Exploring the Integration of WordNet and FrameNet Egoitz Laparra German Rigau Montse Cuadros IXA group, UPV/EHU IXA group, UPV/EHU TALP Research Center, UPC Donostia, Spain Donostia, Spain Barcelona, Spain [email protected] [email protected] [email protected] Abstract However, building large and rich enough predi- cate models for broad–coverage semantic process- This paper presents a novel automatic approach to ing takes a great deal of expensive manual effort partially integrate FrameNet and WordNet. In that involving large research groups during long peri- way we expect to extend FrameNet coverage, to ods of development. Thus, the coverage of cur- enrich WordNet with frame semantic information rently available predicate-argument resources is and possibly to extend FrameNet to languages other still unsatisfactory. For example, (Burchardt et than English. The method uses a knowledge-based al., 2005) or (Shen and Lapata, 2007) indicate the Word Sense Disambiguation algorithm for match- limited coverage of FrameNet as one of the main ing the FrameNet lexical units to WordNet synsets. problems of this resource. In fact, FrameNet1.3 Specifically, we exploit a graph-based Word Sense covers around 10,000 lexical-units while for in- Disambiguation algorithm that uses a large-scale stance, WordNet3.0 contains more than 150,000 knowledge-base derived from existing resources. words. Furthermore, the same effort should be in- We have developed and tested additional versions vested for each different language (Subirats and of this algorithm showing substantial improvements Petruck, 2003). Following the line of previous over state-of-the-art results. Finally, we show some works (Shi and Mihalcea, 2005), (Burchardt et al., examples and figures of the resulting semantic re- 2005), (Johansson and Nugues, 2007), (Pennac- source. chiotti et al., 2008), (Cao et al., 2008), (Tonelli and Pianta, 2009), we empirically study a novel 1 Introduction approach to partially integrate FrameNet (Baker Predicate models such as FrameNet (Baker et et al., 1998) and WordNet (Fellbaum, 1998). al., 1998), VerbNet (Kipper, 2005) or PropBank The method relies on the use of a knowledge- (Palmer et al., 2005) are core resources in most based Word Sense Disambiguation (WSD) algo- advanced NLP tasks, such as Question Answer- rithm that uses a large-scale graph of concepts ing, Textual Entailment or Information Extraction. derived from WordNet (Fellbaum, 1998) and eX- Most of the systems with Natural Language Un- tented WordNet (Mihalcea and Moldovan, 2001). derstanding capabilities require a large and precise The WSD algorithm is applied to coherent group- amount of semantic knowledge at the predicate- ings of words belonging to the same frame. In that argument level. This type of knowledge allows way we expect to extend the coverage of FrameNet to identify the underlying typical participants of (by including from WordNet closely related con- a particular event independently of its realization cepts), to enrich WordNet with frame semantic in- in the text. Thus, using these models, different formation (by porting frame information to Word- linguistic phenomena expressing the same event, net) and possibly to extend FrameNet to languages such as active/passive transformations, verb alter- other than English (by exploiting local wordnets nations and nominalizations can be harmonized aligned to the English WordNet). into a common semantic representation. In fact, WordNet 1 (Fellbaum, 1998) is by far the most lately, several systems have been developed for widely-used knowledge base. In fact, WordNet shallow semantic parsing and semantic role label- is being used world-wide for anchoring differ- ing using these resources (Erk and Pado, 2004), ent types of semantic knowledge including word- (Shi and Mihalcea, 2005), (Giuglea and Moschitti, 2006). 1http://wordnet.princeton.edu/ nets for languages other than English (Atserias JECT or TEACHER that are semantic participants et al., 2004), domain knowledge (Magnini and of the frame and their corresponding LUs (see ex- Cavaglia,` 2000) or ontologies like SUMO (Niles ample below). and Pease, 2001) or the EuroWordNet Top Con- ´ cept Ontology (Alvez et al., 2008). It contains [Bernard Lansky]ST UDENT studied [the piano]SUBJECT manually coded information about English nouns, [with Peter Wallfisch]T EACHER. verbs, adjectives and adverbs and is organized synset around the notion of a . A synset is a set The paper is organized as follows. After this of words with the same part-of-speech that can short introduction, in section 2 we present the be interchanged in a certain context. For exam- graph-based Word Sense Disambiguation algo- <student pupil educatee> ple, , , form a synset rithm and the additional versions studied in this because they can be used to refer to the same work. The evaluation framework and the results concept. A synset is often further described obtained by the different algorithms are presented by a gloss, in this case: ”a learner who is en- and analyzed in section 3. Section 4 shows some rolled in an educational institution” and by ex- examples and figures of the resulting semantic re- plicit semantic relations to other synsets. Each source, and finally, in section 5, we draw some fi- synset represents a concept that are related with nal conclusions and outline future work. an large number of semantic relations, includ- ing hypernymy/hyponymy, meronymy/holonymy, 2 SSI algorithms antonymy, entailment, etc. FrameNet 2 (Baker et al., 1998) is a very rich We have used a version of the Structural Seman- semantic resource that contains descriptions and tic Interconnections algorithm (SSI) called SSI- corpus annotations of English words following the Dijkstra(Cuadros and Rigau, 2008)(Laparra and paradigm of Frame Semantics (Fillmore, 1976). In Rigau, 2009). SSI is a knowledge-based iterative frame semantics, a Frame corresponds to a sce- approach to Word Sense Disambiguation (Navigli nario that involves the interaction of a set of typ- and Velardi, 2005). The original SSI algorithm is ical participants, playing a particular role in the very simple and consists of an initialization step scenario. FrameNet groups words or lexical units and a set of iterative steps. (LUs hereinafter) into coherent semantic classes Given W, an ordered list of words to be dis- or frames, and each frame is further characterized ambiguated, the SSI algorithm performs as fol- by a list of participants or lexical elements (LEs lows. During the initialization step, all monose- hereinafter). Different senses for a word are repre- mous words are included into the set I of already sented in FrameNet by assigning different frames. interpreted words, and the polysemous words are Currently, FrameNet represents more than included in P (all of them pending to be disam- 10,000 LUs and 825 frames. More than 6,100 biguated). At each step, the set I is used to disam- of these LUs also provide linguistically annotated biguate one word of P, selecting the word sense corpus examples. However, only 722 frames have which is closer to the set I of already disam- associated a LU. From those, only 9,360 LUs3 biguated words. Once a sense is disambiguated, where recognized by WordNet (out of 92%) cor- the word sense is removed from P and included responding to only 708 frames. into I. The algorithm finishes when no more pend- ing words remain in P. LUs of a frame can be nouns, verbs, adjectives In order to measure the proximity of one synset and adverbs representing a coherent and closely (of the word to be disambiguated at each step) related set of meanings that can be viewed as a to a set of synsets (those word senses already in- small semantic field. For example, the frame ED- terpreted in I), the original SSI uses an in-house UCATION TEACHING contains LUs referring to knowledge base derived semi-automatically which the teaching activity and their participants. It is integrates a variety of online resources (Navigli, evoked by LUs like student.n, teacher.n, learn.v, 2005). This very rich knowledge-base is used to instruct.v, study.v, etc. The frame also defines core calculate graph distances between synsets. In or- semantic roles (or FEs) such as STUDENT, SUB- der to avoid the exponential explosion of possibil- 2http://framenet.icsi.berkeley.edu/ ities, not all paths are considered. They used a 3Word-frame pairs context-free grammar of relations trained on Sem- Cor to filter-out inappropriate paths and to provide tial guess based on the most probable sense of the weights to the appropriate paths. less ambiguous word of W. For this reason we im- Instead, SSI-Dijkstra uses the Dijkstra algo- plemented two different versions of the basic SSI- rithm to obtain the shortest path distance between Dijkstra algorithm: SSI-Dijkstra-FirstSenses-I a node and some nodes of the whole graph. The (hereinafter FSI) and SSI-Dijkstra-AllSenses-I Dijkstra algorithm is a greedy algorithm that com- (hereinafter ASI). Thus, these two versions per- putes the shortest path distance between one node form as SSI-Dijkstra when W contains monose- an the rest of nodes of a graph. BoostGraph4 li- mous terms, but differently when W contains only brary can be used to compute very efficiently the polysemous words. In fact, FSI and ASI always shortest distance between any two given nodes on provide an interpretation of W. very large graphs. As (Cuadros and Rigau, 2008), While FSI includes in I the sense having min- we also use already available knowledge resources imal cumulated distance to the first senses of the to build a very large connected graph with 99,635 rest of words in W, ASI includes in I the sense nodes (synsets) and 636,077 edges (the set of having minimal cumulated distance to the all the direct relations between synsets gathered from senses of the rest of words in W.