Semantic-Based Multilingual Document Clustering via Tensor Modeling

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 600–609, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Semantic-Based Multilingual Document Clustering via Tensor Modeling Salvatore Romeo, Andrea Tagarelli DIMES, University of Calabria Arcavacata di Rende, Italy sromeo@dimes.unical.it tagarelli@dimes.unical.it Dino Ienco IRSTEA, UMR TETIS Montpellier, France LIRMM Montpellier, France dino.ienco@irstea.fr Abstract A major challenge in document clustering re- search arises from the growing amount of text data written in different languages. Previ- ous approaches depend on language-specific solutions (e.g., bilingual dictionaries, sequen- tial machine translation) to evaluate document similarities, and the required transformations may alter the original document semantics. To cope with this issue we propose a new docu- ment clustering approach for multilingual cor- pora that (i) exploits a large-scale multilingual knowledge base, (ii) takes advantage of the multi-topic nature of the text documents, and (iii) employs a tensor-based model to deal with high dimensionality and sparseness. Results have shown the significance of our approach and its better performance w.r.t. classic docu- ment clustering approaches, in both a balanced and an unbalanced corpus evaluation. 1 Introduction Document clustering research was initially focused on the development of general purpose strategies to group unstructured text data. Recent studies have started de- veloping new methodologies and algorithms that take into account both linguistic and topical characteristics, where the former include the size of the text and the type of language, and the latter focus on the commu- nicative function and targets of the documents. A major challenge in document clustering research arises from the growing amount of text data that are written in different languages, also due to the increased popularity of a number of tools for collaboratively edit- ing through contributors across the world. Multilingual document clustering (MDC) aims to detect clusters in a collection of texts written in different languages. This can aid a variety of applications in cross-lingual infor- mation retrieval, including statistical machine transla- tion and corpora alignment. Existing approaches to MDC can be divided in two broad categories, depending on whether a parallel cor- pus rather than a comparable corpus is used (Kumar et al., 2011c). A parallel corpus is typically comprised of documents with their related translations (Kim et al., 2010). These translations are usually obtained through machine translation techniques based on a se- lected anchor language. Conversely, a comparable cor- pus is a collection of multilingual documents written over the same set of classes (Ni et al., 2011; Yo- gatama and Tanaka-Ishii, 2009) without any restric- tion about translation or perfect correspondence be- tween documents. To mine this kind of corpus, external knowledge is employed to map concepts or terms from a language to another (Kumar et al., 2011c; Kumar et al., 2011a), which enables the extraction of cross- lingual document correlations. In this case, a major issue lies in the definition of a cross-lingual similarity measure that can fit the extracted cross-lingual correla- tions. Also, from a semi-supervised perspective, other works attempt to define must-link constraints to …

Content truncated to 500 words