emnlp emnlp2011 emnlp2011-143 knowledge-graph by maker-knowledge-mining

143 emnlp-2011-Unsupervised Information Extraction with Distributional Prior Knowledge

Source: pdf

Author: Cane Wing-ki Leung ; Jing Jiang ; Kian Ming A. Chai ; Hai Leong Chieu ; Loo-Nin Teow

Abstract: We address the task of automatic discovery of information extraction template from a given text collection. Our approach clusters candidate slot fillers to identify meaningful template slots. We propose a generative model that incorporates distributional prior knowledge to help distribute candidates in a document into appropriate slots. Empirical results suggest that the proposed prior can bring substantial improvements to our task as compared to a K-means baseline and a Gaussian mixture model baseline. Specifically, the proposed prior has shown to be effective when coupled with discriminative features of the candidates.

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 Our approach clusters candidate slot fillers to identify meaningful template slots. [sent-5, score-0.816]

2 We propose a generative model that incorporates distributional prior knowledge to help distribute candidates in a document into appropriate slots. [sent-6, score-0.477]

3 Empirical results suggest that the proposed prior can bring substantial improvements to our task as compared to a K-means baseline and a Gaussian mixture model baseline. [sent-7, score-0.277]

4 For example, the domain in the Sixth Message Understanding Conference (MUC-6, 1995) is management succession, and the pre-defined template consists of the slots position, the person leaving, the person joining, and the organization. [sent-11, score-0.573]

5 These approaches have a commonality: they try to cluster candidate slot fillers, which are often nouns and noun phrases, into slots of the template to be constructed. [sent-22, score-0.867]

6 However, most of them have neglected the following important observation: a single document or text segment tends to cover different slots rather than redundantly fill the same slot. [sent-23, score-0.288]

7 In this paper, we propose a generative model that incorporates this distributional prior knowledge. [sent-25, score-0.268]

8 We define a prior distribution over the possible label assignments in a document or a text segment such that a more diversified label assignment is preferred. [sent-26, score-0.551]

9 We also compare a number of generative models for generating slot fillers and find that the Gaussian mixture model is the best. [sent-28, score-0.795]

10 We then combine the Poissonbased label assignment prior with the Gaussian mixture model to perform slot clustering. [sent-29, score-0.764]

11 We find that compared with a K-means baseline and a Gaussian mixture model baseline, our combined model with the proposed label assignment prior substantially performs better on two of the three data sets we use for evaluation. [sent-30, score-0.429]

12 ec th2o0d1s1 i Ans Nsoactuiartaioln La fonrg Cuaogmep Purtoatcieosnsainlg L,in pgaugies ti 8c1s4–824, our Poisson-based label assignment prior is effective when coupled with good discriminative features. [sent-34, score-0.312]

13 These methods leveraged heavily on the entity types of candidates when assigning them to template slots. [sent-38, score-0.297]

14 This problem is alleviated in our work by exploiting distributional prior knowledge about template slots, which is shown effective when coupled with discriminative features of candidates. [sent-41, score-0.342]

15 Their algorithm assigned a candidate from a document to a cluster based on the candidate’s feature similarity with candidates from other documents only. [sent-48, score-0.334]

16 Our work is based on a different perspective: we model label assignments for all candidates in the same document with a distributional prior that prefers a document to cover more distinct slots. [sent-50, score-0.614]

17 We show empirically that this prior improves slot clustering results greatly in some cases. [sent-51, score-0.558]

18 We propose a generative model with a distributional prior for the unsupervised IE task, where slot fillers correspond to observations in the model, and their labels correspond to hidden variables we want to learn. [sent-59, score-1.0]

19 In contrast, we place constraints on label assignments through a probabilistic prior on the distribution of slots. [sent-65, score-0.337]

20 The proposed prior is simple and easy to interpret in a generative model. [sent-66, score-0.267]

21 The discovered template should contain a set of slots that play different semantic roles in the domain. [sent-75, score-0.403]

22 Delaney Figure 1: An input text from a seminar announcement collection and the discovered IE template. [sent-86, score-0.277]

23 Note that the slots are automatically discovered and the slot names are manually assigned. [sent-87, score-0.619]

24 To construct such a template, we start with identifying candidate slot fillers, hereafter referred to as candidates, from the input text. [sent-88, score-0.399]

25 Then we cluster these candidates with the aim that each cluster will represent a semantically meaningful slot. [sent-89, score-0.331]

26 Figure 1 gives an example of an input text from a collection of seminar announcements and the resulting template discovered from the collection. [sent-90, score-0.486]

27 As we can see, the template contains some semantically meaningful slots such as the start time, end time, location and speaker of a seminar. [sent-91, score-0.456]

28 Moreover, it also contains a slot that covers an irrelevant candidate. [sent-92, score-0.401]

29 We call such slots covering irrelevant candidates garbage slots. [sent-93, score-0.523]

30 We can make two observations on the mapping from candidates to template slots from real data, such as the text in Figure 1. [sent-94, score-0.56]

31 Firstly, a template slot may be filled by more than one candidate from a single document, although this number has been observed to be small. [sent-95, score-0.549]

32 For example, the template slot end time in Figure 1 has two slot fillers: “1 :30 PM” from the semi-structured header and “1 :30 p. [sent-96, score-0.882]

33 Secondly, a document tends to contain candidates that cover different template slots. [sent-99, score-0.359]

34 2 A General Solution Recall that our general solution to the unsupervised IE problem is to cluster candidate slot fillers in order to identify meaningful slots. [sent-103, score-0.794]

35 For the i-th document in our collection, we assume that the number of candidates is known and we draw a label assignment yi according to some distribution parameterized by Λ. [sent-117, score-0.516]

36 D denotes the number of documents in the given collection, ni denotes the number of candidates in the i-th document, and K is the number of slots (clusters). [sent-123, score-0.373]

37 In the next section, we detail two designs of the prior p(yi; Λ), followed by different generative models for the distribution p(xi,j |yi,j ; Θ) in Section 5. [sent-126, score-0.262]

38 4 Label Assignment Prior The label assignment prior, p(yi; Λ), models the generation of labels for candidates in a document. [sent-128, score-0.325]

39 In this section, we first describe a commonly used multinomial prior, and then introduce the proposed Poisson-based prior for the unsupervised IE task. [sent-129, score-0.316]

40 ∑i,j ∑y (2) Figure 2(a) depicts a generative model with this multinomial prior in plate notation. [sent-135, score-0.325]

41 Note that the independence assumption on label assignment in this model does not capture our observation that candidates in a document are likely to cover different semantic roles. [sent-136, score-0.361]

42 2 The Proposed Poisson-based Prior We propose a prior distribution that favors more diverse label assignments. [sent-138, score-0.28]

43 The proposed prior distribution should therefore assign higher probability to a label assignment that covers more distinct slots. [sent-141, score-0.369]

44 Our design thus allows a slot to generate multiple fillers in a document, up to a limited number of times. [sent-143, score-0.633]

45 Thirdly, there may exist candidates that do not belong to slots in the extracted template. [sent-144, score-0.373]

46 Therefore, we introduce a dummy slot or garbage slot to the label set to collect such candidates. [sent-145, score-0.941]

47 Yet, we shall not as- sume any prior/domain knowledge about candidates generated by the garbage slot as they are essentially irrelevant in the given domain. [sent-146, score-0.663]

48 First, we fix the K-th slot (or cluster) in the label set to be the garbage slot. [sent-148, score-0.575]

49 There is no λK for the garbage slot because the number of fillers is not constrained for this slot. [sent-156, score-0.748]

50 This allows all candidates in a document to be generated by the garbage slot. [sent-157, score-0.324]

51 The set of possible label assignments for the ith document is the sample space on which we place the prior distribution p(yi; Λ). [sent-165, score-0.399]

52 For a given yi for the i-th document, let ni,k be the number of candidates in the document that have been assigned to slot k. [sent-167, score-0.661]

53 eflects the lack of prior knowledge on the number of garbage slot fillers. [sent-174, score-0.641]

54 Figure 2(b) depicts the proposed generative model with the Poisson-based prior in plate notation. [sent-175, score-0.267]

55 5 Generating Slot Fillers Different existing generative models can be used to model the generation of a slot filler given a label, that is, p(x|y; Θ). [sent-176, score-0.442]

56 a Wivee Bayes model, ft thhee Bmer fnooru oullir mtasikx,ture model, the Gaussian mixture model, and a locally normalized logistic regression model proposed by Berg-Kirkpatrick et al. [sent-178, score-0.275]

57 Fo∈r a given label y, feature f follows a multinomial distribution parameterized by ψy,f, where ψy,f,v denotes the probability of feature f taking the value v ∈ Vf given label y. [sent-191, score-0.346]

58 For the Bernoulli mixture model, as well as the Gaussian mixture model and the locally normalized logistic regression model in the next subsections, we first convert each observation x into a binary feature vector x ∈ {0, 1}F where F 818 is the number of binary features. [sent-195, score-0.33]

59 3 (5) Gaussian Mixture Model In the Gaussian mixture model, we assume that a given label y generates observations with a multivariate Gaussian distribution N(µy, Σy), where µy ∈ tReF G iasu tshsiea mean rainbdu Σy ∈ R(µF×F is the co- vari∈anc Re matrix of the Gaussian. [sent-200, score-0.278]

60 For the multinomial prior, there are standard closed form solutions for the naive Bayes, the Bernoulli mixture and the Gaussian mixture models. [sent-214, score-0.293]

61 For locally normalized logistic regression, parameters can be learned with a gradient-based M-step as in the multinomial prior setting. [sent-229, score-0.369]

62 7 Experiments In this section, we first describe the data sets we used in our experiments, detailing the target slots and candidates in each data set, as well as features we extract for the candidates. [sent-232, score-0.373]

63 The first data set contains a set of seminar announcements (Freitag and McCallum, 1999), annotated with four slot labels, namely stime (start time), etime (end time), speaker and location. [sent-237, score-0.922]

64 There are 309 seminar announcements with 2262 candidates in this data set. [sent-240, score-0.429]

65 The second data set is a collection of paragraphs describing aviation incidents, taken from the Wikipedia article on “List of accidents and incidents involving commercial aircraft” (Wikipedia, 2009). [sent-241, score-0.381]

66 For evaluation, we manually annotated the paragraphs of incidents from 2006 to 2009 with five slot labels: the flight number (FN), the airline (AL), the aircraft model (AC), the exact location (LO) of the incident (e. [sent-244, score-0.679]

67 We extract from the original data set all sentences that were tagged with a management succession event, and use as candidates all tagged strings in those sentences. [sent-250, score-0.32]

68 To evaluate clustering results, we match each slot in the labeled data to the cluster that gives the best F1-measure when evaluated for the slot. [sent-263, score-0.49]

69 We report the precision (P), recall (R) and F1-measure for individual slot labels, as well as the macro- and micro- average results across all labels for each experiment. [sent-264, score-0.392]

70 As noted, we use a garbage slot to capture irrelevant candidates, thus the value of K is set to the number of target slots plus 1 for each data set. [sent-271, score-0.742]

71 Table 1summarizes the performance of Naive Bayes (NB), the Bernoulli mixture model (BMM), the Gaussian mixture model (GMM), the locally normalized logistic regression (LNLR) model, and Kmeans. [sent-279, score-0.33]

72 Specifically, BMM outperforms Kmeans for aviation incidents, but performs poorly for seminar announcements. [sent-284, score-0.327]

73 Overall speaking, results show that GMM is the best among the four generative models for the distri- (a) Results on seminar announcements. [sent-287, score-0.269]

74 No macro- and micro-average result is reported for NB and BMM as they merged the etime cluster with the stime cluster. [sent-288, score-0.33]

75 Model stime etime speaker location NB BMM GMM LNLR K-means 0. [sent-290, score-0.318]

76 Target slots are airline (AL) , flight number (FN), aircraft model (AC), location (LO) and country (CO). [sent-318, score-0.338]

77 Target slots are person joining (PersonIn), person leaving (PersonOut), organization (Org), and position (Post). [sent-356, score-0.334]

78 Data set Seminar announcements Parameter Value {λk }k4=1 {τk}k4=1 {λk }k5=1 {τk}k5=1 {λk }k4=1 {τk}k4=1 {2}k4=1 {1}k4=1 {1}k5=1 {1}k5=1 {1,2,2,2} {1,2,2,2} {τ}{1} Aviation incidents {τ}{1} Management succession {τ}{1,2,2,2} Table 2: Parameter settings for p(yi ; Λ). [sent-389, score-0.323]

79 Effectiveness of the proposed prior We evaluate the effectiveness of the proposed prior by combining it with GMM. [sent-393, score-0.382]

80 Rsoe cfa {lλl th}at λk specifies the {τk}kK=−11 821 maximum number of candidates that the k-th slot can generate, and its value is observed to be small in real data. [sent-400, score-0.513]

81 τk specifies the expected number of candidates that the k-th slot will generate. [sent-401, score-0.513]

82 The combined model improves over both GMM and K-means for seminar announcements and aviation incidents, as can be seen from the models’ macro- and micro-average performance. [sent-403, score-0.416]

83 The advantages brought by the proposed prior are mainly reflected in slots that are difficult to cluster under GMM and K-means. [sent-404, score-0.509]

84 Taking seminar announcements as an example, GMM and K-means achieve high precision but low recall for stime, and low precision but high recall for etime. [sent-405, score-0.282]

85 Model Metric stime etime speaker location Macro-avg Micro-avg GMM with Prior P R F1 0. [sent-408, score-0.318]

86 For aviation incidents, the advantage of the proposed prior is reflected in the location (LO) and country (CO) slots, which may confuse the various models as they both belong to the entity type location. [sent-589, score-0.369]

87 The proposed prior improves the precision of these two slots greatly by trying to distribute them into appropriate slots in the clustering process. [sent-590, score-0.675]

88 Surprisingly, incorporating the Poisson-based prior into GMM does not seem useful in separating PersonIn and PersonOut slot fillers. [sent-592, score-0.526]

89 ofthem being general context features that might not help characterizing candidates from different slots (e. [sent-612, score-0.373]

90 Note that the “dobj:succeeds” feature in the PersonIn cluster is in fact contributed by PersonOut slot fillers, while the “nsubj:succeeds” feature in the PersonOut cluster is contributed by PersonIn slot fillers. [sent-616, score-0.916]

91 These observations may suggest that the management succession data set lacks strong, discriminative features for all models to effectively distinguish between PersonIn and PersonOut candidates in an unsupervised manner. [sent-619, score-0.393]

92 To conclude, the proposed prior is effective in assigning different but confusing candidate slot fillers into appropriate slots, when there exist reasonable features that can be exploited in the label assignment process. [sent-620, score-1.009]

93 This is evident by the improvements the proposed prior brings to GMM in the seminar announcement and aviation incident data sets. [sent-621, score-0.594]

94 8 Conclusions We propose a generative model that incorporates distributional prior knowledge about template slots in a document for the unsupervised IE task. [sent-622, score-0.742]

95 Specifically, we propose a Poisson-based prior that prefers label assignments to cover more distinct slots in the same document. [sent-623, score-0.537]

96 The proposed prior also allows a slot to generate multiple fillers in a document, up to a certain number of times depending on the domain of interest. [sent-624, score-0.859]

97 models for the task of clustering slot fillers with a multinomial prior, which assumes that labels are generated independently in a document. [sent-645, score-0.78]

98 We then evaluate the effectiveness of the proposed prior by incorporating it into the Gaussian mixture model (GMM), which is shown to be the best among the four existing models in our experiments. [sent-646, score-0.277]

99 Firstly, we assume that some adjustable parameters in the proposed prior can be manually fixed, such as the number of template slots in the output and the maximum numbers of fillers that can be generated by different slots. [sent-649, score-0.864]

100 Secondly, we currently consider in the prior a probability distribution over all possible label assignments for ev- ery document. [sent-652, score-0.337]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('gmm', 0.431), ('slot', 0.366), ('fillers', 0.267), ('slots', 0.226), ('personout', 0.193), ('seminar', 0.193), ('personin', 0.166), ('prior', 0.16), ('template', 0.15), ('candidates', 0.147), ('nil', 0.139), ('aviation', 0.134), ('stime', 0.134), ('ie', 0.126), ('succession', 0.119), ('incidents', 0.115), ('garbage', 0.115), ('bmm', 0.104), ('etime', 0.104), ('label', 0.094), ('cluster', 0.092), ('multinomial', 0.089), ('announcements', 0.089), ('lnlr', 0.089), ('mixture', 0.086), ('yi', 0.086), ('gaussian', 0.084), ('generative', 0.076), ('kk', 0.075), ('nb', 0.071), ('document', 0.062), ('bernoulli', 0.06), ('centriods', 0.059), ('assignment', 0.058), ('assignments', 0.057), ('person', 0.054), ('nsubj', 0.054), ('succeeds', 0.054), ('xf', 0.054), ('management', 0.054), ('kt', 0.051), ('trials', 0.051), ('sekine', 0.05), ('incident', 0.046), ('poisson', 0.046), ('logistic', 0.045), ('filatova', 0.045), ('sudo', 0.045), ('vf', 0.045), ('location', 0.044), ('parameterized', 0.043), ('locally', 0.042), ('paragraphs', 0.04), ('aircraft', 0.038), ('regression', 0.038), ('observations', 0.037), ('unsupervised', 0.036), ('speaker', 0.036), ('bayes', 0.035), ('domain', 0.035), ('irrelevant', 0.035), ('multivariate', 0.035), ('accidents', 0.035), ('lo', 0.034), ('satoshi', 0.033), ('candidate', 0.033), ('normalized', 0.033), ('distributional', 0.032), ('naive', 0.032), ('clustering', 0.032), ('proposed', 0.031), ('unigram', 0.031), ('templates', 0.031), ('commercial', 0.03), ('adjustable', 0.03), ('airline', 0.03), ('announcement', 0.03), ('cane', 0.03), ('centriod', 0.03), ('delaney', 0.03), ('duda', 0.03), ('kiesler', 0.03), ('lunch', 0.03), ('tuynpiger', 0.03), ('ypi', 0.03), ('dobj', 0.029), ('fn', 0.028), ('org', 0.028), ('collection', 0.027), ('discovered', 0.027), ('pm', 0.027), ('labels', 0.026), ('distribution', 0.026), ('em', 0.026), ('rosenfeld', 0.026), ('co', 0.026), ('shinyama', 0.026), ('sara', 0.026), ('edmund', 0.026), ('student', 0.025)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.99999946 143 emnlp-2011-Unsupervised Information Extraction with Distributional Prior Knowledge

Author: Cane Wing-ki Leung ; Jing Jiang ; Kian Ming A. Chai ; Hai Leong Chieu ; Loo-Nin Teow

2 0.08015386 128 emnlp-2011-Structured Relation Discovery using Generative Models

Author: Limin Yao ; Aria Haghighi ; Sebastian Riedel ; Andrew McCallum

Abstract: We explore unsupervised approaches to relation extraction between two named entities; for instance, the semantic bornIn relation between a person and location entity. Concretely, we propose a series of generative probabilistic models, broadly similar to topic models, each which generates a corpus of observed triples of entity mention pairs and the surface syntactic dependency path between them. The output of each model is a clustering of observed relation tuples and their associated textual expressions to underlying semantic relation types. Our proposed models exploit entity type constraints within a relation as well as features on the dependency path between entity mentions. We examine effectiveness of our approach via multiple evaluations and demonstrate 12% error reduction in precision over a state-of-the-art weakly supervised baseline.

3 0.064941436 146 emnlp-2011-Unsupervised Structure Prediction with Non-Parallel Multilingual Guidance

Author: Shay B. Cohen ; Dipanjan Das ; Noah A. Smith

Abstract: We describe a method for prediction of linguistic structure in a language for which only unlabeled data is available, using annotated data from a set of one or more helper languages. Our approach is based on a model that locally mixes between supervised models from the helper languages. Parallel data is not used, allowing the technique to be applied even in domains where human-translated texts are unavailable. We obtain state-of-theart performance for two tasks of structure prediction: unsupervised part-of-speech tagging and unsupervised dependency parsing.

4 0.056974057 39 emnlp-2011-Discovering Morphological Paradigms from Plain Text Using a Dirichlet Process Mixture Model

Author: Markus Dreyer ; Jason Eisner

Abstract: We present an inference algorithm that organizes observed words (tokens) into structured inflectional paradigms (types). It also naturally predicts the spelling of unobserved forms that are missing from these paradigms, and discovers inflectional principles (grammar) that generalize to wholly unobserved words. Our Bayesian generative model of the data explicitly represents tokens, types, inflections, paradigms, and locally conditioned string edits. It assumes that inflected word tokens are generated from an infinite mixture of inflectional paradigms (string tuples). Each paradigm is sampled all at once from a graphical model, whose potential functions are weighted finitestate transducers with language-specific parameters to be learned. These assumptions naturally lead to an elegant empirical Bayes inference procedure that exploits Monte Carlo EM, belief propagation, and dynamic programming. Given 50–100 seed paradigms, adding a 10million-word corpus reduces prediction error for morphological inflections by up to 10%.

5 0.055379499 1 emnlp-2011-A Bayesian Mixture Model for PoS Induction Using Multiple Features

Author: Christos Christodoulopoulos ; Sharon Goldwater ; Mark Steedman

Abstract: In this paper we present a fully unsupervised syntactic class induction system formulated as a Bayesian multinomial mixture model, where each word type is constrained to belong to a single class. By using a mixture model rather than a sequence model (e.g., HMM), we are able to easily add multiple kinds of features, including those at both the type level (morphology features) and token level (context and alignment features, the latter from parallel corpora). Using only context features, our system yields results comparable to state-of-the art, far better than a similar model without the one-class-per-type constraint. Using the additional features provides added benefit, and our final system outperforms the best published results on most of the 25 corpora tested.

6 0.055138692 14 emnlp-2011-A generative model for unsupervised discovery of relations and argument classes from clinical texts

7 0.055129405 129 emnlp-2011-Structured Sparsity in Structured Prediction

8 0.053718209 116 emnlp-2011-Robust Disambiguation of Named Entities in Text

9 0.051929757 145 emnlp-2011-Unsupervised Semantic Role Induction with Graph Partitioning

10 0.048285007 87 emnlp-2011-Lexical Generalization in CCG Grammar Induction for Semantic Parsing

11 0.048180409 61 emnlp-2011-Generating Aspect-oriented Multi-Document Summarization with Event-aspect model

12 0.046609841 28 emnlp-2011-Closing the Loop: Fast, Interactive Semi-Supervised Annotation With Queries on Features and Instances

13 0.045551352 119 emnlp-2011-Semantic Topic Models: Combining Word Distributional Statistics and Dictionary Definitions

14 0.045394186 29 emnlp-2011-Collaborative Ranking: A Case Study on Entity Linking

15 0.043229658 114 emnlp-2011-Relation Extraction with Relation Topics

16 0.042418614 98 emnlp-2011-Named Entity Recognition in Tweets: An Experimental Study

17 0.042242941 21 emnlp-2011-Bayesian Checking for Topic Models

18 0.041673318 70 emnlp-2011-Identifying Relations for Open Information Extraction

19 0.041635066 96 emnlp-2011-Multilayer Sequence Labeling

20 0.039783996 93 emnlp-2011-Minimum Imputed-Risk: Unsupervised Discriminative Training for Machine Translation

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, 0.148), (1, -0.067), (2, -0.083), (3, -0.017), (4, -0.023), (5, 0.006), (6, -0.013), (7, -0.064), (8, -0.075), (9, 0.064), (10, -0.017), (11, 0.005), (12, -0.074), (13, -0.005), (14, 0.021), (15, 0.008), (16, 0.003), (17, 0.004), (18, 0.033), (19, 0.039), (20, -0.107), (21, 0.049), (22, 0.034), (23, -0.039), (24, 0.079), (25, 0.024), (26, -0.022), (27, 0.034), (28, 0.089), (29, 0.065), (30, 0.029), (31, -0.126), (32, -0.152), (33, -0.072), (34, 0.096), (35, -0.037), (36, 0.029), (37, -0.02), (38, -0.015), (39, 0.068), (40, -0.067), (41, -0.023), (42, -0.111), (43, -0.157), (44, 0.068), (45, 0.063), (46, 0.095), (47, -0.227), (48, -0.052), (49, -0.152)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.91830873 143 emnlp-2011-Unsupervised Information Extraction with Distributional Prior Knowledge

Author: Cane Wing-ki Leung ; Jing Jiang ; Kian Ming A. Chai ; Hai Leong Chieu ; Loo-Nin Teow

2 0.49087894 79 emnlp-2011-Lateen EM: Unsupervised Training with Multiple Objectives, Applied to Dependency Grammar Induction

Author: Valentin I. Spitkovsky ; Hiyan Alshawi ; Daniel Jurafsky

Abstract: We present new training methods that aim to mitigate local optima and slow convergence in unsupervised training by using additional imperfect objectives. In its simplest form, lateen EM alternates between the two objectives of ordinary “soft” and “hard” expectation maximization (EM) algorithms. Switching objectives when stuck can help escape local optima. We find that applying a single such alternation already yields state-of-the-art results for English dependency grammar induction. More elaborate lateen strategies track both objectives, with each validating the moves proposed by the other. Disagreements can signal earlier opportunities to switch or terminate, saving iterations. De-emphasizing fixed points in these ways eliminates some guesswork from tuning EM. An evaluation against a suite of unsupervised dependency parsing tasks, for a vari- ety of languages, showed that lateen strategies significantly speed up training of both EM algorithms, and improve accuracy for hard EM.

3 0.48900068 39 emnlp-2011-Discovering Morphological Paradigms from Plain Text Using a Dirichlet Process Mixture Model

Author: Markus Dreyer ; Jason Eisner

4 0.45139921 70 emnlp-2011-Identifying Relations for Open Information Extraction

Author: Anthony Fader ; Stephen Soderland ; Oren Etzioni

Abstract: Open Information Extraction (IE) is the task of extracting assertions from massive corpora without requiring a pre-specified vocabulary. This paper shows that the output of state-ofthe-art Open IE systems is rife with uninformative and incoherent extractions. To overcome these problems, we introduce two simple syntactic and lexical constraints on binary relations expressed by verbs. We implemented the constraints in the REVERB Open IE system, which more than doubles the area under the precision-recall curve relative to previous extractors such as TEXTRUNNER and WOEpos. More than 30% of REVERB’s extractions are at precision 0.8 or higher— compared to virtually none for earlier systems. The paper concludes with a detailed analysis of REVERB’s errors, suggesting directions for future work.1 1 Introduction and Motivation Typically, Information Extraction (IE) systems learn an extractor for each target relation from labeled training examples (Kim and Moldovan, 1993; Riloff, 1996; Soderland, 1999). This approach to IE does not scale to corpora where the number of target relations is very large, or where the target relations cannot be specified in advance. Open IE solves this problem by identifying relation phrases—phrases that denote relations in English sentences (Banko et al., 2007). The automatic identification of rela1The source code for REVERB is available at reverb . cs .washingt on .edu/ http : // 1535 tion phrases enables the extraction of arbitrary relations from sentences, obviating the restriction to a pre-specified vocabulary. Open IE systems have achieved a notable measure of success on massive, open-domain corpora drawn from the Web, Wikipedia, and elsewhere. (Banko et al., 2007; Wu and Weld, 2010; Zhu et al., 2009). The output of Open IE systems has been used to support tasks like learning selectional preferences (Ritter et al., 2010), acquiring common sense knowledge (Lin et al., 2010), and recognizing entailment (Schoen- mackers et al., 2010; Berant et al., 2011). In addition, Open IE extractions have been mapped onto existing ontologies (Soderland et al., 2010). We have observed that two types of errors are frequent in the output of Open IE systems such as TEXTRUNNER and WOE: incoherent extractions and uninformative extractions. Incoherent extractions are cases where the extracted relation phrase has no meaningful interpretation (see Table 1 for examples). Incoherent extractions arise because the learned extractor makes a sequence of decisions about whether to include each word in the relation phrase, often resulting in incomprehensible predictions. To solve this problem, we introduce a syntactic constraint: every multi-word relation phrase must begin with a verb, end with a preposition, and be a contiguous sequence of words in the sentence. Thus, the identification of a relation phrase is made in one fell swoop instead of on the basis of multiple, word-by-word decisions. Uninformative extractions are extractions that omit critical information. For example, consider the sentence “Faust made a deal with the devil.” PreviProce dEindgisnb oufr tgh e, 2 S0c1o1tl Canodn,f eUrKen,c Jeuol yn 2 E7m–3p1ir,ic 2a0l1 M1.e ?tc ho2d0s1 in A Nsasotucira tlio Lnan fogru Cagoem Ppruotcaetisosninagl, L pinag uesis 1ti5c3s5–1545, ous Open IE systems return the uninformative (Faust, made, a deal) instead of (Faust, made a deal with, the devil). This type of error is caused by improper handling of relation phrases that are expressed by a combination of a verb with a noun, such as light verb constructions (LVCs). An LVC is a multi-word expression composed of a verb and a noun, with the noun carrying the semantic content of the predicate (Grefenstette and Teufel, 1995; Stevenson et al., 2004; Allerton, 2002). Table 2 illustrates the wide range of relations expressed this way, which are not captured by existing open extractors. Our syntactic constraint leads the extractor to include nouns in the relation phrase, solving this problem. Although the syntactic constraint significantly reduces incoherent and uninformative extractions, it allows overly-specific relation phrases such as is offering only modest greenhouse gas reduction targets at. To avoid overly-specific relation phrases, we introduce an intuitive lexical constraint: a binary relation phrase ought to appear with at least a minimal number of distinct argument pairs in a large corpus. In summary, this paper articulates two simple but surprisingly powerful constraints on how binary relationships are expressed via verbs in English sentences, and implements them in the REVERB Open IE system. We release REVERB and the data used in our experiments to the research community. The rest of the paper is organized as follows. Section 2 analyzes previous work. Section 3 defines our constraints precisely. Section 4 describes REVERB, our implementation of the constraints. Section 5 reports on our experimental results. Section 6 concludes with a summary and discussion of future work. 2 Previous Work Open IE systems like TEXTRUNNER (Banko et al., 2007), WOEpos, and WOEparse (Wu and Weld, 2010) focus on extracting binary relations of the form (arg1, relation phrase, arg2) from text. These systems all use the following three-step method: 1. Label: Sentences are automatically labeled with extractions using heuristics or distant supervision. 1536 tTSblaoenh rgdpe atyoneMgdrm.uoahcietrsdkcaes1lci4neatodrwnsea.tlrhoianfctsehndNaterufplndergtcoli.sn hkcetrIwnoca stlhceindrstbeoanmgltiRonserplatdion Table 1: Examples of incoherent extractions. Incoherent extractions make up approximately 13% of TEXTRUNNER’s output, 15% of WOEpos’s output, and 30% of WOEparse’s output. tighmosatvdkeihgmosa tvdnkeaicbp lkaroedpucthmsalitonbw,yigtoanhivst,oekamfh,cedotanlusetkroahlntPpr,ohgv.aeDomfvrt,.isonue,atdkhowcainstmdygaevifncromtnaeg oinf Table 2: Examples of uninformative relations (left) and their completions (right). Uninformative relations occur in approximately 4% of WOEparse’s output, 6% of WOEpos’s output, and 7% of TEXTRUNNER’s output. 2. Learn: A relation phrase extractor is learned using a sequence-labeling (e.g., CRF). graphical model 3. Extract: the system takes a sentence as input, identifies a candidate pair of NP arguments (arg1, arg2) from the sentence, and then uses the learned extractor to label each word between the two arguments as part of the relation phrase or not. The extractor is applied to the successive sentences in the corpus, and the resulting extractions are collected. This method faces several challenges. First, the training phase requires a large number of labeled training examples (e.g., 200, 000 heuristicallylabeled sentences for TEXTRUNNER and 300, 000 for WOE). Heuristic labeling of examples obviates hand labeling but results in noisy labels and distorts the distribution of examples. Second, the extraction step is posed as a sequence-labeling problem, where each word is assigned its own label. Because each assignment is uncertain, the likelihood that the extracted relation phrase is flawed increases with the length of the sequence. Finally, the extractor chooses an extraction’s arguments heuristically, and cannot backtrack over this choice. This is problematic when a word that belongs in the relation phrase is chosen as an argument (for example, deal from the “made a deal with” sentence). Because of the feature sets utilized in previous work, the learned extractors ignore both “holistic” aspects of the relation phrase (e.g., is it contiguous?) as well as lexical aspects (e.g., how many instances of this relation are there?). Thus, as we show in Section 5, systems such as TEXTRUNNER are unable to learn the constraints embedded in REVERB. Of course, a learning system, utilizing a different hypothesis space, and an appropriate set of training examples, could potentially learn and refine the constraints in REVERB. This is a topic for future work, which we consider in Section 6. The first Open IE system was TEXTRUNNER (Banko et al., 2007), which used a Naive Bayes model with unlexicalized POS and NP-chunk features, trained using examples heuristically generated from the Penn Treebank. Subsequent work showed that utilizing a linear-chain CRF (Banko and Etzioni, 2008) or Markov Logic Network (Zhu et al., 2009) can lead to improved extraction. The WOE systems introduced by Wu and Weld make use of Wikipedia as a source of training data for their extractors, which leads to further improvements over TEXTRUNNER (Wu and Weld, 2010). Wu and Weld also show that dependency parse features result in a dramatic increase in precision and recall over shallow linguistic features, but at the cost of extraction speed. Other approaches to large-scale IE have included Preemptive IE (Shinyama and Sekine, 2006), OnDemand IE (Sekine, 2006), and weak supervision for IE (Mintz et al., 2009; Hoffmann et al., 2010). Preemptive IE and On-Demand IE avoid relationspecific extractors, but rely on document and entity clustering, which is too costly for Web-scale IE. Weakly supervised methods use an existing ontology to generate training data for learning relationspecific extractors. While this allows for learning relation-specific extractors at a larger scale than what was previously possible, the extractions are still restricted to a specific ontology. Many systems have used syntactic patterns based on verbs to extract relation phrases, usually rely1537 ing on a full dependency parse of the input sentence (Lin and Pantel, 2001 ; Stevenson, 2004; Specia and Motta, 2006; Kathrin Eichler and Neumann, 2008). Our work differs from these approaches by focusing on relation phrase patterns expressed in terms of POS tags and NP chunks, instead of full parse trees. Banko and Etzioni (Banko and Etzioni, 2008) showed that a small set of POS-tag patterns cover a large fraction of relationships in English, but never incorporated the patterns into an extractor. This paper reports on a substantially improved model of binary relation phrases, which increases the recall of the Banko-Etzioni model (see Section 3.3). Further, while previous work in Open IE has mainly focused on syntactic patterns for relation extraction, we introduce a lexical constraint that boosts precision and recall. Finally, Open IE is closely related to semantic role labeling (SRL) (Punyakanok et al., 2008; Toutanova et al., 2008) in that both tasks extract relations and arguments from sentences. However, SRL systems traditionally rely on syntactic parsers, which makes them susceptible to parser errors and substantially slower than Open IE systems such as REVERB. This difference is particularly important when operating on the Web corpus due to its size and heterogeneity. Finally, SRL requires hand-constructed semantic resources like Propbank and Framenet (Martha and Palmer, 2002; Baker et al., 1998) as input. In contrast, Open IE systems require no relation-specific training data. ReVerb, in particular, relies on its explicit lexical and syntactic constraints, which have no correlate in SRL systems. For a more detailed comparison of SRL and Open IE, see (Christensen et al., 2010). 3 Constraints on Relation Phrases In this section we introduce two constraints on relation phrases: a syntactic constraint and a lexical constraint. 3.1 Syntactic Constraint The syntactic constraint serves two purposes. First, it eliminates incoherent extractions, and second, it reduces uninformative extractions by capturing relation phrases expressed by a verb-noun combination, including light verb constructions. PVW= v(p enr oVbupn |a prVta iPcdrtljiec|?lVaedaW|dv in|∗?fpP.rmonar|kder)t Figure 1: A simple part-of-speech-based regular expression reduces the number of incoherent extractions like was central torpedo and covers relations expressed via light verb constructions like gave a talk at. The syntactic constraint requires the relation phrase to match the POS tag pattern shown in Figure 1. The pattern limits relation phrases to be either a verb (e.g., invented), a verb followed immediately by a preposition (e.g., located in), or a verb followed by nouns, adjectives, or adverbs ending in a preposition (e.g., has atomic weight of). Ifthere are multiple possible matches in a sentence for a single verb, the longest possible match is chosen. Finally, if the pattern matches multiple adjacent sequences, we merge them into a single relation phrase (e.g., wants to extend). This refinement enables the model to readily handle relation phrases containing multiple verbs. A consequence ofthis pattern is that the relation phrase must be a contiguous span of words in the sentence. The syntactic constraint eliminates the incoherent relation phrases returned by existing systems. example, given the sentence For Extendicare agreed to buy Arbor Health Care for about US $432 million in cash and assumed debt. TEXTRUNNER returns the extraction (Arbor Health Care, for assumed, debt). The phrase for assumed is clearly not a valid relation phrase: it begins with a preposition and splices together two distant words in the sentence. The syntactic constraint prevents this type of error by simply restricting relation phrases to match the pattern in Figure 1. The syntactic constraint reduces uninformative extractions by capturing relation phrases expressed via LVCs. For example, the POS pattern matched against the sentence “Faust made a deal with the Devil,” would result in the relation phrase made a deal with, instead of the uninformative made. Finally, we require the relation phrase to appear between its two arguments in the sentence. This is a common constraint that has been implicitly enforced in other open extractors. 1538 3.2 Lexical Constraint While the syntactic constraint greatly reduces uninformative extractions, it can sometimes match relation phrases that are so specific that they have only a few possible instances, even in a Web-scale corpus. Consider the sentence: The Obama administration is offering only modest greenhouse gas reduction targets at the conference. The POS pattern will match the phrase: is offering only modest greenhouse gas reduction targets at (1) Thus, there are phrases that satisfy the syntactic constraint, but are not relational. To overcome this limitation, we introduce a lexical constraint that is used to separate valid relation phrases from overspecified relation phrases, like the example in (1). The constraint is based on the intuition that a valid relation phrase should take many distinct arguments in a large corpus. The phrase in (1) is specific to the argument pair (Obama administration, conference), so it is unlikely to represent a bona fide relation. We describe the implementation details of the lexical constraint in Section 4. 3.3 Limitations Our constraints represent an idealized model of relation phrases in English. This raises the question: How much recall is lost due to the constraints? To address this question, we analyzed Wu and Weld’s set of 300 sentences from a set of random Web pages, manually identifying all verb-based relationships between noun phrase pairs. This resulted in a set of 327 relation phrases. For each relation phrase, we checked whether it satisfies our constraints. We found that 85% of the relation phrases do satisfy the constraints. Of the remaining 15%, we identified some of the common cases where the constraints were violated, summarized in Table 3. Many of the example relation phrases shown in Table 3 involve long-range dependencies between words in the sentence. These types of dependencies are not easily representable using a pattern over POS tags. A deeper syntactic analysis of the input sentence would provide a much more general language for modeling relation phrases. For example, one could create a model of relations expressed in Table 3: Approximately 85% of the binary verbal relation phrases in a sample of Web sentences satisfy our constraints. terms of dependency parse features that would capture the non-contiguous relation phrases in Table 3. Previous work has shown that dependency paths do indeed boost the recall of relation extraction systems (Wu and Weld, 2010; Mintz et al., 2009). While using dependency path features allows for a more flexible model of relations, it significantly increases pro- cessing time, which is problematic for Web-scale extraction. Further, we have found that this increased recall comes at the cost of lower precision on Web text (see Section 5). The results in Table 3 are similar to Banko and Etzioni’s findings that a set of eight POS patterns cover a large fraction of binary verbal relation phrases. However, their analysis was based on a set of sentences known to contain either a company acquisition or birthplace relationship, while our results are on a random sample of Web sentences. We applied Banko and Etzioni’s verbal patterns to our random sample of 300 Web sentences, and found that they cover approximately 69% of the relation phrases in the corpus. The gap in recall between this and the 85% shown in Table 3 is largely due to LVC relation phrases (made a deal with) and phrases containing multiple verbs (refuses to return to), which their patterns do not cover. In sum, our model is by no means complete. However, we have empirically shown that the majority of binary verbal relation phrases in a sample of Web sentences are captured by our model. By focusing on this subset of language, our model can 1539 be used to perform Open IE at significantly higher precision than before. 4 REVERB This section introduces REVERB, a novel open extractor based on the constraints defined in the previous section. REVERB first identifies relation phrases that satisfy the syntactic and lexical constraints, and then finds a pair of NP arguments for each identified relation phrase. The resulting extractions are then assigned a confidence score using a logistic regression classifier. This algorithm differs in three important ways from previous methods (Section 2). First, the relation phrase is identified “holistically” rather than word-by-word. Second, potential phrases are filtered based on statistics over a large corpus (the implementation of our lexical constraint). Finally, REVERB is “relation first” rather than “arguments first”, which enables it to avoid a common error made by previous methods—confusing a noun in the relation phrase for an argument, e.g. the noun deal in made a deal with. 4.1 Extraction Algorithm REVERB takes as input a POS-tagged and NPchunked sentence and returns a set of (x, r, y) extraction triples.2 Given an input sentence s, REVERB uses the following extraction algorithm: 1. Relation Extraction: For each verb v in s, find the longest sequence of words rv such that (1) rv starts at v, (2) rv satisfies the syntactic constraint, and (3) rv satisfies the lexical constraint. If any pair of matches are adjacent or overlap in s, merge them into a single match. 2. Argument Extraction: For each relation phrase r identified in Step 1, find the nearest noun phrase x to the left of r in s such that x is not a relative pronoun, WHO-adverb, or existential “there”. Find the nearest noun phrase y to the right of r in s. If such an (x, y) pair could be found, return (x, r, y) as an extraction. We check whether a candidate relation phrase rv satisfies the syntactic constraint by matching it against the regular expression in Figure 1. 2REVERB uses OpenNLP for POS tagging and NP chunking: http : / / opennlp . s ource forge . net / To determine whether rv satisfies the lexical constraint, we use a large dictionary D of relation phrases that are known to take many distinct arguments. In an offline step, we construct D by finding all matches of the POS pattern in a corpus of 500 million Web sentences. For each matching relation phrase, we heuristically identify its arguments (as in Step 2 above). We set D to be the set of all relation phrases that take at least k distinct argument pairs in the set of extractions. In order to allow for minor variations in relation phrases, we normalize each relation phrase by removing inflection, auxiliary verbs, adjectives, and adverbs. Based on experiments on a held-out set of sentences, we found that a value of k = 20 works well for filtering out overspecified relations. This results in a set of approximately 1.7 million distinct normalized relation phrases, which are stored in memory at extraction time. As an example of the extraction algorithm in action, consider the following input sentence: Hudson was born in Hampstead, which is a suburb of London. Step 1 of the algorithm identifies three relation phrases that satisfy the syntactic and lexical constraints: was, born in, and is a suburb of. The first two phrases are adjacent in the sentence, so they are merged into the single relation phrase was born in. Step 2 then finds an argument pair for each relation phrase. For was born in, the nearest NPs are (Hudson, Hampstead). For is a suburb of, the extractor skips over the NP which and chooses the argument pair (Hampstead, London). The final output is e1: (Hudson, was born in, Hampstead) e2: (Hampstead, is a suburb of, London). 4.2 Confidence Function The extraction algorithm in the previous section has high recall, but low precision. Like with previous open extractors, we want way to trade recall for precision by tuning a confidence threshold. We use a logistic regression classifier to assign a confidence score to each extraction, which uses the features shown in Table 4. All of these features are efficiently computable and relation independent. We trained the confidence function by manually labeling the extractions from a set of 1, 000 sentences from the Web and Wikipedia as correct or incorrect. 1540 Table 4: REVERB uses these features to assign a confidence score to an extraction (x, r, y) from a sentence s using a logistic regression classifier. Previous open extractors require labeled training data to learn a model of relations, which is then used to extract relation phrases from text. In contrast, REVERB uses a specified model of relations for extraction, and requires labeled data only for assigning confidence scores to its extractions. Learning a confidence function is a much simpler task than learning a full model of relations, using two orders of magnitude fewer training examples than TEXTRUNNER or WOE. 4.3 TEXTRUNNER-R The model of relation phrases used by REVERB is specified, but could a TEXTRUNNER-like system learn this model from training data? While it is difficult to answer such a question for all possible permutations of features sets, training examples, and learning biases, we demonstrate that TEXTRUNNER itself cannot learn REVERB’s model even when re-trained using the output of REVERB as labeled training data. The resulting system, TEXTRUNNER-R, uses the same feature representation as TEXTRUNNER, but different parameters, and a different set of training examples. To generate positive instances, we ran REVERB on the Penn Treebank, which is the same dataset that TEXTRUNNER is trained on. To generate negative instances from a sentence, we took each noun phrase pair in the sentence that does not appear as arguments in a REVERB extraction. This process resulted in a set of 67, 562 positive instances, and 356, 834 negative instances. We then passed these labeled examples to TEXTRUNNER’s training procedure, which learns a linear-chain CRF using closedclass features like POS tags, capitalization, punctuation, etc.TEXTRUNNER-R uses the argument-first extraction algorithm described in Section 2. 5 Experiments We compare REVERB to the following systems: • • REVERB¬lex - The REVERB system described iRn the previous section, but without the lexical constraint. REVERB¬lex uses the same confidence function as REVERB. TEXTRUNNER - Banko and Etzioni’s 2008 extractor, which uses a second order linear-chain CRF trained on extractions heuristically generated from the Penn Treebank. TEXTRUNNER uses shallow linguistic features in its CRF, which come from the same POS tagger and NPchunker that REVERB uses. • • • TEXTRUNNER-R - Our modification to TEXTRUNNER, which uses the same extraction code, but with a model of relations trained on REVERB extractions. WOEpos - Wu and Weld’s modification to TEXTRUNNER, which uses a model of relations learned from extractions heuristically generated from Wikipedia. WOEparse - Wu and Weld’s parser-based extractor, which uses a large dictionary of dependency path patterns learned from heuristic extractions generated from Wikipedia. Each system is given a set of sentences as input, and returns a set of binary extractions as output. We created a test set of 500 sentences sampled from the Web, using Yahoo’s random link service.3 After run3http : / /random . yahoo .com/bin/ryl 1541 rAaneUudeCrv aPdRCAPreu0 0 . 543210 REV RB EV RBWOET XT-WOET XT¬¬lleexx ppaarrssee RRUUNNNNEERR--RR ppooss RRUUNNNNEERR Figure 2: REVERB outperforms state-of-the-art open extractors, with an AUC more than twice that of TEXTRUNNER or WOEpos, and 38% higher than WOEparse. CCoommppaarriissoonn ooff RREEVVEERRBB--BBaasseedd SSyysstteemmss RReeccaall l Figure 3: The lexical constraint gives REVERB a boost in precision and recall over REVERB¬lex. TEXTRUNNER-R is unable to learn the model used by REVERB, which results in lower precision and recall. ning each extractor over the input sentences, two human judges independently evaluated each extraction as correct or incorrect. The judges reached agreement on 86% of the extractions, with an agreement score of κ = 0.68. We report results on the subset of the data where the two judges concur. Thejudges labeled uninformative extractions conservatively. That is, if critical information was dropped from the relation phrase but included in the second argument, it is labeled correct. For example, both the extractions (Ackerman, is a professor of, biology) and (Ackerman, is, a professor of biology) are considered correct. Each system returns confidence scores for its extractions. For a given threshold, we can measure the precision and recall of the output. Precision is the fraction of returned extractions that are correct. Recall is the fraction of correct extractions in EExxttrraaccttiioonnss RReeccaall l Figure 4: REVERB achieves significantly higher precision than state-of-the-art Open IE systems, and comparable recall to WOEparse. the corpus that are returned. We use the total number of extractions labeled as correct by the judges as our measure of recall for the corpus. In order to avoid double-counting, we treat extractions that differ superficially (e.g., different punctuation or dropping inessential modifiers) as a single extraction. We compute a precision-recall curve by varying the confidence threshold, and then compute the area under the curve (AUC). 5.1 Results Figure 2 shows the AUC of each system. REVERB achieves an AUC that is 30% higher than WOEparse and is more than double the AUC of WOEpos or TEXTRUNNER. The lexical constraint provides a significant boost in performance, with REVERB achieving an AUC 23% higher than REVERB¬lex. REVERB proves to be a useful source of training data, with TEXTRUNNER-R having an AUC 71% higher than TEXTRUNNER and performing on par with WOEpos. From the training data, TEXTRUNNER-R was able to learn a model that predicts contiguous relation phrases, but still returned incoherent relation phrases (e.g., starting with a preposition) and overspecified relation phrases. These errors are due to TEXTRUNNER-R overfitting the training data and not having access to the lexical constraint. Figure 3 shows the precision-recall curves of the systems introduced in this paper. TEXTRUNNER-R has much lower precision than REVERB and 1542 RReellaattiioonnss OOnnllyy RReeccaall l Figure 5: On the subtask of identifying relations phrases, REVERB is able to achieve even higher precision and recall than other systems. REVERB¬lex at all levels of recall. The lexical constraint gives REVERB a boost in precision over REVERB¬lex, reducing overspecified extractions from 20% of REVERB¬lex’s output to 1% of REVERB’s. The lexical constraint also boosts recall over REVERB¬lex, since REVERB is able to find a correct relation phrase where REVERB¬lex finds an overspecified one. Figure 4 shows the precision-recall curves of REVERB and the external systems. REVERB has much higher precision than the other systems at nearly all levels of recall. In particular, more than 30% of REVERB’s extractions are at precision 0.8 or higher, compared to virtually none for the other systems. WOEparse achieves a slightly higher recall than REVERB (0.62 versus 0.64), but at the cost of lower precision. In order to highlight the role of the relational model of each system, we also evaluate their performance on the subtask of extracting just the relation phrases from the input text. Figure 5 shows the precision-recall curves for each system on the relation phrase-only evaluation. In this case, REVERB has both higher precision and recall than the other systems. REVERB’s biggest improvement came from the elimination of incoherent extractions. Incoherent extractions were a large fraction of the errors made by previous systems, accounting for approximately 13% of TEXTRUNNER’s extractions, 15% of WOEpos’s, and 30% of WOEparse’s. Uninformative 162857% IOCNmoRt-vhparenE -yrsVc,paoiEtenlRrvcetailBfgu oev-tdinoIur bsgecplroPahtOliroeStsni/c ,phEiunrxhcaoktsreincgtieoransgument Table 5: The majority of the incorrect extractions returned by REVERB are due to errors in argument extraction. extractions had a smaller effect on other systems’ precision, accounting for 4% of WOEparse’s extractions, 5% of WOEpos’s, and 7% of TEXTRUNNER’s, while only appearing in 1% of REVERB’s extractions. REVERB’s reduction in uninformative extractions resulted in a boost in recall, capturing many LVC relation phrases missed by other systems (like those shown in Table 2). To test the systems’ speed, we ran each extractor on a set of 100, 000 sentences using a Pentium 4 machine with 4GB of RAM. The processing times were 16 minutes for REVERB, 21 minutes for TEXTRUNNER, 21 minutes for WOEpos, and 11 hours for WOEparse. The times for REVERB, TEXTRUNNER, and WOEpos are all approximately the same, since they all use the same POS-tagging and NP-chunking software. WOEparse processes each sentence with a dependency parser, resulting in much longer processing time. 5.2 REVERB Error Analysis To better understand the limitations of REVERB, we performed a detailed analysis of its errors in precision (incorrect extractions returned by REVERB) and its errors in recall (correct extractions that REVERB missed). Table 5 summarizes the types of incorrect extractions that REVERB returns. We found that 65% of the incorrect extractions returned by REVERB were cases where a relation phrase was correctly identified, but the argument-finding heuristics failed. The remaining errors were cases where REVERB extracted an incorrect relation phrase. One common mistake that REVERB made was extracting a relation phrase that expresses an n-ary relationship via a ditransitive verb. For example, given the sentence 1543 521873% ICRPdOeolEunSaVt/dciEofhneRuodBnftiakl-dmneMrgnotdiersfoyresudpctroebEcyrxieftlcrxa icertlgaounimcosne tsrain Table 6: The majority of extractions that were missed by REVERB were cases where the correct relation phrase was found, but the arguments were not correctly identified. “I gave him 15 photographs,” REVERB extracts (I, gave, him). These errors are due to the fact that REVERB only models binary relations. Table 6 summarizes the correct extractions that were extracted by other systems and were not extracted by REVERB. As with the false positive extractions, the majority of false negatives (52%) were due to the argument-finding heuristics choosing the wrong arguments, or failing to extract all possible arguments (in the case of coordinating conjunctions). Other sources of failure were due to the lexical constraint either failing to filter out an overspecified relation phrase or filtering out a valid relation phrase. These errors hurt both precision and recall, since each case results in the extractor overlooking a correct relation phrase and choosing another. 5.3 Evaluation At Scale Section 5.1 shows that REVERB outperforms existing Open IE systems when evaluated on a sample of sentences. Previous work has shown that the frequency of an extraction in a large corpus is useful for assessing the correctness of extractions (Downey et al., 2005). Thus, it is possible a priori that REVERB’s gains over previous systems will diminish when extraction frequency is taken into account. In fact, we found that REVERB’s advantage over TEXTRUNNER when run at scale is qualitatively similar to its advantage on single sentences. We ran both REVERB and TEXTRUNNER on Banko and Etzioni’s corpus of 500 million Web sentences and examined the effect of redundancy on precision. As Downey’s work predicts, precision increased in both systems for extractions found multiple times, compared with extractions found only once. However, REVERB had higher precision than TEXTRUNNER at all frequency thresholds. In fact, REVERB’s frequency 1 extractions had a precision of 0.75, which TEXTRUNNER could not approach even with frequency 10 extractions, which had a precision of 0.34. Thus, REVERB is able to return more correct extractions at a higher precision than TEXTRUNNER, even when redundancy is taken into account. 6 Conclusions and Future Work The paper’s contributions are as follows: • We have identified and analyzed the problems oWfe ein hcaovhee riedennt tainfide dun ainndfo armnaaltyivzeed de txhtera pctrioobnlse mfors Open IE systems, and shown their prevalence for systems such as TEXTRUNNER and WOE. • We articulated general, easy-to-enforce consWtreain atrst on binary, nveerrabl,-b aesaseyd rteo-laetniofonr phrases in English that ameliorate these problems and yield richer and more informative relations (see, for example, Table 2). • Based on these constraints, we designed, implemented, haensde e cvoanlustartaeind tsh,e w wReE dVeEsRigBn eedx,tr iamc-tor, which substantially outperforms previous Open IE systems in both recall and precision. • We make REVERB and the data used in our experiments available to the research community.4 In future work, we plan to explore utilizing our constraints to improve the performance of learned CRF models. Roth et al. have shown how to incorporate constraints into CRF learners (Roth and Yih, 2005). It is natural, then, to consider whether the combination of heuristically labeled training examples, CRF learning, and our constraints will result in superior performance. The error analysis in Section 5.2 also suggests natural directions for future work. For instance, since many of REVERB’s errors are due to incorrect arguments, improved methods for argument extraction are in order. Acknowledgments We would like to thank Mausam, Dan Weld, Yoav Artzi, Luke Zettlemoyer, members of the KnowItAll 4http : / / reverb . cs .washingt on . edu 1544 group, and the anonymous reviewers for their helpful comments. This research was supported in part by NSF grant IIS-0803481, ONR grant N00014-08- 1-0431, and DARPA contract FA8750-09-C-0179, and carried out at the University of Washington’s Turing Center. References David J. Allerton. 2002. Stretched Verb Constructions in English. Routledge Studies in Germanic Linguistics. Routledge (Taylor and Francis), New York. Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The berkeley framenet project. In Proceedings of the 1 international conference on Computational 7th linguistics, pages 86–90. Michele Banko and Oren Etzioni. 2008. The tradeoffs between open and traditional relation extraction. In Proceedings of ACL-08: HLT, pages 28–36, Columbus, Ohio, June. Association for Computational Linguistics. Michele Banko, Michael J. Cafarella, Stephen Soderland, Matt Broadhead, and Oren Etzioni. 2007. Open information extraction from the web. In In the Proceedings of the 20th International Joint Conference on Artificial Intelligence, pages 2670–2676, January. Jonathan Berant, Ido Dagan, and Jacob Goldberger. 2011. Global learning of typed entailment rules. In Proceedings of ACL, Portland, OR. Janara Christensen, Mausam, Stephen Soderland, and Oren Etzioni. 2010. Semantic role labeling for open information extraction. In Proceedings of the NAACL HLT 2010 First International Workshop on Formalisms and Methodology for Learning by Reading, FAM-LbR ’ 10, pages 52–60, Stroudsburg, PA, USA. Association for Computational Linguistics. Doug Downey, Oren Etzioni, and Stephen Soderland. 2005. A probabilistic model of redundancy in information extraction. In IJCAI, pages 1034–1041. Gregory Grefenstette and Simone Teufel. 1995. Corpusbased method for automatic identification of support verbs for nominalizations. In Proceedings of the seventh conference on European chapter of the Association for Computational Linguistics, pages 98–103, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc. Raphael Hoffmann, Congle Zhang, and Daniel S. Weld. 2010. Learning 5000 relational extractors. In Proceedings ofthe 48thAnnual Meeting ofthe Association for Computational Linguistics, ACL ’ 10, pages 286– 295, Stroudsburg, PA, USA. Association for Computational Linguistics. Holmer Hemsen Kathrin Eichler and Gnter Neumann. 2008. Unsupervised relation extraction from web documents. In LREC. http://www.lrecconf.org/proceedings/lrec2008/. J. Kim and D. Moldovan. 1993. Acquisition of semantic patterns for information extraction from corpora. In Procs. of Ninth IEEE Conference on Artificial Intelligence for Applications, pages 171–176. Dekang Lin and Patrick Pantel. 2001. DIRT-Discovery of Inference Rules from Text. In Proceedings of ACM Conference on Knowledge Discovery and Data Mining(KDD-01), pages pp. 323–328. Thomas Lin, Mausam, and Oren Etzioni. 2010. Identifying Functional Relations in Web Text. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1266–1276, Cambridge, MA, October. Association for Computational Linguistics. Paul Kingsbury Martha and Martha Palmer. 2002. From treebank to propbank. In In Proceedings of LREC2002. Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In ACL-IJCNLP ’09: Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP: Volume 2, pages 1003–101 1, Morristown, NJ, USA. Association for Computational Linguistics. V. Punyakanok, D. Roth, and W. Yih. 2008. The importance of syntactic parsing and inference in semantic role labeling. Computational Linguistics, 34(2). E. Riloff. 1996. Automatically constructing extraction patterns from untagged text. In Procs. of the Thirteenth National Conference on Artificial Intelligence (AAAI-96), pages 1044–1049. Alan Ritter, Mausam, and Oren Etzioni. 2010. A Latent Dirichlet Allocation Method for Selectional Preferences. In ACL. Dan Roth and Wen-tau Yih. 2005. Integer linear programming inference for conditional random fields. In Proceedings of the 22nd international conference on Machine learning, ICML ’05, pages 736–743, New York, NY, USA. ACM. Stefan Schoenmackers, Oren Etzioni, Daniel S. Weld, and Jesse Davis. 2010. Learning first-order horn clauses from web text. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, EMNLP ’ 10, pages 1088–1098, Stroudsburg, PA, USA. Association for Computational Linguistics. Satoshi Sekine. 2006. On-demand information extraction. In Proceedings of the COLING/ACL on Main 1545 conference poster sessions, pages 73 1–738, Morristown, NJ, USA. Association for Computational Linguistics. Yusuke Shinyama and Satoshi Sekine. 2006. Preemptive Information Extraction using Unrestricted Relation Discovery. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 304–3 11, New York City, USA, June. Association for Computational Linguistics. Stephen Soderland, Brendan Roof, Bo Qin, Shi Xu, Mausam, and Oren Etzioni. 2010. Adapting open information extraction to domain-specific relations. AI Magazine, 3 1(3):93–102. S. Soderland. 1999. Learning Information Extraction Rules for Semi-Structured and Free Text. Machine Learning, 34(1-3):233–272. Lucia Specia and Enrico Motta. 2006. M.: A hybrid approach for extracting semantic relations from texts. In In. Proceedings of the 2 nd Workshop on Ontology Learning and Population, pages 57–64. Suzanne Stevenson, Afsaneh Fazly, and Ryan North. 2004. Statistical measures of the semi-productivity of light verb constructions. In 2nd ACL Workshop on Multiword Expressions, pages 1–8. M. Stevenson. 2004. An unsupervised WordNet-based algorithm for relation extraction. In Proceedings of the “Beyond Named Entity ” workshop at the Fourth International Conference on Language Resources and Evalutaion (LREC’04). Kristina Toutanova, Aria Haghighi, and Christopher D. Manning. 2008. A global joint model for semantic role labeling. Computational Linguistics, 34(2): 161– 191. Fei Wu and Daniel S. Weld. 2010. Open information ex- traction using Wikipedia. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, ACL ’ 10, pages 118–127, Morristown, NJ, USA. Association for Computational Linguistics. Jun Zhu, Zaiqing Nie, Xiaojiang Liu, Bo Zhang, and Ji-Rong Wen. 2009. StatSnowball: a statistical approach to extracting entity relationships. In WWW ’09: Proceedings of the 18th international conference on Worldwide web, pages 101–1 10, New York, NY, USA. ACM.

5 0.37728947 128 emnlp-2011-Structured Relation Discovery using Generative Models

Author: Limin Yao ; Aria Haghighi ; Sebastian Riedel ; Andrew McCallum

6 0.36928216 96 emnlp-2011-Multilayer Sequence Labeling

7 0.35798168 129 emnlp-2011-Structured Sparsity in Structured Prediction

8 0.34195554 93 emnlp-2011-Minimum Imputed-Risk: Unsupervised Discriminative Training for Machine Translation

9 0.32279122 14 emnlp-2011-A generative model for unsupervised discovery of relations and argument classes from clinical texts

10 0.30740243 114 emnlp-2011-Relation Extraction with Relation Topics

11 0.30604357 87 emnlp-2011-Lexical Generalization in CCG Grammar Induction for Semantic Parsing

12 0.29047301 73 emnlp-2011-Improving Bilingual Projections via Sparse Covariance Matrices

13 0.287469 104 emnlp-2011-Personalized Recommendation of User Comments via Factor Models

14 0.2863178 131 emnlp-2011-Syntactic Decision Tree LMs: Random Selection or Intelligent Design?

15 0.26749587 29 emnlp-2011-Collaborative Ranking: A Case Study on Entity Linking

16 0.26630607 86 emnlp-2011-Lexical Co-occurrence, Statistical Significance, and Word Association

17 0.26253858 55 emnlp-2011-Exploiting Syntactic and Distributional Information for Spelling Correction with Web-Scale N-gram Models

18 0.25904587 67 emnlp-2011-Hierarchical Verb Clustering Using Graph Factorization

19 0.25880814 23 emnlp-2011-Bootstrapped Named Entity Recognition for Product Attribute Extraction

20 0.25201619 148 emnlp-2011-Watermarking the Outputs of Structured Prediction with an application in Statistical Machine Translation.

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(15, 0.015), (23, 0.115), (27, 0.354), (36, 0.021), (37, 0.026), (45, 0.077), (53, 0.025), (54, 0.028), (57, 0.028), (62, 0.033), (64, 0.027), (66, 0.03), (69, 0.014), (79, 0.028), (82, 0.041), (87, 0.014), (90, 0.014), (96, 0.04)]

similar papers list:

simIndex simValue paperId paperTitle

1 0.77070409 144 emnlp-2011-Unsupervised Learning of Selectional Restrictions and Detection of Argument Coercions

Author: Kirk Roberts ; Sanda Harabagiu

Abstract: Metonymic language is a pervasive phenomenon. Metonymic type shifting, or argument type coercion, results in a selectional restriction violation where the argument’s semantic class differs from the class the predicate expects. In this paper we present an unsupervised method that learns the selectional restriction of arguments and enables the detection of argument coercion. This method also generates an enhanced probabilistic resolution of logical metonymies. The experimental results indicate substantial improvements the detection of coercions and the ranking of metonymic interpretations.

same-paper 2 0.73042935 143 emnlp-2011-Unsupervised Information Extraction with Distributional Prior Knowledge

Author: Cane Wing-ki Leung ; Jing Jiang ; Kian Ming A. Chai ; Hai Leong Chieu ; Loo-Nin Teow

3 0.70985538 147 emnlp-2011-Using Syntactic and Semantic Structural Kernels for Classifying Definition Questions in Jeopardy!

Author: Alessandro Moschitti ; Jennifer Chu-carroll ; Siddharth Patwardhan ; James Fan ; Giuseppe Riccardi

Abstract: The last decade has seen many interesting applications of Question Answering (QA) technology. The Jeopardy! quiz show is certainly one of the most fascinating, from the viewpoints of both its broad domain and the complexity of its language. In this paper, we study kernel methods applied to syntactic/semantic structures for accurate classification of Jeopardy! definition questions. Our extensive empirical analysis shows that our classification models largely improve on classifiers based on word-language models. Such classifiers are also used in the state-of-the-art QA pipeline constituting Watson, the IBM Jeopardy! system. Our experiments measuring their impact on Watson show enhancements in QA accuracy and a consequent increase in the amount of money earned in game-based evaluation.

4 0.41101116 107 emnlp-2011-Probabilistic models of similarity in syntactic context

Author: Diarmuid O Seaghdha ; Anna Korhonen

Abstract: This paper investigates novel methods for incorporating syntactic information in probabilistic latent variable models of lexical choice and contextual similarity. The resulting models capture the effects of context on the interpretation of a word and in particular its effect on the appropriateness of replacing that word with a potentially related one. Evaluating our techniques on two datasets, we report performance above the prior state of the art for estimating sentence similarity and ranking lexical substitutes.

5 0.40841135 128 emnlp-2011-Structured Relation Discovery using Generative Models

Author: Limin Yao ; Aria Haghighi ; Sebastian Riedel ; Andrew McCallum

6 0.40650907 1 emnlp-2011-A Bayesian Mixture Model for PoS Induction Using Multiple Features

7 0.4050265 28 emnlp-2011-Closing the Loop: Fast, Interactive Semi-Supervised Annotation With Queries on Features and Instances

8 0.40476483 108 emnlp-2011-Quasi-Synchronous Phrase Dependency Grammars for Machine Translation

9 0.40232292 59 emnlp-2011-Fast and Robust Joint Models for Biomedical Event Extraction

10 0.40185279 123 emnlp-2011-Soft Dependency Constraints for Reordering in Hierarchical Phrase-Based Translation

11 0.40050918 136 emnlp-2011-Training a Parser for Machine Translation Reordering

12 0.40031448 98 emnlp-2011-Named Entity Recognition in Tweets: An Experimental Study

13 0.39859393 46 emnlp-2011-Efficient Subsampling for Training Complex Language Models

14 0.39775619 58 emnlp-2011-Fast Generation of Translation Forest for Large-Scale SMT Discriminative Training

15 0.39729697 8 emnlp-2011-A Model of Discourse Predictions in Human Sentence Processing

16 0.39689422 53 emnlp-2011-Experimental Support for a Categorical Compositional Distributional Model of Meaning

17 0.39672777 79 emnlp-2011-Lateen EM: Unsupervised Training with Multiple Objectives, Applied to Dependency Grammar Induction

18 0.39614934 22 emnlp-2011-Better Evaluation Metrics Lead to Better Machine Translation

19 0.39577964 68 emnlp-2011-Hypotheses Selection Criteria in a Reranking Framework for Spoken Language Understanding

20 0.39540094 61 emnlp-2011-Generating Aspect-oriented Multi-Document Summarization with Event-aspect model