LIPN: Introducing a new Geographical Context Similarity Measure and a Statistical Similarity Measure based on the Bhattacharyya coefficient

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 400–405, Dublin, Ireland, August 23-24, 2014. LIPN: Introducing a new Geographical Context Similarity Measure and a Statistical Similarity Measure Based on the Bhattacharyya Coefficient Davide Buscaldi, Jorge J. Garc´ıa Flores, Joseph Le Roux, Nadi Tomeh Laboratoire d’Informatique de Paris Nord, CNRS (UMR 7030) Universit´e Paris 13, Sorbonne Paris Cit´e, Villetaneuse, France {buscaldi,jgflores,joseph.le-roux,nadi.tomeh}@lipn.univ-paris13.fr Bel´em Priego Sanchez Laboratoire LDI (Lexique, Dictionnaires, Informatique) Universit´e Paris 13, Sorbonne Paris Cit´e, Villetaneuse, France LKE, FCC, BUAP, San Manuel, Puebla, Mexico belemps@gmail.com Abstract This paper describes the system used by the LIPN team in the task 10, Multilin- gual Semantic Textual Similarity, at Sem- Eval 2014, in both the English and Span- ish sub-tasks. The system uses a sup- port vector regression model, combining different text similarity measures as fea- tures. With respect to our 2013 partici- pation, we included a new feature to take into account the geographical context and a new semantic distance based on the Bhattacharyya distance calculated on co- occurrence distributions derived from the Spanish Google Books n-grams dataset. 1 Introduction After our participation at SemEval 2013 with LIPN-CORE (Buscaldi et al., 2013) we found that geography has an important role in discriminating the semantic similarity of sentences (especially in the case of newswire). If two events happened in a different location, their semantic relatedness is usually low, no matter if the events are the same. Therefore, we worked on a similarity measure able to capture the similarity between the geographic contexts of two sentences. We tried also to rein- force the semantic similarity features by introduc- ing a new measure that calculates word similari- ties on co-occurrence distributions extracted from Google Books bigrams. This measure was intro- duced only for the Spanish runs, due to time con- straints. The regression model used to integrate the features was the ν-Support Vector Regression This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence de- tails:http://creativecommons.org/licenses/by/4.0/ model (ν-SVR) (Sch¨olkopf et al., 1999) imple- mentation provided by LIBSVM (Chang and Lin, 2011), with a radial basis function kernel with the standard parameters (ν = 0.5). We describe all the measures in Section 2; the results obtained by the system are detailed in Section 3. 2 Similarity Measures In this section we describe the measures used as features in our system. The description of mea- sures already used in our 2013 participation is less detailed than the description of the new ones. Ad- ditional details on the measures may be found in (Buscaldi et al., 2013). When POS tagging and NE recognition were required, we used the Stan- ford CoreNLP1 for English and FreeLing2 3.1 for Spanish. 2.1 WordNet-based Conceptual Similarity This measure has been introduced in order to mea- sure similarities between concepts with respect to an ontology. The similarity is calculated as fol- lows: first of all, words in sentences p and q are lemmatised and mapped …

Content truncated to 500 words