TALN at SemEval-2016 Task 14: Semantic Taxonomy Enrichment Via Sense-Based Embeddings

Proceedings of SemEval-2016, pages 1332–1336, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics TALN at SemEval-2016 Task 14: Semantic Taxonomy Enrichment Via Sense-Based Embeddings Luis Espinosa-Anke, Horacio Saggion and Francesco Ronzano TALN - Department of Information and Communication Technologies Universitat Pompeu Fabra Carrer T`anger 122-134, 08018 Barcelona (Spain) {luis.espinosa, horacio.saggion, francesco.ronzano}@upf.edu Abstract This paper describes the participation of the TALN team in SemEval-2016 Task 14: Se- mantic Taxonomy Enrichment. The purpose of the task is to find the best point of attach- ment in WordNet for a set of Out of Vocabu- lary (OOV) terms. These may come, to name a few, from domain specific glossaries, slang or typical jargon from Internet forums and cha- trooms. Our contribution takes as input an OOV term, its part of speech and its associ- ated definition, and generates a set of WordNet synset candidates derived from modelling the term’s definition as a sense embedding repre- sentation. We leverage a BabelNet-based vec- tor space representation, which allows us to map the algorithm’s prediction to WordNet. Our approach is designed to be generic and fitting to any domain, without exploiting, for instance, HTML markup in source web pages. Our system performs above the median of all submitted systems, and rivals in performance a powerful baseline based on extracting the first word of the definition with the same part- of-speech as the OOV term. 1 Introduction Semantic knowledge, expressed in terms of concepts and relations holding among them, is an essential enabling component of NLP applications (Jurgens and Pilehvar, 2015). One of the best known se- mantic repository is WordNet (Miller et al., 1990), a manually created semantic network with a coverage of over 200k English senses and 155k word forms. However, as knowledge domains advance and ex- pand, novel terms beyond the coverage of Word- Net are coined on a regular basis. Thus, the need for automatic approaches which are able to gather information from unstructured online sources and structure them is in high demand. In this context, WordNet has become the reference semantic net- work in previous attempts to formalize novel knowl- edge. The SemEval-2016 Task 14: Semantic Taxon- omy Enrichment (Jurgens and Pilehvar, 2016) aims at providing an experimental ground on the task of, given an Out-of-Vocabulary (OOV) term, and an as- sociated definition and part of speech, find the best point of attachment in WordNet (its most similar synset). This is a challenging problem because an OOV term may be defined without any explicit men- tion to its closest WordNet synset. For instance, for the OOV term (from training data) “lectionary”, the associated definition is “The book that contains all the readings from the Scriptures for use in the cel- ebration of the liturgy”, and the best point of at- tachment is sacred text#n#01. However, if one was to simply obtain the first head word of the defini- tion and retrieve its first sense, the result would be book#n#01, with a score of 0.125/1 according to the official evaluation …

Content truncated to 500 words