A Knowledge-based Representation for Cross-Language Document Retrieval and Categorization

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 414–423, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics A Knowledge-based Representation for Cross-Language Document Retrieval and Categorization Marc Franco-Salvador1,2, Paolo Rosso2 and Roberto Navigli1 1 Department of Computer Science Sapienza Universit`a di Roma, Italy {francosalvador,navigli}@di.uniroma1.it 2 Natural Language Engineering Lab - PRHLT Research Center Universitat Polit`ecnica de Val`encia, Spain {mfranco,prosso}@dsic.upv.es Abstract Current approaches to cross-language doc- ument retrieval and categorization are based on discriminative methods which represent documents in a low-dimensional vector space. In this paper we pro- pose a shift from the supervised to the knowledge-based paradigm and provide a document similarity measure which draws on BabelNet, a large multilingual knowl- edge resource. Our experiments show state-of-the-art results in cross-lingual document retrieval and categorization. 1 Introduction The huge amount of text that is available on- line is becoming ever increasingly multilingual, providing an additional wealth of useful informa- tion. Most of this information, however, is not eas- ily accessible to the majority of users because of language barriers which hamper the cross-lingual search and retrieval of knowledge. Today’s search engines would benefit greatly from effective techniques for the cross-lingual re- trieval of valuable information that can satisfy a user’s needs by not only providing (Landauer and Littman, 1994) and translating (Munteanu and Marcu, 2005) relevant results into different lan- guages, but also by reranking the results in a lan- guage of interest on the basis of the importance of search results in other languages. Vector-based models are typically used in the literature for representing documents both in monolingual and cross-lingual settings (Manning et al., 2008). However, because of the large size of the vocabulary, having each term as a compo- nent of the vector makes the document represen- tation very sparse. To address this issue several approaches to dimensionality reduction have been proposed, such as Principal Component Analysis (Jolliffe, 1986), Latent Semantic Indexing (Hull, 1994), Latent Dirichlet Allocation (LDA) (Blei et al., 2003) and variants thereof, which project these vectors into a lower-dimensional vector space. In order to enable multilinguality, the vectors of com- parable documents written in different languages are concatenated, making up the document ma- trix which is then reduced using linear projection (Platt et al., 2010; Yih et al., 2011). However, to do so, comparable documents are needed as train- ing. Additionally, the lower dimensional represen- tations are not of easy interpretation. The availability of wide-coverage lexical knowledge resources extracted automatically from Wikipedia, such as DBPedia (Bizer et al., 2009), YAGO (Hoffart et al., 2013) and BabelNet (Navigli and Ponzetto, 2012a), has considerably boosted research in several areas, especially where multilinguality is a concern (Hovy et al., 2013). Among these latter are cross-language plagiarism detection (Potthast et al., 2011; Franco-Salvador et al., 2013), multilingual semantic relatedness (Navigli and Ponzetto, 2012b; Nastase and Strube, 2013) and semantic alignment (Navigli and Ponzetto, 2012a; Matuschek and Gurevych, 2013). One main advantage of knowledge-based methods is that they provide a human-readable, …

Content truncated to 500 words