acl acl2012 acl2012-30 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Emily Pitler
Abstract: Prepositions and conjunctions are two of the largest remaining bottlenecks in parsing. Across various existing parsers, these two categories have the lowest accuracies, and mistakes made have consequences for downstream applications. Prepositions and conjunctions are often assumed to depend on lexical dependencies for correct resolution. As lexical statistics based on the training set only are sparse, unlabeled data can help ameliorate this sparsity problem. By including unlabeled data features into a factorization of the problem which matches the representation of prepositions and conjunctions, we achieve a new state-of-the-art for English dependencies with 93.55% correct attachments on the current standard. Furthermore, conjunctions are attached with an accuracy of 90.8%, and prepositions with an accuracy of 87.4%.
Reference: text
sentIndex sentText sentNum sentScore
1 edu Abstract Prepositions and conjunctions are two of the largest remaining bottlenecks in parsing. [sent-3, score-0.48]
2 Prepositions and conjunctions are often assumed to depend on lexical dependencies for correct resolution. [sent-5, score-0.484]
3 As lexical statistics based on the training set only are sparse, unlabeled data can help ameliorate this sparsity problem. [sent-6, score-0.222]
4 By including unlabeled data features into a factorization of the problem which matches the representation of prepositions and conjunctions, we achieve a new state-of-the-art for English dependencies with 93. [sent-7, score-0.838]
5 Furthermore, conjunctions are attached with an accuracy of 90. [sent-9, score-0.508]
6 1 Introduction Prepositions and conjunctions are two large remaining bottlenecks in parsing. [sent-12, score-0.48]
7 Machine translation is sensitive to parsing errors involving prepositions and conjunctions, because in some languages different attachment decisions in the parse of the source language sentence produce different translations. [sent-14, score-0.463]
8 , 2003) which uses a different postposition for different attachments; conjunction mis768 takes can cause word ordering mistakes when translating into Chinese (Huang, 1983). [sent-16, score-0.22]
9 Prepositions and conjunctions are often assumed to depend on lexical dependencies for correct resolution (Jurafsky and Martin, 2008). [sent-17, score-0.484]
10 Unlabeled data has been shown to improve the accuracy of conjunctions within complex noun phrases (Pitler et al. [sent-23, score-0.523]
11 However, it has so far been less effective within full parsing while first-order web-scale counts noticeably improved overall parsing in Bansal and Klein (201 1), the accuracy on conjunctions actually decreased when the web-scale features were added (Table 4 in that paper). [sent-26, score-0.657]
12 In this paper we show that unlabeled data can help prepositions and conjunctions, provided that the dependency representation is compatible with how the parsing problem is decomposed for learning and inference. [sent-27, score-0.658]
13 By incorporating unlabeled data into factorizations which capture the relevant dependencies for prepositions and conjunctions, we produce a parser for English which has an unlabeled attachment accuracy of 93. [sent-28, score-1.141]
14 5%, over an 18% reduction in error — over the best previously published parser (Bansal and Klein, 2011) on the current standard for dependency parsing. [sent-29, score-0.187]
15 The best model for conjunctions atProce dJienjgus, R ofep thueb 5lic0t hof A Knonruea ,l M 8-e1e4ti Jnugly o f2 t0h1e2 A. [sent-30, score-0.435]
16 5% reduction in error over MSTParser), and the best model for prepositions with 87. [sent-34, score-0.319]
17 We describe the dependency representations of prepositions and conjunctions in Section 2. [sent-37, score-0.847]
18 We discuss the implications of these representations for how learning and inference for parsing are decomposed (Section 3) and how unlabeled data may be used (Section 4). [sent-38, score-0.291]
19 2 Dependency Representations A dependency tree is a rooted, directed tree (or arborescence), in which the vertices are the words in the sentence plus an artificial root node, and each edge (h, m) represents a directed dependency relation from the head h to the modifier m. [sent-40, score-0.355]
20 Dependency parsing requires a conversion from these constituency trees to dependency trees. [sent-44, score-0.419]
21 The presence or absence of these noun phrase internal annotations interacts with constituency-todependency conversion program in ways which have effects on conjunctions and prepositions. [sent-47, score-0.658]
22 Figures 1 and 2 show how conjunctions and prepositions, respectively, are represented after the two different conversion processes. [sent-69, score-0.609]
23 2% of prepositions in the development set have a different parent under the two conversion types. [sent-72, score-0.599]
24 These representational differences have serious implications for how well various factorizations will be able to capture these two phenomena. [sent-73, score-0.225]
25 3 Implications of Representations on the Scope of Factorization Parsing requires a) learning to score potential parse trees, and b) given a particular scoring function, finding the highest scoring tree according to that function. [sent-74, score-0.264]
26 The conjunction is bolded, the left conjunct (in the linear order of the sentence) is underlined, and the right conjunct is italicized. [sent-77, score-0.334]
27 Four possible factorizations are: sin- gle edges (edge-based), pairs of edges which share a parent (siblings), pairs of edges where the child of one is the parent of the other (grandparents), and triples of edges where the child of one is the parent of two others (grandparent+sibling). [sent-79, score-0.916]
28 In this section, we discuss these factorizations and their relevance to conjunction and preposition representations. [sent-80, score-0.524]
29 Under edge-based scoring, the conjunction would be scored along with neither of its conjuncts in 1(a). [sent-83, score-0.274]
30 In Figure 1(c), the conjunction is scored along with its right conjunct only; in figure 1(e) along with its left conjunct only. [sent-84, score-0.334]
31 Furthermore, the conjunction is connected with an edge to either zero or one of its two arguments; at least one of the arguments is completely ignored in terms of scoring the conjunction. [sent-86, score-0.382]
32 This overloads the meaning of an edge; an edge indicates both a head-modifier relationship and a conjunction relationship. [sent-88, score-0.25]
33 For example, compare the two natural phrases dogs and cats and really nice. [sent-89, score-0.229]
34 dogs and cats are a good pair to conjoin, but cats is not a good modifier for dogs, so there is a tension when scoring an edge like (dogs, cats) : it should get a high score when actually indicating a conjunction and low otherwise. [sent-90, score-0.7]
35 In Figures 1(b), 1(d) and 1(f), the conjunction participates in a directed edge with each of the conjuncts. [sent-94, score-0.284]
36 Thus, in edge-based scoring, at least under Conversion 2 neither of the conjuncts is being ignored; however, the factorization scores each edge independently, so how compatible these two conjuncts are with each other cannot be included in the scoring of a tree. [sent-95, score-0.654]
37 Prepositions: For all of the examples in Figure 2, there is a directed edge from the head of the phrase that the preposition modifies to the preposition. [sent-96, score-0.316]
38 Differences in head finding rules account for the differences in preposition representations. [sent-97, score-0.216]
39 In the second example, the first conversion scheme chooses yesterday as the head of the overall NP, resulting in the edge yesterday→ of, while the second convertshioen e sdcgheem yees ignores temporal phrases cwohnedn finding the head, resulting in the more semantically meaningful opening→of. [sent-98, score-0.366]
40 t Staicmhielsa ltoy ,th ine pronoun ewxhaomsep lein, the first conversion scheme, while it attaches to the noun plans in the second. [sent-100, score-0.253]
41 With edge-based scoring, the object is not accessible when scoring where the preposition should attach, and PP-attachment is known to depend on the object of the preposition (Hindle and Rooth, 1993). [sent-101, score-0.482]
42 2 Sibling Scoring Another alternative factorization is to score siblings as well as parent-child edges (McDonald and Pereira, 2006). [sent-103, score-0.422]
43 Under this factorization, two of the three examples in Conversion 1 (and none of the examples in Conversion 2) in Figure 1 now include the conjunction and both conjuncts in the same score (Figures 1(c) and 1(e)). [sent-105, score-0.274]
44 The scoring for headmodifier dependencies and conjunction dependencies are again being overloaded: (debt, notes, and) and (debt, and, other) are both sibling parts in Figure 1(c), yet only one of them represents a conjunc- tion. [sent-106, score-0.656]
45 The position of the conjunction in the sibling is not enough to determine whether one is scoring a true conjunction relation or just the conjunction and a different sibling; in 1(c) the conjunction is on the right of its sibling argument, while in 1(e) the conjunction is on the left. [sent-107, score-1.536]
46 For none of the other preposition or conjunction examples does a sibling factorization bring more of the arguments into the scope of what is scored along with the preposition/conjunction. [sent-108, score-0.893]
47 Sibling scoring may have some benefit in that prepositions/conjunctions should have only one argument, so for prepositions (under both conversions) and conjunctions (under Conversion 2), the model can learn to disprefer the existence of any siblings and thus enforce choosing a single child. [sent-109, score-0.963]
48 Under Conversion 1, this factorization is particularly appropriate for prepositions, but would be unlikely to help conjunctions, which have no children. [sent-113, score-0.302]
49 4 Using Unlabeled Data Effectively Associations from unlabeled data have the potential to improve both conjunctions and prepositions. [sent-118, score-0.601]
50 We predict that web counts which include both conjuncts (for conjunctions), or which include both the attachment site and the object of a preposition (for prepositions) will lead to the largest improvements. [sent-119, score-0.418]
51 For the phrase dogs and cats, edge-based counts would measure the associations between dogs and and, and and and cats, but never any web counts that include both dogs and cats. [sent-120, score-0.5]
52 For the phrase ate spaghetti with a fork, edge-based scoring would not use any web counts involving both ate and fork. [sent-121, score-0.199]
53 The phrases trading and transacting versus trading and what provide an example of the difference between associations and counts. [sent-123, score-0.26]
54 The phrase trading and what has a higher count than the phrase trading and transacting, but trading and transacting are more highly associated. [sent-124, score-0.291]
55 In this paper, we use point-wise mutual information (PMI) to measure the strength of associations of words participating in potential conjunctions or prepositions. [sent-125, score-0.486]
56 For example, the choice of yesterday as the head of opening of trading here yesterday in Figure 2(c) or whose in 2(e) may make cluster-based features less useful than if the semantic heads were chosen (opening and plans, respectively). [sent-138, score-0.3]
57 5 Experiments The previous section motivated the use of unlabeled data for attaching prepositions and conjunctions. [sent-139, score-0.54]
58 We have also hypothesized that these features will be most effective when the data representation and the learning representation both capture relevant properties of prepositions and conjunctions. [sent-140, score-0.375]
59 We predict that Conversion 2 and a factorization which includes grand-parent scoring will achieve the highest performance. [sent-141, score-0.408]
60 In this section, we investigate the impact of unlabeled data on parsing accuracy using the two conversions and using each of the factorizations described in Section 3. [sent-142, score-0.482]
61 (2008), which includes features over all edges (h, m), grand-parent triples (h, m, c), and parent sibling triples (h, m, s). [sent-147, score-0.509]
62 For conjunctions, we only do this for triples of both conjunct and the conjunction (and if the conjunction is and or or and the two potential conjuncts are the same coarse grained part-of-speech). [sent-155, score-0.579]
63 For prepositions, we consider only cases in which the parent is a noun or a verb and the child is a noun (this corresponds to the cases considered by Hindle and Rooth (1993) and others). [sent-156, score-0.247]
64 For the scope of this paper we use only the above counts related to prepositions and conjunctions. [sent-159, score-0.427]
65 We augment the feature set used with the web-counts-based features relevant to prepositions and conjunctions and the cluster-based features. [sent-162, score-0.754]
66 The set of potential edges was pruned using the marginals produced by a first-order parser trained using exponentiated gradient descent (Collins et al. [sent-169, score-0.222]
67 We train the full parser for 15 iterations of averaged perceptron training (Collins, 2002), choose the iteration with the best unlabeled attachment score (UAS) on the development set, and apply the model after that iteration to the test set. [sent-171, score-0.378]
68 edu / nlp / dpo 3 / 773 We also ran MSTParser (McDonald and Pereira, 2006), the Berkeley constituency parser (Petrov and Klein, 2007), and the unmodified dpo3 Model 1 (Koo and Collins, 2010) using Conversion 2 (the current recommendations) for comparison. [sent-175, score-0.214]
69 The Berkeley parser was trained on the constituency trees of the PTB patched with Vadas and Curran (2007), and then the predicted parses were converted using pennconverter. [sent-177, score-0.318]
70 6 Results and Discussion Table 1 shows the unlabeled attachment scores, complete sentence exact match accuracies, and the accuracies of conjunctions and prepositions under Conversion 2. [sent-178, score-1.041]
71 7 The incorporation of the unlabeled data features (clusters and web counts) into the dpo3 parser yields a significantly better parser than dpo3 alone (93. [sent-179, score-0.418]
72 A parser which uses only grandparents (referred to as Model 0 in Koo and Collins (2010)) may therefore be preferable, as it contains far fewer parameters than a third-order parser. [sent-185, score-0.247]
73 While the grandparent factorization and the sib- ling factorization (Sib) are both “second-order” parsers, scoring up to two edges (involving three words) simultaneously, their results are quite different, with the sibling factorization scoring much worse. [sent-186, score-1.595]
74 This is particularly notable in the conjunction case, where the sibling model is over 5% absolute worse in accuracy than the grandparent model. [sent-187, score-0.657]
75 2 Impact of Unlabeled Data The unlabeled data features improved the already state-of-the-art dpo3 parser in UAS, complete sentence accuracy, conjunctions, and prepositions. [sent-197, score-0.292]
76 9 Overall, the results in Table 1 show that while the inclusion of unlabeled data improves parser performance, increasing the size of factorization matters even more. [sent-199, score-0.568]
77 Ablation experiments showed that cluster features have a larger impact on overall UAS, while count features have a larger impact on prepositions and conjunctions. [sent-200, score-0.377]
78 3 Comparison with Other Parsers The resulting dpo3+Unlabeled parser is significantly better than both versions of MSTParser and the Berkeley parser converted to dependencies across all four evaluations. [sent-202, score-0.332]
79 The MSTParser uses sibling scoring, so it is unsurprising that it performs less well on the new conversion. [sent-206, score-0.242]
80 While the converted constituency parser is not as good on dependencies as MSTParser overall, note that it is over a percent and a half better than MSTParser on attaching conjunctions (85. [sent-207, score-0.784]
81 The dependencies arising from the Berkeley constituency trees have higher conjunction accuracies than either the edge-based or sibling-based dpo3+Unlabeled parser. [sent-212, score-0.394]
82 However, once grandparents are included in the factorization, the dpo3+Unlabeled is significantly better at attaching conjunctions than the constituency parser, attaching conjunctions with an accuracy over 90%. [sent-213, score-1.228]
83 Therefore, some of the disadvantages of dependency parsing compared with constituency parsing can be compensated for with larger factorizations. [sent-214, score-0.291]
84 4 Impact of Data Representation Tables 2 and 3 show the results of the dpo3+Unlabeled parser for conjunctions and prepositions, respectively, under the two different conversions. [sent-221, score-0.561]
85 The data representation has an impact on which factorizations perform best. [sent-222, score-0.247]
86 Under Conversion 1, conjunctions are more accurate under a sibling parser than a grandparent parser, while the Prepositions Bolded items are the best in each column, or not significantly different (sign test, p < . [sent-223, score-0.995]
87 Conjunctions show a much stronger need for higher order factorizations than prepositions do. [sent-226, score-0.509]
88 This is not too surprising, as prepositions have more of a selectional preference than conjunctions, and so the preposition itself is more informative about where it should attach. [sent-227, score-0.469]
89 While prepositions do improve with larger factorizations, the improvement beyond edge-based is not significant for Conversion 2. [sent-228, score-0.319]
90 One hypothesis for why Conversion 1shows more of an improvement is that the wider scope leads to the semantic head being included; in Conversion 2, the semantic head is chosen as the parent of the preposition, so the wider scope is less necessary. [sent-229, score-0.32]
91 In the PP-attachment classification task, the two choices for where the preposition attaches are the previous verb or the previous noun, and the preposition itself has a noun object. [sent-233, score-0.379]
92 The ones that do attach to the preceeding noun or verb (not necessarily the preceeding word) and have a noun object (2323 prepositions) are attached by the dpo3+Unlabeled grandparent-scoring parser with 92. [sent-234, score-0.447]
93 Local attachments are more accurate prepositions are attached with 94. [sent-237, score-0.404]
94 8% accuracy if the correct parent is the immediately preceeding word (2364 cases) and only 79. [sent-238, score-0.214]
95 The preference is not necessarily for low — 775 attachments though: the prepositions whose parent is not the preceeding word are attached more accurately if the parent is the root word (usually corresponding to the main verb) of the sentence (90. [sent-240, score-0.685]
96 7 Conclusion Features derived from unlabeled data (clusters and web counts) significantly improve a state-of-the-art dependency parser for English. [sent-243, score-0.353]
97 We showed how well various factorizations are able to take advantage of these unlabeled data features, focusing our analysis on conjunctions and prepositions. [sent-244, score-0.791]
98 Including grandparents in the factorization increases the accuracy of conjunctions over 5% absolute over edgebased or sibling-based scoring. [sent-245, score-0.871]
99 The representation of the data is extremely important for how the problem should be factored–under the old Penn2Malt dependency representation, a sibling parser was more accurate than a grandparent parser. [sent-246, score-0.649]
100 Under the new pennconverter standard, a grandparent parser is significantly better than a sibling parser, and there is no significant improvement when including both. [sent-248, score-0.635]
wordName wordTfidf (topN-words)
[('conjunctions', 0.435), ('prepositions', 0.319), ('factorization', 0.276), ('sibling', 0.242), ('grandparent', 0.192), ('factorizations', 0.19), ('conjunction', 0.184), ('conversion', 0.174), ('koo', 0.168), ('unlabeled', 0.166), ('preposition', 0.15), ('scoring', 0.132), ('parser', 0.126), ('grandparents', 0.121), ('parent', 0.106), ('dogs', 0.105), ('mstparser', 0.103), ('conjuncts', 0.09), ('cats', 0.09), ('uas', 0.088), ('constituency', 0.088), ('vadas', 0.086), ('attachment', 0.086), ('trading', 0.082), ('siblings', 0.077), ('conjunct', 0.075), ('pennconverter', 0.075), ('preceeding', 0.069), ('sib', 0.069), ('edges', 0.069), ('counts', 0.067), ('edge', 0.066), ('head', 0.066), ('collins', 0.065), ('dependency', 0.061), ('regime', 0.061), ('bolded', 0.06), ('yesterday', 0.06), ('parsing', 0.058), ('attaching', 0.055), ('bansal', 0.053), ('debt', 0.052), ('rooth', 0.052), ('attachments', 0.051), ('pitler', 0.051), ('associations', 0.051), ('dependencies', 0.049), ('noun', 0.049), ('triples', 0.046), ('bergsma', 0.046), ('weischedel', 0.046), ('hindle', 0.046), ('bottlenecks', 0.045), ('transacting', 0.045), ('ptb', 0.044), ('child', 0.043), ('curran', 0.042), ('mcdonald', 0.042), ('parsers', 0.041), ('scope', 0.041), ('accuracy', 0.039), ('church', 0.038), ('trees', 0.038), ('pmi', 0.036), ('mistakes', 0.036), ('accuracies', 0.035), ('maltparser', 0.035), ('implications', 0.035), ('patched', 0.035), ('attached', 0.034), ('directed', 0.034), ('really', 0.034), ('penn', 0.033), ('nugues', 0.033), ('nice', 0.033), ('modifier', 0.033), ('figures', 0.033), ('opening', 0.032), ('representations', 0.032), ('converted', 0.031), ('ameliorate', 0.03), ('attaches', 0.03), ('nonprojective', 0.03), ('np', 0.03), ('impact', 0.029), ('nivre', 0.029), ('representation', 0.028), ('johansson', 0.028), ('sign', 0.028), ('exponentiated', 0.027), ('grand', 0.027), ('marcus', 0.027), ('help', 0.026), ('yamada', 0.026), ('compensated', 0.026), ('attach', 0.026), ('pages', 0.026), ('argument', 0.025), ('object', 0.025), ('berkeley', 0.025)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000004 30 acl-2012-Attacking Parsing Bottlenecks with Unlabeled Data and Relevant Factorizations
Author: Emily Pitler
Abstract: Prepositions and conjunctions are two of the largest remaining bottlenecks in parsing. Across various existing parsers, these two categories have the lowest accuracies, and mistakes made have consequences for downstream applications. Prepositions and conjunctions are often assumed to depend on lexical dependencies for correct resolution. As lexical statistics based on the training set only are sparse, unlabeled data can help ameliorate this sparsity problem. By including unlabeled data features into a factorization of the problem which matches the representation of prepositions and conjunctions, we achieve a new state-of-the-art for English dependencies with 93.55% correct attachments on the current standard. Furthermore, conjunctions are attached with an accuracy of 90.8%, and prepositions with an accuracy of 87.4%.
2 0.17302477 109 acl-2012-Higher-order Constituent Parsing and Parser Combination
Author: Xiao Chen ; Chunyu Kit
Abstract: This paper presents a higher-order model for constituent parsing aimed at utilizing more local structural context to decide the score of a grammar rule instance in a parse tree. Experiments on English and Chinese treebanks confirm its advantage over its first-order version. It achieves its best F1 scores of 91.86% and 85.58% on the two languages, respectively, and further pushes them to 92.80% and 85.60% via combination with other highperformance parsers.
3 0.15039136 213 acl-2012-Utilizing Dependency Language Models for Graph-based Dependency Parsing Models
Author: Wenliang Chen ; Min Zhang ; Haizhou Li
Abstract: Most previous graph-based parsing models increase decoding complexity when they use high-order features due to exact-inference decoding. In this paper, we present an approach to enriching high-orderfeature representations for graph-based dependency parsing models using a dependency language model and beam search. The dependency language model is built on a large-amount of additional autoparsed data that is processed by a baseline parser. Based on the dependency language model, we represent a set of features for the parsing model. Finally, the features are efficiently integrated into the parsing model during decoding using beam search. Our approach has two advantages. Firstly we utilize rich high-order features defined over a view of large scope and additional large raw corpus. Secondly our approach does not increase the decoding complexity. We evaluate the proposed approach on English and Chinese data. The experimental results show that our new parser achieves the best accuracy on the Chinese data and comparable accuracy with the best known systems on the English data.
4 0.14930266 106 acl-2012-Head-driven Transition-based Parsing with Top-down Prediction
Author: Katsuhiko Hayashi ; Taro Watanabe ; Masayuki Asahara ; Yuji Matsumoto
Abstract: This paper presents a novel top-down headdriven parsing algorithm for data-driven projective dependency analysis. This algorithm handles global structures, such as clause and coordination, better than shift-reduce or other bottom-up algorithms. Experiments on the English Penn Treebank data and the Chinese CoNLL-06 data show that the proposed algorithm achieves comparable results with other data-driven dependency parsing algorithms.
5 0.13995771 5 acl-2012-A Comparison of Chinese Parsers for Stanford Dependencies
Author: Wanxiang Che ; Valentin Spitkovsky ; Ting Liu
Abstract: Stanford dependencies are widely used in natural language processing as a semanticallyoriented representation, commonly generated either by (i) converting the output of a constituent parser, or (ii) predicting dependencies directly. Previous comparisons of the two approaches for English suggest that starting from constituents yields higher accuracies. In this paper, we re-evaluate both methods for Chinese, using more accurate dependency parsers than in previous work. Our comparison of performance and efficiency across seven popular open source parsers (four constituent and three dependency) shows, by contrast, that recent higher-order graph-based techniques can be more accurate, though somewhat slower, than constituent parsers. We demonstrate also that n-way jackknifing is a useful technique for producing automatic (rather than gold) partof-speech tags to train Chinese dependency parsers. Finally, we analyze the relations produced by both kinds of parsing and suggest which specific parsers to use in practice.
6 0.11685137 175 acl-2012-Semi-supervised Dependency Parsing using Lexical Affinities
7 0.11021246 170 acl-2012-Robust Conversion of CCG Derivations to Phrase Structure Trees
8 0.10649288 90 acl-2012-Extracting Narrative Timelines as Temporal Dependency Structures
9 0.10414038 87 acl-2012-Exploiting Multiple Treebanks for Parsing with Quasi-synchronous Grammars
10 0.10240566 4 acl-2012-A Comparative Study of Target Dependency Structures for Statistical Machine Translation
11 0.099251501 127 acl-2012-Large-Scale Syntactic Language Modeling with Treelets
12 0.079505861 71 acl-2012-Dependency Hashing for n-best CCG Parsing
13 0.077927694 189 acl-2012-Syntactic Annotations for the Google Books NGram Corpus
14 0.076831594 95 acl-2012-Fast Syntactic Analysis for Statistical Language Modeling via Substructure Sharing and Uptraining
15 0.072779857 45 acl-2012-Capturing Paradigmatic and Syntagmatic Lexical Relations: Towards Accurate Chinese Part-of-Speech Tagging
16 0.071492516 172 acl-2012-Selective Sharing for Multilingual Dependency Parsing
17 0.070123568 25 acl-2012-An Exploration of Forest-to-String Translation: Does Translation Help or Hurt Parsing?
18 0.066890009 64 acl-2012-Crosslingual Induction of Semantic Roles
19 0.061248414 83 acl-2012-Error Mining on Dependency Trees
20 0.059757624 122 acl-2012-Joint Evaluation of Morphological Segmentation and Syntactic Parsing
topicId topicWeight
[(0, -0.184), (1, 0.005), (2, -0.2), (3, -0.129), (4, -0.072), (5, -0.084), (6, -0.001), (7, -0.049), (8, 0.048), (9, 0.015), (10, 0.1), (11, 0.085), (12, -0.001), (13, 0.014), (14, -0.014), (15, 0.0), (16, -0.005), (17, 0.002), (18, -0.115), (19, 0.001), (20, -0.062), (21, -0.082), (22, 0.048), (23, -0.048), (24, 0.001), (25, -0.033), (26, 0.035), (27, 0.005), (28, 0.029), (29, 0.047), (30, -0.089), (31, -0.076), (32, 0.032), (33, -0.059), (34, -0.081), (35, -0.006), (36, -0.008), (37, -0.02), (38, 0.011), (39, 0.078), (40, 0.062), (41, -0.033), (42, 0.03), (43, 0.012), (44, -0.038), (45, 0.085), (46, 0.072), (47, -0.039), (48, -0.053), (49, -0.024)]
simIndex simValue paperId paperTitle
same-paper 1 0.95477927 30 acl-2012-Attacking Parsing Bottlenecks with Unlabeled Data and Relevant Factorizations
Author: Emily Pitler
Abstract: Prepositions and conjunctions are two of the largest remaining bottlenecks in parsing. Across various existing parsers, these two categories have the lowest accuracies, and mistakes made have consequences for downstream applications. Prepositions and conjunctions are often assumed to depend on lexical dependencies for correct resolution. As lexical statistics based on the training set only are sparse, unlabeled data can help ameliorate this sparsity problem. By including unlabeled data features into a factorization of the problem which matches the representation of prepositions and conjunctions, we achieve a new state-of-the-art for English dependencies with 93.55% correct attachments on the current standard. Furthermore, conjunctions are attached with an accuracy of 90.8%, and prepositions with an accuracy of 87.4%.
2 0.77078849 213 acl-2012-Utilizing Dependency Language Models for Graph-based Dependency Parsing Models
Author: Wenliang Chen ; Min Zhang ; Haizhou Li
Abstract: Most previous graph-based parsing models increase decoding complexity when they use high-order features due to exact-inference decoding. In this paper, we present an approach to enriching high-orderfeature representations for graph-based dependency parsing models using a dependency language model and beam search. The dependency language model is built on a large-amount of additional autoparsed data that is processed by a baseline parser. Based on the dependency language model, we represent a set of features for the parsing model. Finally, the features are efficiently integrated into the parsing model during decoding using beam search. Our approach has two advantages. Firstly we utilize rich high-order features defined over a view of large scope and additional large raw corpus. Secondly our approach does not increase the decoding complexity. We evaluate the proposed approach on English and Chinese data. The experimental results show that our new parser achieves the best accuracy on the Chinese data and comparable accuracy with the best known systems on the English data.
3 0.76977009 5 acl-2012-A Comparison of Chinese Parsers for Stanford Dependencies
Author: Wanxiang Che ; Valentin Spitkovsky ; Ting Liu
Abstract: Stanford dependencies are widely used in natural language processing as a semanticallyoriented representation, commonly generated either by (i) converting the output of a constituent parser, or (ii) predicting dependencies directly. Previous comparisons of the two approaches for English suggest that starting from constituents yields higher accuracies. In this paper, we re-evaluate both methods for Chinese, using more accurate dependency parsers than in previous work. Our comparison of performance and efficiency across seven popular open source parsers (four constituent and three dependency) shows, by contrast, that recent higher-order graph-based techniques can be more accurate, though somewhat slower, than constituent parsers. We demonstrate also that n-way jackknifing is a useful technique for producing automatic (rather than gold) partof-speech tags to train Chinese dependency parsers. Finally, we analyze the relations produced by both kinds of parsing and suggest which specific parsers to use in practice.
4 0.756549 106 acl-2012-Head-driven Transition-based Parsing with Top-down Prediction
Author: Katsuhiko Hayashi ; Taro Watanabe ; Masayuki Asahara ; Yuji Matsumoto
Abstract: This paper presents a novel top-down headdriven parsing algorithm for data-driven projective dependency analysis. This algorithm handles global structures, such as clause and coordination, better than shift-reduce or other bottom-up algorithms. Experiments on the English Penn Treebank data and the Chinese CoNLL-06 data show that the proposed algorithm achieves comparable results with other data-driven dependency parsing algorithms.
5 0.73863226 175 acl-2012-Semi-supervised Dependency Parsing using Lexical Affinities
Author: Seyed Abolghasem Mirroshandel ; Alexis Nasr ; Joseph Le Roux
Abstract: Treebanks are not large enough to reliably model precise lexical phenomena. This deficiency provokes attachment errors in the parsers trained on such data. We propose in this paper to compute lexical affinities, on large corpora, for specific lexico-syntactic configurations that are hard to disambiguate and introduce the new information in a parser. Experiments on the French Treebank showed a relative decrease ofthe error rate of 7. 1% Labeled Accuracy Score yielding the best parsing results on this treebank.
6 0.72324175 87 acl-2012-Exploiting Multiple Treebanks for Parsing with Quasi-synchronous Grammars
7 0.71920377 109 acl-2012-Higher-order Constituent Parsing and Parser Combination
8 0.66889155 4 acl-2012-A Comparative Study of Target Dependency Structures for Statistical Machine Translation
9 0.61081278 95 acl-2012-Fast Syntactic Analysis for Statistical Language Modeling via Substructure Sharing and Uptraining
10 0.55567104 71 acl-2012-Dependency Hashing for n-best CCG Parsing
11 0.55384547 11 acl-2012-A Feature-Rich Constituent Context Model for Grammar Induction
12 0.54021466 122 acl-2012-Joint Evaluation of Morphological Segmentation and Syntactic Parsing
13 0.52087408 172 acl-2012-Selective Sharing for Multilingual Dependency Parsing
14 0.50747502 90 acl-2012-Extracting Narrative Timelines as Temporal Dependency Structures
15 0.49123979 189 acl-2012-Syntactic Annotations for the Google Books NGram Corpus
16 0.48521563 127 acl-2012-Large-Scale Syntactic Language Modeling with Treelets
17 0.48410925 75 acl-2012-Discriminative Strategies to Integrate Multiword Expression Recognition and Parsing
18 0.48372701 83 acl-2012-Error Mining on Dependency Trees
19 0.44246882 34 acl-2012-Automatically Learning Measures of Child Language Development
20 0.43666217 170 acl-2012-Robust Conversion of CCG Derivations to Phrase Structure Trees
topicId topicWeight
[(14, 0.276), (26, 0.026), (28, 0.038), (30, 0.027), (37, 0.055), (39, 0.034), (49, 0.014), (71, 0.048), (74, 0.034), (82, 0.051), (84, 0.02), (85, 0.04), (90, 0.101), (92, 0.049), (94, 0.022), (99, 0.071)]
simIndex simValue paperId paperTitle
same-paper 1 0.76642799 30 acl-2012-Attacking Parsing Bottlenecks with Unlabeled Data and Relevant Factorizations
Author: Emily Pitler
Abstract: Prepositions and conjunctions are two of the largest remaining bottlenecks in parsing. Across various existing parsers, these two categories have the lowest accuracies, and mistakes made have consequences for downstream applications. Prepositions and conjunctions are often assumed to depend on lexical dependencies for correct resolution. As lexical statistics based on the training set only are sparse, unlabeled data can help ameliorate this sparsity problem. By including unlabeled data features into a factorization of the problem which matches the representation of prepositions and conjunctions, we achieve a new state-of-the-art for English dependencies with 93.55% correct attachments on the current standard. Furthermore, conjunctions are attached with an accuracy of 90.8%, and prepositions with an accuracy of 87.4%.
2 0.68306947 66 acl-2012-DOMCAT: A Bilingual Concordancer for Domain-Specific Computer Assisted Translation
Author: Ming-Hong Bai ; Yu-Ming Hsieh ; Keh-Jiann Chen ; Jason S. Chang
Abstract: In this paper, we propose a web-based bilingual concordancer, DOMCAT , for domain-specific computer assisted translation. Given a multi-word expression as a query, the system involves retrieving sentence pairs from a bilingual corpus, identifying translation equivalents of the query in the sentence pairs (translation spotting) and ranking the retrieved sentence pairs according to the relevance between the query and the translation equivalents. To provide high-precision translation spotting for domain-specific translation tasks, we exploited a normalized correlation method to spot the translation equivalents. To ranking the retrieved sentence pairs, we propose a correlation function modified from the Dice coefficient 1 for assessing the correlation between the query and the translation equivalents. The performances of the translation spotting module and the ranking module are evaluated in terms of precision-recall measures and coverage rate respectively. 1
3 0.49447459 191 acl-2012-Temporally Anchored Relation Extraction
Author: Guillermo Garrido ; Anselmo Penas ; Bernardo Cabaleiro ; Alvaro Rodrigo
Abstract: Although much work on relation extraction has aimed at obtaining static facts, many of the target relations are actually fluents, as their validity is naturally anchored to a certain time period. This paper proposes a methodological approach to temporally anchored relation extraction. Our proposal performs distant supervised learning to extract a set of relations from a natural language corpus, and anchors each of them to an interval of temporal validity, aggregating evidence from documents supporting the relation. We use a rich graphbased document-level representation to generate novel features for this task. Results show that our implementation for temporal anchoring is able to achieve a 69% of the upper bound performance imposed by the relation extraction step. Compared to the state of the art, the overall system achieves the highest precision reported.
4 0.48905933 4 acl-2012-A Comparative Study of Target Dependency Structures for Statistical Machine Translation
Author: Xianchao Wu ; Katsuhito Sudoh ; Kevin Duh ; Hajime Tsukada ; Masaaki Nagata
Abstract: This paper presents a comparative study of target dependency structures yielded by several state-of-the-art linguistic parsers. Our approach is to measure the impact of these nonisomorphic dependency structures to be used for string-to-dependency translation. Besides using traditional dependency parsers, we also use the dependency structures transformed from PCFG trees and predicate-argument structures (PASs) which are generated by an HPSG parser and a CCG parser. The experiments on Chinese-to-English translation show that the HPSG parser’s PASs achieved the best dependency and translation accuracies. 1
5 0.48563778 12 acl-2012-A Graph-based Cross-lingual Projection Approach for Weakly Supervised Relation Extraction
Author: Seokhwan Kim ; Gary Geunbae Lee
Abstract: Although researchers have conducted extensive studies on relation extraction in the last decade, supervised approaches are still limited because they require large amounts of training data to achieve high performances. To build a relation extractor without significant annotation effort, we can exploit cross-lingual annotation projection, which leverages parallel corpora as external resources for supervision. This paper proposes a novel graph-based projection approach and demonstrates the merits of it by using a Korean relation extraction system based on projected dataset from an English-Korean parallel corpus.
6 0.48330575 5 acl-2012-A Comparison of Chinese Parsers for Stanford Dependencies
7 0.47967479 146 acl-2012-Modeling Topic Dependencies in Hierarchical Text Categorization
8 0.47840509 80 acl-2012-Efficient Tree-based Approximation for Entailment Graph Learning
9 0.47808921 187 acl-2012-Subgroup Detection in Ideological Discussions
10 0.47730571 71 acl-2012-Dependency Hashing for n-best CCG Parsing
11 0.47656342 106 acl-2012-Head-driven Transition-based Parsing with Top-down Prediction
12 0.47636575 214 acl-2012-Verb Classification using Distributional Similarity in Syntactic and Semantic Structures
13 0.47479859 63 acl-2012-Cross-lingual Parse Disambiguation based on Semantic Correspondence
15 0.47095534 201 acl-2012-Towards the Unsupervised Acquisition of Discourse Relations
16 0.47018608 84 acl-2012-Estimating Compact Yet Rich Tree Insertion Grammars
17 0.46978894 72 acl-2012-Detecting Semantic Equivalence and Information Disparity in Cross-lingual Documents
18 0.4692933 206 acl-2012-UWN: A Large Multilingual Lexical Knowledge Base
19 0.46846944 40 acl-2012-Big Data versus the Crowd: Looking for Relationships in All the Right Places
20 0.46744943 175 acl-2012-Semi-supervised Dependency Parsing using Lexical Affinities