acl acl2013 acl2013-357 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Yuan Zhang ; Regina Barzilay ; Amir Globerson
Abstract: In this paper, we consider the problem of cross-formalism transfer in parsing. We are interested in parsing constituencybased grammars such as HPSG and CCG using a small amount of data specific for the target formalism, and a large quantity of coarse CFG annotations from the Penn Treebank. While all of the target formalisms share a similar basic syntactic structure with Penn Treebank CFG, they also encode additional constraints and semantic features. To handle this apparent discrepancy, we design a probabilistic model that jointly generates CFG and target formalism parses. The model includes features of both parses, allowing trans- fer between the formalisms, while preserving parsing efficiency. We evaluate our approach on three constituency-based grammars CCG, HPSG, and LFG, augmented with the Penn Treebank-1. Our experiments show that across all three formalisms, the target parsers significantly benefit from the coarse annotations.1 —
Reference: text
sentIndex sentText sentNum sentScore
1 Abstract In this paper, we consider the problem of cross-formalism transfer in parsing. [sent-3, score-0.141]
2 We are interested in parsing constituencybased grammars such as HPSG and CCG using a small amount of data specific for the target formalism, and a large quantity of coarse CFG annotations from the Penn Treebank. [sent-4, score-0.495]
3 While all of the target formalisms share a similar basic syntactic structure with Penn Treebank CFG, they also encode additional constraints and semantic features. [sent-5, score-0.273]
4 To handle this apparent discrepancy, we design a probabilistic model that jointly generates CFG and target formalism parses. [sent-6, score-0.272]
5 Our experiments show that across all three formalisms, the target parsers significantly benefit from the coarse annotations. [sent-9, score-0.266]
6 Moreover, the ongoing process of developing new formalisms is intrinsic to linguistic research. [sent-11, score-0.196]
7 l The standard solution to this bottleneck has relied on manually crafted transformation rules that map readily available syntactic annotations (e. [sent-21, score-0.144]
8 Designing these transformation rules is a major undertaking which requires multiple correction cycles and a deep understanding of the underlying grammar formalisms. [sent-23, score-0.155]
9 Instead of using manually-crafted transformation rules, this approach relies on a small amount of annotations in the target formalism. [sent-27, score-0.176]
10 All of these formalisms share a similar basic syntactic structure with Penn Treebank CFG. [sent-33, score-0.196]
11 However, the target formalisms also encode additional constraints and semantic features. [sent-34, score-0.273]
12 For instance, Penn Treebank annotations do not make an explicit distinction between complement and adjunct, while all the above grammars mark these 291 Proce dingSsof oifa, th Beu 5l1gsarti Aan,An u aglu Mste 4e-ti9n2g 0 o1f3 t. [sent-35, score-0.17]
13 Figure 1n ssh iotw ass derivations in the three target formalisms we consider, as well as a CFG derivation. [sent-41, score-0.353]
14 We can see that the derivations of these formalisms share the same basic structure, while the formalism-specific information is mainly encoded in the lexical entries and node labels. [sent-42, score-0.315]
15 To enable effective transfer the model has to identify shared structural components between the formalisms despite the apparent differences. [sent-43, score-0.364]
16 To this end, our model jointly parses the two corpora according to the corresponding annotations, enabling transfer via parameter sharing. [sent-45, score-0.221]
17 In particular, we augment each target tree node with hidden variables that capture the connection to the coarse annotations. [sent-46, score-0.361]
18 Specifically, each node in the target tree has two labels: an entry which is specific to the target formalism, and a latent label containing a value from the Penn Treebank tagset, such as NP (see Figure 2). [sent-47, score-0.226]
19 This design enables us to represent three types of features: the target formalismspecific features, the coarse formalism features, and features that connect the two. [sent-48, score-0.534]
20 This modeling approach makes it possible to perform transfer to a range of target formalisms, without manually drafting formalism-specific rules. [sent-49, score-0.218]
21 As a source of coarse annotations, we use the Penn Treebank. [sent-51, score-0.189]
22 2 Our results clearly demonstrate that for all three formalisms, parsing accuracy can be improved by training with additional coarse annotations. [sent-52, score-0.248]
23 To achieve similar performance in the absence of coarse annotations, the parser has to be trained on about 1,500 sentences, namely three times what is needed when using coarse annotations. [sent-58, score-0.404]
24 Similar results are — 2While the Penn Treebank-2 contains richer annotations, we decided to use the Penn Treebank-1 to demonstrate the feasibility of transfer from coarse annotations. [sent-59, score-0.33]
25 2 Related Work Our work belongs to a broader class of research on transfer learning in parsing. [sent-68, score-0.141]
26 , 2006), and crossformalism transfer (Hockenmaier and Steedman, 2002; Miyao et al. [sent-77, score-0.173]
27 There have been several attempts to map annotations in coarse grammars like CFG to annotations in richer grammar, like HPSG, LFG, or CCG. [sent-82, score-0.458]
28 These conversion rules are typically utilized in two ways: (1) to create a new treebank which is consequently used to train a parser for the target formalism (Hockenmaier and Steedman, 2002; Clark and Curran, 2003; Miyao et al. [sent-85, score-0.444]
29 , 2005; Miyao and Tsujii, 2008), (2) to translate the output of a CFG parser into the target formalism (Cahill et al. [sent-86, score-0.298]
30 By nature, the mapping rules are formalism spe292 cific and therefore not transferable. [sent-91, score-0.24]
31 For instance, Hockenmaier and Steedman (2002) made thousands of POS and constituent modifications to the Penn Treebank to facilitate transfer to CCG. [sent-93, score-0.141]
32 More importantly, in some transfer scenarios, deterministic rules are not sufficient, due to the high ambiguity inherent in the mapping. [sent-94, score-0.208]
33 Therefore, our work considers an alternative set-up for crossformalism transfer where a small amount of an- notations in the target formalism is used as an alternative to using deterministic rules. [sent-95, score-0.467]
34 The limitation of deterministic transfer rules has been recognized in prior work (Riezler et al. [sent-96, score-0.208]
35 In contrast to this method, we neither require a parser for the target formalism, nor manual rules for partial mapping. [sent-100, score-0.148]
36 Consequently, our method can be applied to many different target grammar formalisms without significant engineering effort for each one. [sent-101, score-0.357]
37 The utility of coarse-grained treebanks is determined by the degree of structural overlap with the target formalism. [sent-102, score-0.126]
38 3 The Learning Problem Recall that our goal is to learn how to parse the target formalisms while using two annotated sources: a small set of sentences annotated in the target formalism (e. [sent-103, score-0.545]
39 , CCG), and a large set of sentences with coarse annotations. [sent-105, score-0.189]
40 For simplicity we focus on the CCG formalism in what follows. [sent-107, score-0.195]
41 The shadowed labels correspond to the CFG derivation yCFG, whereas the other labels correspond to the CCG derivation yCCG. [sent-128, score-0.22]
42 Also shown are features that are turned on for this joint derivation (see Section 6). [sent-130, score-0.135]
43 CFG parses in the Penn Treebank are not binary, and we therefore binarize them, as explained in Section 5. [sent-141, score-0.133]
44 Second, we assume that any yCFG and yCCG jointly generated must share the same derivation tree structure. [sent-143, score-0.143]
45 Since both formalisms are constituency-based, their trees are expected to describe the same constituents. [sent-145, score-0.247]
46 For the distribution over yCCG we do precisely this, namely use: pCCG(yCCG|S;θ) =X pjoint(y|S;θ) (2) yXCFG For the distribution over yCFG we could have marginalized pjoint over yCCG. [sent-160, score-0.162]
47 5 Implementation This section introduces important implementation details, including supertagging, feature forest pruning and binarization methods. [sent-185, score-0.204]
48 1 Supertagging When parsing a target formalism tree, one needs to associate each word with a lexical entry. [sent-188, score-0.331]
49 However, a common problem for lexicalized grammars is that the forest size is too large. [sent-201, score-0.139]
50 For the target formalism, a common practice is to prune the forest using the supertagger (Clark and Curran, 2007; Miyao, 2006). [sent-204, score-0.184]
51 In our implementation, we applied all pruning techniques, because the forest is a combination of CFG and target grammar formalisms (e. [sent-205, score-0.453]
52 3 Binarization We assume that the derivation tree in the target formalism is in a normal form, which is indeed the case for the treebanks we consider. [sent-209, score-0.462]
53 As mentioned in Section 4, we would also like to work with binarized CFG derivations, such that all trees are in normal form and it is easy to construct features that link the two (see Section 6). [sent-210, score-0.135]
54 The procedure is based on the available target formalism parses in the training corpus, which are binarized. [sent-212, score-0.352]
55 In what follows, we describe derivations using the POS of the head words of the corresponding node in the tree. [sent-214, score-0.152]
56 This makes it possible to transfer binarization rules between formalisms. [sent-215, score-0.272]
57 Suppose we want to learn the binarization rule of the following derivation in CFG: NN → (DT JJ NN) (4) We now look for binary derivations with these POS in the target formalism corpus, and take the most common binarization form. [sent-216, score-0.634]
58 We also experiment with using fixed binarization rules such as left/right branching, instead of learning them. [sent-218, score-0.131]
59 In a derivation tree, the formalism-specific information is mainly encoded in the lexical entries and the applied grammar rules, rather than the tree structures. [sent-222, score-0.227]
60 In our model, as mentioned in Section 1, the features are also defined to enable information transfer between coarse and rich formalisms. [sent-225, score-0.355]
61 In this section, we first introduce how different types of feature templates are designed, and then show an example of how the features help transfer the syntactic structure information. [sent-226, score-0.233]
62 Note that the same feature templates are used for all the target grammar formalisms. [sent-227, score-0.228]
63 Recall that our y contains both the CFG and CCG parses, and that these use the same derivation tree structure. [sent-228, score-0.143]
64 Each feature will consider either the CFG derivation, the CCG derivation or these two derivations jointly. [sent-229, score-0.212]
65 We define the following feature templates: fbinary for binary derivations, funary for unary derivations, and froot for the root nodes. [sent-238, score-0.118]
66 It can be seen that some features depend only on the CFG derivations (i. [sent-242, score-0.128]
67 However, we limit the features to be designed locally in a derivation in order to run inside-outside efficiently. [sent-250, score-0.135]
68 In order to apply the same feature templates to other target formalisms, we only need to assign the atomic features r and hl with the formalismspecific values. [sent-254, score-0.248]
69 Penn Treebank CCGbank wVrPiteVPletNPersVP (Se[dactVl]\PN P )S/[NdPcl]\NPNaPp NlePs fCFG(y,S) : VP fCFG(y,S) : VP,NP VP VP,NP fCCG(y,S) : S[dcl]\NP (S[dcl]\NP)/NP,NP fjoint(y,S) : VP, S[dcl]\NP (VP, (S[dcl]\NP)/NP), (NP, NP) Figure 3: Example of transfer between CFG and CCG formalisms. [sent-256, score-0.141]
70 Figure 3 gives an example in CCG of how features help transfer the syntactic information from Penn Treebank and learn the correspondence to the formalism-specific information. [sent-257, score-0.166]
71 Datasets: As a source of coarse annotations, we use the Penn Treebank-1 (Marcus et al. [sent-271, score-0.189]
72 This metric is commonly used to measure parsing quality for the formalisms considered in this paper. [sent-293, score-0.255]
73 The detailed definition of this measure as applied for each formalism is provided in (Clark and Cur- ran, 2003; Miyao and Tsujii, 2008; Cahill et al. [sent-294, score-0.195]
74 Training without CFG Data: To assess the impact of coarse data in the experiments below, we also consider the model trained only on formalism-specific annotations. [sent-307, score-0.189]
75 In this set-up, the model reduces to a normal loglinear model for the target formalism. [sent-309, score-0.129]
76 8 Experiment and Analysis Impact of Coarse Annotations on Target Formalism: To analyze the effectiveness of annotation transfer, we fix the number of annotated trees in the target formalism and vary the amount of coarse annotations available to the algorithm during training. [sent-317, score-0.611]
77 As Figure 4 shows, CFG data boosts parsing ac- curacy for all the target formalisms. [sent-319, score-0.136]
78 2% in labeled dependency F-score for HPSG formalism when 15,000 CFG trees are used. [sent-321, score-0.288]
79 Moreover, increasing the number of coarse annotations used in training leads to further improvement on different evaluation metrics. [sent-322, score-0.288]
80 Tradeoff between Target and Coarse Annotations: We also assess the relative contribution of coarse annotations when the size of annotated training corpus in the target formalism varies. [sent-325, score-0.56]
81 In this set of experiments, we fix the number of CFG trees to 15,000 and vary the number of target annotations from 500 to 4,000. [sent-326, score-0.227]
82 Figure 5 shows the relative contribution of formalism-specific annotations compared to that of the coarse annotations. [sent-327, score-0.288]
83 For instance, Figure 5a shows that the parsing performance achieved using 2,000 CCG sentences can be achieved using approximately 500 CCG sentences when coarse annotations are available for training. [sent-328, score-0.347]
84 More generally, the result convincingly demonstrates that coarse annotations are helpful for all the sizes offormalism-specific training data. [sent-329, score-0.288]
85 Figure 5 also illustrates a slightly different characteristics of transfer performance between two evaluation metrics. [sent-331, score-0.141]
86 The unlabeled PARSEVAL score (Figure 5d-f) mainly re- lies on the coarse structural information. [sent-334, score-0.241]
87 On the other hand, predicate-argument dependency Fscore (Figure 5a-c) also relies on the target grammar information. [sent-335, score-0.203]
88 Comparison to State-of-the-art Parsers: We would also like to demonstrate that the above gains of our transfer model are achieved using an adequate formalism-specific parser. [sent-342, score-0.141]
89 They show that we correctly learn a frequent derivation in the target formalism and CFG. [sent-353, score-0.382]
90 Our model correctly learns that a syntactic derivation with children VP and NP is very likely to be mapped to the derivation (head comp)→ ([NhViN],[N. [sent-355, score-0.248]
91 9 Conclusions We present a method for cross-formalism transfer in parsing. [sent-358, score-0.141]
92 Our model utilizes coarse syntactic annotations to supplement a small number of formalism-specific trees for training on constituency-based grammars. [sent-359, score-0.339]
93 Our experimental results show that across a range of such formalisms, the model significantly benefits from the coarse annotations. [sent-360, score-0.189]
94 Long-distance dependency resolution in automatically acquired wide-coverage pcfg-based lfg approximations. [sent-381, score-0.269]
95 Statistical french dependency parsing: treebank conversion and first results. [sent-392, score-0.143]
96 Widecoverage efficient statistical parsing with ccg and log-linear models. [sent-414, score-0.454]
97 Automatic adaptation of annotation standards for dependency parsing: using projected treebank as source corpus. [sent-446, score-0.167]
98 Building a large annotated corpus of english: The penn treebank. [sent-462, score-0.132]
99 Corpus-oriented grammar development for acquiring a head-driven phrase structure grammar from the penn treebank. [sent-485, score-0.3]
100 Towards holistic grammar engineering and testing–grafting treebank maintenance into the grammar revision cycle. [sent-499, score-0.269]
wordName wordTfidf (topN-words)
[('cfg', 0.495), ('ccg', 0.395), ('lfg', 0.227), ('hpsg', 0.222), ('formalisms', 0.196), ('formalism', 0.195), ('ycfg', 0.192), ('coarse', 0.189), ('transfer', 0.141), ('penn', 0.132), ('miyao', 0.129), ('cahill', 0.117), ('yccg', 0.112), ('derivation', 0.11), ('treebank', 0.101), ('annotations', 0.099), ('fcfg', 0.096), ('pjoint', 0.096), ('binarization', 0.086), ('grammar', 0.084), ('derivations', 0.08), ('parses', 0.08), ('target', 0.077), ('pcfg', 0.075), ('hockenmaier', 0.073), ('grammars', 0.071), ('forest', 0.068), ('vp', 0.067), ('parseval', 0.065), ('fccg', 0.064), ('steedman', 0.061), ('parsing', 0.059), ('dcl', 0.052), ('trees', 0.051), ('clark', 0.05), ('kaplan', 0.049), ('np', 0.048), ('fjoint', 0.048), ('formalismspecific', 0.048), ('templates', 0.045), ('tsujii', 0.045), ('rules', 0.045), ('riezler', 0.044), ('curran', 0.043), ('pccg', 0.042), ('dependency', 0.042), ('pollard', 0.039), ('supertagging', 0.039), ('supertagger', 0.039), ('node', 0.039), ('aoife', 0.037), ('charniak', 0.035), ('tracy', 0.035), ('yusuke', 0.034), ('binarized', 0.034), ('head', 0.033), ('chart', 0.033), ('king', 0.033), ('tree', 0.033), ('crossformalism', 0.032), ('fbinary', 0.032), ('froot', 0.032), ('funary', 0.032), ('gropp', 0.032), ('hp', 0.031), ('atomic', 0.031), ('candito', 0.028), ('hli', 0.028), ('pruning', 0.028), ('children', 0.028), ('explained', 0.027), ('hw', 0.027), ('structural', 0.027), ('ronald', 0.027), ('loglinear', 0.027), ('mcdonald', 0.027), ('blitzer', 0.026), ('undertaking', 0.026), ('binarize', 0.026), ('nn', 0.026), ('parser', 0.026), ('features', 0.025), ('normal', 0.025), ('unlabeled', 0.025), ('designing', 0.024), ('adaptation', 0.024), ('jj', 0.024), ('crouch', 0.023), ('sy', 0.023), ('oepen', 0.023), ('depend', 0.023), ('connection', 0.023), ('distribution', 0.022), ('deterministic', 0.022), ('treebanks', 0.022), ('sag', 0.022), ('feature', 0.022), ('marginalized', 0.022), ('grandparent', 0.022), ('passing', 0.022)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000002 357 acl-2013-Transfer Learning for Constituency-Based Grammars
Author: Yuan Zhang ; Regina Barzilay ; Amir Globerson
Abstract: In this paper, we consider the problem of cross-formalism transfer in parsing. We are interested in parsing constituencybased grammars such as HPSG and CCG using a small amount of data specific for the target formalism, and a large quantity of coarse CFG annotations from the Penn Treebank. While all of the target formalisms share a similar basic syntactic structure with Penn Treebank CFG, they also encode additional constraints and semantic features. To handle this apparent discrepancy, we design a probabilistic model that jointly generates CFG and target formalism parses. The model includes features of both parses, allowing trans- fer between the formalisms, while preserving parsing efficiency. We evaluate our approach on three constituency-based grammars CCG, HPSG, and LFG, augmented with the Penn Treebank-1. Our experiments show that across all three formalisms, the target parsers significantly benefit from the coarse annotations.1 —
2 0.33362374 199 acl-2013-Integrating Multiple Dependency Corpora for Inducing Wide-coverage Japanese CCG Resources
Author: Sumire Uematsu ; Takuya Matsuzaki ; Hiroki Hanaoka ; Yusuke Miyao ; Hideki Mima
Abstract: This paper describes a method of inducing wide-coverage CCG resources for Japanese. While deep parsers with corpusinduced grammars have been emerging for some languages, those for Japanese have not been widely studied, mainly because most Japanese syntactic resources are dependency-based. Our method first integrates multiple dependency-based corpora into phrase structure trees and then converts the trees into CCG derivations. The method is empirically evaluated in terms of the coverage of the obtained lexi- con and the accuracy of parsing.
3 0.31071728 372 acl-2013-Using CCG categories to improve Hindi dependency parsing
Author: Bharat Ram Ambati ; Tejaswini Deoskar ; Mark Steedman
Abstract: We show that informative lexical categories from a strongly lexicalised formalism such as Combinatory Categorial Grammar (CCG) can improve dependency parsing of Hindi, a free word order language. We first describe a novel way to obtain a CCG lexicon and treebank from an existing dependency treebank, using a CCG parser. We use the output of a supertagger trained on the CCGbank as a feature for a state-of-the-art Hindi dependency parser (Malt). Our results show that using CCG categories improves the accuracy of Malt on long distance dependencies, for which it is known to have weak rates of recovery.
4 0.23384713 347 acl-2013-The Role of Syntax in Vector Space Models of Compositional Semantics
Author: Karl Moritz Hermann ; Phil Blunsom
Abstract: Modelling the compositional process by which the meaning of an utterance arises from the meaning of its parts is a fundamental task of Natural Language Processing. In this paper we draw upon recent advances in the learning of vector space representations of sentential semantics and the transparent interface between syntax and semantics provided by Combinatory Categorial Grammar to introduce Combinatory Categorial Autoencoders. This model leverages the CCG combinatory operators to guide a non-linear transformation of meaning within a sentence. We use this model to learn high dimensional embeddings for sentences and evaluate them in a range of tasks, demonstrating that the incorporation of syntax allows a concise model to learn representations that are both effective and general.
5 0.20598443 4 acl-2013-A Context Free TAG Variant
Author: Ben Swanson ; Elif Yamangil ; Eugene Charniak ; Stuart Shieber
Abstract: We propose a new variant of TreeAdjoining Grammar that allows adjunction of full wrapping trees but still bears only context-free expressivity. We provide a transformation to context-free form, and a further reduction in probabilistic model size through factorization and pooling of parameters. This collapsed context-free form is used to implement efficient gram- mar estimation and parsing algorithms. We perform parsing experiments the Penn Treebank and draw comparisons to TreeSubstitution Grammars and between different variations in probabilistic model design. Examination of the most probable derivations reveals examples of the linguistically relevant structure that our variant makes possible.
6 0.16750666 144 acl-2013-Explicit and Implicit Syntactic Features for Text Classification
7 0.14928752 270 acl-2013-ParGramBank: The ParGram Parallel Treebank
8 0.1189202 165 acl-2013-General binarization for parsing and translation
9 0.11664642 98 acl-2013-Cross-lingual Transfer of Semantic Role Labeling Models
10 0.11368058 204 acl-2013-Iterative Transformation of Annotation Guidelines for Constituency Parsing
11 0.11073338 80 acl-2013-Chinese Parsing Exploiting Characters
12 0.1099234 368 acl-2013-Universal Dependency Annotation for Multilingual Parsing
13 0.10261991 112 acl-2013-Dependency Parser Adaptation with Subtrees from Auto-Parsed Target Domain Data
14 0.097321212 261 acl-2013-Nonparametric Bayesian Inference and Efficient Parsing for Tree-adjoining Grammars
15 0.095100798 343 acl-2013-The Effect of Higher-Order Dependency Features in Discriminative Phrase-Structure Parsing
16 0.094535753 226 acl-2013-Learning to Prune: Context-Sensitive Pruning for Syntactic MT
17 0.091817327 313 acl-2013-Semantic Parsing with Combinatory Categorial Grammars
18 0.086297199 275 acl-2013-Parsing with Compositional Vector Grammars
19 0.083280571 348 acl-2013-The effect of non-tightness on Bayesian estimation of PCFGs
20 0.082627073 208 acl-2013-Joint Inference for Heterogeneous Dependency Parsing
topicId topicWeight
[(0, 0.183), (1, -0.111), (2, -0.168), (3, -0.001), (4, -0.219), (5, -0.007), (6, 0.075), (7, -0.04), (8, 0.105), (9, -0.069), (10, 0.047), (11, 0.031), (12, 0.209), (13, -0.003), (14, -0.138), (15, -0.011), (16, 0.122), (17, 0.12), (18, -0.258), (19, 0.09), (20, 0.107), (21, -0.182), (22, -0.149), (23, -0.011), (24, 0.065), (25, -0.019), (26, -0.024), (27, 0.042), (28, 0.019), (29, -0.025), (30, 0.079), (31, 0.002), (32, -0.013), (33, 0.12), (34, -0.041), (35, -0.022), (36, 0.012), (37, 0.044), (38, -0.026), (39, 0.05), (40, 0.021), (41, 0.102), (42, -0.058), (43, 0.025), (44, 0.049), (45, -0.031), (46, 0.04), (47, -0.034), (48, 0.061), (49, -0.009)]
simIndex simValue paperId paperTitle
same-paper 1 0.91865361 357 acl-2013-Transfer Learning for Constituency-Based Grammars
Author: Yuan Zhang ; Regina Barzilay ; Amir Globerson
Abstract: In this paper, we consider the problem of cross-formalism transfer in parsing. We are interested in parsing constituencybased grammars such as HPSG and CCG using a small amount of data specific for the target formalism, and a large quantity of coarse CFG annotations from the Penn Treebank. While all of the target formalisms share a similar basic syntactic structure with Penn Treebank CFG, they also encode additional constraints and semantic features. To handle this apparent discrepancy, we design a probabilistic model that jointly generates CFG and target formalism parses. The model includes features of both parses, allowing trans- fer between the formalisms, while preserving parsing efficiency. We evaluate our approach on three constituency-based grammars CCG, HPSG, and LFG, augmented with the Penn Treebank-1. Our experiments show that across all three formalisms, the target parsers significantly benefit from the coarse annotations.1 —
2 0.88063842 199 acl-2013-Integrating Multiple Dependency Corpora for Inducing Wide-coverage Japanese CCG Resources
Author: Sumire Uematsu ; Takuya Matsuzaki ; Hiroki Hanaoka ; Yusuke Miyao ; Hideki Mima
Abstract: This paper describes a method of inducing wide-coverage CCG resources for Japanese. While deep parsers with corpusinduced grammars have been emerging for some languages, those for Japanese have not been widely studied, mainly because most Japanese syntactic resources are dependency-based. Our method first integrates multiple dependency-based corpora into phrase structure trees and then converts the trees into CCG derivations. The method is empirically evaluated in terms of the coverage of the obtained lexi- con and the accuracy of parsing.
3 0.87079471 372 acl-2013-Using CCG categories to improve Hindi dependency parsing
Author: Bharat Ram Ambati ; Tejaswini Deoskar ; Mark Steedman
Abstract: We show that informative lexical categories from a strongly lexicalised formalism such as Combinatory Categorial Grammar (CCG) can improve dependency parsing of Hindi, a free word order language. We first describe a novel way to obtain a CCG lexicon and treebank from an existing dependency treebank, using a CCG parser. We use the output of a supertagger trained on the CCGbank as a feature for a state-of-the-art Hindi dependency parser (Malt). Our results show that using CCG categories improves the accuracy of Malt on long distance dependencies, for which it is known to have weak rates of recovery.
4 0.63007915 347 acl-2013-The Role of Syntax in Vector Space Models of Compositional Semantics
Author: Karl Moritz Hermann ; Phil Blunsom
Abstract: Modelling the compositional process by which the meaning of an utterance arises from the meaning of its parts is a fundamental task of Natural Language Processing. In this paper we draw upon recent advances in the learning of vector space representations of sentential semantics and the transparent interface between syntax and semantics provided by Combinatory Categorial Grammar to introduce Combinatory Categorial Autoencoders. This model leverages the CCG combinatory operators to guide a non-linear transformation of meaning within a sentence. We use this model to learn high dimensional embeddings for sentences and evaluate them in a range of tasks, demonstrating that the incorporation of syntax allows a concise model to learn representations that are both effective and general.
5 0.60712206 313 acl-2013-Semantic Parsing with Combinatory Categorial Grammars
Author: Yoav Artzi ; Nicholas FitzGerald ; Luke Zettlemoyer
Abstract: unkown-abstract
6 0.46906716 4 acl-2013-A Context Free TAG Variant
7 0.449541 270 acl-2013-ParGramBank: The ParGram Parallel Treebank
8 0.4273473 311 acl-2013-Semantic Neighborhoods as Hypergraphs
9 0.42161852 204 acl-2013-Iterative Transformation of Annotation Guidelines for Constituency Parsing
10 0.41415545 165 acl-2013-General binarization for parsing and translation
11 0.41275239 261 acl-2013-Nonparametric Bayesian Inference and Efficient Parsing for Tree-adjoining Grammars
12 0.39707834 335 acl-2013-Survey on parsing three dependency representations for English
13 0.39126715 144 acl-2013-Explicit and Implicit Syntactic Features for Text Classification
14 0.38805449 57 acl-2013-Arguments and Modifiers from the Learner's Perspective
15 0.37259924 275 acl-2013-Parsing with Compositional Vector Grammars
16 0.36350676 94 acl-2013-Coordination Structures in Dependency Treebanks
17 0.35940826 176 acl-2013-Grounded Unsupervised Semantic Parsing
18 0.35603526 343 acl-2013-The Effect of Higher-Order Dependency Features in Discriminative Phrase-Structure Parsing
19 0.35491782 28 acl-2013-A Unified Morpho-Syntactic Scheme of Stanford Dependencies
20 0.35376319 368 acl-2013-Universal Dependency Annotation for Multilingual Parsing
topicId topicWeight
[(0, 0.048), (6, 0.03), (11, 0.052), (14, 0.296), (24, 0.026), (26, 0.057), (28, 0.012), (35, 0.049), (42, 0.117), (48, 0.032), (53, 0.013), (70, 0.063), (88, 0.028), (90, 0.017), (95, 0.063)]
simIndex simValue paperId paperTitle
same-paper 1 0.83430922 357 acl-2013-Transfer Learning for Constituency-Based Grammars
Author: Yuan Zhang ; Regina Barzilay ; Amir Globerson
Abstract: In this paper, we consider the problem of cross-formalism transfer in parsing. We are interested in parsing constituencybased grammars such as HPSG and CCG using a small amount of data specific for the target formalism, and a large quantity of coarse CFG annotations from the Penn Treebank. While all of the target formalisms share a similar basic syntactic structure with Penn Treebank CFG, they also encode additional constraints and semantic features. To handle this apparent discrepancy, we design a probabilistic model that jointly generates CFG and target formalism parses. The model includes features of both parses, allowing trans- fer between the formalisms, while preserving parsing efficiency. We evaluate our approach on three constituency-based grammars CCG, HPSG, and LFG, augmented with the Penn Treebank-1. Our experiments show that across all three formalisms, the target parsers significantly benefit from the coarse annotations.1 —
2 0.69844663 314 acl-2013-Semantic Roles for String to Tree Machine Translation
Author: Marzieh Bazrafshan ; Daniel Gildea
Abstract: We experiment with adding semantic role information to a string-to-tree machine translation system based on the rule extraction procedure of Galley et al. (2004). We compare methods based on augmenting the set of nonterminals by adding semantic role labels, and altering the rule extraction process to produce a separate set of rules for each predicate that encompass its entire predicate-argument structure. Our results demonstrate that the second approach is effective in increasing the quality of translations.
3 0.67750573 13 acl-2013-A New Syntactic Metric for Evaluation of Machine Translation
Author: Melania Duma ; Cristina Vertan ; Wolfgang Menzel
Abstract: Machine translation (MT) evaluation aims at measuring the quality of a candidate translation by comparing it with a reference translation. This comparison can be performed on multiple levels: lexical, syntactic or semantic. In this paper, we propose a new syntactic metric for MT evaluation based on the comparison of the dependency structures of the reference and the candidate translations. The dependency structures are obtained by means of a Weighted Constraints Dependency Grammar parser. Based on experiments performed on English to German translations, we show that the new metric correlates well with human judgments at the system level. 1
4 0.65940583 31 acl-2013-A corpus-based evaluation method for Distributional Semantic Models
Author: Abdellah Fourtassi ; Emmanuel Dupoux
Abstract: Evaluation methods for Distributional Semantic Models typically rely on behaviorally derived gold standards. These methods are difficult to deploy in languages with scarce linguistic/behavioral resources. We introduce a corpus-based measure that evaluates the stability of the lexical semantic similarity space using a pseudo-synonym same-different detection task and no external resources. We show that it enables to predict two behaviorbased measures across a range of parameters in a Latent Semantic Analysis model.
5 0.65636349 137 acl-2013-Enlisting the Ghost: Modeling Empty Categories for Machine Translation
Author: Bing Xiang ; Xiaoqiang Luo ; Bowen Zhou
Abstract: Empty categories (EC) are artificial elements in Penn Treebanks motivated by the government-binding (GB) theory to explain certain language phenomena such as pro-drop. ECs are ubiquitous in languages like Chinese, but they are tacitly ignored in most machine translation (MT) work because of their elusive nature. In this paper we present a comprehensive treatment of ECs by first recovering them with a structured MaxEnt model with a rich set of syntactic and lexical features, and then incorporating the predicted ECs into a Chinese-to-English machine translation task through multiple approaches, including the extraction of EC-specific sparse features. We show that the recovered empty categories not only improve the word alignment quality, but also lead to significant improvements in a large-scale state-of-the-art syntactic MT system.
6 0.55764896 80 acl-2013-Chinese Parsing Exploiting Characters
7 0.5335896 335 acl-2013-Survey on parsing three dependency representations for English
8 0.53171515 155 acl-2013-Fast and Accurate Shift-Reduce Constituent Parsing
9 0.52733397 4 acl-2013-A Context Free TAG Variant
10 0.52019441 144 acl-2013-Explicit and Implicit Syntactic Features for Text Classification
11 0.51938337 57 acl-2013-Arguments and Modifiers from the Learner's Perspective
12 0.51613438 204 acl-2013-Iterative Transformation of Annotation Guidelines for Constituency Parsing
13 0.5156849 303 acl-2013-Robust multilingual statistical morphological generation models
14 0.51548874 367 acl-2013-Universal Conceptual Cognitive Annotation (UCCA)
15 0.51409549 132 acl-2013-Easy-First POS Tagging and Dependency Parsing with Beam Search
16 0.51160234 56 acl-2013-Argument Inference from Relevant Event Mentions in Chinese Argument Extraction
17 0.51135343 205 acl-2013-Joint Apposition Extraction with Syntactic and Semantic Constraints
18 0.51073039 275 acl-2013-Parsing with Compositional Vector Grammars
19 0.51043105 343 acl-2013-The Effect of Higher-Order Dependency Features in Discriminative Phrase-Structure Parsing
20 0.50629354 38 acl-2013-Additive Neural Networks for Statistical Machine Translation