LIPN-CORE: Semantic Text Similarity using n-grams, WordNet, Syntactic Analysis, ESA and Information Retrieval based Features

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 162–168, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics LIPN-CORE: Semantic Text Similarity using n-grams, WordNet, Syntactic Analysis, ESA and Information Retrieval based Features Davide Buscaldi, Joseph Le Roux, Jorge J. Garc´ıa Flores Laboratoire d’Informatique de Paris Nord, CNRS, (UMR 7030) Universit´e Paris 13, Sorbonne Paris Cit´e, F-93430, Villetaneuse, France {buscaldi,joseph.le-roux,jgflores} @lipn.univ-paris13.fr Adrian Popescu CEA, LIST, Vision & Content Engineering Laboratory F-91190 Gif-sur-Yvette, France adrian.popescu@cea.fr Abstract This paper describes the system used by the LIPN team in the Semantic Textual Similarity task at *SEM 2013. It uses a support vector re- gression model, combining different text sim- ilarity measures that constitute the features. These measures include simple distances like Levenshtein edit distance, cosine, Named En- tities overlap and more complex distances like Explicit Semantic Analysis, WordNet-based similarity, IR-based similarity, and a similar- ity measure based on syntactic dependencies. 1 Introduction The Semantic Textual Similarity task (STS) at *SEM 2013 requires systems to grade the degree of similarity between pairs of sentences. It is closely related to other well known tasks in NLP such as tex- tual entailment, question answering or paraphrase detection. However, as noticed in (B¨ar et al., 2012), the major difference is that STS systems must give a graded, as opposed to binary, answer. One of the most successful systems in *SEM 2012 STS, (B¨ar et al., 2012), managed to grade pairs of sentences accurately by combining focused mea- sures, either simple ones based on surface features (ie n-grams), more elaborate ones based on lexical semantics, or measures requiring external corpora such as Explicit Semantic Analysis, into a robust measure by using a log-linear regression model. The LIPN-CORE system is built upon this idea of combining simple measures with a regression model to obtain a robust and accurate measure of tex- tual similarity, using the individual measures as fea- tures for the global system. These measures include simple distances like Levenshtein edit distance, co- sine, Named Entities overlap and more complex dis- tances like Explicit Semantic Analysis, WordNet- based similarity, IR-based similarity, and a similar- ity measure based on syntactic dependencies. The paper is organized as follows. Measures are presented in Section 2. Then the regression model, based on Support Vector Machines, is described in Section 3. Finally we discuss the results of the sys- tem in Section 4. 2 Text Similarity Measures 2.1 WordNet-based Conceptual Similarity (Proxigenea) First of all, sentences p and q are analysed in or- der to extract all the included WordNet synsets. For each WordNet synset, we keep noun synsets and put into the set of synsets associated to the sentence, Cp and Cq, respectively. If the synsets are in one of the other POS categories (verb, adjective, adverb) we look for their derivationally related forms in order to find a related noun synset: if there is one, we put this synsets in Cp (or Cq). For instance, the word …

Content truncated to 500 words