emnlp emnlp2011 emnlp2011-62 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Jette Viethen ; Robert Dale ; Markus Guhe
Abstract: Traditional computational approaches to referring expression generation operate in a deliberate manner, choosing the attributes to be included on the basis of their ability to distinguish the intended referent from its distractors. However, work in psycholinguistics suggests that speakers align their referring expressions with those used previously in the discourse, implying less deliberate choice and more subconscious reuse. This raises the question as to which is a more accurate characterisation of what people do. Using a corpus of dialogues containing 16,358 referring expressions, we explore this question via the generation of subsequent references in shared visual scenes. We use a machine learning approach to referring expression generation and demonstrate that incorporating features that correspond to the computational tradition does not match human referring behaviour as well as using features corresponding to the process of alignment. The results support the view that the traditional model of referring expression generation that is widely assumed in work on natural language generation may not in fact be correct; our analysis may also help explain the oft-observed redundancy found in humanproduced referring expressions.
Reference: text
sentIndex sentText sentNum sentScore
1 However, work in psycholinguistics suggests that speakers align their referring expressions with those used previously in the discourse, implying less deliberate choice and more subconscious reuse. [sent-12, score-0.661]
2 Using a corpus of dialogues containing 16,358 referring expressions, we explore this question via the generation of subsequent references in shared visual scenes. [sent-14, score-0.885]
3 We use a machine learning approach to referring expression generation and demonstrate that incorporating features that correspond to the computational tradition does not match human referring behaviour as well as using features corresponding to the process of alignment. [sent-15, score-1.202]
4 The results support the view that the traditional model of referring expression generation that is widely assumed in work on natural language generation may not in fact be correct; our analysis may also help explain the oft-observed redundancy found in humanproduced referring expressions. [sent-16, score-1.208]
5 1 Introduction Computational work on referring expression generation (REG) has an extensive history, and a wide variety of algorithms have been proposed, dealing with various facets of what is recognised to be a complex problem. [sent-17, score-0.633]
6 Almost all of this work sees the task 1158 as being concerned with choosing those attributes of an intended referent that distinguish it from the other entities with which it might be confused (see, for example, Dale (1989), Dale and Reiter (1995), Krahmer et al. [sent-18, score-0.332]
7 Independently, an alternative way of thinking about reference has arisen within the psycholinguistics com- munity: there is now a long tradition of work that explores how a dialogue participant’s forms of reference are influenced by those previously used for a given entity. [sent-20, score-0.366]
8 Most recently, this line of work has been discussed in terms of the notions of alignment (Pickering and Garrod, 2004) and conceptual pacts (Clark and Wilkes-Gibbs, 1986; Brennan and Clark, 1996). [sent-21, score-0.272]
9 Using a large corpus ofreferring expressions in task-oriented dialogues, this paper presents a machine learning approach that allows us to combine features corresponding to the two perspectives. [sent-23, score-0.131]
10 Our results show that models based on the alignment perspective outperform models based on traditional REG considerations, as well as a number of simpler baselines. [sent-24, score-0.168]
11 In Section 3, we describe the iMAP Corpus and the referring expressions it contains. [sent-27, score-0.559]
12 1 The Algorithmic Approach We use the term algorithmic approach here to refer to the perspective that is common to the considerable body of work within computational linguistics on the problem of referring expression generation developed over the last 20 years. [sent-33, score-0.748]
13 This work has focused on the design of algorithms which take into account the context of reference in order to decide what properties of an entity should be mentioned in order to distinguish that entity from others with which it might be confused. [sent-35, score-0.123]
14 Early work was concerned with subsequent reference in discourse, inspired by Grosz and Sidner’s (1986) observations on how the attentional structure of a discourse made particular referents accessible at any given point. [sent-36, score-0.305]
15 More recently, attention has shifted to initial reference in visual domains, driven in large part by the availability of the TUNA dataset and the shared tasks that make use of it (Gatt et al. [sent-37, score-0.318]
16 Scenarios that require the generation ofreferences in multi-turn dialogues that concern visual scenes are likely to be among the first where we can expect computational approaches to referring expression generation to be practically useful. [sent-40, score-1.034]
17 Surprisingly, however, the more recent work on initial reference in visual domains and the earlier work on subsequent reference in discourse remain somewhat distinct and separate from each other, despite much the same algorithms having been used in both. [sent-41, score-0.574]
18 There is very little work that brings these two strands together by looking at both initial and subsequent references in dialogues that concern visual scenes. [sent-42, score-0.346]
19 However, their approach was concerned with choosing the type of reference to use (definite or indefinite, pronominal, bare or 1159 modified head noun), and not with the content of the reference; and their data set consisted of only 1242 referring expressions. [sent-45, score-0.73]
20 2 The Alignment Approach Meanwhile, starting with the early work of Carroll (1980), a quite distinct strand of research in psycholinguistics has explored how a speaker’s form of reference to an entity is impacted by the way that entity has been previously referred to in the discourse or dialogue. [sent-47, score-0.222]
21 The general idea behind what we will call the alignment approach is that a conversational participant will often adopt the same semantic, syntactic and lexical alternatives as the other party in a dialogue. [sent-48, score-0.178]
22 With respect to reference in particular, speakers are said to form conceptual pacts in their use of language (Clark and Wilkes-Gibbs, 1986; Brennan and Clark, 1996). [sent-50, score-0.296]
23 Recent work by Goudbeek and Krahmer (2010) supports the view that subconscious alignment does indeed take place at the level of content selection for referring expressions. [sent-52, score-0.69]
24 The participants in their study were more likely to use a dispreferred attribute to describe a target referent if this attribute had recently been used in a description by a confederate. [sent-53, score-0.403]
25 There is some work within natural language generation that attempts to model the process of alignment (Buschmeier et al. [sent-54, score-0.209]
26 , 2009; Janarthanam and Lemon, 2009), but this is predominantly concerned with what we might think of as the ‘lexical perspective’ , focussing on lexical choice rather than the selection of appropriate semantic content for distinguishing descriptions. [sent-55, score-0.177]
27 3 Combined Models This paper is not the first to look at how the algorithmic approach and the alignment approach might be integrated in REG. [sent-57, score-0.194]
28 However, their data set consists of only 393 referring expressions, compared to our 16,358, and these expressions had functions other than identification; most importantly, the entities referred to were not part of a shared visual scene as is the case in our data. [sent-60, score-0.754]
29 Gupta and Stent (2005) instantiated Dale and Reiter’s (1995) Incremental Algorithm with a preference ordering that favours the attributes that were used in the previous mention of the same referent. [sent-61, score-0.258]
30 In a second variant, they even require these attributes to be included in a subsequent reference. [sent-62, score-0.259]
31 Differently from most other work on REG, they extended the task to include ordering of the attributes in the surface form. [sent-63, score-0.188]
32 They therefore create a special evaluation metric that takes ordering into account, which makes it hard to compare the performance they report to that of any system that is not concerned with attribute ordering, such as ours. [sent-64, score-0.16]
33 Their evaluation set was also considerably smaller than ours: they used 1294 and 471 referring expressions from two different corpora, compared to our test set of 4947 referring expressions. [sent-65, score-1.021]
34 , 2007) is a collection of 256 dialogues between 32 participantpairs who contributed 8 dialogues each. [sent-70, score-0.16]
35 Both participants had a map of the same environment, but one participant’s map showed a route winding its way between the landmarks on the map; see Figure 1. [sent-71, score-0.248]
36 In addition to these inherent attributes of the landmarks, participants used spatial relations to other items on the map. [sent-84, score-0.282]
37 Each referring expression in the corpus is annotated with a unique identifier corresponding to the landmark that it describes and the semantic values of the attributes that it contains. [sent-85, score-0.84]
38 For each landmark R referred to in a dialogue, we view the sequence of references to this landmark as a coreference chain, notated hR1, R2 , . [sent-87, score-0.192]
39 From the corpus as a whole we extracted 34,127 referring expressions in 9558 chains. [sent-92, score-0.559]
40 References may be contributed to a chain by either speaker, and can be arbitrarily far apart: in the data, 4201 references are in the utterance immediately following the preceding reference in the chain, but the distance between references in a chain can be as high as 423 utterances. [sent-95, score-0.215]
41 We removed from the data any annotation that was not concerned with the four landmark attributes, type, colour, relation, or the landmark’s other distinguishing attribute. [sent-96, score-0.177]
42 Ordinal numbers that were annotated as the use of the number attribute were re-tagged as spatial relations, as these usually described the position of the target within a line of landmarks. [sent-98, score-0.213]
43 As a result of the removal of annotations not pertaining to the use of the four landmark attributes, 2785 referring expressions had no annotation left; we removed these instances from the final data set. [sent-99, score-0.655]
44 We also do not attempt to replicate the remaining 5552 plural referring expressions or the 3062 pro- Content Pattern Count 1161 Proportion hotheri589336. [sent-100, score-0.559]
45 1 nouns found in the However, we do include all of these instances in the feature extraction step, on the assumption that they might impact on the content of subsequent references. [sent-117, score-0.167]
46 Similarly, we filter out 6369 initial references after we have extracted features from them, since we focus here on the generation of subsequent reference only. [sent-118, score-0.305]
47 The remaining 16,358 referring expressions form the data which we use in our experiments. [sent-119, score-0.559]
48 Contrary to findings from other corpora, in which colour was used much more frequently (Gatt, 2007; Viethen and Dale, 2008), the colour attribute was used in only 26. [sent-120, score-0.491]
49 This is probably due to the often low reliability of colour in this task caused by the ink stains. [sent-122, score-0.245]
50 The proportion of referring expressions mentioning the target’s type might, at 38. [sent-123, score-0.598]
51 16% of the referring expressions, comparable to other corpora in the literature. [sent-129, score-0.462]
52 We can think of each referring expression as be- ing a linguistic realisation of a content pattern: this is the collection of attributes that are used in that instance. [sent-131, score-0.84]
53 The attributes can be derived from the property-level annotation given in the corpus. [sent-132, score-0.188]
54 So, for example, if a particular reference appears as the noun phrase the blue penguin, annotated semantically as hblue, penguini, then the corresponding ctiocnatlelynt pattern i,sp hcolour, kindi . [sent-133, score-0.164]
55 O thuer a ciomrr eiss ptoo replicate etnhet content pattern orf, keinacdhi referring expression in the corpus. [sent-134, score-0.693]
56 1 The Two Perspectives Our task is defined simply as follows: for each subsequent reference R in the corpus, can we predict the content pattern that will be used in that reference? [sent-137, score-0.331]
57 The alignment approach, on the other hand, can be summarised thus: Speakers align the forms of reference they use to be similar or identical to references that have been used before. [sent-140, score-0.255]
58 In particular, once a form of reference to the intended referent has been established, they tend to re-use that form of reference, or perhaps an abbreviated version of it. [sent-141, score-0.218]
59 The alignment approach would appear to be preferable on the grounds of computational cost: we would expect that retrieving a previously-used referring expression, or parts thereof, generally requires less computation than building a new referring expression from scratch. [sent-142, score-1.15]
60 shape of the ink blot(s) on the IF’s map Lmprop Features other Att type of the other attribute of the target [att] Value value for each att of target [att] Difference was att of target different between the two maps? [sent-144, score-0.994]
61 2Unfortunately, determining what counts as a change ofcontext, especially in visual scenes, is fraught with difficulty. [sent-154, score-0.195]
62 TradREG Features (Visual) Count Vis Distractors Prop Vis Same [att] number of visual distractors proportion of visual distractors with same att Dist Closest distance to the closest visual distractor Closest Same [att] has the closest distractor the same att? [sent-155, score-1.275]
63 Dist Closest Same [att] distance to the closest distractor of same att as target Cl Same type Same [att] has the closest distractor of the same type also the same att? [sent-156, score-0.572]
64 TradREG Features (Discourse) Count Intervening LMs number of other LMs mentioned since the last mention of the target Prop Intervening [att] proportion of intervening LMs for which att was used AND which have the same att as target Table 3: The TradREG feature set. [sent-157, score-0.894]
65 2 Features The number of factors that can be hypothesised as having an impact on the form of a referring expression in a dialogic setting associated with a visual do- main is very large. [sent-160, score-0.751]
66 Instead, we here capture a wide range of factors as features that can be used by a machine learning algorithm to automatically induce from the data a classifier that predicts for a given set of features the attributes that should be used in a referring expression. [sent-162, score-0.718]
67 Map Features capture design characteristics of the maps the current dialogue is about; Speaker Features capture the identity and role of the participants; and LMprop Features capture the inherent visual properties of the target referent. [sent-165, score-0.325]
68 Most importantly for our present considerations, 3In these tables, att is an abbreviatory variable that is instantiated once for each of the four attributes type, colour, relation, and the other distinguishing attribute of the landmark. [sent-167, score-0.65]
69 The abbreviation LM stands for landmark 1163 Alignment Features (Recency) Last Men Speaker Same who made the last mention of target? [sent-168, score-0.219]
70 Last Mention [att] was att used in the last mention of target? [sent-169, score-0.442]
71 Alignment Features (Frequency) Count [att] Dial how often has att been used in the dialogue? [sent-171, score-0.319]
72 Count [att] LM how often has att been used for target? [sent-172, score-0.319]
73 Quartile quartile of the dialogue the RE was uttered in Dial No number of dialogues already completed +1 Mention No number of previous mentions of target +1 Table 4: The Alignment feature set. [sent-173, score-0.273]
74 In addition to the main prediction class content pattern, the split was stratified for Speaker ID and Quartile to ensure that training and test set contained the same proportion of descriptions from each speaker and each quartile of the dialogues. [sent-177, score-0.358]
75 We used the J48 algorithm implemented in the Weka toolkit (Witten and Frank, 2005) to train decision trees with the task of judging, based on the given features, which content pattern should be used. [sent-178, score-0.184]
76 It generates the same content pattern that was used in the previous mention of the target referent. [sent-182, score-0.254]
77 MajorityClass generates the content pattern most commonly used in the training set. [sent-183, score-0.137]
78 This gives some measure of the overlap between two referring expressions, assigning a partial score if the two sets share attributes but are not identical. [sent-189, score-0.65]
79 Table 5 compares the performances of the three baselines and the decision trees based on the five feature subsets for each of the individual attributes and for the combined content pattern; note that the HeadNounOnly and RepeatLast baselines do not make attribute-specific predictions. [sent-211, score-0.331]
80 The table shows that the learned systems outperform all three baselines for the individual attributes as well as for the combined content pattern. [sent-212, score-0.284]
81 Also consistent with the results of the three individual feature sets, dropping the Ind features hurts performance more than dropping the TradREG features, but less than dropping the Alignment features. [sent-219, score-0.193]
82 These results suggest that considerations at the heart of traditional REG approaches do not play as important a role as those postulated by alignmentbased models for the selection of semantic content for subsequent referring expressions. [sent-222, score-0.723]
83 Even in the arguably much simpler non-dialogic domains of the REG competitions concerned with pure content selection, the best performing system achieved only 53% Accuracy (see Gatt et al. [sent-230, score-0.145]
84 ferent set of attributes from those included by the human speaker; however, the Accuracy score also counts as incorrect any set that only partly overlaps with the reference found in the test set. [sent-244, score-0.311]
85 A DICE score that is equal to the Accuracy score would mean that each referring expression was either reproduced perfectly, or that a set of attributes was chosen that did not overlap with the original one at all. [sent-246, score-0.744]
86 Interestingly, a large number of the referring expressions produced by the model trained only on TradREG features are subsets of the human reference. [sent-249, score-0.593]
87 This indicates that human speakers tend to include more attributes than are strictly speaking necessary to distinguish the landmark. [sent-250, score-0.221]
88 52 7 Table 7: Comparison of the predictions for the combined content pattern between the models trained on mutually exclusive feature sets. [sent-258, score-0.137]
89 the apparent redundancy that human-produced referring expressions contain. [sent-259, score-0.559]
90 Table 7 lists for those pairings of our learned models which were based on mutually exclusive feature sets how many referring expressions both models predicted correctly, how many both failed to predict, and how many were predicted correctly by either of the two models. [sent-261, score-0.559]
91 1166 7 Conclusions Using the largest corpus of referring expressions to date, we have shown how both the traditional computational view of REG and the alternative psycholinguistic alignment approach can be captured via a large set of features for machine learning. [sent-271, score-0.807]
92 First, we have demonstrated that a model using all these features to predict content patterns in subsequent references in shared visual scenes delivers an Accuracy of 58. [sent-274, score-0.445]
93 Second, our error analysis showed that the main reason for the low performance of a model based on traditional algorithmic features is that it often chooses too few attributes. [sent-278, score-0.132]
94 The fact that the model based on the alignment features does not make this mistake so frequently suggests that it may be the psycholinguistic considerations incorporated in our alignment features that lead people to add those additional attributes. [sent-279, score-0.436]
95 Finally, while the different models make the same correct predictions about the content of referring expressions in many cases, there are also a considerable number of cases where the models based on either the traditional algorithmic features (10. [sent-280, score-0.787]
96 Computational interpretations of the Gricean maxims in the generation of referring expressions. [sent-306, score-0.539]
97 Learning lexical alignment policies for generating referring 1167 expressions for spoken dialogue systems. [sent-346, score-0.774]
98 Learning content selection rules for generating object descriptions in dialogue. [sent-357, score-0.133]
99 Graphs and Booleans: On the generation of referring expressions. [sent-389, score-0.539]
100 The use of spatial relations in referring expression generation. [sent-395, score-0.611]
wordName wordTfidf (topN-words)
[('referring', 0.462), ('att', 0.319), ('tradreg', 0.269), ('visual', 0.195), ('colour', 0.19), ('attributes', 0.188), ('reg', 0.15), ('alignment', 0.132), ('speaker', 0.123), ('reference', 0.123), ('ind', 0.122), ('attribute', 0.111), ('hother', 0.111), ('landmarks', 0.111), ('typei', 0.111), ('dice', 0.102), ('expressions', 0.097), ('content', 0.096), ('landmark', 0.096), ('gatt', 0.095), ('referent', 0.095), ('expression', 0.094), ('dale', 0.092), ('dist', 0.092), ('dialogue', 0.083), ('dialogues', 0.08), ('pacts', 0.079), ('viethen', 0.079), ('generation', 0.077), ('res', 0.074), ('subsequent', 0.071), ('mention', 0.07), ('distractor', 0.063), ('distractors', 0.063), ('hcolour', 0.063), ('krahmer', 0.063), ('quartile', 0.063), ('utts', 0.063), ('algorithmic', 0.062), ('discourse', 0.062), ('conceptual', 0.061), ('considerations', 0.058), ('spatial', 0.055), ('dial', 0.055), ('grosz', 0.055), ('ink', 0.055), ('dropping', 0.053), ('last', 0.053), ('scenes', 0.049), ('concerned', 0.049), ('map', 0.049), ('allfeatures', 0.048), ('birds', 0.048), ('brennan', 0.048), ('combinedpattern', 0.048), ('emiel', 0.048), ('imap', 0.048), ('jette', 0.048), ('lmprop', 0.048), ('pickering', 0.048), ('reiter', 0.048), ('relationi', 0.048), ('repeatlast', 0.048), ('stoia', 0.048), ('decision', 0.047), ('target', 0.047), ('participant', 0.046), ('psycholinguistic', 0.046), ('chain', 0.046), ('pattern', 0.041), ('closest', 0.04), ('lm', 0.04), ('proportion', 0.039), ('behaviour', 0.039), ('participants', 0.039), ('clark', 0.038), ('descriptions', 0.037), ('psycholinguistics', 0.037), ('sidner', 0.037), ('pamela', 0.037), ('traditional', 0.036), ('utterances', 0.036), ('features', 0.034), ('speakers', 0.033), ('jordan', 0.033), ('perspectives', 0.032), ('distinguishing', 0.032), ('buildings', 0.032), ('bunt', 0.032), ('buschmeier', 0.032), ('deemter', 0.032), ('deliberate', 0.032), ('dordrecht', 0.032), ('fork', 0.032), ('gardent', 0.032), ('garrod', 0.032), ('goudbeek', 0.032), ('headnounonly', 0.032), ('inked', 0.032), ('janarthanam', 0.032)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000005 62 emnlp-2011-Generating Subsequent Reference in Shared Visual Scenes: Computation vs Re-Use
Author: Jette Viethen ; Robert Dale ; Markus Guhe
Abstract: Traditional computational approaches to referring expression generation operate in a deliberate manner, choosing the attributes to be included on the basis of their ability to distinguish the intended referent from its distractors. However, work in psycholinguistics suggests that speakers align their referring expressions with those used previously in the discourse, implying less deliberate choice and more subconscious reuse. This raises the question as to which is a more accurate characterisation of what people do. Using a corpus of dialogues containing 16,358 referring expressions, we explore this question via the generation of subsequent references in shared visual scenes. We use a machine learning approach to referring expression generation and demonstrate that incorporating features that correspond to the computational tradition does not match human referring behaviour as well as using features corresponding to the process of alignment. The results support the view that the traditional model of referring expression generation that is widely assumed in work on natural language generation may not in fact be correct; our analysis may also help explain the oft-observed redundancy found in humanproduced referring expressions.
2 0.16413039 56 emnlp-2011-Exploring Supervised LDA Models for Assigning Attributes to Adjective-Noun Phrases
Author: Matthias Hartung ; Anette Frank
Abstract: This paper introduces an attribute selection task as a way to characterize the inherent meaning of property-denoting adjectives in adjective-noun phrases, such as e.g. hot in hot summer denoting the attribute TEMPERATURE, rather than TASTE. We formulate this task in a vector space model that represents adjectives and nouns as vectors in a semantic space defined over possible attributes. The vectors incorporate latent semantic information obtained from two variants of LDA topic models. Our LDA models outperform previous approaches on a small set of 10 attributes with considerable gains on sparse representations, which highlights the strong smoothing power of LDA models. For the first time, we extend the attribute selection task to a new data set with more than 200 classes. We observe that large-scale attribute selection is a hard problem, but a subset of attributes performs robustly on the large scale as well. Again, the LDA models outperform the VSM baseline.
3 0.090404876 3 emnlp-2011-A Correction Model for Word Alignments
Author: J. Scott McCarley ; Abraham Ittycheriah ; Salim Roukos ; Bing Xiang ; Jian-ming Xu
Abstract: Models of word alignment built as sequences of links have limited expressive power, but are easy to decode. Word aligners that model the alignment matrix can express arbitrary alignments, but are difficult to decode. We propose an alignment matrix model as a correction algorithm to an underlying sequencebased aligner. Then a greedy decoding algorithm enables the full expressive power of the alignment matrix formulation. Improved alignment performance is shown for all nine language pairs tested. The improved alignments also improved translation quality from Chinese to English and English to Italian.
4 0.066383854 38 emnlp-2011-Data-Driven Response Generation in Social Media
Author: Alan Ritter ; Colin Cherry ; William B. Dolan
Abstract: Ottawa, Ontario, K1A 0R6 Co l . Cherry@ nrc-cnrc . gc . ca in Redmond, WA 98052 bi l ldol @mi cro so ft . com large corpus of status-response pairs found on Twitter to create a system that responds to Twitter status We present a data-driven approach to generating responses to Twitter status posts, based on phrase-based Statistical Machine Translation. We find that mapping conversational stimuli onto responses is more difficult than translating between languages, due to the wider range of possible responses, the larger fraction of unaligned words/phrases, and the presence of large phrase pairs whose alignment cannot be further decomposed. After addressing these challenges, we compare approaches based on SMT and Information Retrieval in a human evaluation. We show that SMT outperforms IR on this task, and its output is preferred over actual human responses in 15% of cases. As far as we are aware, this is the first work to investigate the use of phrase-based SMT to directly translate a linguistic stimulus into an appropriate response.
5 0.065184109 34 emnlp-2011-Corpus-Guided Sentence Generation of Natural Images
Author: Yezhou Yang ; Ching Teo ; Hal Daume III ; Yiannis Aloimonos
Abstract: We propose a sentence generation strategy that describes images by predicting the most likely nouns, verbs, scenes and prepositions that make up the core sentence structure. The input are initial noisy estimates of the objects and scenes detected in the image using state of the art trained detectors. As predicting actions from still images directly is unreliable, we use a language model trained from the English Gigaword corpus to obtain their estimates; together with probabilities of co-located nouns, scenes and prepositions. We use these estimates as parameters on a HMM that models the sentence generation process, with hidden nodes as sentence components and image detections as the emissions. Experimental results show that our strategy of combining vision and language produces readable and de- , scriptive sentences compared to naive strategies that use vision alone.
6 0.061265372 10 emnlp-2011-A Probabilistic Forest-to-String Model for Language Generation from Typed Lambda Calculus Expressions
7 0.057431303 23 emnlp-2011-Bootstrapped Named Entity Recognition for Product Attribute Extraction
8 0.056877859 116 emnlp-2011-Robust Disambiguation of Named Entities in Text
9 0.053334374 76 emnlp-2011-Language Models for Machine Translation: Original vs. Translated Texts
10 0.051994149 84 emnlp-2011-Learning the Information Status of Noun Phrases in Spoken Dialogues
11 0.050841171 8 emnlp-2011-A Model of Discourse Predictions in Human Sentence Processing
12 0.049791183 105 emnlp-2011-Predicting Thread Discourse Structure over Technical Web Forums
13 0.049275283 58 emnlp-2011-Fast Generation of Translation Forest for Large-Scale SMT Discriminative Training
14 0.048764504 128 emnlp-2011-Structured Relation Discovery using Generative Models
15 0.048627079 142 emnlp-2011-Unsupervised Discovery of Discourse Relations for Eliminating Intra-sentence Polarity Ambiguities
16 0.048367124 131 emnlp-2011-Syntactic Decision Tree LMs: Random Selection or Intelligent Design?
17 0.046988986 57 emnlp-2011-Extreme Extraction - Machine Reading in a Week
18 0.04628437 1 emnlp-2011-A Bayesian Mixture Model for PoS Induction Using Multiple Features
19 0.045756675 113 emnlp-2011-Relation Acquisition using Word Classes and Partial Patterns
20 0.039614853 82 emnlp-2011-Learning Local Content Shift Detectors from Document-level Information
topicId topicWeight
[(0, 0.151), (1, -0.047), (2, -0.045), (3, -0.062), (4, -0.032), (5, -0.032), (6, 0.002), (7, 0.05), (8, -0.017), (9, -0.139), (10, -0.04), (11, -0.024), (12, -0.058), (13, 0.028), (14, -0.13), (15, 0.003), (16, 0.154), (17, 0.046), (18, -0.055), (19, -0.009), (20, 0.086), (21, 0.049), (22, -0.061), (23, -0.197), (24, 0.036), (25, 0.016), (26, 0.311), (27, -0.048), (28, 0.022), (29, -0.091), (30, -0.107), (31, 0.211), (32, 0.093), (33, -0.123), (34, 0.112), (35, -0.174), (36, 0.03), (37, 0.04), (38, -0.019), (39, -0.054), (40, 0.093), (41, 0.15), (42, 0.087), (43, 0.016), (44, 0.043), (45, 0.126), (46, -0.048), (47, 0.003), (48, 0.0), (49, -0.068)]
simIndex simValue paperId paperTitle
same-paper 1 0.94270414 62 emnlp-2011-Generating Subsequent Reference in Shared Visual Scenes: Computation vs Re-Use
Author: Jette Viethen ; Robert Dale ; Markus Guhe
Abstract: Traditional computational approaches to referring expression generation operate in a deliberate manner, choosing the attributes to be included on the basis of their ability to distinguish the intended referent from its distractors. However, work in psycholinguistics suggests that speakers align their referring expressions with those used previously in the discourse, implying less deliberate choice and more subconscious reuse. This raises the question as to which is a more accurate characterisation of what people do. Using a corpus of dialogues containing 16,358 referring expressions, we explore this question via the generation of subsequent references in shared visual scenes. We use a machine learning approach to referring expression generation and demonstrate that incorporating features that correspond to the computational tradition does not match human referring behaviour as well as using features corresponding to the process of alignment. The results support the view that the traditional model of referring expression generation that is widely assumed in work on natural language generation may not in fact be correct; our analysis may also help explain the oft-observed redundancy found in humanproduced referring expressions.
2 0.52419412 56 emnlp-2011-Exploring Supervised LDA Models for Assigning Attributes to Adjective-Noun Phrases
Author: Matthias Hartung ; Anette Frank
Abstract: This paper introduces an attribute selection task as a way to characterize the inherent meaning of property-denoting adjectives in adjective-noun phrases, such as e.g. hot in hot summer denoting the attribute TEMPERATURE, rather than TASTE. We formulate this task in a vector space model that represents adjectives and nouns as vectors in a semantic space defined over possible attributes. The vectors incorporate latent semantic information obtained from two variants of LDA topic models. Our LDA models outperform previous approaches on a small set of 10 attributes with considerable gains on sparse representations, which highlights the strong smoothing power of LDA models. For the first time, we extend the attribute selection task to a new data set with more than 200 classes. We observe that large-scale attribute selection is a hard problem, but a subset of attributes performs robustly on the large scale as well. Again, the LDA models outperform the VSM baseline.
3 0.51302892 23 emnlp-2011-Bootstrapped Named Entity Recognition for Product Attribute Extraction
Author: Duangmanee Putthividhya ; Junling Hu
Abstract: We present a named entity recognition (NER) system for extracting product attributes and values from listing titles. Information extraction from short listing titles present a unique challenge, with the lack of informative context and grammatical structure. In this work, we combine supervised NER with bootstrapping to expand the seed list, and output normalized results. Focusing on listings from eBay’s clothing and shoes categories, our bootstrapped NER system is able to identify new brands corresponding to spelling variants and typographical errors of the known brands, as well as identifying novel brands. Among the top 300 new brands predicted, our system achieves 90.33% precision. To output normalized attribute values, we explore several string comparison algorithms and found n-gram substring matching to work well in practice.
4 0.45509505 3 emnlp-2011-A Correction Model for Word Alignments
Author: J. Scott McCarley ; Abraham Ittycheriah ; Salim Roukos ; Bing Xiang ; Jian-ming Xu
Abstract: Models of word alignment built as sequences of links have limited expressive power, but are easy to decode. Word aligners that model the alignment matrix can express arbitrary alignments, but are difficult to decode. We propose an alignment matrix model as a correction algorithm to an underlying sequencebased aligner. Then a greedy decoding algorithm enables the full expressive power of the alignment matrix formulation. Improved alignment performance is shown for all nine language pairs tested. The improved alignments also improved translation quality from Chinese to English and English to Italian.
5 0.44916549 34 emnlp-2011-Corpus-Guided Sentence Generation of Natural Images
Author: Yezhou Yang ; Ching Teo ; Hal Daume III ; Yiannis Aloimonos
Abstract: We propose a sentence generation strategy that describes images by predicting the most likely nouns, verbs, scenes and prepositions that make up the core sentence structure. The input are initial noisy estimates of the objects and scenes detected in the image using state of the art trained detectors. As predicting actions from still images directly is unreliable, we use a language model trained from the English Gigaword corpus to obtain their estimates; together with probabilities of co-located nouns, scenes and prepositions. We use these estimates as parameters on a HMM that models the sentence generation process, with hidden nodes as sentence components and image detections as the emissions. Experimental results show that our strategy of combining vision and language produces readable and de- , scriptive sentences compared to naive strategies that use vision alone.
6 0.30129755 8 emnlp-2011-A Model of Discourse Predictions in Human Sentence Processing
7 0.28354013 38 emnlp-2011-Data-Driven Response Generation in Social Media
8 0.27068707 131 emnlp-2011-Syntactic Decision Tree LMs: Random Selection or Intelligent Design?
9 0.25499275 10 emnlp-2011-A Probabilistic Forest-to-String Model for Language Generation from Typed Lambda Calculus Expressions
10 0.24967034 85 emnlp-2011-Learning to Simplify Sentences with Quasi-Synchronous Grammar and Integer Programming
11 0.24738657 32 emnlp-2011-Computing Logical Form on Regulatory Texts
12 0.24418658 84 emnlp-2011-Learning the Information Status of Noun Phrases in Spoken Dialogues
13 0.2316135 1 emnlp-2011-A Bayesian Mixture Model for PoS Induction Using Multiple Features
14 0.21751665 76 emnlp-2011-Language Models for Machine Translation: Original vs. Translated Texts
15 0.20530097 37 emnlp-2011-Cross-Cutting Models of Lexical Semantics
16 0.20451035 58 emnlp-2011-Fast Generation of Translation Forest for Large-Scale SMT Discriminative Training
17 0.19800243 82 emnlp-2011-Learning Local Content Shift Detectors from Document-level Information
18 0.1762747 105 emnlp-2011-Predicting Thread Discourse Structure over Technical Web Forums
19 0.17469473 11 emnlp-2011-A Simple Word Trigger Method for Social Tag Suggestion
20 0.17466889 53 emnlp-2011-Experimental Support for a Categorical Compositional Distributional Model of Meaning
topicId topicWeight
[(9, 0.397), (15, 0.02), (23, 0.085), (36, 0.029), (37, 0.035), (45, 0.058), (53, 0.024), (54, 0.031), (57, 0.023), (62, 0.018), (64, 0.018), (66, 0.038), (79, 0.054), (82, 0.017), (87, 0.022), (96, 0.021), (98, 0.026)]
simIndex simValue paperId paperTitle
same-paper 1 0.67845345 62 emnlp-2011-Generating Subsequent Reference in Shared Visual Scenes: Computation vs Re-Use
Author: Jette Viethen ; Robert Dale ; Markus Guhe
Abstract: Traditional computational approaches to referring expression generation operate in a deliberate manner, choosing the attributes to be included on the basis of their ability to distinguish the intended referent from its distractors. However, work in psycholinguistics suggests that speakers align their referring expressions with those used previously in the discourse, implying less deliberate choice and more subconscious reuse. This raises the question as to which is a more accurate characterisation of what people do. Using a corpus of dialogues containing 16,358 referring expressions, we explore this question via the generation of subsequent references in shared visual scenes. We use a machine learning approach to referring expression generation and demonstrate that incorporating features that correspond to the computational tradition does not match human referring behaviour as well as using features corresponding to the process of alignment. The results support the view that the traditional model of referring expression generation that is widely assumed in work on natural language generation may not in fact be correct; our analysis may also help explain the oft-observed redundancy found in humanproduced referring expressions.
2 0.32656568 8 emnlp-2011-A Model of Discourse Predictions in Human Sentence Processing
Author: Amit Dubey ; Frank Keller ; Patrick Sturt
Abstract: This paper introduces a psycholinguistic model of sentence processing which combines a Hidden Markov Model noun phrase chunker with a co-reference classifier. Both models are fully incremental and generative, giving probabilities of lexical elements conditional upon linguistic structure. This allows us to compute the information theoretic measure of surprisal, which is known to correlate with human processing effort. We evaluate our surprisal predictions on the Dundee corpus of eye-movement data show that our model achieve a better fit with human reading times than a syntax-only model which does not have access to co-reference information.
3 0.31963807 1 emnlp-2011-A Bayesian Mixture Model for PoS Induction Using Multiple Features
Author: Christos Christodoulopoulos ; Sharon Goldwater ; Mark Steedman
Abstract: In this paper we present a fully unsupervised syntactic class induction system formulated as a Bayesian multinomial mixture model, where each word type is constrained to belong to a single class. By using a mixture model rather than a sequence model (e.g., HMM), we are able to easily add multiple kinds of features, including those at both the type level (morphology features) and token level (context and alignment features, the latter from parallel corpora). Using only context features, our system yields results comparable to state-of-the art, far better than a similar model without the one-class-per-type constraint. Using the additional features provides added benefit, and our final system outperforms the best published results on most of the 25 corpora tested.
4 0.31723961 35 emnlp-2011-Correcting Semantic Collocation Errors with L1-induced Paraphrases
Author: Daniel Dahlmeier ; Hwee Tou Ng
Abstract: We present a novel approach for automatic collocation error correction in learner English which is based on paraphrases extracted from parallel corpora. Our key assumption is that collocation errors are often caused by semantic similarity in the first language (L1language) of the writer. An analysis of a large corpus of annotated learner English confirms this assumption. We evaluate our approach on real-world learner data and show that L1-induced paraphrases outperform traditional approaches based on edit distance, homophones, and WordNet synonyms.
5 0.31635708 53 emnlp-2011-Experimental Support for a Categorical Compositional Distributional Model of Meaning
Author: Edward Grefenstette ; Mehrnoosh Sadrzadeh
Abstract: Modelling compositional meaning for sentences using empirical distributional methods has been a challenge for computational linguists. We implement the abstract categorical model of Coecke et al. (2010) using data from the BNC and evaluate it. The implementation is based on unsupervised learning of matrices for relational words and applying them to the vectors of their arguments. The evaluation is based on the word disambiguation task developed by Mitchell and Lapata (2008) for intransitive sentences, and on a similar new experiment designed for transitive sentences. Our model matches the results of its competitors . in the first experiment, and betters them in the second. The general improvement in results with increase in syntactic complexity showcases the compositional power of our model.
6 0.31518754 136 emnlp-2011-Training a Parser for Machine Translation Reordering
7 0.31496868 68 emnlp-2011-Hypotheses Selection Criteria in a Reranking Framework for Spoken Language Understanding
8 0.31439364 120 emnlp-2011-Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions
9 0.31437308 22 emnlp-2011-Better Evaluation Metrics Lead to Better Machine Translation
10 0.31408697 108 emnlp-2011-Quasi-Synchronous Phrase Dependency Grammars for Machine Translation
11 0.31369111 123 emnlp-2011-Soft Dependency Constraints for Reordering in Hierarchical Phrase-Based Translation
12 0.31351992 87 emnlp-2011-Lexical Generalization in CCG Grammar Induction for Semantic Parsing
13 0.31208527 111 emnlp-2011-Reducing Grounded Learning Tasks To Grammatical Inference
14 0.31181571 28 emnlp-2011-Closing the Loop: Fast, Interactive Semi-Supervised Annotation With Queries on Features and Instances
15 0.3117784 54 emnlp-2011-Exploiting Parse Structures for Native Language Identification
16 0.31009609 132 emnlp-2011-Syntax-Based Grammaticality Improvement using CCG and Guided Search
17 0.30970722 98 emnlp-2011-Named Entity Recognition in Tweets: An Experimental Study
18 0.30867395 85 emnlp-2011-Learning to Simplify Sentences with Quasi-Synchronous Grammar and Integer Programming
19 0.30824015 38 emnlp-2011-Data-Driven Response Generation in Social Media
20 0.30820775 39 emnlp-2011-Discovering Morphological Paradigms from Plain Text Using a Dirichlet Process Mixture Model