acl acl2010 acl2010-185 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Fei Wu ; Daniel S. Weld
Abstract: Information-extraction (IE) systems seek to distill semantic relations from naturallanguage text, but most systems use supervised learning of relation-specific examples and are thus limited by the availability of training data. Open IE systems such as TextRunner, on the other hand, aim to handle the unbounded number of relations found on the Web. But how well can these open systems perform? This paper presents WOE, an open IE system which improves dramatically on TextRunner’s precision and recall. The key to WOE’s performance is a novel form of self-supervised learning for open extractors using heuris— tic matches between Wikipedia infobox attribute values and corresponding sentences to construct training data. Like TextRunner, WOE’s extractor eschews lexicalized features and handles an unbounded set of semantic relations. WOE can operate in two modes: when restricted to POS tag features, it runs as quickly as TextRunner, but when set to use dependency-parse features its precision and recall rise even higher.
Reference: text
sentIndex sentText sentNum sentScore
1 This paper presents WOE, an open IE system which improves dramatically on TextRunner’s precision and recall. [sent-6, score-0.122]
2 The key to WOE’s performance is a novel form of self-supervised learning for open extractors using heuris— tic matches between Wikipedia infobox attribute values and corresponding sentences to construct training data. [sent-7, score-0.426]
3 Like TextRunner, WOE’s extractor eschews lexicalized features and handles an unbounded set of semantic relations. [sent-8, score-0.161]
4 WOE can operate in two modes: when restricted to POS tag features, it runs as quickly as TextRunner, but when set to use dependency-parse features its precision and recall rise even higher. [sent-9, score-0.186]
5 , 1998) used labeled examples ofthe course s-t aught-by relation to in- duce rules for identifying additional instances of the relation. [sent-14, score-0.1]
6 , 2007) and the “preemptive IE” in (Shinyama and Sekine, 2006), aims to handle an unbounded number of relations and run quickly enough to process Webscale corpora. [sent-20, score-0.092]
7 Domain independence is achieved by extracting the relation name as well as its two arguments. [sent-21, score-0.095]
8 Most open IE systems use selfsupervised learning, in which automatic heuristics generate labeled data for training the extractor. [sent-22, score-0.104]
9 Specifically, WOE generates relation-specific training examples by matching Infobox1 attribute values to corresponding sentences (as done in Kylin (Wu and Weld, 2007) and Luchs (Hoffmann et al. [sent-25, score-0.23]
10 WOE can operate in two modes: when restricted to shallow features like part-of-speech (POS) tags, it runs as quickly as Textrunner, but when set to use dependency-parse features its precision and recall rise even higher. [sent-27, score-0.28]
11 We present a thorough experimental evaluation, making the following contributions: • We present WOE, a new approach to open IE tWhaet uses Wikipedia for self-supervised learn1An infobox is a set of tuples summarizing the key at- tributes of the subject in a Wikipedia article. [sent-28, score-0.259]
12 For example, the infobox in the article on “Sweden” contains attributes like Capital, Population and GDP. [sent-29, score-0.178]
13 — 2 Problem Definition An open information extractor is a function from a document, d, to a set of triples, {harg1 , rel , arg2i}, where the args are noun phrases and rel isi a wtehxetureal t fragment irend nicoautning an implicit, semantic relation between the two noun phrases. [sent-38, score-0.344]
14 The extractor should produce one triple for every relation stated explicitly in the text, but is not required to infer implicit facts. [sent-39, score-0.202]
15 We wish to learn an open extractor without direct supervision, i. [sent-44, score-0.171]
16 As output, WOE produces an unlexicalized and relation-independent open extractor. [sent-48, score-0.119]
17 Our objective is an extractor which generalizes beyond Wikipedia, handling other corpora such as the general Web. [sent-49, score-0.098]
18 1 Preprocessor The preprocessor converts the raw Wikipedia text into a sequence of sentences, attaches NLP annotations, and builds synonym sets for key entities. [sent-54, score-0.168]
19 Sentence Splitting: The preprocessor first renders each Wikipedia article into HTML, then splits the article into sentences using OpenNLP. [sent-57, score-0.247]
20 NLP Annotation: As we discuss fully in Section 4 (Experiments), we consider several variations of our system; one version, WOEparse, uses parser-based features, while another, WOEpos, uses shallow features like POS tags, which may be more quickly computed. [sent-58, score-0.124]
21 Depending on which version is being trained, the preprocessor uses OpenNLP to supply POS tags and NP-chunk annotations or uses the Stanford Parser to create a dependency parse. [sent-59, score-0.171]
22 Compiling Synonyms: As a final step, the preprocessor builds sets of synonyms to help the matcher find sentences that correspond to infobox relations. [sent-61, score-0.571]
23 Additionally, attribute values are often described differently within the infobox than they are in surrounding text. [sent-63, score-0.218]
24 Following (Wu and Weld, 2007; Nakayama and Nishio, 2008), the preprocessor uses Wikipedia redirection pages and back119 ward links to automatically construct synonym sets. [sent-65, score-0.142]
25 2 Matcher The matcher constructs training data for the learner component by heuristically matching attribute-value pairs from Wikipedia articles containing infoboxes with corresponding sentences in the article. [sent-71, score-0.513]
26 Given the article on “Stanford University,” for example, the matcher should associate he st abl i shed, 18 9 1 with the sentence “The i university was fdo,u1nd8e9d1 i n w 1i8th91 t by . [sent-72, score-0.323]
27 ”e cGeiv “eTnh a Wikipedia page with an infobox, the matcher iterates through all its attributes looking for a unique sentence that contains references to both the subject of the article and the attribute value; these noun phrases will be annotated arg1 and arg2 in the training set. [sent-75, score-0.471]
28 The matcher considers a sentence to contain the attribute value if the value or its synonym is present. [sent-76, score-0.399]
29 Matching Primary Entities: In order to match shorthand terms like “MIT” with more complete names, the matcher uses an ordered set of heuristics like those of (Wu and Weld, 2007; Nguyen et al. [sent-78, score-0.273]
30 , 2007): • Full match: strings matching the full name of tFhuel entity are tsreilnegcste md. [sent-79, score-0.101]
31 • Patterns of “the ”: The matcher first iPdaetntetrinfiess o tfhe “ type otfy tphee> entity (e. [sent-84, score-0.278]
32 • The most frequent pronoun: The matcher assumes othstat f treheq uaerntitcl per’os nmouosnt: frequent pronoun denotes the primary entity, e. [sent-93, score-0.242]
33 When there are multiple matches to the primary entity in a sentence, the matcher picks the one which is closest to the matched infobox attribute value in the parser dependency graph. [sent-97, score-0.562]
34 Matching Sentences: The matcher seeks a unique sentence to match the attribute value. [sent-98, score-0.363]
35 To produce the best training set, the matcher performs three filterings. [sent-99, score-0.242]
36 First, it skips the attribute completely when multiple sentences mention the value or its synonym. [sent-100, score-0.132]
37 Second, it rejects the sentence if the subject and/or attribute value are not heads of the noun phrases containing them. [sent-101, score-0.18]
38 Third, it discards the sentence if the subject and the attribute value do not appear in the same clause (or in parent/child clauses) in the parse tree. [sent-102, score-0.179]
39 Since Wikipedia’s Wikimarkup language is semantically ambiguous, parsing infoboxes is surprisingly complex. [sent-103, score-0.092]
40 Fortunately, DBpedia (Auer and Lehmann, 2007) provides a cleaned set of infoboxes from 1,027,744 articles. [sent-104, score-0.092]
41 The matcher uses this data for attribute values, generating a training dataset with a total of 301,962 labeled sentences. [sent-105, score-0.331]
42 3 Learning Extractors We learn two kinds of extractors, one (WOEparse) using features from dependency-parse trees and the other (WOEpos) limited to shallow features like POS tags. [sent-107, score-0.13]
43 WOEparse uses a pattern learner to classify whether the shortest dependency path between two noun phrases indicates a semantic relation. [sent-108, score-0.209]
44 Neither extractor uses individual words or lexical information for features. [sent-110, score-0.098]
45 1 Extraction with Parser Features Despite some evidence that parser-based features have limited utility in IE (Jiang and Zhai, 2007), we hoped dependency paths would improve precision on long sentences. [sent-113, score-0.118]
46 As 120 noted in (de Marneffe and Manning, 2008), this collapsed format often yields simplified patterns which are useful for relation extraction. [sent-116, score-0.102]
47 Fso ar any pair oVf , ,to aknedn Es, such as “Dan” and “Berkeley”, we use the shortest connecting path to represent the possible relation between them: Dan born Berkeley − −n −s −u −bj − −pa −s −s → ←p −r −e −p − i −n − We call such a path a corePath. [sent-120, score-0.235]
48 In order to capture the meaning of the relation, the learner augments the corePath into a tree by adding all adverbial and adjectival modifiers as well as dependencies like “neg” and “auxpass”. [sent-123, score-0.089]
49 Building a Database of Patterns: For each of the 301,962 sentences selected and annotated by the matcher, the learner generates a corePath between the tokens denoting the subject and the infobox attribute value. [sent-125, score-0.34]
50 Further, all Noun POS tags and “PRP” are abstracted to “N”, all Verb POS tags to “V”, all Adverb POS tags to “RB” and all Adjective POS tags to “J”. [sent-131, score-0.128]
51 In total, WOE builds a database (named DBp) of 15,333 distinct patterns and each pattern p is associated with a frequency the number ofmatching − −n −s −u −bj − −pa −s −s → − −n −s −u −bj − −pa −s −s → — ←p −r −e −p − i −n − sentences containing p. [sent-135, score-0.136]
52 Specifically, 185 patterns have fp ≥ 100 and 1929 patterns have fp ≥ 5. [sent-136, score-0.122]
53 Learning a Pattern Classifier: Given the large number of patterns in DBp, we assume few valid open extraction patterns are left behind. [sent-137, score-0.205]
54 Take the previous sentence “Dan was not born in Berkeley” for example. [sent-139, score-0.086]
55 121 P/R Curve on WSJ P/R Curve on Web recall recall iecprnos84. [sent-150, score-0.088]
56 Since high speed can be crucial when processing Web-scale corpora, we additionally learn a CRF extractor WOEpos based on shallow features like POS-tags. [sent-170, score-0.228]
57 In both cases, however, we gen- erate training data from Wikipedia by matching sentences with infoboxes, while TextRunner used a small set of hand-written rules to label training examples from the Penn Treebank. [sent-171, score-0.141]
58 We use the same matching sentence set behind DBp to generate positive examples for WOEpos. [sent-172, score-0.13]
59 Specifically, for each matching sentence, we label the subject and infobox attribute value as arg1 and arg2 to serve as the ends of a linear CRF chain. [sent-173, score-0.316]
60 This is because the parser features help to handle complicated and longdistance relations in difficult sentences. [sent-207, score-0.104]
61 42 triples per sentence on average, while WOEpos outputs 1. [sent-209, score-0.094]
62 Note that we measure TextRunner’s precision & recall differently than (Banko et al. [sent-212, score-0.093]
63 Specifically, we compute the precision & recall based on all extractions, while Banko et al. [sent-214, score-0.093]
64 counted only concrete triples where arg1 is a proper noun, arg2 is a proper noun or date, and 122 Figure 3: WOEposachieves an F-measure, which is between 18% and 34% better than TextRunner’s. [sent-215, score-0.088]
65 Our experiments show that focussing on concrete triples generally improves precision at the expense of recall. [sent-219, score-0.111]
66 4 Of course, one can apply a concreteness filter to any open extractor in order to trade recall for precision. [sent-220, score-0.215]
67 Figure 4: WOEparse’s F-measure decreases more slowly with sentence length than WOEpos and TextRunner, due to its better handling of difficult sentences using parser features. [sent-244, score-0.138]
68 This is because long sentences tend to have complicated and long-distance relations which are difficult for shallow features to capture. [sent-254, score-0.172]
69 This is mainly because parser features are more useful for handling difficult sentences and they help WOEparse to maintain a good recall with only moderate loss of precision. [sent-259, score-0.156]
70 Sentence Length We also tested the extraction speed of different extractors. [sent-261, score-0.098]
71 2 Self-supervision with Wikipedia Results in Better Training Data In this section, we consider how the process of matching Wikipedia infobox values to corresponding sentences results in better training data than the hand-written rules used by TextRunner. [sent-273, score-0.237]
72 Specifically, positive and/or negative examples are selected by TextRunner’s hand-written rules (tr for short), by WOE’s heuristic of matching sentences with infoboxes (w for short), or randomly (r for short). [sent-275, score-0.259]
73 In particular, “+w” results in 221,205 positive examples based on the matching sentence set6. [sent-277, score-0.13]
74 All extractors are trained using about the same number of positive and negative examples. [sent-278, score-0.092]
75 The CRF extractors are trained using the same learning algorithm and feature selection as TextRunner. [sent-280, score-0.092]
76 Most likely, this is because TextRunner’s heuristics rely on parse trees to label training examples, 6This number is smaller than the total number of corePaths (259,046) because we require arg1 to appear before arg2 in a sentence — as specified by TextRunner. [sent-283, score-0.088]
77 We can see that filtering PP attachments (PPa) gives a large precision boost with a noticeable loss in recall; enforcing a lexical ordering of relation arguments (1≺2) yields a smaller improvement in precision with small loss in recall. [sent-288, score-0.196]
78 , 2005) utilize WordNet to learn dependency path patterns for extracting the hypernym relation from text. [sent-314, score-0.205]
79 These works focus 124 P/R Curve on WSJ P/R Curve on Web P/R Curve on Wikipedia recall Figure 6: Matching sentences with Wikipedia infoboxes results in better training data than the hand- written rules used by TextRunner. [sent-317, score-0.179]
80 Figure 7: Filtering prepositional phrase attachments (PPa) shows a strong boost to precision, and we see a smaller boost from enforcing a lexical ordering of relation arguments (1≺2). [sent-318, score-0.098]
81 on identifying general relations such as class attributes, while open IE aims to extract relation posed the “preemptive IE” framework to avoid relation-specificity (Shinyama and Sekine, 2006). [sent-327, score-0.175]
82 They applied a similar heuristic by matching Freebase tuples with unstructured sentences (Wikipedia articles in their experiments) to create features for learning relation extractors. [sent-333, score-0.261]
83 Matching Freebase with arbitrary sentences instead of matching Wikipedia infobox with corresponding Wikipedia articles will potentially increase the size of matched sentences at a cost of accuracy. [sent-334, score-0.28]
84 The KNext system (Durme and Schubert, 2008) performs open knowledge extraction via significant heuristics. [sent-343, score-0.135]
85 (Nakayama and Nishio, 2008) parse selected Wikipedia sentences and perform extraction over the phrase structure trees based on several handcrafted patterns. [sent-349, score-0.13]
86 , 2008) which has the same spirit of matching Wikipedia sentences with infoboxes to learn CRF extractors. [sent-351, score-0.2]
87 Deep features are derived from parse trees with the hope of training better extractors (Zhang et al. [sent-356, score-0.153]
88 Jiang and Zhai (Jiang and Zhai, 2007) did a systematic exploration of the feature space for relation extraction on the ACE corpus. [sent-358, score-0.129]
89 Their results showed limited advantage of parser features over shallow features for IE. [sent-359, score-0.163]
90 However, our results imply that ab- stracted dependency path features are highly informative for open IE. [sent-360, score-0.184]
91 The situation is different when features are completely unlexicalized in open IE. [sent-363, score-0.155]
92 Second, as they noted, many relations defined in the ACE corpus are short-range relations which are easier for shallow features to capture. [sent-364, score-0.164]
93 6 Conclusion This paper introduces WOE, a new approach to open IE that uses self-supervised learning over unlexicalized features, based on a heuristic match between Wikipedia infoboxes and corresponding text. [sent-367, score-0.237]
94 WOE can run in two modes: a CRF extractor (WOEpos) trained with shallow features like POS tags; a pattern classfier (WOEparse) learned from dependency path patterns. [sent-368, score-0.299]
95 Our experiments uncovered two sources of WOE’s strong performance: 1) the Wikipedia heuristic is responsible for the bulk of WOE’s improved accuracy, but 2) dependency-parse features are highly informative when performing unlexicalized extraction. [sent-370, score-0.108]
96 consider all sentences containing both the subject and object of a Freebase record as matching sentences (Mintz et al. [sent-376, score-0.184]
97 We are also interested in merging lexicalized and open extraction methods; the use of some domain-specific lexical features might help to improve WOE’s practical performance, but the best way to do this is unclear. [sent-378, score-0.171]
98 Exploiting syntactic and semantic information for relation extraction from wikipedia. [sent-498, score-0.129]
99 Turning web text and search queries into factual knowledge: Hierarchical class attribute extraction. [sent-502, score-0.089]
100 A re-examination of dependency path kernels for relation extraction. [sent-530, score-0.142]
wordName wordTfidf (topN-words)
[('woeparse', 0.423), ('woe', 0.408), ('textrunner', 0.349), ('woepos', 0.302), ('matcher', 0.242), ('wikipedia', 0.236), ('infobox', 0.129), ('preprocessor', 0.106), ('ie', 0.099), ('extractor', 0.098), ('ppa', 0.093), ('extractors', 0.092), ('infoboxes', 0.092), ('corepaths', 0.091), ('dbp', 0.091), ('attribute', 0.089), ('open', 0.073), ('crf', 0.073), ('relation', 0.067), ('freebase', 0.066), ('mintz', 0.066), ('wsj', 0.065), ('matching', 0.065), ('triples', 0.062), ('extraction', 0.062), ('corepath', 0.06), ('expandpath', 0.06), ('nakayama', 0.06), ('shallow', 0.058), ('born', 0.054), ('article', 0.049), ('precision', 0.049), ('jiang', 0.047), ('unlexicalized', 0.046), ('learner', 0.046), ('banko', 0.046), ('weld', 0.046), ('davidov', 0.045), ('wu', 0.045), ('curve', 0.045), ('bj', 0.044), ('zhai', 0.044), ('recall', 0.044), ('dependencies', 0.043), ('hoffmann', 0.043), ('stanford', 0.043), ('sentences', 0.043), ('path', 0.042), ('auer', 0.04), ('preemptive', 0.04), ('triple', 0.037), ('entity', 0.036), ('synonym', 0.036), ('speed', 0.036), ('features', 0.036), ('relations', 0.035), ('pos', 0.035), ('patterns', 0.035), ('snow', 0.034), ('bunescu', 0.034), ('berkeley', 0.034), ('shinyama', 0.034), ('examples', 0.033), ('subject', 0.033), ('dependency', 0.033), ('parser', 0.033), ('washington', 0.033), ('pattern', 0.032), ('tags', 0.032), ('sentence', 0.032), ('wrongly', 0.032), ('prep', 0.032), ('heuristics', 0.031), ('attachments', 0.031), ('modes', 0.031), ('dip', 0.03), ('nishio', 0.03), ('quickly', 0.03), ('decreases', 0.03), ('shortest', 0.03), ('extracting', 0.028), ('tr', 0.027), ('runs', 0.027), ('rel', 0.027), ('unbounded', 0.027), ('dbpedia', 0.026), ('lehmann', 0.026), ('amherst', 0.026), ('auxpass', 0.026), ('kylin', 0.026), ('raphael', 0.026), ('webkb', 0.026), ('heuristic', 0.026), ('fp', 0.026), ('noun', 0.026), ('builds', 0.026), ('parse', 0.025), ('synonyms', 0.025), ('heuristically', 0.025), ('tuples', 0.024)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000001 185 acl-2010-Open Information Extraction Using Wikipedia
Author: Fei Wu ; Daniel S. Weld
Abstract: Information-extraction (IE) systems seek to distill semantic relations from naturallanguage text, but most systems use supervised learning of relation-specific examples and are thus limited by the availability of training data. Open IE systems such as TextRunner, on the other hand, aim to handle the unbounded number of relations found on the Web. But how well can these open systems perform? This paper presents WOE, an open IE system which improves dramatically on TextRunner’s precision and recall. The key to WOE’s performance is a novel form of self-supervised learning for open extractors using heuris— tic matches between Wikipedia infobox attribute values and corresponding sentences to construct training data. Like TextRunner, WOE’s extractor eschews lexicalized features and handles an unbounded set of semantic relations. WOE can operate in two modes: when restricted to POS tag features, it runs as quickly as TextRunner, but when set to use dependency-parse features its precision and recall rise even higher.
2 0.21689536 159 acl-2010-Learning 5000 Relational Extractors
Author: Raphael Hoffmann ; Congle Zhang ; Daniel S. Weld
Abstract: Many researchers are trying to use information extraction (IE) to create large-scale knowledge bases from natural language text on the Web. However, the primary approach (supervised learning of relation-specific extractors) requires manually-labeled training data for each relation and doesn’t scale to the thousands of relations encoded in Web text. This paper presents LUCHS, a self-supervised, relation-specific IE system which learns 5025 relations more than an order of magnitude greater than any previous approach with an average F1 score of 61%. Crucial to LUCHS’s performance is an automated system for dynamic lexicon learning, which allows it to learn accurately from heuristically-generated training data, which is often noisy and sparse. — —
3 0.15994649 261 acl-2010-Wikipedia as Sense Inventory to Improve Diversity in Web Search Results
Author: Celina Santamaria ; Julio Gonzalo ; Javier Artiles
Abstract: Is it possible to use sense inventories to improve Web search results diversity for one word queries? To answer this question, we focus on two broad-coverage lexical resources of a different nature: WordNet, as a de-facto standard used in Word Sense Disambiguation experiments; and Wikipedia, as a large coverage, updated encyclopaedic resource which may have a better coverage of relevant senses in Web pages. Our results indicate that (i) Wikipedia has a much better coverage of search results, (ii) the distribution of senses in search results can be estimated using the internal graph structure of the Wikipedia and the relative number of visits received by each sense in Wikipedia, and (iii) associating Web pages to Wikipedia senses with simple and efficient algorithms, we can produce modified rankings that cover 70% more Wikipedia senses than the original search engine rankings. 1 Motivation The application of Word Sense Disambiguation (WSD) to Information Retrieval (IR) has been subject of a significant research effort in the recent past. The essential idea is that, by indexing and matching word senses (or even meanings) , the retrieval process could better handle polysemy and synonymy problems (Sanderson, 2000). In practice, however, there are two main difficulties: (i) for long queries, IR models implicitly perform disambiguation, and thus there is little room for improvement. This is the case with most standard IR benchmarks, such as TREC (trec.nist.gov) or CLEF (www.clef-campaign.org) ad-hoc collections; (ii) for very short queries, disambiguation j ul io @ l i uned . e s j avart s . @bec . uned . e s may not be possible or even desirable. This is often the case with one word and even two word queries in Web search engines. In Web search, there are at least three ways of coping with ambiguity: • • • Promoting diversity in the search results (Clarke negt al., 2008): given th seea query s”uolatssis”, the search engine may try to include representatives for different senses of the word (such as the Oasis band, the Organization for the Advancement of Structured Information Standards, the online fashion store, etc.) among the top results. Search engines are supposed to handle diversity as one of the multiple factors that influence the ranking. Presenting the results as a set of (labelled) cPlruessteenrtsi nragth tehre eth reansu as a a rsan ake sde lti ostf (Carpineto et al., 2009). Complementing search results with search suggestions (e.g. e”oaracshis band”, ”woitahsis s fashion store”) that serve to refine the query in the intended way (Anick, 2003). All of them rely on the ability of the search engine to cluster search results, detecting topic similarities. In all of them, disambiguation is implicit, a side effect of the process but not its explicit target. Clustering may detect that documents about the Oasis band and the Oasis fashion store deal with unrelated topics, but it may as well detect a group of documents discussing why one of the Oasis band members is leaving the band, and another group of documents about Oasis band lyrics; both are different aspects of the broad topic Oasis band. A perfect hierarchical clustering should distinguish between the different Oasis senses at a first level, and then discover different topics within each of the senses. Is it possible to use sense inventories to improve search results for one word queries? To answer 1357 Proce dingUsp opfs thaela 4, 8Stwhe Adnen u,a 1l1- M16e Jtiunlgy o 2f0 t1h0e. A ?c s 2o0c1ia0ti Aosnso focria Ctio nm fpourta Ctoiomnpault Laitniognuaislt Licisn,g puaigsetisc 1s357–136 , this question, we will focus on two broad-coverage lexical resources of a different nature: WordNet (Miller et al., 1990), as a de-facto standard used in Word Sense Disambiguation experiments and many other Natural Language Processing research fields; and Wikipedia (www.wikipedia.org), as a large coverage and updated encyclopedic resource which may have a better coverage of relevant senses in Web pages. Our hypothesis is that, under appropriate conditions, any of the above mechanisms (clustering, search suggestions, diversity) might benefit from an explicit disambiguation (classification of pages in the top search results) using a wide-coverage sense inventory. Our research is focused on four relevant aspects of the problem: 1. Coverage: Are Wikipedia/Wordnet senses representative of search results? Otherwise, trying to make a disambiguation in terms of a fixed sense inventory would be meaningless. 2. If the answer to (1) is positive, the reverse question is also interesting: can we estimate search results diversity using our sense inven- tories? 3. Sense frequencies: knowing sense frequencies in (search results) Web pages is crucial to have a usable sense inventory. Is it possible to estimate Web sense frequencies from currently available information? 4. Classification: The association of Web pages to word senses must be done with some unsupervised algorithm, because it is not possible to hand-tag training material for every possible query word. Can this classification be done accurately? Can it be effective to promote diversity in search results? In order to provide an initial answer to these questions, we have built a corpus consisting of 40 nouns and 100 Google search results per noun, manually annotated with the most appropriate Wordnet and Wikipedia senses. Section 2 describes how this corpus has been created, and in Section 3 we discuss WordNet and Wikipedia coverage of search results according to our testbed. As this initial results clearly discard Wordnet as a sense inventory for the task, the rest of the paper mainly focuses on Wikipedia. In Section 4 we estimate search results diversity from our testbed, finding that the use of Wikipedia could substantially improve diversity in the top results. In Section 5 we use the Wikipedia internal link structure and the number of visits per page to estimate relative frequencies for Wikipedia senses, obtaining an estimation which is highly correlated with actual data in our testbed. Finally, in Section 6 we discuss a few strategies to classify Web pages into word senses, and apply the best classifier to enhance diversity in search results. The paper concludes with a discussion of related work (Section 7) and an overall discussion of our results in Section 8. 2 Test Set 2.1 Set of Words The most crucial step in building our test set is choosing the set of words to be considered. We are looking for words which are susceptible to form a one-word query for a Web search engine, and therefore we should focus on nouns which are used to denote one or more named entities. At the same time we want to have some degree of comparability with previous research on Word Sense Disambiguation, which points to noun sets used in Senseval/SemEval evaluation campaigns1 . Our budget for corpus annotation was enough for two persons-month, which limited us to handle 40 nouns (usually enough to establish statistically significant differences between WSD algorithms, although obviously limited to reach solid figures about the general behaviour of words in the Web). With these arguments in mind, we decided to choose: (i) 15 nouns from the Senseval-3 lexical sample dataset, which have been previously employed by (Mihalcea, 2007) in a related experiment (see Section 7); (ii) 25 additional words which satisfy two conditions: they are all ambiguous, and they are all names for music bands in one of their senses (not necessarily the most salient). The Senseval set is: {argument, arm, atmosphere, bank, degree, difference, disc, irmm-, age, paper, party, performance, plan, shelter, sort, source}. The bands set is {amazon, apple, camel, cell, columbia, cream, foreigner, fox, genesis, jaguar, oasis, pioneer, police, puma, rainbow, shell, skin, sun, tesla, thunder, total, traffic, trapeze, triumph, yes}. Fpoerz e,a trchiu noun, we looked up all its possible senses in WordNet 3.0 and in Wikipedia (using 1http://senseval.org 1358 Table 1: Coverage of Search Results: Wikipedia vs. WordNet Wikiped#ia documents # senses WordNe#t documents Senseval setava2il4a2b/1le0/u0sedassign8e7d7 to (5 s9o%me) senseavai9la2b/5le2/usedassigne6d96 to (4 s6o%m)e sense # senses BaTnodtsa lset868420//21774421323558 ((5546%%))17780/3/9911529995 (2 (342%%)) Wikipedia disambiguation pages). Wikipedia has an average of 22 senses per noun (25.2 in the Bands set and 16. 1in the Senseval set), and Wordnet a much smaller figure, 4.5 (3. 12 for the Bands set and 6.13 for the Senseval set). For a conventional dictionary, a higher ambiguity might indicate an excess of granularity; for an encyclopaedic resource such as Wikipedia, however, it is just an indication of larger coverage. Wikipedia en- tries for camel which are not in WordNet, for instance, include the Apache Camel routing and mediation engine, the British rock band, the brand of cigarettes, the river in Cornwall, and the World World War I fighter biplane. 2.2 Set of Documents We retrieved the 150 first ranked documents for each noun, by submitting the nouns as queries to a Web search engine (Google). Then, for each document, we stored both the snippet (small description of the contents of retrieved document) and the whole HTML document. This collection of documents contain an implicit new inventory of senses, based on Web search, as documents retrieved by a noun query are associated with some sense of the noun. Given that every document in the top Web search results is supposed to be highly relevant for the query word, we assume a ”one sense per document” scenario, although we allow annotators to assign more than one sense per document. In general this assumption turned out to be correct except in a few exceptional cases (such as Wikipedia disambiguation pages): only nine docu- ments received more than one WordNet sense, and 44 (1. 1% of all annotated pages) received more than one Wikipedia sense. 2.3 Manual Annotation We implemented an annotation interface which stored all documents and a short description for every Wordnet and Wikipedia sense. The annotators had to decide, for every document, whether there was one or more appropriate senses in each of the dictionaries. They were instructed to provide annotations for 100 documents per name; if an URL in the list was corrupt or not available, it had to be discarded. We provided 150 documents per name to ensure that the figure of 100 usable documents per name could be reached without problems. Each judge provided annotations for the 4,000 documents in the final data set. In a second round, they met and discussed their independent annotations together, reaching a consensus judgement for every document. 3 Coverage of Web Search Results: Wikipedia vs Wordnet Table 1 shows how Wikipedia and Wordnet cover the senses in search results. We report each noun subset separately (Senseval and bands subsets) as well as aggregated figures. The most relevant fact is that, unsurprisingly, Wikipedia senses cover much more search results (56%) than Wordnet (32%). If we focus on the top ten results, in the bands subset (which should be more representative of plausible web queries) Wikipedia covers 68% of the top ten documents. This is an indication that it can indeed be useful for promoting diversity or help clustering search results: even if 32% of the top ten documents are not covered by Wikipedia, it is still a representative source of senses in the top search results. We have manually examined all documents in the top ten results that are not covered by Wikipedia: a majority of the missing senses consists of names of (generally not well-known) companies (45%) and products or services (26%); the other frequent type (12%) of non annotated doc- ument is disambiguation pages (from Wikipedia and also from other dictionaries). It is also interesting to examine the degree of overlap between Wikipedia and Wordnet senses. Being two different types of lexical resource, they might have some degree of complementarity. Table 2 shows, however, that this is not the case: most of the (annotated) documents either fit Wikipedia senses (26%) or both Wikipedia and Wordnet (29%), and just 3% fit Wordnet only. 1359 Table 2: Overlap between Wikipedia and Wordnet in Search Results # documents annotated with Senseval setWikipe60di7a ( &40 W%o)rdnetWi2k7ip0e (d1i8a% on)lyWo8r9d (n6e%t o)nly534no (3n6e%) BaTnodtsa slet1517729 ( (2239%%))1708566 (3 (216%%))12176 ( (13%%))11614195 ( (4415%%)) Therefore, Wikipedia seems to extend the coverage of Wordnet rather than providing complementary sense information. If we wanted to extend the coverage of Wikipedia, the best strategy seems to be to consider lists ofcompanies, products and services, rather than complementing Wikipedia with additional sense inventories. 4 Diversity in Google Search Results Once we know that Wikipedia senses are a representative subset of actual Web senses (covering more than half of the documents retrieved by the search engine), we can test how well search results respect diversity in terms of this subset of senses. Table 3 displays the number of different senses found at different depths in the search results rank, and the average proportion of total senses that they represent. These results suggest that diversity is not a major priority for ranking results: the top ten results only cover, in average, 3 Wikipedia senses (while the average number of senses listed in Wikipedia is 22). When considering the first 100 documents, this number grows up to 6.85 senses per noun. Another relevant figure is the frequency of the most frequent sense for each word: in average, 63% of the pages in search results belong to the most frequent sense of the query word. This is roughly comparable with most frequent sense figures in standard annotated corpora such as Semcor (Miller et al., 1993) and the Senseval/Semeval data sets, which suggests that diversity may not play a major role in the current Google ranking algorithm. Of course this result must be taken with care, because variability between words is high and unpredictable, and we are using only 40 nouns for our experiment. But what we have is a positive indication that Wikipedia could be used to improve diversity or cluster search results: potentially the first top ten results could cover 6.15 different senses in average (see Section 6.5), which would be a substantial growth. 5 Sense Frequency Estimators for Wikipedia Wikipedia disambiguation pages contain no systematic information about the relative importance of senses for a given word. Such information, however, is crucial in a lexicon, because sense distributions tend to be skewed, and knowing them can help disambiguation algorithms. We have attempted to use two estimators of expected sense distribution: • • Internal relevance of a word sense, measured as incoming alinnckes o ffo ar wthoer U seRnLs o, fm a given sense in Wikipedia. External relevance of a word sense, measured as ttheren naulm rebleevr aonfc vei osifts a f woro trhde s eUnRsLe, mofe a given sense (as reported in http://stats.grok.se). The number of internal incoming links is expected to be relatively stable for Wikipedia articles. As for the number of visits, we performed a comparison of the number of visits received by the bands noun subset in May, June and July 2009, finding a stable-enough scenario with one notorious exception: the number of visits to the noun Tesla raised dramatically in July, because July 10 was the anniversary of the birth of Nicola Tesla, and a special Google logo directed users to the Wikipedia page for the scientist. We have measured correlation between the relative frequencies derived from these two indicators and the actual relative frequencies in our testbed. Therefore, for each noun w and for each sense wi, we consider three values: (i) proportion of documents retrieved for w which are manually assigned to each sense wi; (ii) inlinks(wi) : relative amount of incoming links to each sense wi; and (iii) visits(wi) : relative number of visits to the URL for each sense wi. We have measured the correlation between these three values using a linear regression correlation coefficient, which gives a correlation value of .54 for the number of visits and of .71 for the number of incoming links. Both estimators seem 1360 Table 3: Diversity in Search Results according to Wikipedia F ir s t 12570 docsBave6n425.rd9854a6 s8get#snSe 65sien43. v68a3s27elarcthesTu6543l.o t5083as5lBvaen.r3d2a73s81gectovrSaegnso. 4f32v615aWlsiketpdaTs.3oe249tn01asle to be positively correlated with real relative frequencies in our testbed, with a strong preference for the number of links. We have experimented with weighted combinations of both indicators, using weights of the form (k, 1 k) , k ∈ {0, 0.1, 0.2 . . . 1}, reaching a maxi(mk,a1l c−okrre),lkati ∈on { 0of, .07.13, f0o.r2 t.h.e. following weights: − freq(wi) = 0.9∗inlinks(wi) +0. 1∗visits(wi) (1) This weighted estimator provides a slight advantage over the use of incoming links only (.73 vs .71). Overall, we have an estimator which has a strong correlation with the distribution of senses in our testbed. In the next section we will test its utility for disambiguation purposes. 6 Association of Wikipedia Senses to Web Pages We want to test whether the information provided by Wikipedia can be used to classify search results accurately. Note that we do not want to consider approaches that involve a manual creation of training material, because they can’t be used in practice. Given a Web page p returned by the search engine for the query w, and the set of senses w1 . . . wn listed in Wikipedia, the task is to assign the best candidate sense to p. We consider two different techniques: • A basic Information Retrieval approach, wAhe breas tche I dfoocrmumateionnts Ranetdr tvhael Wikipedia pages are represented using a Vector Space Model (VSM) and compared with a standard cosine measure. This is a basic approach which, if successful, can be used efficiently to classify search results. An approach based on a state-of-the-art supervised oWacShD b system, extracting training examples automatically from Wikipedia content. We also compute two baselines: • • • A random assignment of senses (precision is computed as itghnem ienvnter osfe oenfs tehse ( pnruemcibsieorn o isf senses, for every test case). A most frequent sense heuristic which uses our eosstitm fraetiqoune otf s sense frequencies acnhd u assigns the same sense (the most frequent) to all documents. Both are naive baselines, but it must be noted that the most frequent sense heuristic is usually hard to beat for unsupervised WSD algorithms in most standard data sets. We now describe each of the two main approaches in detail. 6.1 VSM Approach For each word sense, we represent its Wikipedia page in a (unigram) vector space model, assigning standard tf*idf weights to the words in the document. idf weights are computed in two different ways: 1. Experiment VSM computes inverse document frequencies in the collection of retrieved documents (for the word being considered). 2. Experiment VSM-GT uses the statistics provided by the Google Terabyte collection (Brants and Franz, 2006), i.e. it replaces the collection of documents with statistics from a representative snapshot of the Web. 3. Experiment VSM-mixed combines statistics from the collection and from the Google Terabyte collection, following (Chen et al., 2009). The document p is represented in the same vector space as the Wikipedia senses, and it is compared with each of the candidate senses wi via the cosine similarity metric (we have experimented 1361 with other similarity metrics such as χ2, but differences are irrelevant). The sense with the highest similarity to p is assigned to the document. In case of ties (which are rare), we pick the first sense in the Wikipedia disambiguation page (which in practice is like a random decision, because senses in disambiguation pages do not seem to be ordered according to any clear criteria). We have also tested a variant of this approach which uses the estimation of sense frequencies presented above: once the similarities are computed, we consider those cases where two or more senses have a similar score (in particular, all senses with a score greater or equal than 80% of the highest score). In that cases, instead of using the small similarity differences to select a sense, we pick up the one which has the largest frequency according to our estimator. We have applied this strategy to the best performing system, VSM-GT, resulting in experiment VSM-GT+freq. 6.2 WSD Approach We have used TiMBL (Daelemans et al., 2001), a state-of-the-art supervised WSD system which uses Memory-Based Learning. The key, in this case, is how to extract learning examples from the Wikipedia automatically. For each word sense, we basically have three sources of examples: (i) occurrences of the word in the Wikipedia page for the word sense; (ii) occurrences of the word in Wikipedia pages pointing to the page for the word sense; (iii) occurrences of the word in external pages linked in the Wikipedia page for the word sense. After an initial manual inspection, we decided to discard external pages for being too noisy, and we focused on the first two options. We tried three alternatives: • • • TiMBL-core uses only the examples found Tini MtheB page rfoer u tshees sense being atrmaipneleds. TiMBL-inlinks uses the examples found in Wikipedia pages pointing etxoa mthep sense being trained. TiMBL-all uses both sources of examples. In order to classify a page p with respect to the senses for a word w, we first disambiguate all occurrences of w in the page p. Then we choose the sense which appears most frequently in the page according to TiMBL results. In case of ties we pick up the first sense listed in the Wikipedia disambiguation page. We have also experimented with a variant of the approach that uses our estimation of sense frequencies, similarly to what we did with the VSM approach. In this case, (i) when there is a tie between two or more senses (which is much more likely than in the VSM approach), we pick up the sense with the highest frequency according to our estimator; and (ii) when no sense reaches 30% of the cases in the page to be disambiguated, we also resort to the most frequent sense heuristic (among the candidates for the page). This experiment is called TiMBL-core+freq (we discarded ”inlinks” and ”all” versions because they were clearly worse than ”core”). 6.3 Classification Results Table 4 shows classification results. The accuracy of systems is reported as precision, i.e. the number of pages correctly classified divided by the total number of predictions. This is approximately the same as recall (correctly classified pages divided by total number of pages) for our systems, because the algorithms provide an answer for every page containing text (actual coverage is 94% because some pages only contain text as part of an image file such as photographs and logotypes). Table 4: Classification Results Experiment Precision random most frequent sense (estimation) .19 .46 TiMBL-core TiMBL-inlinks TiMBL-all TiMBL-core+freq .60 .50 .58 .67 VSM VSM-GT VSM-mixed VSM-GT+freq .67 .68 .67 .69 All systems are significantly better than the random and most frequent sense baselines (using p < 0.05 for a standard t-test). Overall, both approaches (using TiMBL WSD machinery and using VSM) lead to similar results (.67 vs. .69), which would make VSM preferable because it is a simpler and more efficient approach. Taking a 1362 Figure 1: Precision/Coverage curves for VSM-GT+freq classification algorithm closer look at the results with TiMBL, there are a couple of interesting facts: • There is a substantial difference between using only examples itaalke dnif fferroemnc tehe b Wikipedia Web page for the sense being trained (TiMBL-core, .60) and using examples from the Wikipedia pages pointing to that page (TiMBL-inlinks, .50). Examples taken from related pages (even if the relationship is close as in this case) seem to be too noisy for the task. This result is compatible with findings in (Santamar ı´a et al., 2003) using the Open Directory Project to extract examples automatically. • Our estimation of sense frequencies turns oOuutr rto e tbiem very helpful sfeor f cases wcihesere t our TiMBL-based algorithm cannot provide an answer: precision rises from .60 (TiMBLcore) to .67 (TiMBL-core+freq). The difference is statistically significant (p < 0.05) according to the t-test. As for the experiments with VSM, the variations tested do not provide substantial improvements to the baseline (which is .67). Using idf frequencies obtained from the Google Terabyte corpus (instead of frequencies obtained from the set of retrieved documents) provides only a small improvement (VSM-GT, .68), and adding the estimation of sense frequencies gives another small improvement (.69). Comparing the baseline VSM with the optimal setting (VSM-GT+freq), the difference is small (.67 vs .69) but relatively robust (p = 0.066 according to the t-test). Remarkably, the use of frequency estimations is very helpful for the WSD approach but not for the SVM one, and they both end up with similar performance figures; this might indicate that using frequency estimations is only helpful up to certain precision ceiling. 6.4 Precision/Coverage Trade-off All the above experiments are done at maximal coverage, i.e., all systems assign a sense for every document in the test collection (at least for every document with textual content). But it is possible to enhance search results diversity without annotating every document (in fact, not every document can be assigned to a Wikipedia sense, as we have discussed in Section 3). Thus, it is useful to investigate which is the precision/coverage trade-off in our dataset. We have experimented with the best performing system (VSM-GT+freq), introducing a similarity threshold: assignment of a document to a sense is only done if the similarity of the document to the Wikipedia page for the sense exceeds the similarity threshold. We have computed precision and coverage for every threshold in the range [0.00 −0.90] (beyond 0e.v9e0ry coverage was null) anngde represented 0th] e(b breeysuolntds in Figure 1 (solid line). The graph shows that we 1363 can classify around 20% of the documents with a precision above .90, and around 60% of the documents with a precision of .80. Note that we are reporting disambiguation results using a conventional WSD test set, i.e., one in which every test case (every document) has been manually assigned to some Wikipedia sense. But in our Web Search scenario, 44% of the documents were not assigned to any Wikipedia sense: in practice, our classification algorithm would have to cope with all this noise as well. Figure 1 (dotted line) shows how the precision/coverage curve is affected when the algorithm attempts to disambiguate all documents retrieved by Google, whether they can in fact be assigned to a Wikipedia sense or not. At a coverage of 20%, precision drops approximately from .90 to .70, and at a coverage of 60% it drops from .80 to .50. We now address the question of whether this performance is good enough to improve search re- sults diversity in practice. 6.5 Using Classification to Promote Diversity We now want to estimate how the reported classification accuracy may perform in practice to enhance diversity in search results. In order to provide an initial answer to this question, we have re-ranked the documents for the 40 nouns in our testbed, using our best classifier (VSM-GT+freq) and making a list of the top-ten documents with the primary criterion of maximising the number of senses represented in the set, and the secondary criterion of maximising the similarity scores of the documents to their assigned senses. The algorithm proceeds as follows: we fill each position in the rank (starting at rank 1), with the document which has the highest similarity to some of the senses which are not yet represented in the rank; once all senses are represented, we start choosing a second representative for each sense, following the same criterion. The process goes on until the first ten documents are selected. We have also produced a number of alternative rankings for comparison purposes: clustering (centroids): this method applies eHriiengrarc (hciecnatlr Agglomerative Clustering which proved to be the most competitive clustering algorithm in a similar task (Artiles et al., 2009) to the set of search results, forcing the algorithm to create ten clusters. The centroid of each cluster is then selected Table 5: Enhancement of Search Results Diversity • – – rank@10 # senses coverage Original rank2.8049% Wikipedia 4.75 77% clustering (centroids) 2.50 42% clustering (top ranked) 2.80 46% random 2.45 43% upper bound6.1597% as one of the top ten documents in the new rank. • clustering (top ranked): Applies the same clustering algorithm, db u)t: tAhpisp lti emse t tehe s top ranked document (in the original Google rank) of each cluster is selected. • • random: Randomly selects ten documents frraonmd otmhe: :se Rt aofn dreomtrielyve sde lreecstuslts te. upper bound: This is the maximal diversity tuhpapt can o beu nodb:tai Tnheids iins our mteasxtbiemda. lN doivteer tshitayt coverage is not 100%, because some words have more than ten meanings in Wikipedia and we are only considering the top ten documents. All experiments have been applied on the full set of documents in the testbed, including those which could not be annotated with any Wikipedia sense. Coverage is computed as the ratio of senses that appear in the top ten results compared to the number of senses that appear in all search results. Results are presented in Table 5. Note that diversity in the top ten documents increases from an average of 2.80 Wikipedia senses represented in the original search engine rank, to 4.75 in the modified rank (being 6.15 the upper bound), with the coverage of senses going from 49% to 77%. With a simple VSM algorithm, the coverage of Wikipedia senses in the top ten results is 70% larger than in the original ranking. Using Wikipedia to enhance diversity seems to work much better than clustering: both strategies to select a representative from each cluster are unable to improve the diversity of the original ranking. Note, however, that our evaluation has a bias towards using Wikipedia, because only Wikipedia senses are considered to estimate diversity. Of course our results do not imply that the Wikipedia modified rank is better than the original 1364 Google rank: there are many other factors that influence the final ranking provided by a search engine. What our results indicate is that, with simple and efficient algorithms, Wikipedia can be used as a reference to improve search results diversity for one-word queries. 7 Related Work Web search results clustering and diversity in search results are topics that receive an increasing attention from the research community. Diversity is used both to represent sub-themes in a broad topic, or to consider alternative interpretations for ambiguous queries (Agrawal et al., 2009), which is our interest here. Standard IR test collections do not usually consider ambiguous queries, and are thus inappropriate to test systems that promote diversity (Sanderson, 2008); it is only recently that appropriate test collections are being built, such as (Paramita et al., 2009) for image search and (Artiles et al., 2009) for person name search. We see our testbed as complementary to these ones, and expect that it can contribute to foster research on search results diversity. To our knowledge, Wikipedia has not explicitly been used before to promote diversity in search results; but in (Gollapudi and Sharma, 2009), it is used as a gold standard to evaluate diversification algorithms: given a query with a Wikipedia disambiguation page, an algorithm is evaluated as promoting diversity when different documents in the search results are semantically similar to different Wikipedia pages (describing the alternative senses of the query). Although semantic similarity is measured automatically in this work, our results confirm that this evaluation strategy is sound, because Wikipedia senses are indeed representative of search results. (Clough et al., 2009) analyses query diversity in a Microsoft Live Search, using click entropy and query reformulation as diversity indicators. It was found that at least 9.5% - 16.2% of queries could benefit from diversification, although no correlation was found between the number of senses of a word in Wikipedia and the indicators used to discover diverse queries. This result does not discard, however, that queries where applying diversity is useful cannot benefit from Wikipedia as a sense inventory. In the context of clustering, (Carmel et al., 2009) successfully employ Wikipedia to enhance automatic cluster labeling, finding that Wikipedia labels agree with manual labels associated by humans to a cluster, much more than with signif- icant terms that are extracted directly from the text. In a similar line, both (Gabrilovich and Markovitch, 2007) and (Syed et al., 2008) provide evidence suggesting that categories of Wikipedia articles can successfully describe common concepts in documents. In the field of Natural Language Processing, there has been successful attempts to connect Wikipedia entries to Wordnet senses: (RuizCasado et al., 2005) reports an algorithm that provides an accuracy of 84%. (Mihalcea, 2007) uses internal Wikipedia hyperlinks to derive sensetagged examples. But instead of using Wikipedia directly as sense inventory, Mihalcea then manually maps Wikipedia senses into Wordnet senses (claiming that, at the time of writing the paper, Wikipedia did not consistently report ambiguity in disambiguation pages) and shows that a WSD system based on acquired sense-tagged examples reaches an accuracy well beyond an (informed) most frequent sense heuristic. 8 Conclusions We have investigated whether generic lexical resources can be used to promote diversity in Web search results for one-word, ambiguous queries. We have compared WordNet and Wikipedia and arrived to a number of conclusions: (i) unsurprisingly, Wikipedia has a much better coverage of senses in search results, and is therefore more appropriate for the task; (ii) the distribution of senses in search results can be estimated using the internal graph structure of the Wikipedia and the relative number of visits received by each sense in Wikipedia, and (iii) associating Web pages to Wikipedia senses with simple and efficient algorithms, we can produce modified rankings that cover 70% more Wikipedia senses than the original search engine rankings. We expect that the testbed created for this research will complement the - currently short - set of benchmarking test sets to explore search results diversity and query ambiguity. Our testbed is publicly available for research purposes at http://nlp.uned.es. Our results endorse further investigation on the use of Wikipedia to organize search results. Some limitations of our research, however, must be 1365 noted: (i) the nature of our testbed (with every search result manually annotated in terms of two sense inventories) makes it too small to extract solid conclusions on Web searches (ii) our work does not involve any study of diversity from the point of view of Web users (i.e. when a Web query addresses many different use needs in practice); research in (Clough et al., 2009) suggests that word ambiguity in Wikipedia might not be related with diversity of search needs; (iii) we have tested our classifiers with a simple re-ordering of search results to test how much diversity can be improved, but a search results ranking depends on many other factors, some of them more crucial than diversity; it remains to be tested how can we use document/Wikipedia associations to improve search results clustering (for instance, providing seeds for the clustering process) and to provide search suggestions. Acknowledgments This work has been partially funded by the Spanish Government (project INES/Text-Mess) and the Xunta de Galicia. References R. Agrawal, S. Gollapudi, A. Halverson, and S. Leong. 2009. Diversifying Search Results. In Proc. of WSDM’09. ACM. P. Anick. 2003. Using Terminological Feedback for Web Search Refinement : a Log-based Study. In Proc. ACM SIGIR 2003, pages 88–95. ACM New York, NY, USA. J. Artiles, J. Gonzalo, and S. Sekine. 2009. WePS 2 Evaluation Campaign: overview of the Web People Search Clustering Task. In 2nd Web People Search Evaluation Workshop (WePS 2009), 18th WWW Conference. 2009. T. Brants and A. Franz. 2006. Web 1T 5-gram, version 1. Philadelphia: Linguistic Data Consortium. D. Carmel, H. Roitman, and N. Zwerdling. 2009. Enhancing Cluster Labeling using Wikipedia. In Pro- ceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, pages 139–146. ACM. C. Carpineto, S. Osinski, G. Romano, and Dawid Weiss. 2009. A Survey of Web Clustering Engines. ACM Computing Surveys, 41(3). Y. Chen, S. Yat Mei Lee, and C. Huang. 2009. PolyUHK: A Robust Information Extraction System for Web Personal Names. In Proc. WWW’09 (WePS2 Workshop). ACM. C. Clarke, M. Kolla, G. Cormack, O. Vechtomova, A. Ashkan, S. B ¨uttcher, and I. MacKinnon. 2008. Novelty and Diversity in Information Retrieval Evaluation. In Proc. SIGIR ’08, pages 659–666. ACM. P. Clough, M. Sanderson, M. Abouammoh, S. Navarro, and M. Paramita. 2009. Multiple Approaches to Analysing Query Diversity. In Proc. of SIGIR 2009. ACM. W. Daelemans, J. Zavrel, K. van der Sloot, and A. van den Bosch. 2001 . TiMBL: Tilburg Memory Based Learner, version 4.0, Reference Guide. Technical report, University of Antwerp. E. Gabrilovich and S. Markovitch. 2007. Computing Semantic Relatedness using Wikipedia-based Explicit Semantic Analysis. In Proceedings of The 20th International Joint Conference on Artificial Intelligence (IJCAI), Hyderabad, India. S. Gollapudi and A. Sharma. 2009. An Axiomatic Approach for Result Diversification. In Proc. WWW 2009, pages 381–390. ACM New York, NY, USA. R. Mihalcea. 2007. Using Wikipedia for Automatic Word Sense Disambiguation. In Proceedings of NAACL HLT, volume 2007. G. Miller, C. R. Beckwith, D. Fellbaum, Gross, and K. Miller. 1990. Wordnet: An on-line lexical database. International Journal of Lexicograph, 3(4). G.A Miller, C. Leacock, R. Tengi, and Bunker R. T. 1993. A Semantic Concordance. In Proceedings of the ARPA WorkShop on Human Language Technology. San Francisco, Morgan Kaufman. M. Paramita, M. Sanderson, and P. Clough. 2009. Diversity in Photo Retrieval: Overview of the ImageCLEFPhoto task 2009. CLEF working notes, 2009. M. Ruiz-Casado, E. Alfonseca, and P. Castells. 2005. Automatic Assignment of Wikipedia Encyclopaedic Entries to Wordnet Synsets. Advances in Web Intelligence, 3528:380–386. M. Sanderson. 2000. Retrieving with Good Sense. Information Retrieval, 2(1):49–69. M. Sanderson. 2008. Ambiguous Queries: Test Collections Need More Sense. In Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval, pages 499–506. ACM New York, NY, USA. C. Santamar ı´a, J. Gonzalo, and F. Verdejo. 2003. Automatic Association of Web Directories to Word Senses. Computational Linguistics, 29(3):485–502. Z. S. Syed, T. Finin, and Joshi. A. 2008. Wikipedia as an Ontology for Describing Documents. In Proc. ICWSM’08. 1366
4 0.14187323 109 acl-2010-Experiments in Graph-Based Semi-Supervised Learning Methods for Class-Instance Acquisition
Author: Partha Pratim Talukdar ; Fernando Pereira
Abstract: Graph-based semi-supervised learning (SSL) algorithms have been successfully used to extract class-instance pairs from large unstructured and structured text collections. However, a careful comparison of different graph-based SSL algorithms on that task has been lacking. We compare three graph-based SSL algorithms for class-instance acquisition on a variety of graphs constructed from different domains. We find that the recently proposed MAD algorithm is the most effective. We also show that class-instance extraction can be significantly improved by adding semantic information in the form of instance-attribute edges derived from an independently developed knowledge base. All of our code and data will be made publicly available to encourage reproducible research in this area.
5 0.11361976 113 acl-2010-Extraction and Approximation of Numerical Attributes from the Web
Author: Dmitry Davidov ; Ari Rappoport
Abstract: We present a novel framework for automated extraction and approximation of numerical object attributes such as height and weight from the Web. Given an object-attribute pair, we discover and analyze attribute information for a set of comparable objects in order to infer the desired value. This allows us to approximate the desired numerical values even when no exact values can be found in the text. Our framework makes use of relation defining patterns and WordNet similarity information. First, we obtain from the Web and WordNet a list of terms similar to the given object. Then we retrieve attribute values for each term in this list, and information that allows us to compare different objects in the list and to infer the attribute value range. Finally, we combine the retrieved data for all terms from the list to select or approximate the requested value. We evaluate our method using automated question answering, WordNet enrichment, and comparison with answers given in Wikipedia and by leading search engines. In all of these, our framework provides a significant improvement.
6 0.10463199 156 acl-2010-Knowledge-Rich Word Sense Disambiguation Rivaling Supervised Systems
7 0.10232564 44 acl-2010-BabelNet: Building a Very Large Multilingual Semantic Network
8 0.10097478 160 acl-2010-Learning Arguments and Supertypes of Semantic Relations Using Recursive Patterns
9 0.097482964 125 acl-2010-Generating Templates of Entity Summaries with an Entity-Aspect Model and Pattern Mining
10 0.086057857 218 acl-2010-Structural Semantic Relatedness: A Knowledge-Based Method to Named Entity Disambiguation
11 0.078190409 15 acl-2010-A Semi-Supervised Key Phrase Extraction Approach: Learning from Title Phrases through a Document Semantic Network
12 0.076504894 181 acl-2010-On Learning Subtypes of the Part-Whole Relation: Do Not Mix Your Seeds
13 0.073172942 166 acl-2010-Learning Word-Class Lattices for Definition and Hypernym Extraction
14 0.070010431 10 acl-2010-A Latent Dirichlet Allocation Method for Selectional Preferences
15 0.06665998 111 acl-2010-Extracting Sequences from the Web
16 0.065673381 200 acl-2010-Profiting from Mark-Up: Hyper-Text Annotations for Guided Parsing
17 0.06483268 124 acl-2010-Generating Image Descriptions Using Dependency Relational Patterns
18 0.060620878 184 acl-2010-Open-Domain Semantic Role Labeling by Modeling Word Spans
19 0.059858661 250 acl-2010-Untangling the Cross-Lingual Link Structure of Wikipedia
20 0.057518173 132 acl-2010-Hierarchical Joint Learning: Improving Joint Parsing and Named Entity Recognition with Non-Jointly Labeled Data
topicId topicWeight
[(0, -0.179), (1, 0.076), (2, -0.017), (3, 0.01), (4, 0.099), (5, 0.004), (6, 0.105), (7, 0.035), (8, -0.026), (9, 0.048), (10, -0.036), (11, -0.092), (12, -0.128), (13, -0.067), (14, 0.04), (15, 0.102), (16, -0.017), (17, 0.265), (18, 0.015), (19, 0.007), (20, -0.121), (21, 0.06), (22, -0.092), (23, -0.0), (24, -0.112), (25, 0.003), (26, 0.106), (27, -0.114), (28, -0.105), (29, 0.018), (30, -0.074), (31, -0.043), (32, 0.02), (33, 0.01), (34, 0.067), (35, 0.135), (36, 0.039), (37, -0.116), (38, 0.055), (39, 0.011), (40, -0.062), (41, 0.013), (42, -0.034), (43, 0.083), (44, 0.033), (45, -0.033), (46, 0.173), (47, 0.085), (48, -0.027), (49, -0.072)]
simIndex simValue paperId paperTitle
same-paper 1 0.91878521 185 acl-2010-Open Information Extraction Using Wikipedia
Author: Fei Wu ; Daniel S. Weld
Abstract: Information-extraction (IE) systems seek to distill semantic relations from naturallanguage text, but most systems use supervised learning of relation-specific examples and are thus limited by the availability of training data. Open IE systems such as TextRunner, on the other hand, aim to handle the unbounded number of relations found on the Web. But how well can these open systems perform? This paper presents WOE, an open IE system which improves dramatically on TextRunner’s precision and recall. The key to WOE’s performance is a novel form of self-supervised learning for open extractors using heuris— tic matches between Wikipedia infobox attribute values and corresponding sentences to construct training data. Like TextRunner, WOE’s extractor eschews lexicalized features and handles an unbounded set of semantic relations. WOE can operate in two modes: when restricted to POS tag features, it runs as quickly as TextRunner, but when set to use dependency-parse features its precision and recall rise even higher.
2 0.88246459 159 acl-2010-Learning 5000 Relational Extractors
Author: Raphael Hoffmann ; Congle Zhang ; Daniel S. Weld
Abstract: Many researchers are trying to use information extraction (IE) to create large-scale knowledge bases from natural language text on the Web. However, the primary approach (supervised learning of relation-specific extractors) requires manually-labeled training data for each relation and doesn’t scale to the thousands of relations encoded in Web text. This paper presents LUCHS, a self-supervised, relation-specific IE system which learns 5025 relations more than an order of magnitude greater than any previous approach with an average F1 score of 61%. Crucial to LUCHS’s performance is an automated system for dynamic lexicon learning, which allows it to learn accurately from heuristically-generated training data, which is often noisy and sparse. — —
3 0.67176038 113 acl-2010-Extraction and Approximation of Numerical Attributes from the Web
Author: Dmitry Davidov ; Ari Rappoport
Abstract: We present a novel framework for automated extraction and approximation of numerical object attributes such as height and weight from the Web. Given an object-attribute pair, we discover and analyze attribute information for a set of comparable objects in order to infer the desired value. This allows us to approximate the desired numerical values even when no exact values can be found in the text. Our framework makes use of relation defining patterns and WordNet similarity information. First, we obtain from the Web and WordNet a list of terms similar to the given object. Then we retrieve attribute values for each term in this list, and information that allows us to compare different objects in the list and to infer the attribute value range. Finally, we combine the retrieved data for all terms from the list to select or approximate the requested value. We evaluate our method using automated question answering, WordNet enrichment, and comparison with answers given in Wikipedia and by leading search engines. In all of these, our framework provides a significant improvement.
4 0.57446134 111 acl-2010-Extracting Sequences from the Web
Author: Anthony Fader ; Stephen Soderland ; Oren Etzioni
Abstract: Classical Information Extraction (IE) systems fill slots in domain-specific frames. This paper reports on SEQ, a novel open IE system that leverages a domainindependent frame to extract ordered sequences such as presidents of the United States or the most common causes of death in the U.S. SEQ leverages regularities about sequences to extract a coherent set of sequences from Web text. SEQ nearly doubles the area under the precision-recall curve compared to an extractor that does not exploit these regularities.
5 0.48998669 250 acl-2010-Untangling the Cross-Lingual Link Structure of Wikipedia
Author: Gerard de Melo ; Gerhard Weikum
Abstract: Wikipedia articles in different languages are connected by interwiki links that are increasingly being recognized as a valuable source of cross-lingual information. Unfortunately, large numbers of links are imprecise or simply wrong. In this paper, techniques to detect such problems are identified. We formalize their removal as an optimization task based on graph repair operations. We then present an algorithm with provable properties that uses linear programming and a region growing technique to tackle this challenge. This allows us to transform Wikipedia into a much more consistent multilingual register of the world’s entities and concepts.
6 0.48463851 109 acl-2010-Experiments in Graph-Based Semi-Supervised Learning Methods for Class-Instance Acquisition
7 0.48234499 181 acl-2010-On Learning Subtypes of the Part-Whole Relation: Do Not Mix Your Seeds
8 0.48219985 197 acl-2010-Practical Very Large Scale CRFs
9 0.47314155 261 acl-2010-Wikipedia as Sense Inventory to Improve Diversity in Web Search Results
10 0.44883928 68 acl-2010-Conditional Random Fields for Word Hyphenation
11 0.4456335 248 acl-2010-Unsupervised Ontology Induction from Text
12 0.44271228 166 acl-2010-Learning Word-Class Lattices for Definition and Hypernym Extraction
13 0.42961407 125 acl-2010-Generating Templates of Entity Summaries with an Entity-Aspect Model and Pattern Mining
14 0.40408614 160 acl-2010-Learning Arguments and Supertypes of Semantic Relations Using Recursive Patterns
15 0.38733652 44 acl-2010-BabelNet: Building a Very Large Multilingual Semantic Network
16 0.36657602 222 acl-2010-SystemT: An Algebraic Approach to Declarative Information Extraction
17 0.36011627 156 acl-2010-Knowledge-Rich Word Sense Disambiguation Rivaling Supervised Systems
18 0.34381586 15 acl-2010-A Semi-Supervised Key Phrase Extraction Approach: Learning from Title Phrases through a Document Semantic Network
19 0.33942395 124 acl-2010-Generating Image Descriptions Using Dependency Relational Patterns
20 0.33222899 218 acl-2010-Structural Semantic Relatedness: A Knowledge-Based Method to Named Entity Disambiguation
topicId topicWeight
[(16, 0.011), (25, 0.066), (33, 0.292), (39, 0.014), (42, 0.026), (59, 0.1), (72, 0.053), (73, 0.047), (76, 0.017), (78, 0.049), (80, 0.016), (83, 0.08), (84, 0.018), (98, 0.109)]
simIndex simValue paperId paperTitle
1 0.91224027 217 acl-2010-String Extension Learning
Author: Jeffrey Heinz
Abstract: This paper provides a unified, learningtheoretic analysis of several learnable classes of languages discussed previously in the literature. The analysis shows that for these classes an incremental, globally consistent, locally conservative, set-driven learner always exists. Additionally, the analysis provides a recipe for constructing new learnable classes. Potential applications include learnable models for aspects of natural language and cognition.
2 0.82672393 75 acl-2010-Correcting Errors in a Treebank Based on Synchronous Tree Substitution Grammar
Author: Yoshihide Kato ; Shigeki Matsubara
Abstract: This paper proposes a method of correcting annotation errors in a treebank. By using a synchronous grammar, the method transforms parse trees containing annotation errors into the ones whose errors are corrected. The synchronous grammar is automatically induced from the treebank. We report an experimental result of applying our method to the Penn Treebank. The result demonstrates that our method corrects syntactic annotation errors with high precision.
same-paper 3 0.78632647 185 acl-2010-Open Information Extraction Using Wikipedia
Author: Fei Wu ; Daniel S. Weld
Abstract: Information-extraction (IE) systems seek to distill semantic relations from naturallanguage text, but most systems use supervised learning of relation-specific examples and are thus limited by the availability of training data. Open IE systems such as TextRunner, on the other hand, aim to handle the unbounded number of relations found on the Web. But how well can these open systems perform? This paper presents WOE, an open IE system which improves dramatically on TextRunner’s precision and recall. The key to WOE’s performance is a novel form of self-supervised learning for open extractors using heuris— tic matches between Wikipedia infobox attribute values and corresponding sentences to construct training data. Like TextRunner, WOE’s extractor eschews lexicalized features and handles an unbounded set of semantic relations. WOE can operate in two modes: when restricted to POS tag features, it runs as quickly as TextRunner, but when set to use dependency-parse features its precision and recall rise even higher.
4 0.73738611 14 acl-2010-A Risk Minimization Framework for Extractive Speech Summarization
Author: Shih-Hsiang Lin ; Berlin Chen
Abstract: In this paper, we formulate extractive summarization as a risk minimization problem and propose a unified probabilistic framework that naturally combines supervised and unsupervised summarization models to inherit their individual merits as well as to overcome their inherent limitations. In addition, the introduction of various loss functions also provides the summarization framework with a flexible but systematic way to render the redundancy and coherence relationships among sentences and between sentences and the whole document, respectively. Experiments on speech summarization show that the methods deduced from our framework are very competitive with existing summarization approaches. 1
5 0.59172297 53 acl-2010-Blocked Inference in Bayesian Tree Substitution Grammars
Author: Trevor Cohn ; Phil Blunsom
Abstract: Learning a tree substitution grammar is very challenging due to derivational ambiguity. Our recent approach used a Bayesian non-parametric model to induce good derivations from treebanked input (Cohn et al., 2009), biasing towards small grammars composed of small generalisable productions. In this paper we present a novel training method for the model using a blocked Metropolis-Hastings sampler in place of the previous method’s local Gibbs sampler. The blocked sampler makes considerably larger moves than the local sampler and consequently con- verges in less time. A core component of the algorithm is a grammar transformation which represents an infinite tree substitution grammar in a finite context free grammar. This enables efficient blocked inference for training and also improves the parsing algorithm. Both algorithms are shown to improve parsing accuracy.
6 0.58418661 109 acl-2010-Experiments in Graph-Based Semi-Supervised Learning Methods for Class-Instance Acquisition
7 0.57314837 218 acl-2010-Structural Semantic Relatedness: A Knowledge-Based Method to Named Entity Disambiguation
8 0.5690456 128 acl-2010-Grammar Prototyping and Testing with the LinGO Grammar Matrix Customization System
9 0.56886393 211 acl-2010-Simple, Accurate Parsing with an All-Fragments Grammar
10 0.56809223 169 acl-2010-Learning to Translate with Source and Target Syntax
11 0.56724125 160 acl-2010-Learning Arguments and Supertypes of Semantic Relations Using Recursive Patterns
12 0.56564069 162 acl-2010-Learning Common Grammar from Multilingual Corpus
13 0.56334519 181 acl-2010-On Learning Subtypes of the Part-Whole Relation: Do Not Mix Your Seeds
14 0.56284595 248 acl-2010-Unsupervised Ontology Induction from Text
15 0.56244588 71 acl-2010-Convolution Kernel over Packed Parse Forest
16 0.56153607 209 acl-2010-Sentiment Learning on Product Reviews via Sentiment Ontology Tree
17 0.55884856 159 acl-2010-Learning 5000 Relational Extractors
18 0.55763173 255 acl-2010-Viterbi Training for PCFGs: Hardness Results and Competitiveness of Uniform Initialization
19 0.55650109 191 acl-2010-PCFGs, Topic Models, Adaptor Grammars and Learning Topical Collocations and the Structure of Proper Names
20 0.55647695 78 acl-2010-Cross-Language Text Classification Using Structural Correspondence Learning