IRIT: Textual Similarity Combining Conceptual Similarity with an N-Gram Comparison Method

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 552–556, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics IRIT: Textual Similarity Combining Conceptual Similarity with an N-Gram Comparison Method Davide Buscaldi, Ronan Tournier, Nathalie Aussenac-Gilles and Josiane Mothe IRIT 118 Route de Narbonne Toulouse (France) {davide.buscaldi,ronan.tournier}@irit.fr, {nathalie.aussenac,josiane.mothe}@irit.fr Abstract This paper describes the participation of the IRIT team to SemEval 2012 Task 6 (Seman- tic Textual Similarity). The method used con- sists of a n-gram based comparison method combined with a conceptual similarity mea- sure that uses WordNet to calculate the sim- ilarity between a pair of concepts. 1 Introduction The system used for the participation of the IRIT team (composed by members of the research groups SIG and MELODI) to the Semantic Textual Similar- ity (STS) task (Agirre et al., 2012) is based on two sub-modules: • a module that calculates the similarity between sentences using n-gram based similarity; • a module that calculates the similarity between concepts in the two sentences, using a concept similarity measure and WordNet (Miller, 1995) as a resource. In Figure 1, we show the structure of the sys- tem and the connections between the main compo- nents. The input phrases are passed on one hand directly to the n-gram similarity module, and on the other they are annoted with the Stanford POS Tag- ger (Toutanova et al., 2003). All nouns and verbs are extracted from the tagged phrases and WordNet is searched for synsets corresponding to the extracted nouns and nouns associated to the verbs by the de- rived terms relationship. The synsets are the con- cepts used by the conceptual similarity module to Phrases N-gram similarity module POS Tagger Google Web 1T WordNet Concept similarity module Score Geometric Average and Normalisation Concept Extraction Figure 1: Schema of the system. calculate the concept similarity. Each module cal- culates a similarity score using its own method; the final similarity value is calculated as the geometric average between the two scores, multiplied by 5 in order to comply with the task specifications. The n-gram based similarity relies on the idea that two sentences are semantically related if they contain a long enough sub-sequence of non-empty terms. Google Web 1T (Brants and Franz, 2006) has been used to calculate term idf, which is used as a measure of the importance of the terms. The conceptual similarity is based on the idea that, given an ontology, two concepts are semantically similar if their distance from a common ancestor is small enough. We used three different measures: the Wu- Palmer similarity measure (Wu and Palmer, 1994) and two “Proxigenea” measures (Dudognon et al., 2010). In the following we will explain in detail how 552 each similarity module works. 2 N-Gram based Similarity N-gram based similarity is based on the Clustered Keywords Positional Distance (CKPD) model pro- posed in (Buscaldi et al., 2009). This model was originally proposed for passage retrieval in the field of Question Answering (QA), and it has been im- plemented in the JIRS …

Content truncated to 500 words