41 bytes per n-gram. For applications that are tolerant of a certain class of relatively innocuous errors (where unseen n-grams may be accepted as rare n-grams), we can reduce the latter cost to below 1byte per n-gram.
same-paper 3 0.87758344 95 emnlp-2010-SRL-Based Verb Selection for ESL
Author: Xiaohua Liu ; Bo Han ; Kuan Li ; Stephan Hyeonjun Stiller ; Ming Zhou
Abstract: In this paper we develop an approach to tackle the problem of verb selection for learners of English as a second language (ESL) by using features from the output of Semantic Role Labeling (SRL). Unlike existing approaches to verb selection that use local features such as n-grams, our approach exploits semantic features which explicitly model the usage context of the verb. The verb choice highly depends on its usage context which is not consistently captured by local features. We then combine these semantic features with other local features under the generalized perceptron learning framework. Experiments on both indomain and out-of-domain corpora show that our approach outperforms the baseline and achieves state-of-the-art performance. 1
4 0.43830261 54 emnlp-2010-Generating Confusion Sets for Context-Sensitive Error Correction
Author: Alla Rozovskaya ; Dan Roth
Abstract: In this paper, we consider the problem of generating candidate corrections for the task of correcting errors in text. We focus on the task of correcting errors in preposition usage made by non-native English speakers, using discriminative classifiers. The standard approach to the problem assumes that the set of candidate corrections for a preposition consists of all preposition choices participating in the task. We determine likely preposition confusions using an annotated corpus of nonnative text and use this knowledge to produce smaller sets of candidates. We propose several methods of restricting candidate sets. These methods exclude candidate prepositions that are not observed as valid corrections in the annotated corpus and take into account the likelihood of each preposition confusion in the non-native text. We find that restricting candidates to those that are ob- served in the non-native data improves both the precision and the recall compared to the approach that views all prepositions as possible candidates. Furthermore, the approach that takes into account the likelihood of each preposition confusion is shown to be the most effective.
5 0.42990011 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation
Author: Sankaranarayanan Ananthakrishnan ; Rohit Prasad ; David Stallard ; Prem Natarajan
Abstract: Production of parallel training corpora for the development of statistical machine translation (SMT) systems for resource-poor languages usually requires extensive manual effort. Active sample selection aims to reduce the labor, time, and expense incurred in producing such resources, attaining a given performance benchmark with the smallest possible training corpus by choosing informative, nonredundant source sentences from an available candidate pool for manual translation. We present a novel, discriminative sample selection strategy that preferentially selects batches of candidate sentences with constructs that lead to erroneous translations on a held-out development set. The proposed strategy supports a built-in diversity mechanism that reduces redundancy in the selected batches. Simulation experiments on English-to-Pashto and Spanish-to-English translation tasks demon- strate the superiority of the proposed approach to a number of competing techniques, such as random selection, dissimilarity-based selection, as well as a recently proposed semisupervised active learning strategy.
6 0.42069322 92 emnlp-2010-Predicting the Semantic Compositionality of Prefix Verbs
7 0.41874674 56 emnlp-2010-Hashing-Based Approaches to Spelling Correction of Personal Names
8 0.40358678 20 emnlp-2010-Automatic Detection and Classification of Social Events
9 0.38948858 107 emnlp-2010-Towards Conversation Entailment: An Empirical Investigation
10 0.38129556 8 emnlp-2010-A Multi-Pass Sieve for Coreference Resolution
11 0.37706366 19 emnlp-2010-Automatic Analysis of Rhythmic Poetry with Applications to Generation and Translation
12 0.37639302 32 emnlp-2010-Context Comparison of Bursty Events in Web Search and Online Media
13 0.37611634 63 emnlp-2010-Improving Translation via Targeted Paraphrasing
14 0.37468413 68 emnlp-2010-Joint Inference for Bilingual Semantic Role Labeling
15 0.37231946 21 emnlp-2010-Automatic Discovery of Manner Relations and its Applications
16 0.37215692 66 emnlp-2010-Inducing Word Senses to Improve Web Search Result Clustering
17 0.37083712 105 emnlp-2010-Title Generation with Quasi-Synchronous Grammar
18 0.37080812 84 emnlp-2010-NLP on Spoken Documents Without ASR
19 0.364824 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
20 0.36289603 46 emnlp-2010-Evaluating the Impact of Alternative Dependency Graph Encodings on Solving Event Extraction Tasks