UMCC_DLSI_SemSim: Multilingual System for Measuring Semantic Textual Similarity

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 716–721, Dublin, Ireland, August 23-24, 2014. UMCC_DLSI_SemSim: Multilingual System for Measuring Semantic Textual Similarity Alexander Chávez Héctor Dávila DI, University of Matanzas, Cuba. {alexander.chavez, hector.davila}@umcc.cu Yoan Gutiérrez Antonio Fernández-Orquín Andrés Montoyo Rafael Muñoz DLSI, University of Alicante, Spain. {ygutierrez montoyo,rafael}@dlsi.ua.es, antonybr@yahoo.com Abstract In this paper we describe the specifications and results of UMCC_DLSI system, which was involved in Semeval-2014 addressing two subtasks of Semantic Textual Similarity (STS, Task 10, for English and Spanish), and one subtask of Cross-Level Semantic Similarity (Task 3). As a supervised system, it was provided by different types of lexical and semantic features to train a classifier which was used to decide the correct answers for distinct subtasks. These features were obtained applying the Hungarian algorithm over a semantic network to create semantic alignments among words. Regarding the Spanish subtask of Task 10 two runs were submitted, where our Run2 was the best ranked with a general correlation of 0.807. However, for English subtask our best run (Run1 of our 3 runs) reached 16th place of 38 of the official ranking, obtaining a general correlation of 0.682. In terms of Task 3, only addressing Paragraph to Sentence subtask, our best run (Run1 of 2 runs) obtained a correlation value of 0.760 reaching 3rd place of 34. 1 Introduction Many applications of language processing rely on measures of proximity or remoteness of various kinds of linguistic units (words, meanings, sentences, documents). Thus, issues such as disambiguation of meanings, detection of lexical chains, establishing relationships between documents, clustering, etc., require accurate similarity measures. The problem of formalizing and quantifying an intuitive notion of similarity has a long history in philosophy, psychology, artificial intelligence, and through the years has followed many different perspectives (Hirst, 2001). Recent research in the field of Computational Linguistics has emphasized the perspective of semantic relations between two lexemes in a lexical resource, or its inverse, semantic distance. The similarity of sentences is a confidence score that reflects the relationship between the meanings of two sentences. This similarity has been addressed in the literature with terminologies such as affinity, proximity, distance, difference and divergence (Jenhani, et al., 2007). The different applications of text similarity have been separated into a group of similarity tasks: between two long texts, for document classification; between a short text with a long text, for Web search; and between two short texts, for paraphrase recognition, automatic machine translation, etc. (Han, et al., 2013). At present, the calculation of the similarity between texts has been tackled from different points of views. Some have opted for a single measure to capture all the features of texts and other models have been trained with various measures to take text features separately. In this work, we addressed the combination of several measures using a Supervised Machine Learning (SVM) approach. Moreover, we intend to introduce a new approach to calculate textual similarities using a knowledge-based system, which is based on a set of cases …

Content truncated to 500 words