emnlp emnlp2013 emnlp2013-119 knowledge-graph by maker-knowledge-mining

119 emnlp-2013-Learning Distributions over Logical Forms for Referring Expression Generation

Source: pdf

Author: Nicholas FitzGerald ; Yoav Artzi ; Luke Zettlemoyer

Abstract: We present a new approach to referring expression generation, casting it as a density estimation problem where the goal is to learn distributions over logical expressions identifying sets of objects in the world. Despite an extremely large space of possible expressions, we demonstrate effective learning of a globally normalized log-linear distribution. This learning is enabled by a new, multi-stage approximate inference technique that uses a pruning model to construct only the most likely logical forms. We train and evaluate the approach on a new corpus of references to sets of visual objects. Experiments show the approach is able to learn accurate models, which generate over 87% of the expressions people used. Additionally, on the previously studied special case of single object reference, we show a 35% relative error reduction over previous state of the art.

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 edu s , Abstract We present a new approach to referring expression generation, casting it as a density estimation problem where the goal is to learn distributions over logical expressions identifying sets of objects in the world. [sent-3, score-1.409]

2 This learning is enabled by a new, multi-stage approximate inference technique that uses a pruning model to construct only the most likely logical forms. [sent-5, score-0.537]

3 Experiments show the approach is able to learn accurate models, which generate over 87% of the expressions people used. [sent-7, score-0.319]

4 We focus on the task of referring expression generation (REG), where the goal is to produce an expression which uniquely identifies a pre-defined object or set of objects in an environment. [sent-11, score-0.932]

5 Fig1914 ure 1 shows referring expressions provided by human subjects for a set of objects (Figure 1a), demonstrating variation in utterances (Figure 1b) and their corresponding meaning representations (Figure 1c). [sent-13, score-0.744]

6 With these goals in mind, we cast REG as a density estimation problem, where the goal is to learn a distribution over logical forms. [sent-16, score-0.49]

7 For a target set of objects, the number of logical forms that can be used to describe it grows combinatorially with the number of observable properties, such as color and shape. [sent-18, score-0.527]

8 We use a stochastic gradient descent algorithm, where the key challenge is the need to compute feature expectations over all possible logical forms. [sent-22, score-0.442]

9 The correspondence between a s(ezn|tGen,cSe) i fno r1 tbh iasn dsc eitsn ela, beestliemd logical expression in 1c is indicated by the number in parentheses. [sent-94, score-0.574]

10 1 presents a discussion of the space of Pˆ(z|G, possible logical forms. [sent-96, score-0.401]

11 We extend these techniques by modeling the types of plurality and coordination that are prominent in expressions which refer to sets. [sent-99, score-0.483]

12 We also present a new corpus for the task of referring expression generation. [sent-100, score-0.395]

13 Experiments demonstrate highly accurate learned models, able to generate over 87% of the expressions people used. [sent-103, score-0.357]

14 2 Related Work Referring expression generation has been extensively studied in the natural language generation community, dating as far back as SHRDLU (Winograd, 1972). [sent-106, score-0.303]

15 Most work has built on variations of the Incremental Algorithm (Dale and Reiter, 1995), a deterministic algorithm for naming single objects that constructs conjunctive logical expressions. [sent-107, score-0.65]

16 Different approaches have been proposed for generating referring expressions for sets of objects. [sent-111, score-0.577]

17 Further work attempted to resolve the unnaturally long expressions which could be generated by this approach (Gardent, 2002; Horacek, 2004; Gatt and van Deemter, 2007). [sent-113, score-0.403]

18 Much recent work in REG has identified the importance of modeling the variation observed in human-generated referring expressions (Viethen and Dale, 2010; Viethen et al. [sent-120, score-0.541]

19 To the best of our knowledge, this paper presents the first learned probabilistic model for referring expressions defining sets, and is the first effort to treat REG as a density estimation problem. [sent-127, score-0.634]

20 However, most approaches to this problem output bags of con- cepts, while we construct full logical expressions, 1916 allowing our approach to capture complex relations between attributes. [sent-132, score-0.401]

21 Our use of logical forms follows this line of work, while extending it to handle plurality and coordination, as described in Section 4. [sent-138, score-0.454]

22 In addition, lambda calculus was shown to enable effective natural language generation from logical forms (White and Rajkumar, 2009; Lu and Ng, 2011). [sent-140, score-0.564]

23 3 Technical Overview Task Let Z be a set of logical expressions that select a target Zse bte o af objects gGic ainl a wproesrlsdi sntsat eth aSt, as formally defined in Section 5. [sent-142, score-0.958]

24 oFboarb example, i bnu tthioen referring expressions ∈do Zm. [sent-145, score-0.541]

25 The world state S includes the 11 objects in the image, where each object is assigned color (yellow, green . [sent-154, score-0.493]

26 Model and Inference We model P(z|S, G) as a globally normalized log-linear model, using fGe)at uarses a of the logical form z, and its execution with respect to S and G. [sent-166, score-0.489]

27 These sentences are automatically labelled with logical forms with a learned semantic parser, providing a stand-in for manually labeled data (see Section 7). [sent-174, score-0.439]

28 Previous approaches would map this sentence to the same logical expression as the singular “The red cube”, ignoring the semantic distinction encoded by the plural. [sent-185, score-0.703]

29 he, ti-type expressions are therefore functions fromh sets to a teru etxhp-vrealsuseio. [sent-187, score-0.355]

30 red(x) will be true for all sets which contain only objects for which the value red is true. [sent-192, score-0.368]

31 We define semantic plurality in terms of two special collective predicates: sg for singular and plu for plural. [sent-197, score-0.399]

32 h eT sheg plu predicate returns true for sets that contain two or more objects. [sent-203, score-0.305]

33 We also model three kinds of determiners, functional-type hhe, ti, ei-type expressions which sfeunlecctti a single s hehte f,rtoim,e tih-tey power-set represented by their he, ti-type argument. [sent-204, score-0.319]

34 Finally, the indefinite determiner “a” is modeled with the logical constant A, which picks a singleton set by implicitly introducing an epxiciskste anti sailn quantifier (Artzi alnicdZettlemoyer, 2013b). [sent-213, score-0.44]

35 2 Visual Domain Objects in our scenes are labeled with attribute values for four attribute types: color (7 values, such as red, green), shape (9 values, such as cube, sphere), type (16 values, such as broccoli, apple) and a special ob j ect property, which is true for all objects. [sent-226, score-0.401]

36 5 Model and Inference In this section, we describe our approach to modeling the probability P(z | S, G) of a logical form z ∈ gZ t hthea pt names a yse Pt (ozf objects )G o fin a a owgoicrladl fSo, as dze ∈fin Zed t hina tS neacmtioens 3a. [sent-229, score-0.604]

37 1 Space of Possible Meanings The set Z defines logical expressions that we will cTohnes sideetr Z Zfor d picking tghiec target sesets oGn sin t hsattate w Se . [sent-237, score-0.755]

38 However, the vast majority of such expressions are overly complex and redundant, and would never be used in practice as a referring expression. [sent-241, score-0.541]

39 To avoid this explosion, we limit the type and complexity of the logical expressions that are included in Z. [sent-242, score-0.784]

40 We consider only e-type expressions, csiluncdee they name sets, aindedr f ounrltyhe erm-toypree only einssciloundse, expressions that name the desired target set G. [sent-243, score-0.416]

41 The output is the logical expression after the arrow →, constructed using the inputs as shown. [sent-269, score-0.574]

42 also limit the overall complexity of each z ∈ Z, to acolsnota li nm into tth more trhalaln c Mom logical c oofn esatcanhts z. [sent-270, score-0.465]

43 cWedeu friers tfo odre efinnuem Aj ttoin bge Zth,e i nse ot rodfe arll o e- oamndhe, ti-type expressions Athat contain exactly j logicheal, constants. [sent-272, score-0.319]

44 , ∞, by repeatedly adding new constants to expressions in Aj0 for j0 < j. [sent-276, score-0.319]

45 Intuitively, Aj is the set of all complexity j expressions tvhealty can be used as subexpressions for higher complexity entires in our final set Z. [sent-277, score-0.447]

46 MZj of all correct expressions up to a mZa =xim ∪um complexity of M. [sent-283, score-0.383]

47 Together, the pruning model and global model define the distribution | G, S; Π) over z ∈ Z, conditioned on the wPo(rzld | st Ga,teS ;Sθ a,Πnd) target set G, and parameterized by both the parameters of the global model and the parameters Π = {π1 , . [sent-302, score-0.289]

48 4 Features We use three kinds of features: logical expression structure features, situated features and a complexity feature. [sent-307, score-0.673]

49 In order to avoid overly specific features, the attribute value predicates in the logical expressions are replaced with their attribute type (ie. [sent-309, score-0.878]

50 In addition, tthhee special tceon tystpaent (si sg aendd → plu are ignored witihoenn, computing features. [sent-311, score-0.314]

51 In the following description of our features, all examples are computed for the logical expression ι(λx. [sent-312, score-0.574]

52 , Structure Features We use binary features that account for the presence of certain structures in the logical form, allowing the model to learn common usage patterns. [sent-314, score-0.401]

53 For example, the head predicate of the expression “λx. [sent-317, score-0.299]

54 • • • Head-Predicate Bigrams and Trigrams head-predicate bigrams are adnedfine Tdr itgor abme sthe head predicate of a logical form, and the head predicate of one of its children. [sent-321, score-0.653]

55 Coordination Children - this feature set indicCaotoesr tdhien presence ioldf a cnoo -r thdiinsa fteioantu subexpression (∧, ∨, ∪ or \) and the head expressions soifo anll ( pairs ,a ∪nd o triples odf ithtse c hheiladd expressions. [sent-327, score-0.358]

56 Situated Features These features take into account the evaluation of the logical form z with re- spect to the state S and target set G. [sent-331, score-0.436]

57 • Head Predicate and Coverage - this feature set rienddiiccaattees tnhed h Ceoadve predicate osf every sub-expression of the logical form, combined with a comparison between the execution of the sub-expression and the target set G. [sent-333, score-0.523]

58 Coordination Child Relative Coverage - this fCeoaoturrdei sneatt indicates, f Rore every pair voefr cahgiled - s tuhbisexpressions of coordination expressions in the logical form, the coverage of the child sub- expressions relative to each other. [sent-341, score-1.198]

59 Complexity Features We use a single realnumbered feature to account for the complexity of the logical form. [sent-346, score-0.465]

60 We define the complexity of a logical form to be the number of logical constants used. [sent-347, score-0.866]

61 This feature is only used in the global model, since the pruning model always considers logical expressions of fixed complexity. [sent-349, score-0.898]

62 The algorithm is online, using stochastic gradient descent updates for both the globally scored density estimation model and the learned pruning model. [sent-351, score-0.396]

63 logical expressions Zi, a world state Si, and a target set of objects, Gi, which will be identified by the resulting logical expressions. [sent-356, score-1.156]

64 The output is learned parameters for both the globally scored density estimation model θ, and for the learned pruning models Π. [sent-357, score-0.393]

65 n}, where Zi is a ulitsst: o Tfr logical sfeotrm {(sZ, Si is a w)o :rld i state, . [sent-361, score-0.401]

66 Let be the set of all complexity-M logical expressions, afterA pruning (see Section 5. [sent-371, score-0.537]

67 Let SUB(j, z) be all complexity-j sub-expressions of logical expression z. [sent-373, score-0.574]

68 The positive examples, D+, include those sub-expressions owsihtiivche eshxoaumldp lbese, ,i nD the beam - these are all complexity j sub-expressions of logical expressions in Zi. [sent-405, score-0.784]

69 The negative examples, D−, include all complexity-j expressions acmonpsletrsu,c Dted during beam search, minus those which are in D+. [sent-406, score-0.319]

70 Twenty referring expressions were collected for each scene, a total of 5380 expressions. [sent-412, score-0.541]

71 Of the remaining scenes, the sentences of 30 were labeled with logical forms. [sent-414, score-0.401]

72 A small number of expressions (∼5%) from the labeled initniaulm sbeetr were discarded, e∼it5h%er) b ferocmau tshee they dleidd innoitcorrectly name the target set, or because they used very rare attributes (such as texture, or location) to name the target objects. [sent-416, score-0.451]

73 , 2007) contains 856 descriptions of 20 scenes, and although some of these refer to sets, these sets contain two objects at most. [sent-428, score-0.27]

74 Our goal is to achieve a distribution over logical forms that closely matches the distribution observed from human subjects. [sent-439, score-0.469]

75 This metric is quite strict; small differences in the estimated probabilities over a large number of logical expressions can result in a large error, even if the relative ordering is quite similar. [sent-453, score-0.72]

76 Therefore, we report the percentage of observed logical expressions which the model produces, either giving credit multiple times for duplicates (%dup) or counting each unique logical expression in a scene once (%uniq). [sent-454, score-1.36]

77 Put another way, %dup counts logical expression tokens, whereas %uniq counts types. [sent-455, score-0.574]

78 We also report the proportion of scenes where the most likely log- ical expression according to the model matched the most common one in the data (Top1). [sent-456, score-0.282]

79 Single Object Baseline In order to compare our method against the state of the art for generating referring expressions for single objects, we use the subset of our corpus where the target set is a single object. [sent-457, score-0.576]

80 1922 Table 2: Results on the complete corpus for the complete system (Full GenX), ablating the pruning model (NoPrune) and the different features: without coverage features (NoCOV), without structure features (NoSTRUC) and using only the logical expression HeadExp features (HeadExpOnly). [sent-471, score-0.758]

81 Single Objects Table 1shows the different metrics for generating referring expression for single objects only. [sent-475, score-0.598]

82 In addition, unlike VOA, our system (GenX) produces every logical expression used to refer to single objects in our dataset, including a small number which use negation and equality. [sent-478, score-0.777]

83 9% ofthe unique logical expressions used present in our dataset over 87% when these counts are weighted by their frequency. [sent-481, score-0.72]

84 Using the – global model for pruning instead of an explicitly trained model causes a large drop in performance, demonstrating that our global model is inappropri- Figure 4: Example output of our system for the scene on the right. [sent-485, score-0.286]

85 We show the top 10 expressions (z) from the predicted distribution (Pˆ) compared to the empirical distribution estimated from our labeled data (Q). [sent-486, score-0.42]

86 The bottom section shows the predicted probability of the three expressions which were not in the top 10 of the predicted distribution. [sent-487, score-0.385]

87 Much of the error can be attributed to probability mass assigned to logical expressions which, although not observed in our test data, are reasonable referring expressions. [sent-495, score-0.942]

88 This might be due to the fact that our estimate of the empirical distribution comes from a fairly small sample (20), or other factors which we do not model that make these expressions less likely. [sent-496, score-0.353]

89 We demonstrated that we can learn to produce distributions over logical referring expressions using a globally normalized model. [sent-498, score-1.03]

90 Key to the approach was the use of a learned pruning model to define the space of logical expression that are explicitly enumerated during inference. [sent-499, score-0.748]

91 Experiments demonstrate state-of-the-art performance on single object reference and the first results for learning to name sets of objects, correctly recovering over 87% of the observed logical forms. [sent-500, score-0.564]

92 Although the focus of this paper is on REG, the approach is also applicable to learning distributions over logical meaning representations for many other tasks. [sent-504, score-0.401]

93 Computational interpretations of the gricean maxims in the gener1924 ation of referring expressions. [sent-575, score-0.265]

94 Evaluating algorithms for the generation of referring expressions using a balanced corpus. [sent-595, score-0.606]

95 Charting the potential of description logic for the generation of referring expressions. [sent-679, score-0.287]

96 Generating referring expressions: Boolean extensions of the incremental algorithm. [sent-698, score-0.257]

97 Speaker-dependent variation in content selection for referring expression generation. [sent-720, score-0.395]

98 Graphs and spatial relations in the generation of referring expressions. [sent-726, score-0.287]

99 Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars. [sent-745, score-0.401]

100 Online learning of relaxed CCG grammars for parsing to logical form. [sent-750, score-0.401]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('logical', 0.401), ('expressions', 0.319), ('deemter', 0.23), ('referring', 0.222), ('objects', 0.203), ('plu', 0.182), ('expression', 0.173), ('zettlemoyer', 0.154), ('pruning', 0.136), ('artzi', 0.135), ('sg', 0.132), ('red', 0.129), ('reg', 0.121), ('coordination', 0.111), ('scenes', 0.109), ('green', 0.103), ('yellow', 0.098), ('object', 0.096), ('color', 0.091), ('globally', 0.088), ('orange', 0.088), ('predicate', 0.087), ('reiter', 0.086), ('gi', 0.084), ('van', 0.084), ('cube', 0.081), ('gatt', 0.079), ('sub', 0.078), ('xx', 0.075), ('viethen', 0.072), ('matuszek', 0.069), ('mitchell', 0.069), ('blue', 0.067), ('voa', 0.066), ('scene', 0.066), ('generation', 0.065), ('complexity', 0.064), ('aj', 0.063), ('dale', 0.058), ('everything', 0.058), ('attribute', 0.058), ('mooney', 0.057), ('density', 0.055), ('carpenter', 0.053), ('gx', 0.053), ('plurality', 0.053), ('si', 0.051), ('genx', 0.05), ('rajkumar', 0.05), ('brown', 0.049), ('calculus', 0.049), ('krahmer', 0.049), ('lambda', 0.049), ('mae', 0.049), ('coverage', 0.048), ('zi', 0.048), ('conjunctive', 0.046), ('ob', 0.044), ('circled', 0.043), ('rectangle', 0.043), ('gricean', 0.043), ('predicates', 0.042), ('lu', 0.042), ('global', 0.042), ('gradient', 0.041), ('shape', 0.041), ('head', 0.039), ('ti', 0.039), ('determiner', 0.039), ('visual', 0.039), ('scored', 0.038), ('learned', 0.038), ('cubes', 0.037), ('berg', 0.037), ('sets', 0.036), ('incremental', 0.035), ('instructions', 0.035), ('situated', 0.035), ('target', 0.035), ('distribution', 0.034), ('predicted', 0.033), ('andpattern', 0.033), ('areces', 0.033), ('broccoli', 0.033), ('ctx', 0.033), ('dup', 0.033), ('everye', 0.033), ('exc', 0.033), ('grice', 0.033), ('mza', 0.033), ('oxb', 0.033), ('oxr', 0.033), ('scontras', 0.033), ('uniq', 0.033), ('zitnick', 0.033), ('pj', 0.033), ('collective', 0.032), ('name', 0.031), ('descriptions', 0.031), ('definite', 0.03)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.99999928 119 emnlp-2013-Learning Distributions over Logical Forms for Referring Expression Generation

Author: Nicholas FitzGerald ; Yoav Artzi ; Luke Zettlemoyer

2 0.28555974 185 emnlp-2013-Towards Situated Dialogue: Revisiting Referring Expression Generation

Author: Rui Fang ; Changsong Liu ; Lanbo She ; Joyce Y. Chai

Abstract: In situated dialogue, humans and agents have mismatched capabilities of perceiving the shared environment. Their representations of the shared world are misaligned. Thus referring expression generation (REG) will need to take this discrepancy into consideration. To address this issue, we developed a hypergraph-based approach to account for group-based spatial relations and uncertainties in perceiving the environment. Our empirical results have shown that this approach outperforms a previous graph-based approach with an absolute gain of 9%. However, while these graph-based approaches perform effectively when the agent has perfect knowledge or perception of the environment (e.g., 84%), they perform rather poorly when the agent has imperfect perception of the environment (e.g., 45%). This big performance gap calls for new solutions to REG that can mediate a shared perceptual basis in situated dialogue.

3 0.28293455 164 emnlp-2013-Scaling Semantic Parsers with On-the-Fly Ontology Matching

Author: Tom Kwiatkowski ; Eunsol Choi ; Yoav Artzi ; Luke Zettlemoyer

Abstract: We consider the challenge of learning semantic parsers that scale to large, open-domain problems, such as question answering with Freebase. In such settings, the sentences cover a wide variety of topics and include many phrases whose meaning is difficult to represent in a fixed target ontology. For example, even simple phrases such as ‘daughter’ and ‘number of people living in’ cannot be directly represented in Freebase, whose ontology instead encodes facts about gender, parenthood, and population. In this paper, we introduce a new semantic parsing approach that learns to resolve such ontological mismatches. The parser is learned from question-answer pairs, uses a probabilistic CCG to build linguistically motivated logicalform meaning representations, and includes an ontology matching model that adapts the output logical forms for each target ontology. Experiments demonstrate state-of-the-art performance on two benchmark semantic parsing datasets, including a nine point accuracy improvement on a recent Freebase QA corpus.

4 0.26043057 166 emnlp-2013-Semantic Parsing on Freebase from Question-Answer Pairs

Author: Jonathan Berant ; Andrew Chou ; Roy Frostig ; Percy Liang

Abstract: In this paper, we train a semantic parser that scales up to Freebase. Instead of relying on annotated logical forms, which is especially expensive to obtain at large scale, we learn from question-answer pairs. The main challenge in this setting is narrowing down the huge number of possible logical predicates for a given question. We tackle this problem in two ways: First, we build a coarse mapping from phrases to predicates using a knowledge base and a large text corpus. Second, we use a bridging operation to generate additional predicates based on neighboring predicates. On the dataset ofCai and Yates (2013), despite not having annotated logical forms, our system outperforms their state-of-the-art parser. Additionally, we collected a more realistic and challenging dataset of question-answer pairs and improves over a natural baseline.

5 0.16600871 193 emnlp-2013-Unsupervised Induction of Cross-Lingual Semantic Relations

Author: Mike Lewis ; Mark Steedman

Abstract: Creating a language-independent meaning representation would benefit many crosslingual NLP tasks. We introduce the first unsupervised approach to this problem, learning clusters of semantically equivalent English and French relations between referring expressions, based on their named-entity arguments in large monolingual corpora. The clusters can be used as language-independent semantic relations, by mapping clustered expressions in different languages onto the same relation. Our approach needs no parallel text for training, but outperforms a baseline that uses machine translation on a cross-lingual question answering task. We also show how to use the semantics to improve the accuracy of machine translation, by using it in a simple reranker.

6 0.12796071 78 emnlp-2013-Exploiting Language Models for Visual Recognition

7 0.094457947 44 emnlp-2013-Centering Similarity Measures to Reduce Hubs

8 0.089733578 153 emnlp-2013-Predicting the Resolution of Referring Expressions from User Behavior

9 0.080026507 11 emnlp-2013-A Multimodal LDA Model integrating Textual, Cognitive and Visual Modalities

10 0.076958209 98 emnlp-2013-Image Description using Visual Dependency Representations

11 0.074731745 196 emnlp-2013-Using Crowdsourcing to get Representations based on Regular Expressions

12 0.066860028 62 emnlp-2013-Detection of Product Comparisons - How Far Does an Out-of-the-Box Semantic Role Labeling System Take You?

13 0.065202646 96 emnlp-2013-Identifying Phrasal Verbs Using Many Bilingual Corpora

14 0.064637102 12 emnlp-2013-A Semantically Enhanced Approach to Determine Textual Similarity

15 0.060290433 71 emnlp-2013-Efficient Left-to-Right Hierarchical Phrase-Based Translation with Improved Reordering

16 0.058985651 194 emnlp-2013-Unsupervised Relation Extraction with General Domain Knowledge

17 0.05834214 140 emnlp-2013-Of Words, Eyes and Brains: Correlating Image-Based Distributional Semantic Models with Neural Representations of Concepts

18 0.057269983 106 emnlp-2013-Inducing Document Plans for Concept-to-Text Generation

19 0.057239674 66 emnlp-2013-Dynamic Feature Selection for Dependency Parsing

20 0.055429153 145 emnlp-2013-Optimal Beam Search for Machine Translation

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, -0.215), (1, 0.051), (2, -0.025), (3, 0.085), (4, -0.069), (5, 0.324), (6, -0.127), (7, -0.068), (8, 0.113), (9, 0.12), (10, -0.223), (11, -0.13), (12, 0.274), (13, -0.008), (14, 0.07), (15, -0.004), (16, 0.015), (17, -0.13), (18, -0.118), (19, -0.03), (20, 0.022), (21, -0.137), (22, 0.022), (23, 0.083), (24, -0.041), (25, -0.095), (26, 0.023), (27, -0.18), (28, 0.007), (29, 0.101), (30, 0.108), (31, -0.091), (32, -0.059), (33, -0.073), (34, 0.047), (35, -0.005), (36, -0.063), (37, -0.097), (38, 0.073), (39, -0.06), (40, 0.105), (41, 0.078), (42, 0.073), (43, 0.022), (44, 0.093), (45, -0.065), (46, 0.039), (47, 0.017), (48, 0.047), (49, -0.054)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.96928602 119 emnlp-2013-Learning Distributions over Logical Forms for Referring Expression Generation

Author: Nicholas FitzGerald ; Yoav Artzi ; Luke Zettlemoyer

2 0.76055992 185 emnlp-2013-Towards Situated Dialogue: Revisiting Referring Expression Generation

Author: Rui Fang ; Changsong Liu ; Lanbo She ; Joyce Y. Chai

3 0.73250884 164 emnlp-2013-Scaling Semantic Parsers with On-the-Fly Ontology Matching

Author: Tom Kwiatkowski ; Eunsol Choi ; Yoav Artzi ; Luke Zettlemoyer

4 0.59673256 166 emnlp-2013-Semantic Parsing on Freebase from Question-Answer Pairs

Author: Jonathan Berant ; Andrew Chou ; Roy Frostig ; Percy Liang

5 0.49208111 153 emnlp-2013-Predicting the Resolution of Referring Expressions from User Behavior

Author: Nikos Engonopoulos ; Martin Villalba ; Ivan Titov ; Alexander Koller

Abstract: We present a statistical model for predicting how the user of an interactive, situated NLP system resolved a referring expression. The model makes an initial prediction based on the meaning of the utterance, and revises it continuously based on the user’s behavior. The combined model outperforms its components in predicting reference resolution and when to give feedback.

6 0.44848421 193 emnlp-2013-Unsupervised Induction of Cross-Lingual Semantic Relations

7 0.41627479 44 emnlp-2013-Centering Similarity Measures to Reduce Hubs

8 0.402044 196 emnlp-2013-Using Crowdsourcing to get Representations based on Regular Expressions

9 0.37887728 78 emnlp-2013-Exploiting Language Models for Visual Recognition

10 0.28265116 190 emnlp-2013-Ubertagging: Joint Segmentation and Supertagging for English

11 0.27918974 11 emnlp-2013-A Multimodal LDA Model integrating Textual, Cognitive and Visual Modalities

12 0.26380312 32 emnlp-2013-Automatic Idiom Identification in Wiktionary

13 0.24494006 106 emnlp-2013-Inducing Document Plans for Concept-to-Text Generation

14 0.24290642 191 emnlp-2013-Understanding and Quantifying Creativity in Lexical Composition

15 0.23984373 195 emnlp-2013-Unsupervised Spectral Learning of WCFG as Low-rank Matrix Completion

16 0.22492693 140 emnlp-2013-Of Words, Eyes and Brains: Correlating Image-Based Distributional Semantic Models with Neural Representations of Concepts

17 0.216305 12 emnlp-2013-A Semantically Enhanced Approach to Determine Textual Similarity

18 0.21303278 66 emnlp-2013-Dynamic Feature Selection for Dependency Parsing

19 0.21237306 98 emnlp-2013-Image Description using Visual Dependency Representations

20 0.20818076 43 emnlp-2013-Cascading Collective Classification for Bridging Anaphora Recognition using a Rich Linguistic Feature Set

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(3, 0.028), (9, 0.013), (18, 0.041), (22, 0.034), (30, 0.072), (50, 0.047), (51, 0.201), (53, 0.01), (66, 0.023), (71, 0.022), (75, 0.038), (77, 0.017), (90, 0.012), (95, 0.01), (96, 0.013), (97, 0.321)]

similar papers list:

simIndex simValue paperId paperTitle

1 0.89373028 10 emnlp-2013-A Multi-Teraflop Constituency Parser using GPUs

Author: John Canny ; David Hall ; Dan Klein

Abstract: Constituency parsing with rich grammars remains a computational challenge. Graphics Processing Units (GPUs) have previously been used to accelerate CKY chart evaluation, but gains over CPU parsers were modest. In this paper, we describe a collection of new techniques that enable chart evaluation at close to the GPU’s practical maximum speed (a Teraflop), or around a half-trillion rule evaluations per second. Net parser performance on a 4-GPU system is over 1 thousand length30 sentences/second (1 trillion rules/sec), and 400 general sentences/second for the Berkeley Parser Grammar. The techniques we introduce include grammar compilation, recursive symbol blocking, and cache-sharing.

2 0.83659261 23 emnlp-2013-Animacy Detection with Voting Models

Author: Joshua Moore ; Christopher J.C. Burges ; Erin Renshaw ; Wen-tau Yih

Abstract: Animacy detection is a problem whose solution has been shown to be beneficial for a number of syntactic and semantic tasks. We present a state-of-the-art system for this task which uses a number of simple classifiers with heterogeneous data sources in a voting scheme. We show how this framework can give us direct insight into the behavior of the system, allowing us to more easily diagnose sources of error.

same-paper 3 0.82088041 119 emnlp-2013-Learning Distributions over Logical Forms for Referring Expression Generation

Author: Nicholas FitzGerald ; Yoav Artzi ; Luke Zettlemoyer

4 0.64116335 185 emnlp-2013-Towards Situated Dialogue: Revisiting Referring Expression Generation

Author: Rui Fang ; Changsong Liu ; Lanbo She ; Joyce Y. Chai

5 0.60061979 164 emnlp-2013-Scaling Semantic Parsers with On-the-Fly Ontology Matching

Author: Tom Kwiatkowski ; Eunsol Choi ; Yoav Artzi ; Luke Zettlemoyer

6 0.58533865 86 emnlp-2013-Feature Noising for Log-Linear Structured Prediction

7 0.57975155 11 emnlp-2013-A Multimodal LDA Model integrating Textual, Cognitive and Visual Modalities

8 0.57674295 157 emnlp-2013-Recursive Autoencoders for ITG-Based Translation

9 0.5754202 73 emnlp-2013-Error-Driven Analysis of Challenges in Coreference Resolution

10 0.57336158 53 emnlp-2013-Cross-Lingual Discriminative Learning of Sequence Models with Posterior Regularization

11 0.57011592 51 emnlp-2013-Connecting Language and Knowledge Bases with Embedding Models for Relation Extraction