emnlp emnlp2010 emnlp2010-38 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Terry Koo ; Alexander M. Rush ; Michael Collins ; Tommi Jaakkola ; David Sontag
Abstract: This paper introduces algorithms for nonprojective parsing based on dual decomposition. We focus on parsing algorithms for nonprojective head automata, a generalization of head-automata models to non-projective structures. The dual decomposition algorithms are simple and efficient, relying on standard dynamic programming and minimum spanning tree algorithms. They provably solve an LP relaxation of the non-projective parsing problem. Empirically the LP relaxation is very often tight: for many languages, exact solutions are achieved on over 98% of test sentences. The accuracy of our models is higher than previous work on a broad range of datasets.
Reference: text
sentIndex sentText sentNum sentScore
1 Rush Michael Collins Tommi Jaakkola David Sontag MIT CSAIL, Cambridge, MA 02139, USA {mae st ro , s rush ,mco l ins ,t ommi ,ds ont ag} @ c s ail l . [sent-2, score-0.099]
2 edu Abstract This paper introduces algorithms for nonprojective parsing based on dual decomposition. [sent-4, score-0.792]
3 We focus on parsing algorithms for nonprojective head automata, a generalization of head-automata models to non-projective structures. [sent-5, score-0.522]
4 The dual decomposition algorithms are simple and efficient, relying on standard dynamic programming and minimum spanning tree algorithms. [sent-6, score-0.954]
5 They provably solve an LP relaxation of the non-projective parsing problem. [sent-7, score-0.309]
6 Empirically the LP relaxation is very often tight: for many languages, exact solutions are achieved on over 98% of test sentences. [sent-8, score-0.309]
7 1 Introduction Non-projective dependency parsing is useful for many languages that exhibit non-projective syntactic structures. [sent-10, score-0.242]
8 Unfortunately, the non-projective parsing problem is known to be NP-hard for all but the sim- plest models (McDonald and Satta, 2007). [sent-11, score-0.148]
9 There has been a long history in combinatorial optimization of methods that exploit structure in complex problems, using methods such as dual decomposition or Lagrangian relaxation (Lemar´ echal, 2001). [sent-12, score-0.952]
10 This paper introduces algorithms for nonprojective parsing based on dual decomposition. [sent-14, score-0.792]
11 We focus on parsing algorithms for non-projective head automata, a generalization of the head-automata models of Eisner (2000) and Alshawi (1996) to nonprojective structures. [sent-15, score-0.522]
12 These models include nonprojective dependency parsing models with higherorder (e. [sent-16, score-0.532]
13 , sibling and/or grandparent) dependency relations as a special case. [sent-18, score-0.283]
14 Although decoding offull parse structures with non-projective head automata is intractable, we leverage the observation that key components of the decoding can be efficiently computed using combinatorial algorithms. [sent-19, score-0.872]
15 Decoding for individual head-words can be accomplished using dynamic programming. [sent-21, score-0.072]
16 Decoding for arc-factored models can be accomplished using directed minimum-weight spanning tree (MST) algorithms. [sent-23, score-0.165]
17 The resulting parsing algorithms have the following properties: • • • They are efficient and easy to implement, relying on esyta anrdea redff dynamic programming manedn tM, reSlTy nalggorithms. [sent-24, score-0.308]
18 They provably solve a linear programming (LP) rTehlaexya ptiroonv oabfl tyhe s original decoding problem. [sent-25, score-0.318]
19 Empirically the algorithms very often give an exaEcmt psoirliuctailolyn htoe tahleg decoding problem, vine wanh eicxhcase they also provide a certificate of optimality. [sent-26, score-0.338]
20 In this paper we first give the definition for nonprojective head automata, and describe the parsing algorithm. [sent-27, score-0.458]
21 We describe a generalization to models that include grandparent dependencies. [sent-29, score-0.081]
22 We describe experiments on non-projective parsing for a number of languages, and in particu- lar compare the dual decomposition algorithm to approaches based on general-purpose linear programming (LP) or integer linear programming (ILP) solvers (Martins et al. [sent-31, score-1.114]
23 The method gives exact solutions to the decoding problem, together with a certificate ofoptimality, on over 98% oftest examples for many of the test languages, with parsing times ranging between 0. [sent-34, score-0.547]
24 The method compares favorably to previous work using LP/ILP formulations, both in terms of efficiency, and also in terms of the percentage of exact solutions returned. [sent-39, score-0.214]
25 While the focus of the current paper is on nonprojective dependency parsing, the approach opens up new ways of thinking about parsing algorithms for lexicalized formalisms such as TAG (Joshi and Schabes, 1997), CCG (Steedman, 2000), and projective head automata. [sent-40, score-0.818]
26 (2005) describe MST-based parsing for non-projective dependency parsing models with arc-factored decompositions; McDonald and Pereira (2006) make use of an approximate (hill-climbing) algorithm for parsing with more complex models. [sent-42, score-0.517]
27 McDonald and Pereira (2006) and McDonald and Satta (2007) describe complexity results for nonprojective parsing, showing that parsing for a variety of models is NP-hard. [sent-43, score-0.363]
28 Our algorithm differs in that we do not use general-purpose LP or ILP solvers, instead using an MST solver in combination with dynamic programming; thus we leverage the underlying structure of the problem, thereby deriving more efficient decoding algorithms. [sent-46, score-0.178]
29 Both dual decomposition and Lagrangian relaxation have a long history in combinatorial optimization. [sent-47, score-0.897]
30 Our work was originally inspired by recent work on dual decomposition for inference in graphical models (Wainwright et al. [sent-48, score-0.688]
31 However, the non-projective parsing problem has a very different structure from these models, and the decomposition we use is very dif- ferent in nature from those used in graphical models. [sent-51, score-0.373]
32 Other work has made extensive use of decomposition approaches for efficiently solving LP relaxations for graphical models (e. [sent-52, score-0.436]
33 Methods that incorporate combinatorial solvers within loopy belief propagation (LBP) (Duchi et al. [sent-56, score-0.2]
34 Unlike LBP, our method has strong theoretical guarantees, such as guaranteed convergence and the possibility of a certificate of optimality. [sent-58, score-0.147]
35 (2010) describe dual decomposition approaches for other NLP problems. [sent-60, score-0.657]
36 3 Sibling Models This section describes a particular class of models, sibling models; the next section describes a dualdecomposition algorithm for decoding these models. [sent-61, score-0.319]
37 Consider the dependency parsing problem for a sentence with n words. [sent-62, score-0.215]
38 We define the index set for dependency parsing to be I {(i, j) : i ∈ = {0 . [sent-63, score-0.215]
39 A { dependency = parse ins a jve ∈ctor { y = {y(i, j) : (i, j) ∈ I}, where y(i, j) = v1e cift a dependency )w :ith (i ,hje)ad ∈ w Io}rd, wi haenrde modifier j is in the parse, 0 otherwise. [sent-70, score-0.225]
40 We define Y to be the set of all fwoerll th-feor rmooetd s non-projective dependency parses (i. [sent-72, score-0.098]
41 , the set of directed spanning trees rooted at node 0). [sent-74, score-0.073]
42 Given a function f : Y → R that assigns scores to parse trees, tchtieo optimal parse is y∗= argy∈mYaxf(y) (1) A particularly simple definition of f(y) is f(y) = P(i,j)∈I y(i, j)θ(i, j) where θ(i, j) is the score for dPependency (i, j). [sent-75, score-0.084]
43 In this case the optimal parse tree y∗ can be found efficiently using MST algorithms (McDonald et al. [sent-77, score-0.189]
44 This paper describes algorithms that compute y∗ for more complex definitions of f(y) ; in this section, we focus on algorithms for models that capture interactions between sibling dependencies. [sent-79, score-0.362]
45 n, j i} Hence y|i specifies the set of modifiers to word i; = note that the vectors y|i for i= 0 . [sent-84, score-0.077]
46 We then assume that f(y) takes the form Xn f(y) =Xi=0fi(y|i) (2) Thus f(y) decomposes into a sum of terms, where each fi considers modifiers to the i’th word alone. [sent-88, score-0.177]
47 However for certain definitions of fi, it is possible to efficiently compute argmaxy|i∈Zi fi(y|i) for any value of i, typically using dynamic programming. [sent-90, score-0.114]
48 Later we will see that dual decomposition can effectively use MST inference to rule out ill-formed structures. [sent-109, score-0.657]
49 We now give the main assumption underlying sibling models: Assumption 1(Sibling Decompositions) A model f(y) satisfies the sibling-decomposition assumption if: 1) f(y) = Pin=0 fi(y|i) for some set of functions f0 . [sent-110, score-0.263]
50 These terms are benign for most definitions of fi, in that they do not alter decoding complexity. [sent-123, score-0.162]
51 They will be of direct use in the dual decomposition parsing algorithm. [sent-124, score-0.774]
52 Recall that y|i is a binary vector specifying which words are modifiers to the head-word i. [sent-126, score-0.077]
53 lp to be 1290 the sequence of left modifiers to word iunder y|i, and r1 . [sent-130, score-0.291]
54 In bigram sibling models, we have Xp+1 Xq+1 fi(y|i) = XgL(i,lk−1,lk) + XgR(i,rk−1,rk) kX= X1 Xk=1 where l0 = r0 = START is the initial state, and lp+1 = rq+1 = END is the end state. [sent-136, score-0.219]
55 Example 2: Head Automata Head-automata models constitute a second important model type that satisfy the sibling-decomposition assumption (bigram sibling models are a special case of head automata). [sent-145, score-0.373]
56 The CertS and CertG columns in Table 1give the results for the sibling and G+S models respectively. [sent-148, score-0.243]
57 For all but one setting3 over 95% of the test sentences are decoded exactly, with 99% exactness in many cases. [sent-149, score-0.086]
58 For comparison, we also ran both the singlecommodity flow and multiple-commodity flow LP relaxations of Martins et al. [sent-150, score-0.205]
59 We measure how often these relaxations terminate with an exact solution. [sent-152, score-0.19]
60 The results in Table 2 show that our method gives exact solutions more often than both of these relaxations. [sent-153, score-0.177]
61 re- laxations would have given a higher proportion of integral solutions if their relaxation was used during training. [sent-158, score-0.312]
62 Sib/G+S: Non-projective head automata with sibling or grandparent/sibling interactions, decoded via dual decomposition. [sent-171, score-1.024]
63 CertS/CertG: Percent of test examples for which dual decomposition produced a certificate of optimality, for Sib/G+S. [sent-179, score-0.776]
64 For consistency of timing, test decoding was carried out on identical machines with zero additional load; however, training was conducted on machines with varying hardware and load. [sent-182, score-0.134]
65 (2009), we project fractional solutions to a wellformed spanning tree, as described in that paper. [sent-187, score-0.203]
66 Finally, to better compare the tightness of our LP relaxation to that of earlier work, we consider randomly-generated instances. [sent-188, score-0.132]
67 Table 2 gives results for our model and the LP relaxations ofMartins et al. [sent-189, score-0.107]
68 We again recover exact solutions more often than the Martins et al. [sent-191, score-0.177]
69 Note that with random parameters the percentage of exact solutions is significantly lower, suggesting that the exactness of decoding ofthe trained models is a special case. [sent-193, score-0.425]
70 We speculate that this is due to the high performance of approximate decoding with Z in place of Y aunncdeer o fth aep tprraoinxeimd amteod deeclso fdoirn fi; tihthe training algorYith umnd eders tcherib teradi niend dse mcotidoenl s6 f may have the tendency to make the LP relaxation tight. [sent-194, score-0.303]
71 3 Speed Table 1, columns TimeS and TimeG, shows decoding times for the dual decomposition algorithms. [sent-196, score-0.818]
72 Our method gives significant speed-ups over mteHuofcRadep%Amoa21305 102 %30recomputed4,0gs+ib50 1295 Iterations of Dual Decomposition Figure 3: The average percentage of head automata that must be recomputed on each iteration of dual decomposition on the PTB validation set. [sent-199, score-1.131]
73 4 Lazy Decoding Here we describe an important optimization in the dual decomposition algorithms. [sent-203, score-0.712]
74 DD-5000/DD-250: Dual decomposition with nonprojective head automata, with K = 5000/250. [sent-221, score-0.571]
75 Upper results are for the sibling model, lower results are G+S. [sent-222, score-0.185]
76 Columns give scores for UAS accuracy, percentage of solutions which are integral, and solution speed in seconds per sentence. [sent-223, score-0.193]
77 The last column is the percentage of integral solutions on a random problem of length 10 words. [sent-225, score-0.217]
78 In lazy decoding we immediately set z|(ik) = z(|ik−1) if u(k)(i,j) = u(k−1)(i,j) for all j; this check takes for i= 0 . [sent-228, score-0.18]
79 However, if for some i, O(n) time, and saves us from decoding with the i’th automaton. [sent-232, score-0.134]
80 Figure 3 demonstrates the utility of this method for both sibling automata and G+S automata. [sent-234, score-0.483]
81 5 Early Stopping We also ran experiments varying the value of K— the maximum number of iterations—in the dual decomposition algorithms. [sent-236, score-0.689]
82 47+30S Table 3: UAS of projective and non-projective decoding for the English (PTB) and Czech (PDT) validation sets. [sent-243, score-0.371]
83 P-Sib/P-G+S: Projective versions of Sib/G+S, where the MST component has been replaced with the Eisner (2000) first-order projective parser. [sent-245, score-0.198]
84 29% of these solutions certificate of optimality (y(k) In fact, for this setsame solution as for examples. [sent-250, score-0.299]
85 Because z∗ may not be a well-formed tree with n dependencies, we report precision and recall rather than conventional dependency accuracy. [sent-255, score-0.131]
86 Thus the z∗ approximation is clearly a good one; we suspect that this is one reason for the good convergence results for the method. [sent-259, score-0.093]
87 Table 3 shows results for projective and non-projective parsing using the dual decomposition approach. [sent-262, score-0.972]
88 For Czech data, where nonprojective structures are common, non-projective decoding has clear benefits. [sent-263, score-0.376]
89 In contrast, there is little difference in accuracy between projective and nonprojective decoding on English. [sent-264, score-0.547]
90 8 Conclusions We have described dual decomposition algorithms for non-projective parsing, which leverage existing dynamic programming and MST algorithms. [sent-265, score-0.848]
91 7, the algorithms can be easily modified to consider projective structures by replacing Y with the set of projective trees, and then using finirgst Y-or wdeitrh dependency parsing algorithms einn place of MST decoding. [sent-268, score-0.756]
92 This method could be used to derive parsing algorithms that include higher-order features, as an alternative to specialized dynamic programming algorithms. [sent-269, score-0.308]
93 Eisner (2000) describes extensions of head automata to include word senses; we have not discussed this issue in the current pa- per, but it is simple to develop dual decomposition algorithms for this case, using similar methods to those used for the grandparent models. [sent-270, score-1.164]
94 The general approach should be applicable to other lexicalized syntactic formalisms, and potentially also to decoding in syntax-driven translation. [sent-271, score-0.134]
95 In addition, our dual decomposition approach is well-suited to parallelization. [sent-272, score-0.657]
96 Hheernec te tihs eth lee anrunminbge rra toef tdirmoepss that the dual increases from one iteration to the next. [sent-284, score-0.401]
97 2 The ized now call Use of the γ(i, j) Parameters parsing algorithms both consider a generalproblem that includes γ(i, j) parameters. [sent-286, score-0.176]
98 In our models, f(z) can be written as f0(z) + Pi,j α(i, j)z(i, j) where f0(z) includes only termPs depending on higherorder (non arc-factored features), and α(i, j) are weights that consider the dependency between i and j alone. [sent-289, score-0.138]
99 001, which puts almost all the weight in the head-automata models, but allows weights on spanning tree edges to break ties in MST inference in a sensible way. [sent-294, score-0.106]
100 We suspect this is important in early iterations of the algorithm, when many values for u(i, j) or v(i, j) will be zero, and where with β = 0 many spanning tree solutions would be essentially random, leading to very noisy updates to the u(i, j) and v(i, j) values. [sent-295, score-0.288]
wordName wordTfidf (topN-words)
[('dual', 0.401), ('automata', 0.298), ('decomposition', 0.256), ('martins', 0.232), ('nonprojective', 0.215), ('lp', 0.214), ('projective', 0.198), ('sibling', 0.185), ('mst', 0.179), ('uas', 0.143), ('decoding', 0.134), ('relaxation', 0.132), ('solutions', 0.13), ('certificate', 0.119), ('sontag', 0.119), ('parsing', 0.117), ('mcdonald', 0.115), ('combinatorial', 0.108), ('relaxations', 0.107), ('head', 0.1), ('fi', 0.1), ('rush', 0.099), ('dependency', 0.098), ('argmaxz', 0.093), ('programming', 0.088), ('zi', 0.087), ('lagrangian', 0.079), ('ik', 0.077), ('modifiers', 0.077), ('spanning', 0.073), ('ilp', 0.069), ('jaakkola', 0.066), ('eisner', 0.063), ('argmaxy', 0.06), ('provably', 0.06), ('algorithms', 0.059), ('ptb', 0.058), ('pereira', 0.058), ('optimization', 0.055), ('solvers', 0.054), ('integral', 0.05), ('grandparent', 0.05), ('optimality', 0.05), ('exact', 0.047), ('exactness', 0.046), ('lazy', 0.046), ('lemar', 0.046), ('meshi', 0.046), ('xju', 0.046), ('dynamic', 0.044), ('efficiently', 0.042), ('gr', 0.041), ('decoded', 0.04), ('decompositions', 0.04), ('duchi', 0.04), ('globerson', 0.04), ('higherorder', 0.04), ('komodakis', 0.04), ('lbp', 0.04), ('satta', 0.04), ('validation', 0.039), ('approximation', 0.039), ('integer', 0.038), ('belief', 0.038), ('percentage', 0.037), ('approximate', 0.037), ('terminate', 0.036), ('rq', 0.036), ('linear', 0.036), ('bigram', 0.034), ('darpa', 0.034), ('tree', 0.033), ('flow', 0.033), ('smith', 0.032), ('ran', 0.032), ('models', 0.031), ('wainwright', 0.031), ('formalisms', 0.031), ('riedel', 0.031), ('formulations', 0.031), ('nivre', 0.031), ('pages', 0.03), ('parse', 0.029), ('buchholz', 0.029), ('convergence', 0.028), ('definitions', 0.028), ('parser', 0.028), ('accomplished', 0.028), ('koo', 0.028), ('afrl', 0.028), ('dl', 0.028), ('languages', 0.027), ('structures', 0.027), ('columns', 0.027), ('joshi', 0.026), ('suspect', 0.026), ('iterations', 0.026), ('optimal', 0.026), ('assumption', 0.026), ('give', 0.026)]
simIndex simValue paperId paperTitle
same-paper 1 0.99999976 38 emnlp-2010-Dual Decomposition for Parsing with Non-Projective Head Automata
Author: Terry Koo ; Alexander M. Rush ; Michael Collins ; Tommi Jaakkola ; David Sontag
Abstract: This paper introduces algorithms for nonprojective parsing based on dual decomposition. We focus on parsing algorithms for nonprojective head automata, a generalization of head-automata models to non-projective structures. The dual decomposition algorithms are simple and efficient, relying on standard dynamic programming and minimum spanning tree algorithms. They provably solve an LP relaxation of the non-projective parsing problem. Empirically the LP relaxation is very often tight: for many languages, exact solutions are achieved on over 98% of test sentences. The accuracy of our models is higher than previous work on a broad range of datasets.
2 0.48840532 88 emnlp-2010-On Dual Decomposition and Linear Programming Relaxations for Natural Language Processing
Author: Alexander M Rush ; David Sontag ; Michael Collins ; Tommi Jaakkola
Abstract: This paper introduces dual decomposition as a framework for deriving inference algorithms for NLP problems. The approach relies on standard dynamic-programming algorithms as oracle solvers for sub-problems, together with a simple method for forcing agreement between the different oracles. The approach provably solves a linear programming (LP) relaxation of the global inference problem. It leads to algorithms that are simple, in that they use existing decoding algorithms; efficient, in that they avoid exact algorithms for the full model; and often exact, in that empirically they often recover the correct solution in spite of using an LP relaxation. We give experimental results on two problems: 1) the combination of two lexicalized parsing models; and 2) the combination of a lexicalized parsing model and a trigram part-of-speech tagger.
3 0.11566744 110 emnlp-2010-Turbo Parsers: Dependency Parsing by Approximate Variational Inference
Author: Andre Martins ; Noah Smith ; Eric Xing ; Pedro Aguiar ; Mario Figueiredo
Abstract: We present a unified view of two state-of-theart non-projective dependency parsers, both approximate: the loopy belief propagation parser of Smith and Eisner (2008) and the relaxed linear program of Martins et al. (2009). By representing the model assumptions with a factor graph, we shed light on the optimization problems tackled in each method. We also propose a new aggressive online algorithm to learn the model parameters, which makes use of the underlying variational representation. The algorithm does not require a learning rate parameter and provides a single framework for a wide family of convex loss functions, includ- ing CRFs and structured SVMs. Experiments show state-of-the-art performance for 14 languages.
4 0.085210346 67 emnlp-2010-It Depends on the Translation: Unsupervised Dependency Parsing via Word Alignment
Author: Samuel Brody
Abstract: We reveal a previously unnoticed connection between dependency parsing and statistical machine translation (SMT), by formulating the dependency parsing task as a problem of word alignment. Furthermore, we show that two well known models for these respective tasks (DMV and the IBM models) share common modeling assumptions. This motivates us to develop an alignment-based framework for unsupervised dependency parsing. The framework (which will be made publicly available) is flexible, modular and easy to extend. Using this framework, we implement several algorithms based on the IBM alignment models, which prove surprisingly effective on the dependency parsing task, and demonstrate the potential of the alignment-based approach.
5 0.084691204 46 emnlp-2010-Evaluating the Impact of Alternative Dependency Graph Encodings on Solving Event Extraction Tasks
Author: Ekaterina Buyko ; Udo Hahn
Abstract: In state-of-the-art approaches to information extraction (IE), dependency graphs constitute the fundamental data structure for syntactic structuring and subsequent knowledge elicitation from natural language documents. The top-performing systems in the BioNLP 2009 Shared Task on Event Extraction all shared the idea to use dependency structures generated by a variety of parsers either directly or in some converted manner — and optionally modified their output to fit the special needs of IE. As there are systematic differences between various dependency representations being used in this competition, we scrutinize on different encoding styles for dependency information and their possible impact on solving several IE tasks. After assessing more or less established dependency representations such as the Stanford and CoNLL-X dependen— cies, we will then focus on trimming operations that pave the way to more effective IE. Our evaluation study covers data from a number of constituency- and dependency-based parsers and provides experimental evidence which dependency representations are particularly beneficial for the event extraction task. Based on empirical findings from our study we were able to achieve the performance of 57.2% F-score on the development data set of the BioNLP Shared Task 2009.
6 0.084594987 115 emnlp-2010-Uptraining for Accurate Deterministic Question Parsing
7 0.077080876 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
8 0.075488709 98 emnlp-2010-Soft Syntactic Constraints for Hierarchical Phrase-Based Translation Using Latent Syntactic Distributions
9 0.065672472 106 emnlp-2010-Top-Down Nearly-Context-Sensitive Parsing
10 0.060368109 113 emnlp-2010-Unsupervised Induction of Tree Substitution Grammars for Dependency Parsing
11 0.055618171 116 emnlp-2010-Using Universal Linguistic Knowledge to Guide Grammar Induction
12 0.053443994 42 emnlp-2010-Efficient Incremental Decoding for Tree-to-String Translation
13 0.051045809 65 emnlp-2010-Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification
14 0.049949426 114 emnlp-2010-Unsupervised Parse Selection for HPSG
15 0.049184438 105 emnlp-2010-Title Generation with Quasi-Synchronous Grammar
16 0.046301313 72 emnlp-2010-Learning First-Order Horn Clauses from Web Text
17 0.045857616 118 emnlp-2010-Utilizing Extra-Sentential Context for Parsing
18 0.043030843 71 emnlp-2010-Latent-Descriptor Clustering for Unsupervised POS Induction
19 0.042035777 31 emnlp-2010-Constraints Based Taxonomic Relation Classification
20 0.040343173 64 emnlp-2010-Incorporating Content Structure into Text Analysis Applications
topicId topicWeight
[(0, 0.172), (1, 0.075), (2, 0.266), (3, -0.035), (4, 0.071), (5, 0.206), (6, 0.019), (7, 0.325), (8, 0.006), (9, -0.257), (10, -0.401), (11, 0.245), (12, -0.244), (13, -0.153), (14, 0.015), (15, -0.062), (16, -0.082), (17, 0.084), (18, -0.041), (19, -0.052), (20, -0.097), (21, 0.043), (22, 0.023), (23, -0.011), (24, 0.067), (25, 0.044), (26, -0.023), (27, -0.056), (28, 0.033), (29, 0.038), (30, -0.008), (31, -0.06), (32, -0.049), (33, -0.033), (34, 0.022), (35, 0.033), (36, 0.049), (37, 0.008), (38, -0.044), (39, 0.05), (40, -0.038), (41, 0.068), (42, -0.077), (43, -0.032), (44, -0.059), (45, -0.01), (46, 0.036), (47, -0.012), (48, 0.076), (49, -0.078)]
simIndex simValue paperId paperTitle
same-paper 1 0.96976936 38 emnlp-2010-Dual Decomposition for Parsing with Non-Projective Head Automata
Author: Terry Koo ; Alexander M. Rush ; Michael Collins ; Tommi Jaakkola ; David Sontag
Abstract: This paper introduces algorithms for nonprojective parsing based on dual decomposition. We focus on parsing algorithms for nonprojective head automata, a generalization of head-automata models to non-projective structures. The dual decomposition algorithms are simple and efficient, relying on standard dynamic programming and minimum spanning tree algorithms. They provably solve an LP relaxation of the non-projective parsing problem. Empirically the LP relaxation is very often tight: for many languages, exact solutions are achieved on over 98% of test sentences. The accuracy of our models is higher than previous work on a broad range of datasets.
2 0.92840713 88 emnlp-2010-On Dual Decomposition and Linear Programming Relaxations for Natural Language Processing
Author: Alexander M Rush ; David Sontag ; Michael Collins ; Tommi Jaakkola
Abstract: This paper introduces dual decomposition as a framework for deriving inference algorithms for NLP problems. The approach relies on standard dynamic-programming algorithms as oracle solvers for sub-problems, together with a simple method for forcing agreement between the different oracles. The approach provably solves a linear programming (LP) relaxation of the global inference problem. It leads to algorithms that are simple, in that they use existing decoding algorithms; efficient, in that they avoid exact algorithms for the full model; and often exact, in that empirically they often recover the correct solution in spite of using an LP relaxation. We give experimental results on two problems: 1) the combination of two lexicalized parsing models; and 2) the combination of a lexicalized parsing model and a trigram part-of-speech tagger.
3 0.40763393 110 emnlp-2010-Turbo Parsers: Dependency Parsing by Approximate Variational Inference
Author: Andre Martins ; Noah Smith ; Eric Xing ; Pedro Aguiar ; Mario Figueiredo
Abstract: We present a unified view of two state-of-theart non-projective dependency parsers, both approximate: the loopy belief propagation parser of Smith and Eisner (2008) and the relaxed linear program of Martins et al. (2009). By representing the model assumptions with a factor graph, we shed light on the optimization problems tackled in each method. We also propose a new aggressive online algorithm to learn the model parameters, which makes use of the underlying variational representation. The algorithm does not require a learning rate parameter and provides a single framework for a wide family of convex loss functions, includ- ing CRFs and structured SVMs. Experiments show state-of-the-art performance for 14 languages.
4 0.23049793 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
Author: Guillaume Wisniewski ; Alexandre Allauzen ; Francois Yvon
Abstract: Extant Statistical Machine Translation (SMT) systems are very complex softwares, which embed multiple layers of heuristics and embark very large numbers of numerical parameters. As a result, it is difficult to analyze output translations and there is a real need for tools that could help developers to better understand the various causes of errors. In this study, we make a step in that direction and present an attempt to evaluate the quality of the phrase-based translation model. In order to identify those translation errors that stem from deficiencies in the phrase table (PT), we propose to compute the oracle BLEU-4 score, that is the best score that a system based on this PT can achieve on a reference corpus. By casting the computation of the oracle BLEU-1 as an Integer Linear Programming (ILP) problem, we show that it is possible to efficiently compute accurate lower-bounds of this score, and report measures performed on several standard benchmarks. Various other applications of these oracle decoding techniques are also reported and discussed. 1 Phrase-Based Machine Translation 1.1 Principle A Phrase-Based Translation System (PBTS) consists of a ruleset and a scoring function (Lopez, 2009). The ruleset, represented in the phrase table, is a set of phrase1pairs {(f, e) }, each pair expressing that the source phrase f can ,bee) r}e,w earicthten p (atirra enxslparteedss)i inngto t a target phrase e. Trarsaens flation hypotheses are generated by iteratively rewriting portions of the source sentence as prescribed by the ruleset, until each source word has been consumed by exactly one rule. The order of target words in an hypothesis is uniquely determined by the order in which the rewrite operation are performed. The search space ofthe translation model corresponds to the set of all possible sequences of 1Following the usage in statistical machine translation literature, use “phrase” to denote a subsequence of consecutive words. we 933 rules applications. The scoring function aims to rank all possible translation hypotheses in such a way that the best one has the highest score. A PBTS is learned from a parallel corpus in two independent steps. In a first step, the corpus is aligned at the word level, by using alignment tools such as Gi z a++ (Och and Ney, 2003) and some symmetrisation heuristics; phrases are then extracted by other heuristics (Koehn et al., 2003) and assigned numerical weights. In the second step, the parameters of the scoring function are estimated, typically through Minimum Error Rate training (Och, 2003). Translating a sentence amounts to finding the best scoring translation hypothesis in the search space. Because of the combinatorial nature of this problem, translation has to rely on heuristic search techniques such as greedy hill-climbing (Germann, 2003) or variants of best-first search like multi-stack decoding (Koehn, 2004). Moreover, to reduce the overall complexity of decoding, the search space is typically pruned using simple heuristics. For instance, the state-of-the-art phrase-based decoder Moses (Koehn et al., 2007) considers only a restricted number of translations for each source sequence2 and enforces a distortion limit3 over which phrases can be reordered. As a consequence, the best translation hypothesis returned by the decoder is not always the one with the highest score. 1.2 Typology of PBTS Errors Analyzing the errors of a SMT system is not an easy task, because of the number of models that are combined, the size of these models, and the high complexity of the various decision making processes. For a SMT system, three different kinds of errors can be distinguished (Germann et al., 2004; Auli et al., 2009): search errors, induction errors and model errors. The former corresponds to cases where the hypothesis with the best score is missed by the search procedure, either because of the use of an ap2the 3the option of Moses, defaulting to 20. dl option of Moses, whose default value is 7. tt l ProceMedITin,g Ms oasfs thaceh 2u0se1t0ts C,o UnSfAer,e n9c-e11 on O Ectmobpeir ic 2a0l1 M0.e ?tc ho2d0s10 in A Nsastouciraatlio Lnan fogru Cagoem Ppruotcaetisosninagl, L pinaggeusis 9t3ic3s–943, proximate search method or because of the restrictions of the search space. Induction errors correspond to cases where, given the model, the search space does not contain the reference. Finally, model errors correspond to cases where the hypothesis with the highest score is not the best translation according to the evaluation metric. Model errors encompass several types oferrors that occur during learning (Bottou and Bousquet, 2008)4. Approximation errors are errors caused by the use of a restricted and oversimplistic class of functions (here, finitestate transducers to model the generation of hypotheses and a linear scoring function to discriminate them) to model the translation process. Estimation errors correspond to the use of sub-optimal values for both the phrase pairs weights and the parameters of the scoring function. The reasons behind these errors are twofold: first, training only considers a finite sample of data; second, it relies on error prone alignments. As a result, some “good” phrases are extracted with a small weight, or, in the limit, are not extracted at all; and conversely that some “poor” phrases are inserted into the phrase table, sometimes with a really optimistic score. Sorting out and assessing the impact of these various causes of errors is of primary interest for SMT system developers: for lack of such diagnoses, it is difficult to figure out which components of the system require the most urgent attention. Diagnoses are however, given the tight intertwining among the various component of a system, very difficult to obtain: most evaluations are limited to the computation of global scores and usually do not imply any kind of failure analysis. 1.3 Contribution and organization To systematically assess the impact of the multiple heuristic decisions made during training and decoding, we propose, following (Dreyer et al., 2007; Auli et al., 2009), to work out oracle scores, that is to evaluate the best achievable performances of a PBTS. We aim at both studying the expressive power of PBTS and at providing tools for identifying and quantifying causes of failure. Under standard metrics such as BLEU (Papineni et al., 2002), oracle scores are difficult (if not impossible) to compute, but, by casting the computation of the oracle unigram recall and precision as an Integer Linear Programming (ILP) problem, we show that it is possible to efficiently compute accurate lower-bounds of the oracle BLEU-4 scores and report measurements performed on several standard benchmarks. The main contributions of this paper are twofold. We first introduce an ILP program able to efficiently find the best hypothesis a PBTS can achieve. This program can be easily extended to test various improvements to 4We omit here optimization errors. 934 phrase-base systems or to evaluate the impact of different parameter settings. Second, we present a number of complementary results illustrating the usage of our oracle decoder for identifying and analyzing PBTS errors. Our experimental results confirm the main conclusions of (Turchi et al., 2008), showing that extant PBTs have the potential to generate hypotheses having very high BLEU4 score and that their main bottleneck is their scoring function. The rest of this paper is organized as follows: in Section 2, we introduce and formalize the oracle decoding problem, and present a series of ILP problems of increasing complexity designed so as to deliver accurate lowerbounds of oracle score. This section closes with various extensions allowing to model supplementary constraints, most notably reordering constraints (Section 2.5). Our experiments are reported in Section 3, where we first introduce the training and test corpora, along with a description of our system building pipeline (Section 3. 1). We then discuss the baseline oracle BLEU scores (Section 3.2), analyze the non-reachable parts of the reference translations, and comment several complementary results which allow to identify causes of failures. Section 4 discuss our approach and findings with respect to the existing literature on error analysis and oracle decoding. We conclude and discuss further prospects in Section 5. 2 Oracle Decoder 2.1 The Oracle Decoding Problem Definition To get some insights on the errors of phrasebased systems and better understand their limits, we propose to consider the oracle decoding problem defined as follows: given a source sentence, its reference translation5 and a phrase table, what is the “best” translation hypothesis a system can generate? As usual, the quality of an hypothesis is evaluated by the similarity between the reference and the hypothesis. Note that in the oracle decoding problem, we are only assessing the ability of PBT systems to generate good candidate translations, irrespective of their ability to score them properly. We believe that studying this problem is interesting for various reasons. First, as described in Section 3.4, comparing the best hypothesis a system could have generated and the hypothesis it actually generates allows us to carry on both quantitative and qualitative failure analysis. The oracle decoding problem can also be used to assess the expressive power of phrase-based systems (Auli et al., 2009). Other applications include computing acceptable pseudo-references for discriminative training (Tillmann and Zhang, 2006; Liang et al., 2006; Arun and 5The oracle decoding problem can be extended to the case of multiple references. For the sake of simplicity, we only describe the case of a single reference. Koehn, 2007) or combining machine translation systems in a multi-source setting (Li and Khudanpur, 2009). We have also used oracle decoding to identify erroneous or difficult to translate references (Section 3.3). Evaluation Measure To fully define the oracle decoding problem, a measure of the similarity between a translation hypothesis and its reference translation has to be chosen. The most obvious choice is the BLEU-4 score (Papineni et al., 2002) used in most machine translation evaluations. However, using this metric in the oracle decoding problem raises several issues. First, BLEU-4 is a metric defined at the corpus level and is hard to interpret at the sentence level. More importantly, BLEU-4 is not decomposable6: as it relies on 4-grams statistics, the contribution of each phrase pair to the global score depends on the translation of the previous and following phrases and can not be evaluated in isolation. Because of its nondecomposability, maximizing BLEU-4 is hard; in particular, the phrase-level decomposability of the evaluation × metric is necessary in our approach. To circumvent this difficulty, we propose to evaluate the similarity between a translation hypothesis and a reference by the number of their common words. This amounts to evaluating translation quality in terms of unigram precision and recall, which are highly correlated with human judgements (Lavie et al., ). This measure is closely related to the BLEU-1 evaluation metric and the Meteor (Banerjee and Lavie, 2005) metric (when it is evaluated without considering near-matches and the distortion penalty). We also believe that hypotheses that maximize the unigram precision and recall at the sentence level yield corpus level BLEU-4 scores close the maximal achievable. Indeed, in the setting we will introduce in the next section, BLEU-1 and BLEU-4 are highly correlated: as all correct words of the hypothesis will be compelled to be at their correct position, any hypothesis with a high 1-gram precision is also bound to have a high 2-gram precision, etc. 2.2 Formalizing the Oracle Decoding Problem The oracle decoding problem has already been considered in the case of word-based models, in which all translation units are bound to contain only one word. The problem can then be solved by a bipartite graph matching algorithm (Leusch et al., 2008): given a n m binary matarligxo describing possible t 2r0an08sl)a:ti goinv elinn aks n b×emtw beeinna source words and target words7, this algorithm finds the subset of links maximizing the number of words of the reference that have been translated, while ensuring that each word 6Neither at the sentence (Chiang et al., 2008), nor at the phrase level. 7The (i, j) entry of the matrix is 1if the ith word of the source can be translated by the jth word of the reference, 0 otherwise. 935 is translated only once. Generalizing this approach to phrase-based systems amounts to solving the following problem: given a set of possible translation links between potential phrases of the source and of the target, find the subset of links so that the unigram precision and recall are the highest possible. The corresponding oracle hypothesis can then be easily generated by selecting the target phrases that are aligned with one source phrase, disregarding the others. In addition, to mimic the way OOVs are usually handled, we match identical OOV tokens appearing both in the source and target sentences. In this approach, the unigram precision is always one (every word generated in the oracle hypothesis matches exactly one word in the reference). As a consequence, to find the oracle hypothesis, we just have to maximize the recall, that is the number of words appearing both in the hypothesis and in the reference. Considering phrases instead of isolated words has a major impact on the computational complexity: in this new setting, the optimal segmentations in phrases of both the source and of the target have to be worked out in addition to links selection. Moreover, constraints have to be taken into account so as to enforce a proper segmentation of the source and target sentences. These constraints make it impossible to use the approach of (Leusch et al., 2008) and concur in making the oracle decoding problem for phrase-based models more complex than it is for word-based models: it can be proven, using arguments borrowed from (De Nero and Klein, 2008), that this problem is NP-hard even for the simple unigram precision measure. 2.3 An Integer Program for Oracle Decoding To solve the combinatorial problem introduced in the previous section, we propose to cast it into an Integer Linear Programming (ILP) problem, for which many generic solvers exist. ILP has already been used in SMT to find the optimal translation for word-based (Germann et al., 2001) and to study the complexity of learning phrase alignments (De Nero and Klein, 2008) models. Following the latter reference, we introduce the following variables: fi,j (resp. ek,l) is a binary indicator variable that is true when the phrase contains all spans from betweenword position i to j (resp. k to l) of the source (resp. target) sentence. We also introduce a binary variable, denoted ai,j,k,l, to describe a possible link between source phrase fi,j and target phrase ek,l. These variables are built from the entries of the phrase table according to selection strategies introduced in Section 2.4. In the following, index variables are so that: 0 ≤ i< j ≤ n, in the source sentence and 0 ≤ k < l ≤ m, in the target sentence, where n (resp. m) is the length of the source (resp. target) sentence. Solving the oracle decoding problem then amounts to optimizing the following objective function: mi,j,akx,li,Xj,k,lai,j,k,l· (l − k), (1) under the constraints: X ∀x ∈ J1,mK : ek,l ≤ 1 (2) = (3) 1∀,kn,lK : Xai,j,k,l = fk,l (4) ∀i,j : Xai,j,k,l (5) k,l s.tX. Xk≤x≤l ∀∀xy ∈∈ J11,,mnKK : X i,j s.tX. Xi≤y≤j fi,j 1 Xi,j = ei,j Xk,l The objective function (1) corresponds to the number of target words that are generated. The first set of constraints (2) ensures that each word in the reference e ap- pears in no more than one phrase. Maximizing the objective under these constraints amounts to maximizing the unigram recall. The second set of constraints (3) ensures that each word in the source f is translated exactly once, which guarantees that the search space of the ILP problem is the same as the search space of a phrase-based system. Constraints (4) bind the fk,l and ai,j,k,l variables, ensuring that whenever a link ai,j,k,l is active, the corresponding phrase fk,l is also active. Constraints (5) play a similar role for the reference. The Relaxed Problem Even though it accurately models the search space of a phrase-based decoder, this programs is not really useful as is: due to out-ofvocabulary words or missing entries in the phrase table, the constraint that all source words should be translated yields infeasible problems8. We propose to relax this problem and allow some source words to remain untranslated. This is done by replacing constraints (3) by: ∀y ∈ J1,nK : X i,j s.tX. Xi≤y≤j fi,j ≤ 1 To better ref∀lyec ∈t th J1e, bneKh :avior of phrase-based decoders, which attempt to translate all source words, we also need to modify the objective function as follows: X i,Xj,k,l ai,j,k,l · (l − k) +Xfi,j · (j − i) Xi,j (6) The second term in this new objective ensures that optimal solutions translate as many source words as possible. 8An ILP problem is said to be infeasible when tion violates at least one constraint. every possible solu- 936 The Relaxed-Distortion Problem A last caveat with the Relaxed optimization program is caused by frequently occurring source tokens, such as function words or punctuation signs, which can often align with more than one target word. For lack of taking distortion information into account in our objective function, all these alignments are deemed equivalent, even if some of them are clearly more satisfactory than others. This situation is illustrated on Figure 1. le chat et the cat and le the chien dog Figure 1: Equivalent alignments between “le” and “the”. The dashed lines corresponds to a less interpretable solution. To overcome this difficulty, we propose a last change to the objective function: X i,Xj,k,l ai,j,k,l · (l − k) +Xfi,j · (j − i) X ai,j,k,l|k − i| Xi,j −α (7) i Xk ,l X,j, Compared to the objective function of the relaxed problem (6), we introduce here a supplementary penalty factor which favors monotonous alignments. For each phrase pair, the higher the difference between source and target positions, the higher this penalty. If α is small enough, this extra term allows us to select, among all the optimal alignments of the re l axed problem, the one with the lowest distortion. In our experiments, we set α to min {n, m} to ensure that the penalty factor is always smminall{enr, ,tmha}n tthoe e rneswuarred t fhoart aligning atwltyo single iwso ardlwsa. 2.4 Selecting Indicator Variables In the approach introduced in the previous sections, the oracle decoding problem is solved by selecting, among a set of possible translation links, the ones that yield the solution with the highest unigram recall. We propose two strategies to build this set of possible translation links. In the first one, denoted exact match, an indicator ai,j,k,l is created if there is an entry (f, e) so that f spans from word position ito j in the source and e from word position k to l in the target. In this strategy, the ILP program considers exactly the same ruleset as conventional phrase-based decoders. We also consider an alternative strategy, which could help us to identify errors made during the phrase extraction process. In this strategy, denoted inside match, an indicator ai,j,k,l is created when the following three criteria are met: i) f spans from position ito j of the source; ii) a substring of e, denoted e, spans from position k to l of the reference; iii) (f, e¯) is not an entry of the phrase table. The resulting set of indicator variables thus contains, at least, all the variables used in the exact match strategy. In addition, we license here the use of phrases containing words that do not occur in the reference. In fact, using such solutions can yield higher BLEU scores when the reward for additional correct matches exceeds the cost incurred by wrong predictions. These cases are symptoms of situations where the extraction heuristic failed to extract potentially useful subphrases. 2.5 Oracle Decoding with Reordering Constraints The ILP problem introduced in the previous section can be extended in several ways to describe and test various improvements to phrase-based systems or to evaluate the impact of different parameter settings. This flexibility mainly stems from the possibility offered by our framework to express arbitrary constraints over variables. In this section, we illustrate these possibilities by describing how reordering constraints can easily be considered. As a first example, the Moses decoder uses a distortion limit to constrain the set of possible reorderings. This constraint “enforces (...) that the last word of a phrase chosen for translation cannot be more than d9 words from the leftmost untranslated word in the source” (Lopez, 2009) and is expressed as: ∀aijkl , ai0j0k0l0 s.t. k > k0, aijkl · ai0j0k0l0 · |j − i0 + 1| ≤ d, The maximum distortion limit strategy (Lopez, 2009) is also easily expressed and take the following form (assuming this constraint is parameterized by d): ∀l < m − 1, ai,j,k,l·ai0,j0,l+1,l0 · |i0 − j − 1| 71is%t e6hs.a distortion greater that Moses default distortion limit. alignment decisions enabled by the use of larger training corpora and phrase table. To evaluate the impact ofthe second heuristic, we computed the number of phrases discarded by Moses (be- cause of the default ttl limit) but used in the oracle hypotheses. In the English to French NEWSCO setting, they account for 34.11% of the total number of phrases used in the oracle hypotheses. When the oracle decoder is constrained to use the same phrase table as Moses, its BLEU-4 score drops to 42.78. This shows that filtering the phrase table prior to decoding discards many useful phrase pairs and is seriously limiting the best achievable performance, a conclusion shared with (Auli et al., 2009). Search Errors Search errors can be identified by comparing the score of the best hypothesis found by Moses and the score of the oracle hypothesis. If the score of the oracle hypothesis is higher, then there has been a search error; on the contrary, there has been an estimation error when the score of the oracle hypothesis is lower than the score of the best hypothesis found by Moses. 940 Based on the comparison of the score of Moses hypotheses and of oracle hypotheses for the English to French NEWSCO setting, our preliminary conclusion is that the number of search errors is quite limited: only about 5% of the hypotheses of our oracle decoder are actually getting a better score than Moses solutions. Again, this shows that the scoring function (model error) is one of the main bottleneck of current PBTS. Comparing these hypotheses is nonetheless quite revealing: while Moses mostly selects phrase pairs with high translation scores and generates monotonous alignments, our ILP decoder uses larger reorderings and less probable phrases to achieve better solutions: on average, the reordering score of oracle solutions is −5.74, compared to −76.78 fscoro rMeo osfe osr outputs. iGonivsen is −the5 weight assigned through MERT training to the distortion score, no wonder that these hypotheses are severely penalized. The Impact of Phrase Length The observed outputs do not only depend on decisions made during the search, but also on decisions made during training. One such decision is the specification of maximal length for the source and target phrases. In our framework, evaluating the impact of this decision is simple: it suffices to change the definition of indicator variables so as to consider only alignments between phrases of a given length. In the English-French NEWSCO setting, the most restrictive choice, when only alignments between single words are authorized, yields an oracle BLEU-4 of 48.68; however, authorizing phrases up to length 2 allows to achieve an oracle value of 66.57, very close to the score achieved when considering all extracted phrases (67.77). This is corroborated with a further analysis of our oracle alignments, which use phrases whose average source length is 1.21 words (respectively 1.31 for target words). If many studies have already acknowledged the predomi- nance of “small” phrases in actual translations, our oracle scores suggest that, for this language pair, increasing the phrase length limit beyond 2 or 3 might be a waste of computational resources. 4 Related Work To the best of our knowledge, there are only a few works that try to study the expressive power ofphrase-based machine translation systems or to provide tools for analyzing potential causes of failure. The approach described in (Auli et al., 2009) is very similar to ours: in this study, the authors propose to find and analyze the limits of machine translation systems by studying the reference reachability. A reference is reachable for a given system if it can be exactly generated by this system. Reference reachability is assessed using Moses in forced decoding mode: during search, all hypotheses that deviate from the reference are simply discarded. Even though the main goal of this study was to compare the search space of phrase-based and hierarchical systems, it also provides some insights on the impact of various search parameters in Moses, delivering conclusions that are consistent with our main results. As described in Section 1.2, these authors also propose a typology of the errors of a statistical translation systems, but do not attempt to provide methods for identifying them. The authors of (Turchi et al., 2008) study the learn- ing capabilities of Moses by extensively analyzing learning curves representing the translation performances as a function of the number of examples, and by corrupting the model parameters. Even though their focus is more on assessing the scoring function, they reach conclusions similar to ours: the current bottleneck of translation performances is not the representation power of the PBTS but rather in their scoring functions. Oracle decoding is useful to compute reachable pseudo-references in the context of discriminative training. This is the main motivation of (Tillmann and Zhang, 2006), where the authors compute high BLEU hypotheses by running a conventional decoder so as to maximize a per-sentence approximation of BLEU-4, under a simple (local) reordering model. Oracle decoding has also been used to assess the limitations induced by various reordering constraints in (Dreyer et al., 2007). To this end, the authors propose to use a beam-search based oracle decoder, which computes lower bounds of the best achievable BLEU-4 using dynamic programming techniques over finite-state (for so-called local and IBM constraints) or hierarchically structured (for ITG constraints) sets of hypotheses. Even 941 though the numbers reported in this study are not directly comparable with ours17, it seems that our decoder is not only conceptually much simpler, but also achieves much more optimistic lower-bounds of the oracle BLEU score. The approach described in (Li and Khudanpur, 2009) employs a similar technique, which is to guide a heuristic search in an hypergraph representing possible translation hypotheses with n-gram counts matches, which amounts to decoding with a n-gram model trained on the sole reference translation. Additional tricks are presented in this article to speed-up decoding. Computing oracle BLEU scores is also the subject of (Zens and Ney, 2005; Leusch et al., 2008), yet with a different emphasis. These studies are concerned with finding the best hypotheses in a word graph or in a consensus network, a problem that has various implications for multi-pass decoding and/or system combination techniques. The former reference describes an exponential approximate algorithm, while the latter proves the NPcompleteness of this problem and discuss various heuristic approaches. Our problem is somewhat more complex and using their techniques would require us to built word graphs containing all the translations induced by arbitrary segmentations and permutations of the source sentence. 5 Conclusions In this paper, we have presented a methodology for analyzing the errors of PBTS, based on the computation of an approximation of the BLEU-4 oracle score. We have shown that this approximation could be computed fairly accurately and efficiently using Integer Linear Programming techniques. Our main result is a confirmation of the fact that extant PBTS systems are expressive enough to achieve very high translation performance with respect to conventional quality measurements. The main efforts should therefore strive to improve on the way phrases and hypotheses are scored during training. This gives further support to attempts aimed at designing context-dependent scoring functions as in (Stroppa et al., 2007; Gimpel and Smith, 2008), or at attempts to perform discriminative training of feature-rich models. (Bangalore et al., 2007). We have shown that the examination of difficult-totranslate sentences was an effective way to detect errors or inconsistencies in the reference translations, making our approach a potential aid for controlling the quality or assessing the difficulty of test data. Our experiments have also highlighted the impact of various parameters. Various extensions of the baseline ILP program have been suggested and/or evaluated. In particular, the ILP formalism lends itself well to expressing various constraints that are typically used in conventional PBTS. In 17The best BLEU-4 oracle they achieve on Europarl German to English is approximately 48; but they considered a smaller version of the training corpus and the WMT’06 test set. our future work, we aim at using this ILP framework to systematically assess various search configurations. We plan to explore how replacing non-reachable references with high-score pseudo-references can improve discrim- inative training of PBTS. We are also concerned by determining how tight is our approximation of the BLEU4 score is: to this end, we intend to compute the best BLEU-4 score within the n-best solutions of the oracle decoding problem. Acknowledgments Warm thanks to Houda Bouamor for helping us with the annotation tool. This work has been partly financed by OSEO, the French State Agency for Innovation, under the Quaero program. References Tobias Achterberg. 2007. Constraint Integer Programming. Ph.D. thesis, Technische Universit a¨t Berlin. http : / / opus .kobv .de /tuberl in/vol ltexte / 2 0 0 7 / 16 11/ . Abhishek Arun and Philipp Koehn. 2007. Online learning methods for discriminative training of phrase based statistical machine translation. In Proc. of MT Summit XI, Copenhagen, Denmark. Michael Auli, Adam Lopez, Hieu Hoang, and Philipp Koehn. 2009. A systematic analysis of translation model search spaces. In Proc. of WMT, pages 224–232, Athens, Greece. Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proc. of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan. Srinivas Bangalore, Patrick Haffner, and Stephan Kanthak. 2007. Statistical machine translation through global lexical selection and sentence reconstruction. In Proc. of ACL, pages 152–159, Prague, Czech Republic. L e´on Bottou and Olivier Bousquet. 2008. The tradeoffs oflarge scale learning. In Proc. of NIPS, pages 161–168, Vancouver, B.C., Canada. Chris Callison-Burch, Philipp Koehn, Christof Monz, and Josh Schroeder. 2009. Findings of the 2009 Workshop on Statistical Machine Translation. In Proc. of WMT, pages 1–28, Athens, Greece. David Chiang, Steve DeNeefe, Yee Seng Chan, and Hwee Tou Ng. 2008. Decomposability of translation metrics for improved evaluation and efficient algorithms. In Proc. of ECML, pages 610–619, Honolulu, Hawaii. John De Nero and Dan Klein. 2008. The complexity of phrase alignment problems. In Proc. of ACL: HLT, Short Papers, pages 25–28, Columbus, Ohio. Markus Dreyer, Keith B. Hall, and Sanjeev P. Khudanpur. 2007. Comparing reordering constraints for smt using efficient bleu oracle computation. In NAACL-HLT/AMTA Workshop on Syntax and Structure in Statistical Translation, pages 103– 110, Rochester, New York. 942 Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, and Kenji Yamada. 2001 . Fast decoding and optimal decoding for machine translation. In Proc. of ACL, pages 228–235, Toulouse, France. Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, and Kenji Yamada. 2004. Fast and optimal decoding for machine translation. Artificial Intelligence, 154(1-2): 127– 143. Ulrich Germann. 2003. Greedy decoding for statistical machine translation in almost linear time. In Proc. of NAACL, pages 1–8, Edmonton, Canada. Kevin Gimpel and Noah A. Smith. 2008. Rich source-side context for statistical machine translation. In Proc. of WMT, pages 9–17, Columbus, Ohio. Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proc. of NAACL, pages 48–54, Edmonton, Canada. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris CallisonBurch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proc. of ACL, demonstration session. Philipp Koehn. 2004. Pharaoh: A beam search decoder for phrase-based statistical machine translation models. In Proc. of AMTA, pages 115–124, Washington DC. Shankar Kumar and William Byrne. 2005. Local phrase reordering models for statistical machine translation. In Proc. of HLT, pages 161–168, Vancouver, Canada. Alon Lavie, Kenji Sagae, and Shyamsundar Jayaraman. The significance of recall in automatic metrics for MT evaluation. In In Proc. of AMTA, pages 134–143, Washington DC. Gregor Leusch, Evgeny Matusov, and Hermann Ney. 2008. Complexity of finding the BLEU-optimal hypothesis in a confusion network. In Proc. of EMNLP, pages 839–847, Honolulu, Hawaii. Zhifei Li and Sanjeev Khudanpur. 2009. Efficient extraction of oracle-best translations from hypergraphs. In Proc. of NAACL, pages 9–12, Boulder, Colorado. Percy Liang, Alexandre Bouchard-C oˆt´ e, Dan Klein, and Ben Taskar. 2006. An end-to-end discriminative approach to machine translation. In Proc. of ACL, pages 761–768, Sydney, Australia. Adam Lopez. 2009. Translation as weighted deduction. In Proc. of EACL, pages 532–540, Athens, Greece. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Comput. Linguist. , 29(1): 19–5 1. Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In Proc. of ACL, pages 160–167, Sapporo, Japan. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-jing Zhu. 2002. Bleu: A method for automatic evaluation of machine translation. Technical report, Philadelphia, Pennsylvania. D. Roth and W. Yih. 2005. Integer linear programming inference for conditional random fields. In Proc. of ICML, pages 737–744, Bonn, Germany. Nicolas Stroppa, Antal van den Bosch, and Andy Way. 2007. Exploiting source similarity for smt using context-informed features. In Andy Way and Barbara Proc. of TMI, pages Christoph Tillmann 231–240, Sk¨ ovde, and Tong Zhang. Gawronska, editors, Sweden. 2006. A discriminative global training algorithm for statistical mt. In Proc. of ACL, 721–728, Sydney, Australia. Turchi, Tijl De Bie, and Nello pages Marco Cristianini. 2008. Learn- ing performance of a machine translation system: a statistical and computational analysis. In Proc. of WMT, pages Columbus, Ohio. 35–43, Richard Zens and Hermann Ney. 2005. Word graphs for statistical machine translation. In Proc. of the ACL Workshop on Building and Using Parallel Texts, pages 191–198, Ann Arbor, Michigan. 943
5 0.2212766 67 emnlp-2010-It Depends on the Translation: Unsupervised Dependency Parsing via Word Alignment
Author: Samuel Brody
Abstract: We reveal a previously unnoticed connection between dependency parsing and statistical machine translation (SMT), by formulating the dependency parsing task as a problem of word alignment. Furthermore, we show that two well known models for these respective tasks (DMV and the IBM models) share common modeling assumptions. This motivates us to develop an alignment-based framework for unsupervised dependency parsing. The framework (which will be made publicly available) is flexible, modular and easy to extend. Using this framework, we implement several algorithms based on the IBM alignment models, which prove surprisingly effective on the dependency parsing task, and demonstrate the potential of the alignment-based approach.
6 0.21072508 46 emnlp-2010-Evaluating the Impact of Alternative Dependency Graph Encodings on Solving Event Extraction Tasks
7 0.20556915 115 emnlp-2010-Uptraining for Accurate Deterministic Question Parsing
8 0.20148805 113 emnlp-2010-Unsupervised Induction of Tree Substitution Grammars for Dependency Parsing
9 0.16941954 106 emnlp-2010-Top-Down Nearly-Context-Sensitive Parsing
10 0.15639788 71 emnlp-2010-Latent-Descriptor Clustering for Unsupervised POS Induction
11 0.15601104 114 emnlp-2010-Unsupervised Parse Selection for HPSG
12 0.14916332 42 emnlp-2010-Efficient Incremental Decoding for Tree-to-String Translation
14 0.1406858 2 emnlp-2010-A Fast Decoder for Joint Word Segmentation and POS-Tagging Using a Single Discriminative Model
15 0.14015764 116 emnlp-2010-Using Universal Linguistic Knowledge to Guide Grammar Induction
16 0.13943623 105 emnlp-2010-Title Generation with Quasi-Synchronous Grammar
17 0.13683781 65 emnlp-2010-Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification
18 0.13291922 118 emnlp-2010-Utilizing Extra-Sentential Context for Parsing
19 0.12068239 31 emnlp-2010-Constraints Based Taxonomic Relation Classification
20 0.12006546 60 emnlp-2010-Improved Fully Unsupervised Parsing with Zoomed Learning
topicId topicWeight
[(10, 0.014), (12, 0.019), (14, 0.308), (29, 0.095), (32, 0.019), (52, 0.026), (56, 0.077), (62, 0.199), (66, 0.055), (72, 0.037), (76, 0.032), (79, 0.012), (87, 0.016)]
simIndex simValue paperId paperTitle
same-paper 1 0.78533453 38 emnlp-2010-Dual Decomposition for Parsing with Non-Projective Head Automata
Author: Terry Koo ; Alexander M. Rush ; Michael Collins ; Tommi Jaakkola ; David Sontag
Abstract: This paper introduces algorithms for nonprojective parsing based on dual decomposition. We focus on parsing algorithms for nonprojective head automata, a generalization of head-automata models to non-projective structures. The dual decomposition algorithms are simple and efficient, relying on standard dynamic programming and minimum spanning tree algorithms. They provably solve an LP relaxation of the non-projective parsing problem. Empirically the LP relaxation is very often tight: for many languages, exact solutions are achieved on over 98% of test sentences. The accuracy of our models is higher than previous work on a broad range of datasets.
2 0.64090306 88 emnlp-2010-On Dual Decomposition and Linear Programming Relaxations for Natural Language Processing
Author: Alexander M Rush ; David Sontag ; Michael Collins ; Tommi Jaakkola
Abstract: This paper introduces dual decomposition as a framework for deriving inference algorithms for NLP problems. The approach relies on standard dynamic-programming algorithms as oracle solvers for sub-problems, together with a simple method for forcing agreement between the different oracles. The approach provably solves a linear programming (LP) relaxation of the global inference problem. It leads to algorithms that are simple, in that they use existing decoding algorithms; efficient, in that they avoid exact algorithms for the full model; and often exact, in that empirically they often recover the correct solution in spite of using an LP relaxation. We give experimental results on two problems: 1) the combination of two lexicalized parsing models; and 2) the combination of a lexicalized parsing model and a trigram part-of-speech tagger.
3 0.64039528 32 emnlp-2010-Context Comparison of Bursty Events in Web Search and Online Media
Author: Yunliang Jiang ; Cindy Xide Lin ; Qiaozhu Mei
Abstract: In this paper, we conducted a systematic comparative analysis of language in different contexts of bursty topics, including web search, news media, blogging, and social bookmarking. We analyze (1) the content similarity and predictability between contexts, (2) the coverage of search content by each context, and (3) the intrinsic coherence of information in each context. Our experiments show that social bookmarking is a better predictor to the bursty search queries, but news media and social blogging media have a much more compelling coverage. This comparison provides insights on how the search behaviors and social information sharing behaviors of users are correlated to the professional news media in the context of bursty events.
4 0.54580259 90 emnlp-2010-Positional Language Models for Clinical Information Retrieval
Author: Florian Boudin ; Jian-Yun Nie ; Martin Dawes
Abstract: The PECO framework is a knowledge representation for formulating clinical questions. Queries are decomposed into four aspects, which are Patient-Problem (P), Exposure (E), Comparison (C) and Outcome (O). However, no test collection is available to evaluate such framework in information retrieval. In this work, we first present the construction of a large test collection extracted from systematic literature reviews. We then describe an analysis of the distribution of PECO elements throughout the relevant documents and propose a language modeling approach that uses these distributions as a weighting strategy. In our experiments carried out on a collection of 1.5 million documents and 423 queries, our method was found to lead to an improvement of 28% in MAP and 50% in P@5, as com- pared to the state-of-the-art method.
5 0.53749198 72 emnlp-2010-Learning First-Order Horn Clauses from Web Text
Author: Stefan Schoenmackers ; Jesse Davis ; Oren Etzioni ; Daniel Weld
Abstract: input. Even the entire Web corpus does not explicitly answer all questions, yet inference can uncover many implicit answers. But where do inference rules come from? This paper investigates the problem of learning inference rules from Web text in an unsupervised, domain-independent manner. The SHERLOCK system, described herein, is a first-order learner that acquires over 30,000 Horn clauses from Web text. SHERLOCK embodies several innovations, including a novel rule scoring function based on Statistical Relevance (Salmon et al., 1971) which is effective on ambiguous, noisy and incomplete Web extractions. Our experiments show that inference over the learned rules discovers three times as many facts (at precision 0.8) as the TEXTRUNNER system which merely extracts facts explicitly stated in Web text.
6 0.46288493 110 emnlp-2010-Turbo Parsers: Dependency Parsing by Approximate Variational Inference
7 0.38552356 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
8 0.37960687 105 emnlp-2010-Title Generation with Quasi-Synchronous Grammar
9 0.3759242 107 emnlp-2010-Towards Conversation Entailment: An Empirical Investigation
10 0.3747946 82 emnlp-2010-Multi-Document Summarization Using A* Search and Discriminative Learning
12 0.36747456 67 emnlp-2010-It Depends on the Translation: Unsupervised Dependency Parsing via Word Alignment
13 0.36173186 65 emnlp-2010-Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification
14 0.36006939 94 emnlp-2010-SCFG Decoding Without Binarization
15 0.35829237 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice
16 0.35775846 80 emnlp-2010-Modeling Organization in Student Essays
17 0.35469988 69 emnlp-2010-Joint Training and Decoding Using Virtual Nodes for Cascaded Segmentation and Tagging Tasks
18 0.35465124 116 emnlp-2010-Using Universal Linguistic Knowledge to Guide Grammar Induction
19 0.35439056 98 emnlp-2010-Soft Syntactic Constraints for Hierarchical Phrase-Based Translation Using Latent Syntactic Distributions
20 0.35264313 68 emnlp-2010-Joint Inference for Bilingual Semantic Role Labeling