emnlp emnlp2011 emnlp2011-7 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Vivek Srikumar ; Dan Roth
Abstract: This paper presents a model that extends semantic role labeling. Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. However, sentences express relations via other linguistic phenomena as well. Furthermore, these phenomena interact with each other, thus restricting the structures they articulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. The scarcity of jointly labeled data presents a crucial technical challenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models.
Reference: text
sentIndex sentText sentNum sentScore
1 Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. [sent-2, score-0.331]
2 In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. [sent-5, score-0.776]
3 By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models. [sent-8, score-0.342]
4 In the literature, semantic role extraction has been studied mostly in the context of verb predicates, using the Propbank annotation of Palmer et al. [sent-11, score-0.385]
5 Verb centered semantic role labeling would identify the arguments of the predicate change as (a) The field goal by Brien (A0, the causer of the change), (b) the game (A 1, the thing changing), and (c) in the fourth quarter (temporal modifier). [sent-17, score-0.432]
6 However, this does not tell us that the scorer of the field goal was Brien, which is expressed by the preposition by. [sent-18, score-0.656]
7 In this paper, we propose an extension of the standard semantic role labeling task to include relations expressed by lexical items other than verbs and nominalizations. [sent-20, score-0.448]
8 From the machine learning standpoint, we propose a joint inference scheme to combine existing structure predictors for multiple linguistic phenomena. [sent-26, score-0.307]
9 From an NLP perspective, we motivate the extension of semantic role labeling beyond verbs and nominalizations. [sent-33, score-0.349]
10 We instantiate our joint model for the case of extracting preposition and verb relations together. [sent-34, score-0.917]
11 Our model uses existing systems that identify verb semantic roles and preposition object roles and jointly predicts the output of the two systems in the presence of linguistic constraints that enforce coherence between the predictions. [sent-35, score-1.367]
12 The rest of the paper is organized as follows: We motivate the need for extending semantic role labeling and the necessity for joint inference in Section 2. [sent-38, score-0.533]
13 In Section 3, we describe the component verb SRL and preposition role systems. [sent-39, score-0.901]
14 The usage of the preposition of is different in sentence (3), where it indicates a creatorcreation relationship. [sent-51, score-0.573]
15 In the last sentence, the same preposition tells us that the Kabale district is located in Uganda. [sent-52, score-0.636]
16 However, the same analysis can also be obtained by identifying the sense of the preposition of, which tells us that the subject of the preposition is a nominalization of the underlying verb. [sent-57, score-1.3]
17 A similar redundancy can be observed with analyses of the verb began and the preposition in. [sent-58, score-0.717]
18 1 Preposition Relations Prepositions indicate a relation between the attachment point of the preposition and its object. [sent-77, score-0.573]
19 As we have seen, the same preposition can indicate different types of relations. [sent-78, score-0.573]
20 This sense inventory formed the basis of the SemEval-2007 task of preposition word sense disambiguation of Litkowski and Hargraves (2007). [sent-80, score-0.785]
21 html 131 would be labeled with the sense 8(3) which identifies the object of the preposition as the topic, while the second instance would be labeled as 17(8), which indicates that argument is the day of the occurrence. [sent-84, score-0.984]
22 The preposition sense inventory, while useful to identify the fine grained distinctions between preposition usage, defines a unique sense label for each preposition by indexing the definitions of the prepositions in the Oxford Dictionary of English. [sent-85, score-2.072]
23 For example, in the phrase at noon, the at would be labeled with the sense 2(2), while the preposition in I will see you in an hour will be labeled 4(3). [sent-86, score-0.816]
24 To counter this problem we collapsed preposition senses that are semantically similar to define a new label space, which we refer to as Preposition Roles. [sent-88, score-0.688]
25 We retrained classifiers for preposition sense for the new label space. [sent-89, score-0.768]
26 Before describing the preposition role dataset, we briefly describe the datasets and the features for the sense problem. [sent-90, score-0.84]
27 The best performing system at the SemEval-2007 shared task of preposition sense disambiguation (Ye and Baldwin (2007)) achieves a mean precision of 69. [sent-91, score-0.702]
28 As mentioned earlier, we collapsed the sense labels onto the newly defined preposition role labels. [sent-117, score-0.888]
29 According to this labeling scheme, the first on in our running example will be labeled TOPIC and the second one will 2This dataset does not annotate all prepositions and restricts itself mainly to prepositions that start a Propbank argument. [sent-119, score-0.381]
30 Table 3: Preposition role data statistics for the Penn Treebank preposition dataset. [sent-133, score-0.757]
31 We re-trained the sense disambiguation system to predict preposition roles. [sent-135, score-0.702]
32 We use this system as our independent baseline for preposition role identification. [sent-138, score-0.757]
33 The CoNLL Shared Tasks of 2004 and 2005 (See Carreras and M `arquez 4The mapping from the preposition senses to the roles defines a new dataset and is available for download at http : / / cogcomp . [sent-141, score-0.774]
34 (2008) used global inference to ensure that the predictions across all arguments of the same predicate are coherent. [sent-149, score-0.334]
35 (2008), which we briefly describe here, to serve as our baseline verb semantic role labeler 5. [sent-151, score-0.385]
36 (2008) consists of four stages candidate generation, argument identification, argument classification and inference. [sent-154, score-0.336]
37 iTcahete sinf theraten thcee step produces a combined prediction for all argument candidates of a verb proposition by enforcing global constraints. [sent-159, score-0.482]
38 edu/page / s o ftware 133 learned systems, in this case, the argument identifier and the role classifier. [sent-170, score-0.477]
39 Integer linear programs were used by Roth and Yih (2005) to add gen- eral constraints for inference with conditional random fields. [sent-173, score-0.319]
40 Let viC,abe the Boolean indicator variable that denotes that the ith argument candidate for a predicate is assigned a label a and let ΘiC,a represent the score assigned by the argument classifier for this decision. [sent-177, score-0.565]
41 Similarly, let viI denote the identifier decision for the ith argument candidate of the predicate and ΘiI denote its identifier score. [sent-178, score-0.605]
42 Then, the objective of inference is to maximize the total score of the assignment – vmCa,vxIXΘiC,aviC,a+XΘiIviI Xi,a (1) Xi Here, vC and vI denote all the argument classifier and identifier variables respectively. [sent-179, score-0.579]
43 (2008), we also have the constraints linking the predictions of the identifier and classifier: vvC,i,∅ + vvI,i = 1; ∀v, i. [sent-183, score-0.34]
44 Tleos ttroa itnak tehe v classifiers, we used parse trees from the Charniak and Johnson (2005) parser with the 6The primary advantage of using ILP for inference is that this representation enables us to add arbitrary coherence constraints between the phenomena. [sent-185, score-0.313]
45 As with the preposition roles, we implemented our system using Learning Based Java of Rizzolo and Roth (2010). [sent-190, score-0.573]
46 The problem of predicting preposition roles can be easily transformed into an ILP instance. [sent-203, score-0.679]
47 Let vpR,r denote the decision variable that encodes the prediction that the preposition p is as- signed a role r and let ΘpR,r denote its score. [sent-204, score-0.892]
48 For example, for the SRL argument classification component, the parts of the structure are all the candidates that need to be labeled for a given sentence and the set Zp is the set of all argument labels. [sent-210, score-0.469]
49 Thus, for verb SRL, these would be the constraints defined in the previous section, and for preposition role, the only local constraint would be the constraint (4) defined above. [sent-216, score-0.966]
50 (8) As a technical point, this defines one inference problem per sentence, rather than per predicate as in the verb SRL system of Punyakanok et al. [sent-219, score-0.349]
51 (2007) to study the impact of incorporating crosspredicate constraints for verb SRL. [sent-222, score-0.305]
52 1 Joint inference We consider the problem of jointly predicting several phenomena incorporating linguistic knowledge that enforce consistency between the output labels. [sent-225, score-0.445]
53 This allows us to enforce constraints of the form “If an argument that starts with the preposition ‘at’ is labeled AM-TMP, then the preposition can be labeled either NUMERIC/LEVEL or TEMPORAL. [sent-228, score-1.687]
54 ” This constraint is universally quantified for all arguments that satisfy the precondition of starting with the preposition at. [sent-229, score-0.658]
55 −vp11 +Xvip2 ≥ 0 Xi In the context of the preposition role and verb SRL, we consider constraints between labels for a preposition and SRL argument candidates that begin with that preposition. [sent-231, score-1.904]
56 The joint inference problem can be phrased as that of maximizing the score of the assignment subject to the structural constraints of each phenomenon (Cp) and the joint linguistic constraints (J). [sent-237, score-0.769]
57 (13) Here, v is the vector of inference variables which is obtained by stacking all the inference variables of each phenomena. [sent-244, score-0.316]
58 When we jointly predict verb SRL and preposition role, we have 22 preposition roles (from table 3), one SRL identifier label and 54 SRL argument classifier labels. [sent-250, score-1.924]
59 The research question addressed by the experiments is the following: Given independently trained systems for verb SRL and preposition roles, can their performance be improved using joint inference between the two tasks? [sent-260, score-0.975]
60 In all experiments, we report the F1 measure for the verb SRL performance using the CoNLL 2005 evaluation metric and the accuracy for the preposition role labeling task. [sent-266, score-0.968]
61 1 Data and Constraints For both the verb SRL and preposition roles, we used the first 500 sentences of section 2 of the Penn Treebank corpus to train our scaling parameters. [sent-268, score-0.776]
62 1, we considered joint constraints relating preposition roles to verb argument candidates that start with the preposition. [sent-274, score-1.308]
63 We identified the following types of constraints: (1) For each preposition, the set of invalid verb arguments and preposition roles. [sent-275, score-0.758]
64 (2) For each preposition role, the set of allowed verb argument labels if the role occurred more than ten times in the data, and (3) For each verb argument, the set of allowed preposition roles, similarly with a support of ten. [sent-276, score-1.834]
65 Note that, while the constraints were obtained from jointly labeled data, the constraints could be written down because they encode linguistic intuition about the labels. [sent-277, score-0.54]
66 The following is a constraint extracted from the data, which applies to the preposition with: srlarg(A2) → ∨ ∨ ∨ ∨ ∨ ∨ prep-role(ATTRIBUTE) prep-role(CAUSE) prep-role(INSTRUMENT) prep-role(OBJECTOFVERB) prep-role(PARTWHOLE) prep-role(PARTICIPANT/ACCOMPAINER) prep-role(PROFESSIONALASPECT). [sent-278, score-0.617]
67 This constraint says that if any candidate that starts with with is labeled as an A2, then the preposition can be labeled only with one of the roles on the right hand side. [sent-279, score-0.883]
68 Some of the mined constraints have negated variables to enforce that a role or an argument label should not be allowed. [sent-280, score-0.721]
69 In addition to these constraints that were mined from data, we also enforce the following handwritten constraints: (1) If the role of a verb attached preposition is labeled TEMPORAL, then there should be a verb predicate for which this preposi- tional phrase is labeled AM-TMP. [sent-283, score-1.551]
70 (2) For verb attached prepositions, ifthe preposition is labeled with one of ACTIVITY, ENDCONDITION, INSTRUMENT or PROFESSIONALASPECT, there should be at least one predicate for which the corresponding prepositional phrase is not labeled ∅. [sent-284, score-0.96]
71 For each of the roles in the second constraint, let r denote a role variable that assigns the label to some preposition. [sent-287, score-0.412]
72 Suppose there are n SRL candidates across all verb predicates begin with that preposition, and let s1, s2, · · · , sn denote the SRL variables that assign these, ·c·a·nd ,sidates to the label ∅. [sent-288, score-0.399]
73 For one pipeline, we added the prediction of the baseline preposition role system as an additional feature to both the identifier and the argument classifier for argument candidates that start with a preposition. [sent-291, score-1.378]
74 The verb SRL systems were trained on sections 2-21, while the preposition role classifiers were trained on sections 2-4. [sent-301, score-1.009]
75 For the joint inference system, the scaling parameters were trained on the first 500 sentences of section 2, which were held out. [sent-302, score-0.317]
76 Note that, for a given sentence, even if the joint constraints affect only a few argument candidates directly, they can alter the labels of the other candidates via the “local” SRL constraints. [sent-308, score-0.586]
77 The independent preposition role system incorrectly identifies the to as a LOCATION. [sent-311, score-0.757]
78 The semantic role labeling component identifies the phrase to the cancellation of the planned exchange as the A2 of the verb led. [sent-312, score-0.573]
79 One of the constraints mined from the data prohibits the label LOCATION for the preposition to if the argument it starts is labeled A2. [sent-313, score-1.102]
80 This forces the system to change the preposition label to the correct one, namely ENDCONDITION. [sent-314, score-0.643]
81 Both the independent and the joint systems also label the preposition of as OBJECTOFVERB, which indicates that the phrase the planned exchange is the object of the deverbal noun cancellation. [sent-315, score-0.834]
82 3 Effect of constraints on adaptation Our second experiment compares the performance of the preposition role classifier that has been trained 137 on the SemEval dataset with and without joint constraints. [sent-317, score-1.13]
83 Note that Table 2 in Section 3, shows the drop in performance when applying the preposition sense classifier. [sent-318, score-0.656]
84 We see that the SemEvaltrained preposition role classifier (baseline in the table) achieves an accuracy of 53. [sent-319, score-0.833]
85 Using this classifier jointly with the verb SRL classifier via joint constraints gets an improvement of almost 3 percent in accuracy. [sent-321, score-0.649]
86 ir2 oa92nciyRe)dopler ositn role classifier, when tested on the Treebank dataset with and without joint inference with the verb SRL system. [sent-323, score-0.553]
87 6 Discussion and Related work Roth and Yih (2004) formulated the problem of extracting entities and relations as an integer linear program, allowing them to use global structural constraints at inference time even though the component classifiers were trained independently. [sent-328, score-0.552]
88 In this paper, we use this idea to combine classifiers that were trained for two different tasks on different datasets using constraints to encode linguistic knowledge. [sent-329, score-0.32]
89 (2004) studied verb subcategorization and sense disambiguation of verbs by treating it as a problem of learning with partially labeled structures and proposed to use EM to train the joint model. [sent-332, score-0.497]
90 Meza-Ruiz and Riedel (2009) modeled verb SRL, predicate identification and predicate sense recognition jointly using Markov Logic. [sent-334, score-0.52]
91 (2009) addressed the problem of jointly learning verb SRL and preposition sense using the Penn Treebank annotation that was introduced in that work. [sent-338, score-0.889]
92 Moreover, our model has the advantage that the complexity of the joint parameters is small, hence does not require a large jointly labeled dataset to train the scaling parameters. [sent-340, score-0.331]
93 We consider the tasks verb SRL and preposition roles and combine their predictions to provide a richer semantic annotation of text. [sent-350, score-0.969]
94 (2004), the preposition function analysis of O’Hara and Wiebe (2009) and noun compound analysis as defined by Girju (2007) and Girju et al. [sent-353, score-0.573]
95 7 Conclusion This paper presents a strategy for extending semantic role labeling without the need for extensive retraining or data annotation. [sent-356, score-0.36]
96 While standard semantic role labeling focuses on verb and nominal relations, sentences can express relations using other lexical items also. [sent-357, score-0.507]
97 We instantiate our model using verb semantic role labeling and preposition role labeling and show that, using linguistic constraints between the tasks and minimal joint learning, we can improve the performance of both tasks. [sent-360, score-1.666]
98 Verb sense and subcategorization: Using joint inference to improve performance on complementary tasks. [sent-372, score-0.308]
99 Joint learning of preposition senses and semantic roles of prepositional phrases. [sent-415, score-0.781]
100 The importance of syntactic parsing and inference in semantic role labeling. [sent-537, score-0.363]
wordName wordTfidf (topN-words)
[('preposition', 0.573), ('srl', 0.386), ('role', 0.184), ('argument', 0.168), ('constraints', 0.161), ('verb', 0.144), ('punyakanok', 0.131), ('identifier', 0.125), ('inference', 0.122), ('prepositions', 0.117), ('roth', 0.108), ('roles', 0.106), ('joint', 0.103), ('phenomena', 0.091), ('jointly', 0.089), ('predicate', 0.083), ('sense', 0.083), ('labeled', 0.08), ('classifier', 0.076), ('label', 0.07), ('integer', 0.069), ('labeling', 0.067), ('treebank', 0.067), ('rizzolo', 0.067), ('yih', 0.062), ('litkowski', 0.06), ('pipelines', 0.06), ('inoi', 0.06), ('scaling', 0.059), ('brien', 0.058), ('planned', 0.058), ('semantic', 0.057), ('carreras', 0.056), ('relations', 0.055), ('predictions', 0.054), ('candidates', 0.053), ('enforcing', 0.052), ('retraining', 0.052), ('enforce', 0.052), ('denote', 0.052), ('cogcomp', 0.05), ('gazetteer', 0.05), ('mined', 0.05), ('linguistic', 0.049), ('labels', 0.048), ('penn', 0.047), ('propbank', 0.047), ('disambiguation', 0.046), ('senses', 0.045), ('nombank', 0.045), ('ratinov', 0.045), ('expressed', 0.044), ('predicates', 0.044), ('constraint', 0.044), ('classifiers', 0.042), ('tratz', 0.042), ('dahlmeier', 0.042), ('instantiate', 0.042), ('consistency', 0.042), ('ilp', 0.042), ('java', 0.042), ('verbs', 0.041), ('arguments', 0.041), ('riedel', 0.04), ('maximization', 0.04), ('program', 0.04), ('arquez', 0.039), ('temporal', 0.039), ('kabale', 0.039), ('objectofverb', 0.039), ('professionalaspect', 0.039), ('relativity', 0.039), ('scorer', 0.039), ('identification', 0.038), ('underlying', 0.038), ('meyers', 0.037), ('phenomenon', 0.037), ('linear', 0.036), ('variables', 0.036), ('semeval', 0.035), ('cp', 0.035), ('girju', 0.035), ('tasks', 0.035), ('global', 0.034), ('surdeanu', 0.034), ('hargraves', 0.033), ('cancellation', 0.033), ('phrased', 0.033), ('pz', 0.033), ('trained', 0.033), ('vp', 0.033), ('predictors', 0.033), ('tells', 0.033), ('pipeline', 0.032), ('prediction', 0.031), ('hovy', 0.031), ('palmer', 0.031), ('coherence', 0.03), ('district', 0.03), ('exchange', 0.03)]
simIndex simValue paperId paperTitle
same-paper 1 0.99999768 7 emnlp-2011-A Joint Model for Extended Semantic Role Labeling
Author: Vivek Srikumar ; Dan Roth
Abstract: This paper presents a model that extends semantic role labeling. Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. However, sentences express relations via other linguistic phenomena as well. Furthermore, these phenomena interact with each other, thus restricting the structures they articulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. The scarcity of jointly labeled data presents a crucial technical challenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models.
2 0.19949397 4 emnlp-2011-A Fast, Accurate, Non-Projective, Semantically-Enriched Parser
Author: Stephen Tratz ; Eduard Hovy
Abstract: Dependency parsers are critical components within many NLP systems. However, currently available dependency parsers each exhibit at least one of several weaknesses, including high running time, limited accuracy, vague dependency labels, and lack of nonprojectivity support. Furthermore, no commonly used parser provides additional shallow semantic interpretation, such as preposition sense disambiguation and noun compound interpretation. In this paper, we present a new dependency-tree conversion of the Penn Treebank along with its associated fine-grain dependency labels and a fast, accurate parser trained on it. We explain how a non-projective extension to shift-reduce parsing can be incorporated into non-directional easy-first parsing. The parser performs well when evaluated on the standard test section of the Penn Treebank, outperforming several popular open source dependency parsers; it is, to the best of our knowledge, the first dependency parser capable of parsing more than 75 sentences per second at over 93% accuracy.
3 0.18638131 145 emnlp-2011-Unsupervised Semantic Role Induction with Graph Partitioning
Author: Joel Lang ; Mirella Lapata
Abstract: In this paper we present a method for unsupervised semantic role induction which we formalize as a graph partitioning problem. Argument instances of a verb are represented as vertices in a graph whose edge weights quantify their role-semantic similarity. Graph partitioning is realized with an algorithm that iteratively assigns vertices to clusters based on the cluster assignments of neighboring vertices. Our method is algorithmically and conceptually simple, especially with respect to how problem-specific knowledge is incorporated into the model. Experimental results on the CoNLL 2008 benchmark dataset demonstrate that our model is competitive with other unsupervised approaches in terms of F1 whilst attaining significantly higher cluster purity.
4 0.10835069 75 emnlp-2011-Joint Models for Chinese POS Tagging and Dependency Parsing
Author: Zhenghua Li ; Min Zhang ; Wanxiang Che ; Ting Liu ; Wenliang Chen ; Haizhou Li
Abstract: Part-of-speech (POS) is an indispensable feature in dependency parsing. Current research usually models POS tagging and dependency parsing independently. This may suffer from error propagation problem. Our experiments show that parsing accuracy drops by about 6% when using automatic POS tags instead of gold ones. To solve this issue, this paper proposes a solution by jointly optimizing POS tagging and dependency parsing in a unique model. We design several joint models and their corresponding decoding algorithms to incorporate different feature sets. We further present an effective pruning strategy to reduce the search space of candidate POS tags, leading to significant improvement of parsing speed. Experimental results on Chinese Penn Treebank 5 show that our joint models significantly improve the state-of-the-art parsing accuracy by about 1.5%. Detailed analysis shows that the joint method is able to choose such POS tags that are more helpful and discriminative from parsing viewpoint. This is the fundamental reason of parsing accuracy improvement.
5 0.10710447 59 emnlp-2011-Fast and Robust Joint Models for Biomedical Event Extraction
Author: Sebastian Riedel ; Andrew McCallum
Abstract: Extracting biomedical events from literature has attracted much recent attention. The bestperforming systems so far have been pipelines of simple subtask-specific local classifiers. A natural drawback of such approaches are cascading errors introduced in early stages of the pipeline. We present three joint models of increasing complexity designed to overcome this problem. The first model performs joint trigger and argument extraction, and lends itself to a simple, efficient and exact inference algorithm. The second model captures correlations between events, while the third model ensures consistency between arguments of the same event. Inference in these models is kept tractable through dual decomposition. The first two models outperform the previous best joint approaches and are very competitive with respect to the current state-of-theart. The third model yields the best results reported so far on the BioNLP 2009 shared task, the BioNLP 2011 Genia task and the BioNLP 2011Infectious Diseases task.
6 0.09361773 103 emnlp-2011-Parser Evaluation over Local and Non-Local Deep Dependencies in a Large Corpus
7 0.085036427 144 emnlp-2011-Unsupervised Learning of Selectional Restrictions and Detection of Argument Coercions
8 0.083401553 14 emnlp-2011-A generative model for unsupervised discovery of relations and argument classes from clinical texts
9 0.083099961 128 emnlp-2011-Structured Relation Discovery using Generative Models
10 0.078636825 55 emnlp-2011-Exploiting Syntactic and Distributional Information for Spelling Correction with Web-Scale N-gram Models
11 0.07323537 114 emnlp-2011-Relation Extraction with Relation Topics
12 0.07236398 119 emnlp-2011-Semantic Topic Models: Combining Word Distributional Statistics and Dictionary Definitions
13 0.06898693 70 emnlp-2011-Identifying Relations for Open Information Extraction
14 0.067583658 34 emnlp-2011-Corpus-Guided Sentence Generation of Natural Images
15 0.066415995 102 emnlp-2011-Parse Correction with Specialized Models for Difficult Attachment Types
16 0.066083334 96 emnlp-2011-Multilayer Sequence Labeling
17 0.065958582 127 emnlp-2011-Structured Lexical Similarity via Convolution Kernels on Dependency Trees
18 0.064776987 137 emnlp-2011-Training dependency parsers by jointly optimizing multiple objectives
19 0.064407468 78 emnlp-2011-Large-Scale Noun Compound Interpretation Using Bootstrapping and the Web as a Corpus
20 0.064385638 45 emnlp-2011-Dual Decomposition with Many Overlapping Components
topicId topicWeight
[(0, 0.241), (1, -0.071), (2, -0.155), (3, 0.156), (4, -0.015), (5, -0.045), (6, 0.056), (7, -0.031), (8, 0.073), (9, 0.018), (10, 0.014), (11, -0.064), (12, 0.163), (13, -0.065), (14, -0.041), (15, 0.106), (16, -0.114), (17, -0.084), (18, 0.086), (19, -0.055), (20, -0.076), (21, 0.113), (22, 0.06), (23, 0.08), (24, -0.1), (25, -0.134), (26, 0.007), (27, -0.053), (28, 0.076), (29, 0.042), (30, -0.085), (31, 0.136), (32, -0.092), (33, -0.092), (34, 0.037), (35, 0.184), (36, 0.015), (37, 0.054), (38, 0.139), (39, -0.15), (40, 0.203), (41, -0.092), (42, -0.148), (43, 0.006), (44, 0.081), (45, -0.056), (46, -0.157), (47, -0.126), (48, 0.089), (49, -0.066)]
simIndex simValue paperId paperTitle
same-paper 1 0.97810262 7 emnlp-2011-A Joint Model for Extended Semantic Role Labeling
Author: Vivek Srikumar ; Dan Roth
Abstract: This paper presents a model that extends semantic role labeling. Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. However, sentences express relations via other linguistic phenomena as well. Furthermore, these phenomena interact with each other, thus restricting the structures they articulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. The scarcity of jointly labeled data presents a crucial technical challenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models.
2 0.6300813 145 emnlp-2011-Unsupervised Semantic Role Induction with Graph Partitioning
Author: Joel Lang ; Mirella Lapata
Abstract: In this paper we present a method for unsupervised semantic role induction which we formalize as a graph partitioning problem. Argument instances of a verb are represented as vertices in a graph whose edge weights quantify their role-semantic similarity. Graph partitioning is realized with an algorithm that iteratively assigns vertices to clusters based on the cluster assignments of neighboring vertices. Our method is algorithmically and conceptually simple, especially with respect to how problem-specific knowledge is incorporated into the model. Experimental results on the CoNLL 2008 benchmark dataset demonstrate that our model is competitive with other unsupervised approaches in terms of F1 whilst attaining significantly higher cluster purity.
3 0.4988324 59 emnlp-2011-Fast and Robust Joint Models for Biomedical Event Extraction
Author: Sebastian Riedel ; Andrew McCallum
Abstract: Extracting biomedical events from literature has attracted much recent attention. The bestperforming systems so far have been pipelines of simple subtask-specific local classifiers. A natural drawback of such approaches are cascading errors introduced in early stages of the pipeline. We present three joint models of increasing complexity designed to overcome this problem. The first model performs joint trigger and argument extraction, and lends itself to a simple, efficient and exact inference algorithm. The second model captures correlations between events, while the third model ensures consistency between arguments of the same event. Inference in these models is kept tractable through dual decomposition. The first two models outperform the previous best joint approaches and are very competitive with respect to the current state-of-theart. The third model yields the best results reported so far on the BioNLP 2009 shared task, the BioNLP 2011 Genia task and the BioNLP 2011Infectious Diseases task.
4 0.47691131 4 emnlp-2011-A Fast, Accurate, Non-Projective, Semantically-Enriched Parser
Author: Stephen Tratz ; Eduard Hovy
Abstract: Dependency parsers are critical components within many NLP systems. However, currently available dependency parsers each exhibit at least one of several weaknesses, including high running time, limited accuracy, vague dependency labels, and lack of nonprojectivity support. Furthermore, no commonly used parser provides additional shallow semantic interpretation, such as preposition sense disambiguation and noun compound interpretation. In this paper, we present a new dependency-tree conversion of the Penn Treebank along with its associated fine-grain dependency labels and a fast, accurate parser trained on it. We explain how a non-projective extension to shift-reduce parsing can be incorporated into non-directional easy-first parsing. The parser performs well when evaluated on the standard test section of the Penn Treebank, outperforming several popular open source dependency parsers; it is, to the best of our knowledge, the first dependency parser capable of parsing more than 75 sentences per second at over 93% accuracy.
5 0.44904199 34 emnlp-2011-Corpus-Guided Sentence Generation of Natural Images
Author: Yezhou Yang ; Ching Teo ; Hal Daume III ; Yiannis Aloimonos
Abstract: We propose a sentence generation strategy that describes images by predicting the most likely nouns, verbs, scenes and prepositions that make up the core sentence structure. The input are initial noisy estimates of the objects and scenes detected in the image using state of the art trained detectors. As predicting actions from still images directly is unreliable, we use a language model trained from the English Gigaword corpus to obtain their estimates; together with probabilities of co-located nouns, scenes and prepositions. We use these estimates as parameters on a HMM that models the sentence generation process, with hidden nodes as sentence components and image detections as the emissions. Experimental results show that our strategy of combining vision and language produces readable and de- , scriptive sentences compared to naive strategies that use vision alone.
6 0.44416034 144 emnlp-2011-Unsupervised Learning of Selectional Restrictions and Detection of Argument Coercions
7 0.40223837 70 emnlp-2011-Identifying Relations for Open Information Extraction
8 0.380871 103 emnlp-2011-Parser Evaluation over Local and Non-Local Deep Dependencies in a Large Corpus
9 0.35855368 75 emnlp-2011-Joint Models for Chinese POS Tagging and Dependency Parsing
10 0.35508275 96 emnlp-2011-Multilayer Sequence Labeling
11 0.33971822 111 emnlp-2011-Reducing Grounded Learning Tasks To Grammatical Inference
12 0.3089596 91 emnlp-2011-Literal and Metaphorical Sense Identification through Concrete and Abstract Context
13 0.29779688 45 emnlp-2011-Dual Decomposition with Many Overlapping Components
14 0.29339853 126 emnlp-2011-Structural Opinion Mining for Graph-based Sentiment Representation
15 0.29263344 55 emnlp-2011-Exploiting Syntactic and Distributional Information for Spelling Correction with Web-Scale N-gram Models
16 0.28611892 12 emnlp-2011-A Weakly-supervised Approach to Argumentative Zoning of Scientific Documents
17 0.2824935 81 emnlp-2011-Learning General Connotation of Words using Graph-based Algorithms
18 0.27942187 102 emnlp-2011-Parse Correction with Specialized Models for Difficult Attachment Types
19 0.27690467 53 emnlp-2011-Experimental Support for a Categorical Compositional Distributional Model of Meaning
20 0.260979 127 emnlp-2011-Structured Lexical Similarity via Convolution Kernels on Dependency Trees
topicId topicWeight
[(23, 0.549), (36, 0.029), (37, 0.015), (45, 0.043), (53, 0.013), (54, 0.017), (57, 0.012), (62, 0.015), (64, 0.027), (66, 0.023), (69, 0.015), (79, 0.036), (82, 0.015), (90, 0.014), (96, 0.086), (98, 0.013)]
simIndex simValue paperId paperTitle
same-paper 1 0.99793804 7 emnlp-2011-A Joint Model for Extended Semantic Role Labeling
Author: Vivek Srikumar ; Dan Roth
Abstract: This paper presents a model that extends semantic role labeling. Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. However, sentences express relations via other linguistic phenomena as well. Furthermore, these phenomena interact with each other, thus restricting the structures they articulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. The scarcity of jointly labeled data presents a crucial technical challenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models.
2 0.9914673 42 emnlp-2011-Divide and Conquer: Crowdsourcing the Creation of Cross-Lingual Textual Entailment Corpora
Author: Matteo Negri ; Luisa Bentivogli ; Yashar Mehdad ; Danilo Giampiccolo ; Alessandro Marchetti
Abstract: We address the creation of cross-lingual textual entailment corpora by means of crowdsourcing. Our goal is to define a cheap and replicable data collection methodology that minimizes the manual work done by expert annotators, without resorting to preprocessing tools or already annotated monolingual datasets. In line with recent works emphasizing the need of large-scale annotation efforts for textual entailment, our work aims to: i) tackle the scarcity of data available to train and evaluate systems, and ii) promote the recourse to crowdsourcing as an effective way to reduce the costs of data collection without sacrificing quality. We show that a complex data creation task, for which even experts usually feature low agreement scores, can be effectively decomposed into simple subtasks assigned to non-expert annotators. The resulting dataset, obtained from a pipeline of different jobs routed to Amazon Mechanical Turk, contains more than 1,600 aligned pairs for each combination of texts-hypotheses in English, Italian and German.
3 0.98941845 93 emnlp-2011-Minimum Imputed-Risk: Unsupervised Discriminative Training for Machine Translation
Author: Zhifei Li ; Ziyuan Wang ; Jason Eisner ; Sanjeev Khudanpur ; Brian Roark
Abstract: Discriminative training for machine translation has been well studied in the recent past. A limitation of the work to date is that it relies on the availability of high-quality in-domain bilingual text for supervised training. We present an unsupervised discriminative training framework to incorporate the usually plentiful target-language monolingual data by using a rough “reverse” translation system. Intuitively, our method strives to ensure that probabilistic “round-trip” translation from a target- language sentence to the source-language and back will have low expected loss. Theoretically, this may be justified as (discriminatively) minimizing an imputed empirical risk. Empirically, we demonstrate that augmenting supervised training with unsupervised data improves translation performance over the supervised case for both IWSLT and NIST tasks.
4 0.98699933 135 emnlp-2011-Timeline Generation through Evolutionary Trans-Temporal Summarization
Author: Rui Yan ; Liang Kong ; Congrui Huang ; Xiaojun Wan ; Xiaoming Li ; Yan Zhang
Abstract: We investigate an important and challenging problem in summary generation, i.e., Evolutionary Trans-Temporal Summarization (ETTS), which generates news timelines from massive data on the Internet. ETTS greatly facilitates fast news browsing and knowledge comprehension, and hence is a necessity. Given the collection oftime-stamped web documents related to the evolving news, ETTS aims to return news evolution along the timeline, consisting of individual but correlated summaries on each date. Existing summarization algorithms fail to utilize trans-temporal characteristics among these component summaries. We propose to model trans-temporal correlations among component summaries for timelines, using inter-date and intra-date sen- tence dependencies, and present a novel combination. We develop experimental systems to compare 5 rival algorithms on 6 instinctively different datasets which amount to 10251 documents. Evaluation results in ROUGE metrics indicate the effectiveness of the proposed approach based on trans-temporal information. 1
5 0.98482317 48 emnlp-2011-Enhancing Chinese Word Segmentation Using Unlabeled Data
Author: Weiwei Sun ; Jia Xu
Abstract: This paper investigates improving supervised word segmentation accuracy with unlabeled data. Both large-scale in-domain data and small-scale document text are considered. We present a unified solution to include features derived from unlabeled data to a discriminative learning model. For the large-scale data, we derive string statistics from Gigaword to assist a character-based segmenter. In addition, we introduce the idea about transductive, document-level segmentation, which is designed to improve the system recall for out-ofvocabulary (OOV) words which appear more than once inside a document. Novel features1 result in relative error reductions of 13.8% and 15.4% in terms of F-score and the recall of OOV words respectively.
6 0.92616349 137 emnlp-2011-Training dependency parsers by jointly optimizing multiple objectives
7 0.9169836 58 emnlp-2011-Fast Generation of Translation Forest for Large-Scale SMT Discriminative Training
8 0.91632438 6 emnlp-2011-A Generate and Rank Approach to Sentence Paraphrasing
9 0.89972883 79 emnlp-2011-Lateen EM: Unsupervised Training with Multiple Objectives, Applied to Dependency Grammar Induction
10 0.88664716 136 emnlp-2011-Training a Parser for Machine Translation Reordering
11 0.88369006 126 emnlp-2011-Structural Opinion Mining for Graph-based Sentiment Representation
12 0.87977344 108 emnlp-2011-Quasi-Synchronous Phrase Dependency Grammars for Machine Translation
13 0.87603235 17 emnlp-2011-Active Learning with Amazon Mechanical Turk
14 0.87334281 61 emnlp-2011-Generating Aspect-oriented Multi-Document Summarization with Event-aspect model
15 0.87273622 28 emnlp-2011-Closing the Loop: Fast, Interactive Semi-Supervised Annotation With Queries on Features and Instances
16 0.87029445 59 emnlp-2011-Fast and Robust Joint Models for Biomedical Event Extraction
17 0.86783487 124 emnlp-2011-Splitting Noun Compounds via Monolingual and Bilingual Paraphrasing: A Study on Japanese Katakana Words
18 0.86716646 23 emnlp-2011-Bootstrapped Named Entity Recognition for Product Attribute Extraction
19 0.8667295 89 emnlp-2011-Linguistic Redundancy in Twitter
20 0.86358649 82 emnlp-2011-Learning Local Content Shift Detectors from Document-level Information