INAOE_UPV-CORE: Extracting Word Associations from Document Corpora to estimate Semantic Textual Similarity

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 229–233, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics INAOE_UPV-CORE: Extracting Word Associations from Document Corpora to estimate Semantic Textual Similarity Fernando Sánchez-Vega Manuel Montes-y-Gómez Luis Villaseñor-Pineda Paolo Rosso Laboratorio de Tecnologías del Lenguaje, Instituto Nacional de Astrofísica, Óptica y Electrónica (INAOE), Mexico. Natural Language Engineering Lab., ELiRF, Universitat Politècnica de València, Spain prosso@dsic.upv.es {fer.callotl,mmontesg,villasen} @inaoep.mx Abstract This paper presents three methods to evaluate the Semantic Textual Similarity (STS). The first two methods do not require labeled train- ing data; instead, they automatically extract semantic knowledge in the form of word asso- ciations from a given reference corpus. Two kinds of word associations are considered: co- occurrence statistics and the similarity of word contexts. The third method was done in collaboration with groups from the Universi- ties of Paris 13, Matanzas and Alicante. It uses several word similarity measures as fea- tures in order to construct an accurate predic- tion model for the STS. 1 Introduction Even with the current progress of the natural lan- guage processing, evaluating the semantic text similarity is an extremely challenging task. Due to the existence of multiple semantic relations among words, the measuring of text similarity is a multi- factorial and highly complex task (Turney, 2006). Despite the difficulty of this task, it remains as one of the most attractive research topics for the NLP community. This is because the evaluation of text similarity is commonly used as an internal module in many different tasks, such as, informa- tion retrieval, question answering, document sum- marization, etc. (Resnik, 1999). Moreover, most of these tasks require determining the “semantic” similarity of texts showing stylistic differences or using polysemicwords (Hliaoutakis et al., 2006). The most popular approach to evaluate the se- mantic similarity of words and texts consists in using the semantic knowledge expressed in ontolo- gies (Resnik, 1999); commonly, WorldNet is used for this purpose (Fellbaum, 2005). Unfortunately, despite the great effort that has been the creation of WordNet, it is still far to cover all existing words and senses (Curran, 2003).Therefore, the semantic similarity methods that use this resource tend to reduce their applicability to a restricted domain and to a specific language. We recognize the necessity of having and using manually-constructed semantic-knowledge sources in order to get precise assessments of the semantic similarity of texts, but, in turn, we also consider that it is possible to obtain good estimations of these similarities using less-expensive, and perhaps broader, information sources. In particular our proposal is to automatically extract the semantic knowledge from large amounts of raw data sam- ples i.e. document corpora without labels. In this paper we describe two different strategies to compute the semantic similarity of words from a reference corpus. The first strategy uses word co- occurrence statistics. It determines that two words are associated (in meaning) if they tend to be used together, in the same documents or …

Content truncated to 500 words