acl acl2010 acl2010-25 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Rasoul Samad Zadeh Kaljahi
Abstract: Supervised semantic role labeling (SRL) systems trained on hand-crafted annotated corpora have recently achieved state-of-the-art performance. However, creating such corpora is tedious and costly, with the resulting corpora not sufficiently representative of the language. This paper describes a part of an ongoing work on applying bootstrapping methods to SRL to deal with this problem. Previous work shows that, due to the complexity of SRL, this task is not straight forward. One major difficulty is the propagation of classification noise into the successive iterations. We address this problem by employing balancing and preselection methods for self-training, as a bootstrapping algorithm. The proposed methods could achieve improvement over the base line, which do not use these methods. 1
Reference: text
sentIndex sentText sentNum sentScore
1 my Abstract Supervised semantic role labeling (SRL) systems trained on hand-crafted annotated corpora have recently achieved state-of-the-art performance. [sent-5, score-0.253]
2 This paper describes a part of an ongoing work on applying bootstrapping methods to SRL to deal with this problem. [sent-7, score-0.076]
3 We address this problem by employing balancing and preselection methods for self-training, as a bootstrapping algorithm. [sent-10, score-0.491]
4 The proposed methods could achieve improvement over the base line, which do not use these methods. [sent-11, score-0.082]
5 1 Introduction Semantic role labeling has been an active research field of computational linguistics since its introduction by Gildea and Jurafsky (2002). [sent-12, score-0.194]
6 , 2008) dedicated to semantic role labeling affirm the increasing attention to this field. [sent-16, score-0.253]
7 One important supportive factor of studying supervised statistical SRL has been the existence of hand-annotated semantic corpora for training SRL systems. [sent-17, score-0.147]
8 However, this corpus only exemplifies the semantic role assignment by selecting some illustrative examples for annotation. [sent-20, score-0.238]
9 This evince the infeasibility of building a comprehensive hand-crafted corpus of natural language useful for training a robust semantic role labeler. [sent-27, score-0.202]
10 Semi-supervised methods compensate the scarcity of labeled data by utilizing an additional and much larger amount of unlabeled data via a variety of algorithms. [sent-29, score-0.497]
11 Self-training (Yarowsky, 1995) is a semisupervised algorithm which has been well studied in the NLP area and gained promising result. [sent-30, score-0.074]
12 It iteratively extend its training set by labeling the unlabeled data using a base classifier trained on the labeled data. [sent-31, score-0.806]
13 Although the algo- rithm is theoretically straightforward, it involves a large number of parameters, highly influenced by the specifications of the underlying task. [sent-32, score-0.076]
14 0c S20tu1d0e Ants Roecsiea tirconh f Woror Cksomhop u,t pa tgioensa 9l1 L창€“in9g6u,istics small number of seed words extracted from an online dictionary using a corpus of unannotated English text and gained a comparable accuracy to fully supervised approaches. [sent-38, score-0.149]
15 These studies show that the performance of self- training is tied with its several parameters and the specifications of the underlying task. [sent-43, score-0.12]
16 In SRL field, He and Gildea (2006) used selftraining to address the problem of unseen frames when using FrameNet as the underlying training corpus. [sent-44, score-0.143]
17 They generalized FrameNet frame elements to 15 thematic roles to control the complexity of the process. [sent-45, score-0.092]
18 The improvement gained by the progress of self-training was small and inconsistent. [sent-46, score-0.071]
19 They reported that the NULL label (non-argument) had often dominated other labels in the examples added to the training set. [sent-47, score-0.12]
20 Using Propbank instead of FrameNet, they aimed at increasing the performance of supervised SRL system by exploiting a large amount of unlabeled data (about 7 times more than labeled data). [sent-50, score-0.484]
21 They achieved a minor improvement too and credited it to the relatively poor performance of their base classifier and the insufficiency of the unlabeled data. [sent-52, score-0.593]
22 Architecture: A two-stage pipeline architecture is used, where in the first stage less-probable argument candidates (samples) in the parse tree are pruned, and in the next stage, final arguments are identified and assigned a semantic role. [sent-57, score-0.248]
23 However, for unlabeled data, a preprocessing stage identifies the verb predicates based on the POS tag assigned by the parser. [sent-58, score-0.388]
24 The joint argument identification and classification is chosen to decrease the complexity of self-training process. [sent-59, score-0.087]
25 Features marked with * are used in addition to common features in the literature, due to their impact on the performance in feature selection process. [sent-62, score-0.121]
26 Classifier: We chose a Maximum Entropy classifier for its efficient training time and also its built-in multi-classification capability. [sent-63, score-0.19]
27 Moreover, the probability score that it assigns to labels is useful in selection process in self-training. [sent-64, score-0.156]
28 The size of seed labeled data set L and unlabeled data U, and their ratio are the fundamental parameters in any semi-supervised learning. [sent-84, score-0.576]
29 In addition to performance, efficiency of the classifier (C) is important for self-training, which is computationally expensive. [sent-87, score-0.147]
30 Our classifier is a compromise between performance and efficiency. [sent-88, score-0.147]
31 2008) when trained on the whole labeled training set. [sent-90, score-0.162]
32 Stop criterion (S) can be set to a predetermined number of iterations, finishing all of the unlabeled data, or convergence of the process in terms of improvement. [sent-91, score-0.349]
33 In each iteration, one can label entire the unlabeled data or only a portion of it. [sent-93, score-0.32]
34 In the latter case, a number of unlaleled examples (p) are selected and loaded into a pool (P). [sent-94, score-0.168]
35 The selection can be based on a specific strategy, known as preselection (Abney, 2008) or simply done according to the original order of the unlabeled data. [sent-95, score-0.773]
36 After labeling the p unlabeled data, training set is augmented by adding the newly labeled data. [sent-97, score-0.634]
37 Two main parameters are involved in this step: selection of labeled examples to be added to training set and addition of them to that set. [sent-98, score-0.357]
38 Selection is the crucial point of self-training, in which the propagation of labeling noise into upcoming iterations is the major concern. [sent-99, score-0.2]
39 One can select all of labeled examples, but usually only a number of them (n), known as growth size, based on a quality measure is selected. [sent-100, score-0.223]
40 To prevent poor labelings diminishing the quality of training set, a threshold (t) is set on this confidence score. [sent-102, score-0.147]
41 The selected labeled examples can be retained in unlabeled set to be labeled again in next iterations (delibility) or moved so that they are labeled only once (indelibility). [sent-105, score-0.762]
42 2 Preselection While using a pool can improve the efficiency of the self-training process, there can be two other motivations behind it, concerned with the performance of the process. [sent-108, score-0.126]
43 One idea is that when all data is labeled, since the growth size is often much smaller than the labeled size, a uniform set of examples preferred by the classifier is chosen in each iteration. [sent-109, score-0.457]
44 This leads to a biased classifier like the one discussed in previous section. [sent-110, score-0.147]
45 Limiting the labeling size to a pool and at the same time (pre)selecting divergence examples into it can remedy the problem. [sent-111, score-0.335]
46 The other motivation is originated from the fact that the base classifier is relatively weak due to small seed size, thus its predictions, as the measure of confidence in selection process, may not be reliable. [sent-112, score-0.411]
47 Preselecting a set of unlabeled examples more probable to be correctly labeled by the classifier in initial steps seems to be a useful strategy against this fact. [sent-113, score-0.628]
48 We examine both ideas here, by a random preselection for the first case and a measure of simplicity for the second case. [sent-114, score-0.415]
49 Random preselection is built into our system, since we use randomized 93 training data. [sent-115, score-0.375]
50 As the measure of simplicity, we propose the number of samples extracted from each sentence; that is we sort unlabeled sentences in ascending order based on the number of samples and load the pool from the beginning. [sent-116, score-0.732]
51 Semantic role labeling is a multi-class classification problem with an unbalanced distribution of classes in a given text. [sent-119, score-0.352]
52 For example, the frequency of A1 as the most frequent role in CoNLL training set is 84,917, while the frequency of 21 roles is less than 20. [sent-120, score-0.207]
53 The situation becomes worse when the dominant label NULL (for non-arguments) is added for argument identification purpose in a joint architecture. [sent-121, score-0.087]
54 In previous work, although they used a reduced set of roles (yet not balanced), He and Gildea (2006) and Lee et al. [sent-123, score-0.092]
55 (2007), did not discriminate between roles when selecting highconfidence labeled samples. [sent-124, score-0.278]
56 The former study reports that the majority of labels assigned to samples were NULL and argument labels appeared only in last iterations. [sent-125, score-0.315]
57 To attack this problem, we propose a natural way of balancing, in which instead of labeling and selection based on argument samples, we perform a sentence-based selection and labeling. [sent-126, score-0.451]
58 The idea is that argument roles are distributed over the sentences. [sent-127, score-0.179]
59 As the measure for selecting a labeled sentence, the average of the probabilities assigned by the classifier to all argument samples extracted from the sentence is used. [sent-128, score-0.549]
60 5 Experiments and Results In these experiments, we target two main problems addressed by semi-supervised methods: the performance of the algorithm in exploiting unlabeled data when labeled data is scarce and the domain-generalizability of the algorithm by using an out-of-domain unlabeled data. [sent-129, score-0.759]
61 1 The Data The labeled data are selected from Propbank corpus prepared for CoNLL 2005 shared task. [sent-133, score-0.16]
62 Our learning curve experiments on varying size of labeled data shows that the steepest increase in F1 is achieved by 1/10th of CoNLL training data. [sent-134, score-0.207]
63 Therefore, for training a base classifier as highperformance as possible, while simulating the labeled data scarcity with a reasonably small amount of it, 4000 sentence are selected randomly from the total 39,832 training sentences as seed data (L). [sent-135, score-0.557]
64 These sentences contain 71,400 argument samples covering 38 semantic roles out of 52 roles present in the total training set. [sent-136, score-0.532]
65 We use one unlabeled training set (U) for indomain and another for out-of-domain experiments. [sent-137, score-0.443]
66 The former is the remaining portion of CoNLL training data and contains 35,832 sentences (698,567 samples). [sent-138, score-0.075]
67 We also excluded the biomed section due to its large size to retain the domain balance of the data. [sent-141, score-0.073]
68 First, we do not exclude the argument roles not present in seed data when evaluating the results. [sent-145, score-0.239]
69 2 The Effect of Balanced Selection Figures 2 and 3 depict the results of using unbalanced and balanced selection with WSJ and OANC data respectively. [sent-149, score-0.417]
70 To be comparable with previous work (He and Gildea, 2006), the growth size (n) for unbalanced method is 7000 samples and for balanced method is 350 sentences, since each sentence roughly contains 20 samples. [sent-150, score-0.572]
71 The F1 of base classifier, best-performed classifier, and final classifier are marked. [sent-153, score-0.202]
72 When trained on WSJ unlabeled set, the balanced method outperforms the other in both WSJ (68. [sent-154, score-0.458]
73 A two-tail t-test based on different random selection of training data confirms the statistical significance of this improvement at p<=0. [sent-161, score-0.227]
74 org/OANC 94 F1 BWroSJw tens tet s(Ut )(U) BWrSoJw tens tet s(Bt) (B) 65 57689. [sent-165, score-0.826]
75 7091079 Number of Unlabeled Sentences Figure 2: Balanced (B) and Unbalanced (U) Selection with WSJ Unlabeled Data F1 BWrSoJw tens tet s(Rt) (R) BWrSoJw tens tet s(St) (S) 6558697. [sent-173, score-0.826]
76 However, for both test sets, the best classifier is achieved by the balanced selection (68. [sent-182, score-0.406]
77 Moreover, balanced selection shows a more normal behavior, while the other degrades the performance sharply in the last iterations (due to a swift drop of recall). [sent-189, score-0.374]
78 Consistent with previous work, with unbalanced selection, non-NULL-labeled unlabeled samples are selected only after the middle of the process. [sent-190, score-0.605]
79 But, with the balanced method, selection is more evenly distributed over the roles. [sent-191, score-0.259]
80 A comparison between the results on Brown test set with each of unlabeled sets shows that indomain data generalizes even better than out-ofdomain data (59. [sent-192, score-0.4]
81 One apparent reason is that the classifier cannot accurately label the out-of-domain unlabeled data successively used for training. [sent-196, score-0.467]
82 Furthermore, BWroSJw tens tet s(Ut )(U) BWrSoJw tens tet s(Bt) (B) 65 5697865. [sent-198, score-0.826]
83 3 The Effect of Preselection Figures 4 and 5 show the results of using pool with random and simplicity-based preselection with WSJ and OANC data respectively. [sent-208, score-0.494]
84 The pool size (p) is 2000, and growth size (n) is 1000 sentences. [sent-209, score-0.32]
85 Comparing these figures with the previous figures shows that preselection improves the selftraining trend, so that more unlabeled data can still be useful. [sent-212, score-0.848]
86 This observation was consistent with various random selection of training data. [sent-213, score-0.2]
87 Between the two strategies, simplicity-based method outperforms the random method in both self-training trend and best classifier F1 (68. [sent-214, score-0.248]
88 6 Conclusion and Future Work This work studies the application of self-training in learning semantic role labeling with the use of unlabeled data. [sent-230, score-0.573]
89 We used a balancing method for selecting newly labeled examples for augmenting the training set in each iteration of the selftraining process. [sent-231, score-0.455]
90 The idea was to reduce the effect of unbalanced distribution of semantic roles in training data. [sent-232, score-0.352]
91 We also used a pool and examined two preselection methods for loading unlabeled data into it. [sent-233, score-0.778]
92 These methods showed improvement in both classifier performance and self-training trend. [sent-234, score-0.174]
93 However, using out-of-domain unlabeled data for increasing the domain generalization ability of the system was not more useful than using indomain data. [sent-235, score-0.4]
94 Another major factor that may affect the selftraining behavior here is the poor performance of the base classifier compared to the state-of-theart (see Table 2), which exploits more complicated SRL architecture. [sent-237, score-0.346]
95 Moreover, parameter tuning process shows that other parameters such as pool-size, growth number and probability threshold are very effective. [sent-239, score-0.168]
96 We are currently planning to port this setting to co-training, another bootstrapping algorithm. [sent-241, score-0.076]
97 One direction for future work can be adapting the architecture of the SRL system to better match with the bootstrapping process. [sent-242, score-0.156]
98 Another direction can be adapting bootstrapping parameters to fit the semantic role labeling complexity. [sent-243, score-0.407]
99 Introduction to the CoNLL-2005 shared task: Semantic role labeling. [sent-262, score-0.113]
100 The CoNLL 2008 shared task on joint parsing of syntactic and semantic dependencies. [sent-322, score-0.1]
wordName wordTfidf (topN-words)
[('srl', 0.334), ('preselection', 0.332), ('unlabeled', 0.32), ('oanc', 0.262), ('tet', 0.213), ('tens', 0.2), ('bwrsojw', 0.199), ('unbalanced', 0.158), ('classifier', 0.147), ('balanced', 0.138), ('samples', 0.127), ('pool', 0.126), ('labeling', 0.122), ('selection', 0.121), ('labeled', 0.119), ('gildea', 0.114), ('wsj', 0.114), ('conll', 0.114), ('growth', 0.104), ('selftraining', 0.1), ('roles', 0.092), ('argument', 0.087), ('framenet', 0.085), ('balancing', 0.083), ('punyakanok', 0.081), ('indomain', 0.08), ('surdeanu', 0.076), ('bootstrapping', 0.076), ('marquez', 0.075), ('role', 0.072), ('bwrosjw', 0.066), ('trend', 0.065), ('propbank', 0.062), ('seed', 0.06), ('semantic', 0.059), ('scarcity', 0.058), ('base', 0.055), ('yarowsky', 0.053), ('figures', 0.048), ('abney', 0.047), ('simplicity', 0.047), ('adapting', 0.046), ('size', 0.045), ('specifications', 0.045), ('kingsbury', 0.045), ('supervised', 0.045), ('gained', 0.044), ('poor', 0.044), ('iterations', 0.043), ('degrades', 0.043), ('training', 0.043), ('examples', 0.042), ('mcclosky', 0.042), ('shared', 0.041), ('null', 0.041), ('selecting', 0.038), ('johnson', 0.038), ('bt', 0.037), ('stage', 0.037), ('random', 0.036), ('maxent', 0.036), ('charniak', 0.036), ('jurafsky', 0.035), ('propagation', 0.035), ('ut', 0.035), ('labels', 0.035), ('architecture', 0.034), ('brown', 0.034), ('cardie', 0.033), ('weakly', 0.033), ('carreras', 0.033), ('parameters', 0.032), ('sentences', 0.032), ('rt', 0.032), ('baker', 0.032), ('threshold', 0.032), ('influenced', 0.031), ('assigned', 0.031), ('newly', 0.03), ('semisupervised', 0.03), ('customizing', 0.029), ('attacked', 0.029), ('finishing', 0.029), ('highconfidence', 0.029), ('origination', 0.029), ('rim', 0.029), ('ruleset', 0.029), ('samad', 0.029), ('sharply', 0.029), ('suitability', 0.029), ('tedious', 0.029), ('balance', 0.028), ('confidence', 0.028), ('comprehensive', 0.028), ('improvement', 0.027), ('cl', 0.027), ('lee', 0.027), ('american', 0.027), ('illustrative', 0.027), ('magnified', 0.027)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000002 25 acl-2010-Adapting Self-Training for Semantic Role Labeling
Author: Rasoul Samad Zadeh Kaljahi
Abstract: Supervised semantic role labeling (SRL) systems trained on hand-crafted annotated corpora have recently achieved state-of-the-art performance. However, creating such corpora is tedious and costly, with the resulting corpora not sufficiently representative of the language. This paper describes a part of an ongoing work on applying bootstrapping methods to SRL to deal with this problem. Previous work shows that, due to the complexity of SRL, this task is not straight forward. One major difficulty is the propagation of classification noise into the successive iterations. We address this problem by employing balancing and preselection methods for self-training, as a bootstrapping algorithm. The proposed methods could achieve improvement over the base line, which do not use these methods. 1
2 0.27925622 153 acl-2010-Joint Syntactic and Semantic Parsing of Chinese
Author: Junhui Li ; Guodong Zhou ; Hwee Tou Ng
Abstract: This paper explores joint syntactic and semantic parsing of Chinese to further improve the performance of both syntactic and semantic parsing, in particular the performance of semantic parsing (in this paper, semantic role labeling). This is done from two levels. Firstly, an integrated parsing approach is proposed to integrate semantic parsing into the syntactic parsing process. Secondly, semantic information generated by semantic parsing is incorporated into the syntactic parsing model to better capture semantic information in syntactic parsing. Evaluation on Chinese TreeBank, Chinese PropBank, and Chinese NomBank shows that our integrated parsing approach outperforms the pipeline parsing approach on n-best parse trees, a natural extension of the widely used pipeline parsing approach on the top-best parse tree. Moreover, it shows that incorporating semantic role-related information into the syntactic parsing model significantly improves the performance of both syntactic parsing and semantic parsing. To our best knowledge, this is the first research on exploring syntactic parsing and semantic role labeling for both verbal and nominal predicates in an integrated way. 1
3 0.27047649 184 acl-2010-Open-Domain Semantic Role Labeling by Modeling Word Spans
Author: Fei Huang ; Alexander Yates
Abstract: Most supervised language processing systems show a significant drop-off in performance when they are tested on text that comes from a domain significantly different from the domain of the training data. Semantic role labeling techniques are typically trained on newswire text, and in tests their performance on fiction is as much as 19% worse than their performance on newswire text. We investigate techniques for building open-domain semantic role labeling systems that approach the ideal of a train-once, use-anywhere system. We leverage recently-developed techniques for learning representations of text using latent-variable language models, and extend these techniques to ones that provide the kinds of features that are useful for semantic role labeling. In experiments, our novel system reduces error by 16% relative to the previous state of the art on out-of-domain text.
4 0.23056094 207 acl-2010-Semantics-Driven Shallow Parsing for Chinese Semantic Role Labeling
Author: Weiwei Sun
Abstract: One deficiency of current shallow parsing based Semantic Role Labeling (SRL) methods is that syntactic chunks are too small to effectively group words. To partially resolve this problem, we propose semantics-driven shallow parsing, which takes into account both syntactic structures and predicate-argument structures. We also introduce several new “path” features to improve shallow parsing based SRL method. Experiments indicate that our new method obtains a significant improvement over the best reported Chinese SRL result.
5 0.20540568 216 acl-2010-Starting from Scratch in Semantic Role Labeling
Author: Michael Connor ; Yael Gertner ; Cynthia Fisher ; Dan Roth
Abstract: A fundamental step in sentence comprehension involves assigning semantic roles to sentence constituents. To accomplish this, the listener must parse the sentence, find constituents that are candidate arguments, and assign semantic roles to those constituents. Each step depends on prior lexical and syntactic knowledge. Where do children learning their first languages begin in solving this problem? In this paper we focus on the parsing and argumentidentification steps that precede Semantic Role Labeling (SRL) training. We combine a simplified SRL with an unsupervised HMM part of speech tagger, and experiment with psycholinguisticallymotivated ways to label clusters resulting from the HMM so that they can be used to parse input for the SRL system. The results show that proposed shallow representations of sentence structure are robust to reductions in parsing accuracy, and that the contribution of alternative representations of sentence structure to successful semantic role labeling varies with the integrity of the parsing and argumentidentification stages.
6 0.19948345 238 acl-2010-Towards Open-Domain Semantic Role Labeling
7 0.17331029 94 acl-2010-Edit Tree Distance Alignments for Semantic Role Labelling
8 0.16988064 146 acl-2010-Improving Chinese Semantic Role Labeling with Rich Syntactic Features
9 0.14462577 120 acl-2010-Fully Unsupervised Core-Adjunct Argument Classification
10 0.14171959 212 acl-2010-Simple Semi-Supervised Training of Part-Of-Speech Taggers
11 0.14164947 150 acl-2010-Inducing Domain-Specific Semantic Class Taggers from (Almost) Nothing
12 0.11411516 49 acl-2010-Beyond NomBank: A Study of Implicit Arguments for Nominal Predicates
13 0.10551431 89 acl-2010-Distributional Similarity vs. PU Learning for Entity Set Expansion
14 0.099589683 17 acl-2010-A Structured Model for Joint Learning of Argument Roles and Predicate Senses
15 0.091964394 78 acl-2010-Cross-Language Text Classification Using Structural Correspondence Learning
16 0.082433894 253 acl-2010-Using Smaller Constituents Rather Than Sentences in Active Learning for Japanese Dependency Parsing
17 0.078380428 258 acl-2010-Weakly Supervised Learning of Presupposition Relations between Verbs
18 0.07820414 139 acl-2010-Identifying Generic Noun Phrases
19 0.072724067 24 acl-2010-Active Learning-Based Elicitation for Semi-Supervised Word Alignment
20 0.069464192 57 acl-2010-Bucking the Trend: Large-Scale Cost-Focused Active Learning for Statistical Machine Translation
topicId topicWeight
[(0, -0.24), (1, 0.12), (2, 0.254), (3, 0.15), (4, 0.039), (5, 0.009), (6, -0.17), (7, -0.024), (8, -0.011), (9, 0.119), (10, -0.002), (11, 0.003), (12, -0.034), (13, -0.123), (14, -0.131), (15, -0.006), (16, 0.016), (17, -0.066), (18, -0.014), (19, 0.023), (20, 0.026), (21, 0.068), (22, 0.011), (23, 0.052), (24, -0.131), (25, -0.1), (26, 0.024), (27, 0.11), (28, 0.027), (29, 0.019), (30, 0.015), (31, 0.025), (32, 0.022), (33, 0.084), (34, 0.058), (35, -0.001), (36, 0.001), (37, 0.043), (38, 0.105), (39, 0.057), (40, 0.07), (41, -0.065), (42, -0.001), (43, -0.087), (44, -0.019), (45, -0.063), (46, -0.011), (47, -0.003), (48, -0.028), (49, -0.057)]
simIndex simValue paperId paperTitle
same-paper 1 0.94837248 25 acl-2010-Adapting Self-Training for Semantic Role Labeling
Author: Rasoul Samad Zadeh Kaljahi
Abstract: Supervised semantic role labeling (SRL) systems trained on hand-crafted annotated corpora have recently achieved state-of-the-art performance. However, creating such corpora is tedious and costly, with the resulting corpora not sufficiently representative of the language. This paper describes a part of an ongoing work on applying bootstrapping methods to SRL to deal with this problem. Previous work shows that, due to the complexity of SRL, this task is not straight forward. One major difficulty is the propagation of classification noise into the successive iterations. We address this problem by employing balancing and preselection methods for self-training, as a bootstrapping algorithm. The proposed methods could achieve improvement over the base line, which do not use these methods. 1
2 0.75588 184 acl-2010-Open-Domain Semantic Role Labeling by Modeling Word Spans
Author: Fei Huang ; Alexander Yates
Abstract: Most supervised language processing systems show a significant drop-off in performance when they are tested on text that comes from a domain significantly different from the domain of the training data. Semantic role labeling techniques are typically trained on newswire text, and in tests their performance on fiction is as much as 19% worse than their performance on newswire text. We investigate techniques for building open-domain semantic role labeling systems that approach the ideal of a train-once, use-anywhere system. We leverage recently-developed techniques for learning representations of text using latent-variable language models, and extend these techniques to ones that provide the kinds of features that are useful for semantic role labeling. In experiments, our novel system reduces error by 16% relative to the previous state of the art on out-of-domain text.
3 0.75027651 207 acl-2010-Semantics-Driven Shallow Parsing for Chinese Semantic Role Labeling
Author: Weiwei Sun
Abstract: One deficiency of current shallow parsing based Semantic Role Labeling (SRL) methods is that syntactic chunks are too small to effectively group words. To partially resolve this problem, we propose semantics-driven shallow parsing, which takes into account both syntactic structures and predicate-argument structures. We also introduce several new “path” features to improve shallow parsing based SRL method. Experiments indicate that our new method obtains a significant improvement over the best reported Chinese SRL result.
4 0.73709118 146 acl-2010-Improving Chinese Semantic Role Labeling with Rich Syntactic Features
Author: Weiwei Sun
Abstract: Developing features has been shown crucial to advancing the state-of-the-art in Semantic Role Labeling (SRL). To improve Chinese SRL, we propose a set of additional features, some of which are designed to better capture structural information. Our system achieves 93.49 Fmeasure, a significant improvement over the best reported performance 92.0. We are further concerned with the effect of parsing in Chinese SRL. We empirically analyze the two-fold effect, grouping words into constituents and providing syntactic information. We also give some preliminary linguistic explanations.
5 0.73099142 216 acl-2010-Starting from Scratch in Semantic Role Labeling
Author: Michael Connor ; Yael Gertner ; Cynthia Fisher ; Dan Roth
Abstract: A fundamental step in sentence comprehension involves assigning semantic roles to sentence constituents. To accomplish this, the listener must parse the sentence, find constituents that are candidate arguments, and assign semantic roles to those constituents. Each step depends on prior lexical and syntactic knowledge. Where do children learning their first languages begin in solving this problem? In this paper we focus on the parsing and argumentidentification steps that precede Semantic Role Labeling (SRL) training. We combine a simplified SRL with an unsupervised HMM part of speech tagger, and experiment with psycholinguisticallymotivated ways to label clusters resulting from the HMM so that they can be used to parse input for the SRL system. The results show that proposed shallow representations of sentence structure are robust to reductions in parsing accuracy, and that the contribution of alternative representations of sentence structure to successful semantic role labeling varies with the integrity of the parsing and argumentidentification stages.
6 0.70318878 238 acl-2010-Towards Open-Domain Semantic Role Labeling
7 0.69693053 153 acl-2010-Joint Syntactic and Semantic Parsing of Chinese
8 0.68875515 212 acl-2010-Simple Semi-Supervised Training of Part-Of-Speech Taggers
9 0.63226491 150 acl-2010-Inducing Domain-Specific Semantic Class Taggers from (Almost) Nothing
10 0.55320537 120 acl-2010-Fully Unsupervised Core-Adjunct Argument Classification
11 0.48318714 258 acl-2010-Weakly Supervised Learning of Presupposition Relations between Verbs
12 0.47793278 263 acl-2010-Word Representations: A Simple and General Method for Semi-Supervised Learning
13 0.47161317 253 acl-2010-Using Smaller Constituents Rather Than Sentences in Active Learning for Japanese Dependency Parsing
14 0.46821606 94 acl-2010-Edit Tree Distance Alignments for Semantic Role Labelling
15 0.44092876 161 acl-2010-Learning Better Data Representation Using Inference-Driven Metric Learning
16 0.44031209 49 acl-2010-Beyond NomBank: A Study of Implicit Arguments for Nominal Predicates
17 0.43058571 139 acl-2010-Identifying Generic Noun Phrases
18 0.42971641 55 acl-2010-Bootstrapping Semantic Analyzers from Non-Contradictory Texts
19 0.424187 80 acl-2010-Cross Lingual Adaptation: An Experiment on Sentiment Classifications
20 0.41804174 78 acl-2010-Cross-Language Text Classification Using Structural Correspondence Learning
topicId topicWeight
[(14, 0.016), (25, 0.079), (42, 0.023), (44, 0.018), (59, 0.1), (73, 0.053), (78, 0.08), (80, 0.327), (83, 0.096), (84, 0.02), (98, 0.119)]
simIndex simValue paperId paperTitle
1 0.88453496 19 acl-2010-A Taxonomy, Dataset, and Classifier for Automatic Noun Compound Interpretation
Author: Stephen Tratz ; Eduard Hovy
Abstract: The automatic interpretation of noun-noun compounds is an important subproblem within many natural language processing applications and is an area of increasing interest. The problem is difficult, with disagreement regarding the number and nature of the relations, low inter-annotator agreement, and limited annotated data. In this paper, we present a novel taxonomy of relations that integrates previous relations, the largest publicly-available annotated dataset, and a supervised classification method for automatic noun compound interpretation.
2 0.80640352 33 acl-2010-Assessing the Role of Discourse References in Entailment Inference
Author: Shachar Mirkin ; Ido Dagan ; Sebastian Pado
Abstract: Discourse references, notably coreference and bridging, play an important role in many text understanding applications, but their impact on textual entailment is yet to be systematically understood. On the basis of an in-depth analysis of entailment instances, we argue that discourse references have the potential of substantially improving textual entailment recognition, and identify a number of research directions towards this goal.
same-paper 3 0.78802824 25 acl-2010-Adapting Self-Training for Semantic Role Labeling
Author: Rasoul Samad Zadeh Kaljahi
Abstract: Supervised semantic role labeling (SRL) systems trained on hand-crafted annotated corpora have recently achieved state-of-the-art performance. However, creating such corpora is tedious and costly, with the resulting corpora not sufficiently representative of the language. This paper describes a part of an ongoing work on applying bootstrapping methods to SRL to deal with this problem. Previous work shows that, due to the complexity of SRL, this task is not straight forward. One major difficulty is the propagation of classification noise into the successive iterations. We address this problem by employing balancing and preselection methods for self-training, as a bootstrapping algorithm. The proposed methods could achieve improvement over the base line, which do not use these methods. 1
4 0.73155785 119 acl-2010-Fixed Length Word Suffix for Factored Statistical Machine Translation
Author: Narges Sharif Razavian ; Stephan Vogel
Abstract: Factored Statistical Machine Translation extends the Phrase Based SMT model by allowing each word to be a vector of factors. Experiments have shown effectiveness of many factors, including the Part of Speech tags in improving the grammaticality of the output. However, high quality part of speech taggers are not available in open domain for many languages. In this paper we used fixed length word suffix as a new factor in the Factored SMT, and were able to achieve significant improvements in three set of experiments: large NIST Arabic to English system, medium WMT Spanish to English system, and small TRANSTAC English to Iraqi system. 1
5 0.65200245 121 acl-2010-Generating Entailment Rules from FrameNet
Author: Roni Ben Aharon ; Idan Szpektor ; Ido Dagan
Abstract: Idan Szpektor Ido Dagan Yahoo! Research Department of Computer Science Haifa, Israel Bar-Ilan University idan @ yahoo- inc .com Ramat Gan, Israel dagan @ c s .biu . ac . i l FrameNet is a manually constructed database based on Frame Semantics. It models the semantic Many NLP tasks need accurate knowledge for semantic inference. To this end, mostly WordNet is utilized. Yet WordNet is limited, especially for inference be- tween predicates. To help filling this gap, we present an algorithm that generates inference rules between predicates from FrameNet. Our experiment shows that the novel resource is effective and complements WordNet in terms of rule coverage.
6 0.59752792 1 acl-2010-"Ask Not What Textual Entailment Can Do for You..."
7 0.57769704 127 acl-2010-Global Learning of Focused Entailment Graphs
8 0.57330638 137 acl-2010-How Spoken Language Corpora Can Refine Current Speech Motor Training Methodologies
9 0.57325399 76 acl-2010-Creating Robust Supervised Classifiers via Web-Scale N-Gram Data
10 0.57066256 198 acl-2010-Predicate Argument Structure Analysis Using Transformation Based Learning
11 0.57042038 153 acl-2010-Joint Syntactic and Semantic Parsing of Chinese
12 0.56888247 16 acl-2010-A Statistical Model for Lost Language Decipherment
13 0.56746459 101 acl-2010-Entity-Based Local Coherence Modelling Using Topological Fields
14 0.56433976 184 acl-2010-Open-Domain Semantic Role Labeling by Modeling Word Spans
15 0.56409556 30 acl-2010-An Open-Source Package for Recognizing Textual Entailment
16 0.56400847 112 acl-2010-Extracting Social Networks from Literary Fiction
17 0.56261355 55 acl-2010-Bootstrapping Semantic Analyzers from Non-Contradictory Texts
18 0.5620532 248 acl-2010-Unsupervised Ontology Induction from Text
19 0.56172723 211 acl-2010-Simple, Accurate Parsing with an All-Fragments Grammar
20 0.56124765 160 acl-2010-Learning Arguments and Supertypes of Semantic Relations Using Recursive Patterns