acl acl2012 acl2012-56 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Geoffrey Zweig ; John C. Platt ; Christopher Meek ; Christopher J.C. Burges ; Ainur Yessenalina ; Qiang Liu
Abstract: This paper studies the problem of sentencelevel semantic coherence by answering SATstyle sentence completion questions. These questions test the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena. We tackle the problem with two approaches: methods that use local lexical information, such as the n-grams of a classical language model; and methods that evaluate global coherence, such as latent semantic analysis. We evaluate these methods on a suite of practice SAT questions, and on a recently released sentence completion task based on data taken from five Conan Doyle novels. We find that by fusing local and global information, we can exceed 50% on this task (chance baseline is 20%), and we suggest some avenues for further research.
Reference: text
sentIndex sentText sentNum sentScore
1 Burges Microsoft Research Redmond, WA 98052 Abstract This paper studies the problem of sentencelevel semantic coherence by answering SATstyle sentence completion questions. [sent-4, score-0.542]
2 These questions test the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena. [sent-5, score-0.195]
3 We tackle the problem with two approaches: methods that use local lexical information, such as the n-grams of a classical language model; and methods that evaluate global coherence, such as latent semantic analysis. [sent-6, score-0.182]
4 We evaluate these methods on a suite of practice SAT questions, and on a recently released sentence completion task based on data taken from five Conan Doyle novels. [sent-7, score-0.423]
5 , 2003; Turney, 2008), who used questions from the Test of English as a Foreign Language (TOEFL), Graduate Record Exams (GRE) and English as a Second Language (ESL) exams. [sent-20, score-0.158]
6 Tasks requiring broader competencies include logic puzzles and reading comprehension. [sent-21, score-0.178]
7 Logic puzzles drawn from the Law School Administration Test (LSAT) and the GRE were studied in (Lev et al. [sent-22, score-0.064]
8 , 1999) initiated a long line of research into reading comprehension based on test prep material (Charniak et al. [sent-25, score-0.242]
9 In this paper, we study a new class of problems intermediate in difficulty between the extremes of synonym detection and general question answering - the sentence completion questions found on the Scholastic Aptitude Test (SAT). [sent-29, score-0.631]
10 These questions present a sentence with one or two blanks that need to be filled in. [sent-30, score-0.224]
11 The questions are highly constrained in the sense that all the information necessary is present in the sentence itself without any other context. [sent-34, score-0.224]
12 The first of these examples is relatively simple: the second half of the sentence is a clear description ofthe type ofbehavior characterized by the desired adjective. [sent-36, score-0.066]
13 the contrast between medicine and poison that the correct answer involves a contrast, either useless vs. [sent-40, score-0.136]
14 In general, the questions require a combination of semantic and world knowledge as well as occasional logical reasoning. [sent-44, score-0.303]
15 We study the sentence completion task because we believe it is complex enough to pose a significant challenge, yet structured enough that progress may be possible. [sent-45, score-0.381]
16 As a first step, we have approached the problem from two points-of-view: first by exploiting local sentence structure, and secondly by measuring a novel form of global sentence coherence based on latent semantic analysis. [sent-46, score-0.39]
17 Also in the language modeling vein, but with potentially global context, we evaluate the use of a recurrent neural network language model. [sent-48, score-0.394]
18 In all the language modeling approaches, a model is used to compute a sentence probability with each of the potential completions. [sent-49, score-0.14]
19 To measure global coherence, we propose 602 a novel method based on latent semantic analysis (LSA). [sent-50, score-0.146]
20 We report results on a set of questions taken from a collection of SAT practice exams (Princeton-Review, 2010), and further validate the methods with the recently proposed MSR Sentence Completion Challenge set (Zweig and Burges, 2011). [sent-52, score-0.234]
21 Our paper thus makes the following contributions: First, we present the first published results on the SAT sentence completion task. [sent-53, score-0.381]
22 Secondly, we evaluate the effectiveness of both local n-gram information, and global coherence in the form of a novel LSA-based metric. [sent-54, score-0.11]
23 In contrast to our SAT-inspired task, the original answer is indicated. [sent-64, score-0.093]
24 These operate in two phases: first they find a set of potential replacement words, and then they rank them. [sent-68, score-0.076]
25 That paper also explores the use of Latent Semantic Analysis to measure the degree of similarity between a potential replacement and its context, but the results are poorer than others. [sent-73, score-0.149]
26 The SAT sentence completion sentences do not have this property and thus are more challenging. [sent-75, score-0.381]
27 Related to, but predating the Semeval lexical substitution task are the ESL synonym questions proposed by Turney (2001), and subsequently considered by numerous research groups including Terra and Clarke (2003) and Pado and Lapata (2007). [sent-76, score-0.204]
28 These questions are similar to the SemEval task, but in addition to the original word and the sentence context, the list of options is provided. [sent-77, score-0.224]
29 Other work on standardized tests includes the synonym and antonym tasks mentioned in Section 1, and more recent work on a SAT analogy task introduced by (Turney et al. [sent-79, score-0.126]
30 3 Sentence Completion via Language Modeling Perhaps the most straightforward approach to solving the sentence completion task is to form the complete sentence with each option in turn, and to evaluate its likelihood under a language model. [sent-83, score-0.447]
31 We begin with ngram models; first a classical n-gram backoff model (Chen and Goodman, 1999), and then a recently proposed class-based maximum-entropy n-gram model (Chen, 2009a; Chen, 2009b). [sent-86, score-0.077]
32 Therefore we evaluate the recurrent neural net model of (Mikolov et al. [sent-88, score-0.333]
33 , 2011a), and has the potential to encode sentence-span information in the network hidden-layer activations. [sent-92, score-0.076]
34 All bigrams occurring at least twice were retained in the model, along with all trigrams occurring at least three times. [sent-102, score-0.111]
35 The vocabulary consisted of all words occurring at least 100 times in the data, along with every word in the development or test sets. [sent-103, score-0.127]
36 This resulted in a 124k word vocabulary and 59M n-grams. [sent-104, score-0.055]
37 1), the smaller amount of training data allowed us to use 4-grams and a vocabulary cutoff of 3. [sent-106, score-0.055]
38 Both components are themselves maximum entropy n-gram models in which the probability of a word or class label l given history h is determined by Z1 exp(Pk fk(h, l)). [sent-127, score-0.069]
39 3 Recurrent Neural Net Language Model Many of the questions involve long-range dependencies between words. [sent-130, score-0.158]
40 While n-gram models have no ability to explicitly maintain long-span context, the recently proposed recurrent neural-net model of (Mikolov et al. [sent-131, score-0.18]
41 In this model, a set of neural net activations s(t) is maintained and updated at each sentence position t. [sent-136, score-0.262]
42 These activations encapsulate the sentence history up to the tth word in a realvalued vector which typically has several hundred dimensions. [sent-137, score-0.178]
43 The word at position t is represented as a binary vector w(t) whose length is the vocabulary size, and with a “1” in a position uniquely associated with the word, and “0” elsewhere. [sent-138, score-0.055]
44 Because of the recurrent connections, this model is similar to a nonlinear infinite impulse response (IIR) filter, and has the potential to model long span dependencies. [sent-142, score-0.22]
45 , 1990) is a widely used method for representing words and documents in a low dimensional vector space. [sent-146, score-0.078]
46 The method is based on applying singular value decomposition (SVD) to a matrix W representing the occurrence of words in documents. [sent-147, score-0.077]
47 SVD results in an approximation of W by the product of three matrices, one in which each word is represented as a low-dimensional vector, one in which each document is represented as a low dimensional vector, and a diagonal scaling matrix. [sent-148, score-0.079]
48 The similarity between two words can then be quantified as the cosine-similarity between their respective scaled vectors, and document similarity can be measured likewise. [sent-149, score-0.188]
49 The input is a collection of n documents which are expressed in terms of words from a vocabulary of size m. [sent-153, score-0.096]
50 These documents may be actual documents such as newspaper articles, or simply as in our case notional documents such as sentences. [sent-154, score-0.16]
51 In applications, d << n and d << m; for example one might have a 50, 000 word vocabulary and 1, 000, 000 documents and use a 300 dimensional subspace representation. [sent-160, score-0.17]
52 An important property of SVD is that the rows of US - which represents the words - behave sim- ilarly to the original rows of W, in the sense that the cosine similarity between two rows in US approximates the cosine similarity between the corresponding rows in W. [sent-161, score-0.418]
53 1 Total Word Similarity Perhaps the simplest way of doing sentence completion with LSA is to compute the total similarity of a potential answer a with the rest of the words in the sentence S, and to choose the most related option. [sent-164, score-0.653]
54 sWene ednecfinee S th,e a ntodt atol similarity as: totsim(a,S) = X sim(a,w) Xw∈S When the completion requires two words, total similarity is the sum of the contributions for both words. [sent-165, score-0.461]
55 2 Sentence Reconstruction Recall that LSA approximates a weighted worddocument matrix W as the product of low rank matrices U and V along with a scaling matrix S: W ≈ USVT. [sent-168, score-0.108]
56 Using singular value decomposition, tWhis ≈is UdSonVe so as to minimize the mean square reconstruction error Pij Qi2j where Q = W −USVT. [sent-169, score-0.234]
57 From the basic defPinition ofLSA, eQac =h c Wol u−mUnS SofV W (representing a document) is represented as Wj = USVjT, (1) that is, as a linear combination of the set of basis functions formed by the columns of US, with the combination weights specified in VjT. [sent-170, score-0.098]
58 Moreover, we may take the reconstruction error induced by this representation to be a measure of how consistent the new document is with the original set of documents used to determine U S and V (Bellegarda, 2000). [sent-172, score-0.276]
59 , 1990; Bellegarda, 2000), again with the objective of minimizing the reconstruction error. [sent-175, score-0.193]
60 We can evaluate the reconstruction quality by inserting the result in (1). [sent-177, score-0.193]
61 The reconstruction error is then ||(UUT − I)Wp||2 Note that if all the dimensions are retained, the reconstruction error is zero; in the case that only the highest singular vectors are used, however, it is not. [sent-178, score-0.469]
62 Due to the fact that the sentences vary in length we choose the number of retained singular vectors as a fraction f of the sentence length. [sent-179, score-0.19]
63 If the answer has n words we use the top nf components. [sent-180, score-0.093]
64 This logic makes sense for the sentence completion task as well, motivating us to evaluate it. [sent-186, score-0.417]
65 To do this, we adopt the procedure of (Coccaro and Jurafsky, 1998), using linear interpolation between the n-gram and LSA probabilities: p(w|history) = αpng (w|history) + (1 − α)plsa(w |history) The probability of a word given its history is computed by the LSA model in the following way. [sent-187, score-0.069]
66 Let m be the smallest cosine similarity between h and any word in the vocabulary V : m = minw∈V sim(h, w) . [sent-189, score-0.176]
67 The probability of a word w in the context of history h is given by Plsa(w|h) =Pq∈sVim(s(ihm,w(h), −q) m − m) Since similarity canP P be negative, subtracting the minimum (m) ensures that all the estimated probabilities are between 0 and 1. [sent-190, score-0.142]
68 First, we restrict the set of documents used to those which are “relevant” to a given test set. [sent-198, score-0.078]
69 This is done by requiring that a document contain at least one of the potential answerwords. [sent-199, score-0.082]
70 Secondly, we restrict the vocabulary to the set of words present in the test set. [sent-200, score-0.092]
71 This book contains eleven practice tests, and we used all the sentence completion questions in the first five tests as a development set, and all the questions in the last six tests as the test set. [sent-208, score-0.936]
72 This resulted in sets with 95 and 108 questions respectively. [sent-209, score-0.158]
73 This consists of a set of 1, 040 sentence completion questions based on sentences occurring in five Conan Doyle Sherlock Holmes novels, and is identical in format to the SAT questions. [sent-211, score-0.574]
74 Since there is no publically available collection of SAT questions suitable to training, our methods have all relied on unsupervised data. [sent-214, score-0.158]
75 2 Human Performance To provide human benchmark performance, we asked six native speaking high school students and five graduate students to answer the questions on the development set. [sent-227, score-0.251]
76 Instead, as with LSA, a “relevant” corpus was selected of the sentences which contain at least one answer option from either the Tab3LMRl-eNSogA2dNrtha:-eomPLldeM GrfoTmaD213n96. [sent-235, score-0.093]
77 First, the recurrent neural net has dramatically lower perplexity than the other methods. [sent-248, score-0.384]
78 1 - to use the total cosine similarity between a potential answer and the other words in the sentence - has performed best. [sent-258, score-0.32]
79 proach of using reconstruction error performed very well on the development set, but unremarkably on the test set. [sent-261, score-0.23]
80 For the LSA model, the linear combination has three inputs: the total word similarity, the cosine similarity between the sum of the answer word vectors and the sum of the rest of sentence’s word vectors, and the number of out-of-vocabulary terms in the answer. [sent-266, score-0.305]
81 Each additional language model beyond LSA contributes an additional input: the probability of the sentence under that language model. [sent-267, score-0.066]
82 In this case, the best combination is to blend LSA, the Good-Turing language model, and the recurrent neural network. [sent-283, score-0.325]
83 If we allow 2% of our tests to yield incorrectly false results, then for the SAT data, the combination of the Good-Turing smoothed language model with an LSA-based global similarity model (52% accuracy) is better that the baseline alone (42% accuracy). [sent-287, score-0.25]
84 Secondly, for the Holmes data, we can state that LSA total similarity beats the recurrent neural network, which in turn is better than the baseline n- gram model. [sent-288, score-0.349]
85 Encouragingly, onethird of the errors involve single-word questions which test the dictionary definition of a word. [sent-291, score-0.195]
86 ” At the other end of the difficulty spectrum are questions involving world knowledge and/or logical implications. [sent-297, score-0.201]
87 However, the ability to identify and manipulate logical relationships and embed world knowledge in a manner amenable to logical manipulation may be necessary for a full solution. [sent-301, score-0.086]
88 It is an interesting research question if this could be done implicitly with a machine learning technique, for example recurrent or recursive neural networks. [sent-302, score-0.276]
89 These questions are intriguing because they probe the ability to distinguish semantically coherent sentences from incoherent ones, and yet involve no more context than the single sentence. [sent-304, score-0.158]
90 We find that both local n-gram information and an LSA-based global coherence model do significantly better than chance, and that they can be effectively combined. [sent-305, score-0.11]
91 In Proceedings of the 2000 ANLP/NAACL Workshop on Reading comprehension tests as evaluation for computerbased language understanding sytems - Volume 6, ANLP/NAACL-ReadingComp ’00, pages 1–5. [sent-338, score-0.306]
92 Measuring the similarity between implicit semantic relations from the web. [sent-385, score-0.126]
93 Approximate statistical tests for comparing supervised classification learning algorithms. [sent-403, score-0.08]
94 A solution to Plato’s problem: The latent semantic analysis theory of the acquisition, induction, and representation of knowledge. [sent-444, score-0.098]
95 Empirical evaluation and combination of advanced language modeling techniques. [sent-467, score-0.083]
96 A machine learning approach to answering questions for reading comprehension tests. [sent-486, score-0.409]
97 A rule-based question answering system for reading comprehension tests. [sent-504, score-0.251]
98 In Proceedings of the 2000 ANLP/NAACL Workshop on Reading comprehension tests as evaluationfor computer-based language understanding sytems - Volume 6, ANLP/NAACL-ReadingComp ’00, pages 13– 19. [sent-505, score-0.263]
99 Parsing natural scenes and natural language with recursive neural networks. [sent-518, score-0.096]
100 In Proceedings of the 2000 ANLP/NAACL Workshop on Reading comprehension tests as evaluation for computerbased language understanding sytems - Volume 6, ANLP/NAACL-ReadingComp ’00, pages 28–35. [sent-564, score-0.306]
wordName wordTfidf (topN-words)
[('sat', 0.375), ('lsa', 0.356), ('completion', 0.315), ('mikolov', 0.193), ('reconstruction', 0.193), ('recurrent', 0.18), ('questions', 0.158), ('burges', 0.153), ('holmes', 0.15), ('zweig', 0.149), ('comprehension', 0.127), ('turney', 0.1), ('neural', 0.096), ('answer', 0.093), ('tests', 0.08), ('semeval', 0.079), ('reading', 0.078), ('mohammed', 0.075), ('bellegarda', 0.075), ('coccaro', 0.075), ('similarity', 0.073), ('msr', 0.071), ('history', 0.069), ('deerwester', 0.068), ('sentence', 0.066), ('conan', 0.064), ('doyle', 0.064), ('puzzles', 0.064), ('tomas', 0.064), ('svd', 0.064), ('coherence', 0.062), ('chen', 0.062), ('wi', 0.061), ('net', 0.057), ('sytems', 0.056), ('vocabulary', 0.055), ('semantic', 0.053), ('perplexity', 0.051), ('angeles', 0.05), ('secondly', 0.05), ('combination', 0.049), ('los', 0.049), ('global', 0.048), ('cosine', 0.048), ('unt', 0.048), ('synonym', 0.046), ('answering', 0.046), ('latent', 0.045), ('geoffrey', 0.045), ('rows', 0.044), ('landauer', 0.043), ('activations', 0.043), ('analogies', 0.043), ('benjamini', 0.043), ('burget', 0.043), ('cernocky', 0.043), ('computerbased', 0.043), ('hirschman', 0.043), ('irvine', 0.043), ('jarmasz', 0.043), ('kombrink', 0.043), ('lukas', 0.043), ('medicine', 0.043), ('novels', 0.043), ('psat', 0.043), ('sutskever', 0.043), ('szpakowicz', 0.043), ('terra', 0.043), ('logical', 0.043), ('vectors', 0.042), ('practice', 0.042), ('document', 0.042), ('backoff', 0.041), ('retained', 0.041), ('documents', 0.041), ('singular', 0.041), ('potential', 0.04), ('test', 0.037), ('antonyms', 0.037), ('graeme', 0.037), ('saif', 0.037), ('subspace', 0.037), ('newspaper', 0.037), ('dimensional', 0.037), ('network', 0.036), ('matrix', 0.036), ('logic', 0.036), ('replacement', 0.036), ('classical', 0.036), ('matrices', 0.036), ('ku', 0.036), ('ci', 0.035), ('occurring', 0.035), ('sim', 0.035), ('modeling', 0.034), ('exams', 0.034), ('wp', 0.034), ('clarkson', 0.034), ('dietterich', 0.034), ('emami', 0.034)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000025 56 acl-2012-Computational Approaches to Sentence Completion
Author: Geoffrey Zweig ; John C. Platt ; Christopher Meek ; Christopher J.C. Burges ; Ainur Yessenalina ; Qiang Liu
Abstract: This paper studies the problem of sentencelevel semantic coherence by answering SATstyle sentence completion questions. These questions test the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena. We tackle the problem with two approaches: methods that use local lexical information, such as the n-grams of a classical language model; and methods that evaluate global coherence, such as latent semantic analysis. We evaluate these methods on a suite of practice SAT questions, and on a recently released sentence completion task based on data taken from five Conan Doyle novels. We find that by fusing local and global information, we can exceed 50% on this task (chance baseline is 20%), and we suggest some avenues for further research.
2 0.14635584 145 acl-2012-Modeling Sentences in the Latent Space
Author: Weiwei Guo ; Mona Diab
Abstract: Sentence Similarity is the process of computing a similarity score between two sentences. Previous sentence similarity work finds that latent semantics approaches to the problem do not perform well due to insufficient information in single sentences. In this paper, we show that by carefully handling words that are not in the sentences (missing words), we can train a reliable latent variable model on sentences. In the process, we propose a new evaluation framework for sentence similarity: Concept Definition Retrieval. The new framework allows for large scale tuning and testing of Sentence Similarity models. Experiments on the new task and previous data sets show significant improvement of our model over baselines and other traditional latent variable models. Our results indicate comparable and even better performance than current state of the art systems addressing the problem of sentence similarity.
3 0.13369702 55 acl-2012-Community Answer Summarization for Multi-Sentence Question with Group L1 Regularization
Author: Wen Chan ; Xiangdong Zhou ; Wei Wang ; Tat-Seng Chua
Abstract: We present a novel answer summarization method for community Question Answering services (cQAs) to address the problem of “incomplete answer”, i.e., the “best answer” of a complex multi-sentence question misses valuable information that is contained in other answers. In order to automatically generate a novel and non-redundant community answer summary, we segment the complex original multi-sentence question into several sub questions and then propose a general Conditional Random Field (CRF) based answer summary method with group L1 regularization. Various textual and non-textual QA features are explored. Specifically, we explore four different types of contextual factors, namely, the information novelty and non-redundancy modeling for local and non-local sentence interactions under question segmentation. To further unleash the potential of the abundant cQA features, we introduce the group L1 regularization for feature learning. Experimental results on a Yahoo! Answers dataset show that our proposed method significantly outperforms state-of-the-art methods on cQA summarization task.
4 0.11635908 117 acl-2012-Improving Word Representations via Global Context and Multiple Word Prototypes
Author: Eric Huang ; Richard Socher ; Christopher Manning ; Andrew Ng
Abstract: Unsupervised word representations are very useful in NLP tasks both as inputs to learning algorithms and as extra word features in NLP systems. However, most of these models are built with only local context and one representation per word. This is problematic because words are often polysemous and global context can also provide useful information for learning word meanings. We present a new neural network architecture which 1) learns word embeddings that better capture the semantics of words by incorporating both local and global document context, and 2) accounts for homonymy and polysemy by learning multiple embeddings per word. We introduce a new dataset with human judgments on pairs of words in sentential context, and evaluate our model on it, showing that our model outperforms competitive baselines and other neural language models. 1
5 0.092517287 177 acl-2012-Sentence Dependency Tagging in Online Question Answering Forums
Author: Zhonghua Qu ; Yang Liu
Abstract: Online forums are becoming a popular resource in the state of the art question answering (QA) systems. Because of its nature as an online community, it contains more updated knowledge than other places. However, going through tedious and redundant posts to look for answers could be very time consuming. Most prior work focused on extracting only question answering sentences from user conversations. In this paper, we introduce the task of sentence dependency tagging. Finding dependency structure can not only help find answer quickly but also allow users to trace back how the answer is concluded through user conversations. We use linear-chain conditional random fields (CRF) for sentence type tagging, and a 2D CRF to label the dependency relation between sentences. Our experimental results show that our proposed approach performs well for sentence dependency tagging. This dependency information can benefit other tasks such as thread ranking and answer summarization in online forums.
6 0.087956883 161 acl-2012-Polarity Consistency Checking for Sentiment Dictionaries
7 0.080680296 214 acl-2012-Verb Classification using Distributional Similarity in Syntactic and Semantic Structures
8 0.065280661 127 acl-2012-Large-Scale Syntactic Language Modeling with Treelets
9 0.05926336 143 acl-2012-Mixing Multiple Translation Models in Statistical Machine Translation
10 0.057944801 67 acl-2012-Deciphering Foreign Language by Combining Language Models and Context Vectors
11 0.056863084 199 acl-2012-Topic Models for Dynamic Translation Model Adaptation
12 0.055338874 144 acl-2012-Modeling Review Comments
13 0.054849807 208 acl-2012-Unsupervised Relation Discovery with Sense Disambiguation
14 0.054318387 123 acl-2012-Joint Feature Selection in Distributed Stochastic Learning for Large-Scale Discriminative Training in SMT
15 0.053456098 217 acl-2012-Word Sense Disambiguation Improves Information Retrieval
16 0.052181236 69 acl-2012-Deep Learning for NLP (without Magic)
17 0.051282085 93 acl-2012-Fast Online Lexicon Learning for Grounded Language Acquisition
18 0.050975665 178 acl-2012-Sentence Simplification by Monolingual Machine Translation
19 0.049658131 74 acl-2012-Discriminative Pronunciation Modeling: A Large-Margin, Feature-Rich Approach
20 0.049622666 94 acl-2012-Fast Online Training with Frequency-Adaptive Learning Rates for Chinese Word Segmentation and New Word Detection
topicId topicWeight
[(0, -0.191), (1, 0.057), (2, 0.009), (3, -0.005), (4, -0.007), (5, 0.093), (6, -0.008), (7, 0.024), (8, -0.049), (9, -0.008), (10, -0.004), (11, -0.001), (12, 0.02), (13, 0.07), (14, -0.046), (15, 0.009), (16, 0.097), (17, 0.029), (18, 0.017), (19, -0.013), (20, 0.14), (21, -0.126), (22, -0.03), (23, -0.001), (24, 0.032), (25, -0.006), (26, -0.164), (27, -0.108), (28, -0.031), (29, 0.032), (30, -0.018), (31, -0.141), (32, -0.025), (33, -0.051), (34, 0.109), (35, -0.012), (36, 0.083), (37, -0.098), (38, 0.073), (39, 0.076), (40, 0.067), (41, -0.13), (42, -0.143), (43, -0.111), (44, 0.082), (45, -0.1), (46, 0.022), (47, 0.103), (48, -0.086), (49, -0.072)]
simIndex simValue paperId paperTitle
same-paper 1 0.9145999 56 acl-2012-Computational Approaches to Sentence Completion
Author: Geoffrey Zweig ; John C. Platt ; Christopher Meek ; Christopher J.C. Burges ; Ainur Yessenalina ; Qiang Liu
Abstract: This paper studies the problem of sentencelevel semantic coherence by answering SATstyle sentence completion questions. These questions test the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena. We tackle the problem with two approaches: methods that use local lexical information, such as the n-grams of a classical language model; and methods that evaluate global coherence, such as latent semantic analysis. We evaluate these methods on a suite of practice SAT questions, and on a recently released sentence completion task based on data taken from five Conan Doyle novels. We find that by fusing local and global information, we can exceed 50% on this task (chance baseline is 20%), and we suggest some avenues for further research.
2 0.82141072 145 acl-2012-Modeling Sentences in the Latent Space
Author: Weiwei Guo ; Mona Diab
Abstract: Sentence Similarity is the process of computing a similarity score between two sentences. Previous sentence similarity work finds that latent semantics approaches to the problem do not perform well due to insufficient information in single sentences. In this paper, we show that by carefully handling words that are not in the sentences (missing words), we can train a reliable latent variable model on sentences. In the process, we propose a new evaluation framework for sentence similarity: Concept Definition Retrieval. The new framework allows for large scale tuning and testing of Sentence Similarity models. Experiments on the new task and previous data sets show significant improvement of our model over baselines and other traditional latent variable models. Our results indicate comparable and even better performance than current state of the art systems addressing the problem of sentence similarity.
3 0.66008329 117 acl-2012-Improving Word Representations via Global Context and Multiple Word Prototypes
Author: Eric Huang ; Richard Socher ; Christopher Manning ; Andrew Ng
Abstract: Unsupervised word representations are very useful in NLP tasks both as inputs to learning algorithms and as extra word features in NLP systems. However, most of these models are built with only local context and one representation per word. This is problematic because words are often polysemous and global context can also provide useful information for learning word meanings. We present a new neural network architecture which 1) learns word embeddings that better capture the semantics of words by incorporating both local and global document context, and 2) accounts for homonymy and polysemy by learning multiple embeddings per word. We introduce a new dataset with human judgments on pairs of words in sentential context, and evaluate our model on it, showing that our model outperforms competitive baselines and other neural language models. 1
4 0.60022277 132 acl-2012-Learning the Latent Semantics of a Concept from its Definition
Author: Weiwei Guo ; Mona Diab
Abstract: In this paper we study unsupervised word sense disambiguation (WSD) based on sense definition. We learn low-dimensional latent semantic vectors of concept definitions to construct a more robust sense similarity measure wmfvec. Experiments on four all-words WSD data sets show significant improvement over the baseline WSD systems and LDA based similarity measures, achieving results comparable to state of the art WSD systems.
5 0.59347719 55 acl-2012-Community Answer Summarization for Multi-Sentence Question with Group L1 Regularization
Author: Wen Chan ; Xiangdong Zhou ; Wei Wang ; Tat-Seng Chua
Abstract: We present a novel answer summarization method for community Question Answering services (cQAs) to address the problem of “incomplete answer”, i.e., the “best answer” of a complex multi-sentence question misses valuable information that is contained in other answers. In order to automatically generate a novel and non-redundant community answer summary, we segment the complex original multi-sentence question into several sub questions and then propose a general Conditional Random Field (CRF) based answer summary method with group L1 regularization. Various textual and non-textual QA features are explored. Specifically, we explore four different types of contextual factors, namely, the information novelty and non-redundancy modeling for local and non-local sentence interactions under question segmentation. To further unleash the potential of the abundant cQA features, we introduce the group L1 regularization for feature learning. Experimental results on a Yahoo! Answers dataset show that our proposed method significantly outperforms state-of-the-art methods on cQA summarization task.
6 0.4542025 177 acl-2012-Sentence Dependency Tagging in Online Question Answering Forums
8 0.4304941 11 acl-2012-A Feature-Rich Constituent Context Model for Grammar Induction
9 0.42264205 2 acl-2012-A Broad-Coverage Normalization System for Social Media Language
10 0.38199118 67 acl-2012-Deciphering Foreign Language by Combining Language Models and Context Vectors
11 0.37918252 218 acl-2012-You Had Me at Hello: How Phrasing Affects Memorability
12 0.37812799 156 acl-2012-Online Plagiarized Detection Through Exploiting Lexical, Syntax, and Semantic Information
13 0.35818139 133 acl-2012-Learning to "Read Between the Lines" using Bayesian Logic Programs
14 0.35298088 127 acl-2012-Large-Scale Syntactic Language Modeling with Treelets
15 0.35065183 200 acl-2012-Toward Automatically Assembling Hittite-Language Cuneiform Tablet Fragments into Larger Texts
16 0.35047761 94 acl-2012-Fast Online Training with Frequency-Adaptive Learning Rates for Chinese Word Segmentation and New Word Detection
17 0.34448135 112 acl-2012-Humor as Circuits in Semantic Networks
18 0.34446603 165 acl-2012-Probabilistic Integration of Partial Lexical Information for Noise Robust Haptic Voice Recognition
19 0.34140825 163 acl-2012-Prediction of Learning Curves in Machine Translation
20 0.34128827 76 acl-2012-Distributional Semantics in Technicolor
topicId topicWeight
[(25, 0.373), (26, 0.033), (28, 0.045), (30, 0.021), (37, 0.026), (39, 0.053), (59, 0.026), (74, 0.046), (82, 0.015), (84, 0.022), (85, 0.031), (90, 0.128), (92, 0.053), (94, 0.022), (99, 0.035)]
simIndex simValue paperId paperTitle
1 0.95191491 49 acl-2012-Coarse Lexical Semantic Annotation with Supersenses: An Arabic Case Study
Author: Nathan Schneider ; Behrang Mohit ; Kemal Oflazer ; Noah A. Smith
Abstract: “Lightweight” semantic annotation of text calls for a simple representation, ideally without requiring a semantic lexicon to achieve good coverage in the language and domain. In this paper, we repurpose WordNet’s supersense tags for annotation, developing specific guidelines for nominal expressions and applying them to Arabic Wikipedia articles in four topical domains. The resulting corpus has high coverage and was completed quickly with reasonable inter-annotator agreement.
2 0.93589103 164 acl-2012-Private Access to Phrase Tables for Statistical Machine Translation
Author: Nicola Cancedda
Abstract: Some Statistical Machine Translation systems never see the light because the owner of the appropriate training data cannot release them, and the potential user ofthe system cannot disclose what should be translated. We propose a simple and practical encryption-based method addressing this barrier.
3 0.87338024 1 acl-2012-ACCURAT Toolkit for Multi-Level Alignment and Information Extraction from Comparable Corpora
Author: Marcis Pinnis ; Radu Ion ; Dan Stefanescu ; Fangzhong Su ; Inguna Skadina ; Andrejs Vasiljevs ; Bogdan Babych
Abstract: The lack of parallel corpora and linguistic resources for many languages and domains is one of the major obstacles for the further advancement of automated translation. A possible solution is to exploit comparable corpora (non-parallel bi- or multi-lingual text resources) which are much more widely available than parallel translation data. Our presented toolkit deals with parallel content extraction from comparable corpora. It consists of tools bundled in two workflows: (1) alignment of comparable documents and extraction of parallel sentences and (2) extraction and bilingual mapping of terms and named entities. The toolkit pairs similar bilingual comparable documents and extracts parallel sentences and bilingual terminological and named entity dictionaries from comparable corpora. This demonstration focuses on the English, Latvian, Lithuanian, and Romanian languages.
same-paper 4 0.74847293 56 acl-2012-Computational Approaches to Sentence Completion
Author: Geoffrey Zweig ; John C. Platt ; Christopher Meek ; Christopher J.C. Burges ; Ainur Yessenalina ; Qiang Liu
Abstract: This paper studies the problem of sentencelevel semantic coherence by answering SATstyle sentence completion questions. These questions test the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena. We tackle the problem with two approaches: methods that use local lexical information, such as the n-grams of a classical language model; and methods that evaluate global coherence, such as latent semantic analysis. We evaluate these methods on a suite of practice SAT questions, and on a recently released sentence completion task based on data taken from five Conan Doyle novels. We find that by fusing local and global information, we can exceed 50% on this task (chance baseline is 20%), and we suggest some avenues for further research.
5 0.50694609 44 acl-2012-CSNIPER - Annotation-by-query for Non-canonical Constructions in Large Corpora
Author: Richard Eckart de Castilho ; Sabine Bartsch ; Iryna Gurevych
Abstract: We present CSNIPER (Corpus Sniper), a tool that implements (i) a web-based multiuser scenario for identifying and annotating non-canonical grammatical constructions in large corpora based on linguistic queries and (ii) evaluation of annotation quality by measuring inter-rater agreement. This annotationby-query approach efficiently harnesses expert knowledge to identify instances of linguistic phenomena that are hard to identify by means of existing automatic annotation tools.
6 0.49533117 99 acl-2012-Finding Salient Dates for Building Thematic Timelines
7 0.49352583 206 acl-2012-UWN: A Large Multilingual Lexical Knowledge Base
8 0.48254794 138 acl-2012-LetsMT!: Cloud-Based Platform for Do-It-Yourself Machine Translation
9 0.47942641 72 acl-2012-Detecting Semantic Equivalence and Information Disparity in Cross-lingual Documents
10 0.47863111 214 acl-2012-Verb Classification using Distributional Similarity in Syntactic and Semantic Structures
11 0.47426122 36 acl-2012-BIUTEE: A Modular Open-Source System for Recognizing Textual Entailment
12 0.4733921 161 acl-2012-Polarity Consistency Checking for Sentiment Dictionaries
13 0.47314346 145 acl-2012-Modeling Sentences in the Latent Space
14 0.47311503 219 acl-2012-langid.py: An Off-the-shelf Language Identification Tool
15 0.47112277 217 acl-2012-Word Sense Disambiguation Improves Information Retrieval
16 0.47013667 202 acl-2012-Transforming Standard Arabic to Colloquial Arabic
18 0.4631772 76 acl-2012-Distributional Semantics in Technicolor
19 0.4631418 11 acl-2012-A Feature-Rich Constituent Context Model for Grammar Induction
20 0.46313965 123 acl-2012-Joint Feature Selection in Distributed Stochastic Learning for Large-Scale Discriminative Training in SMT