Measuring Semantic Relatedness using Multilingual Representations

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 20–29, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Measuring Semantic Relatedness using Multilingual Representations Samer Hassan University of North Texas Denton, TX samer@unt.edu Carmen Banea University of North Texas Denton, TX carmenbanea@my.unt.edu Rada Mihalcea University of North Texas Denton, TX rada@cs.unt.edu Abstract This paper explores the hypothesis that se- mantic relatedness may be more reliably in- ferred by using a multilingual space, as com- pared to the typical monolingual representa- tion. Through evaluations using several state- of-the-art semantic relatedness systems, ap- plied on standard datasets, we show that a multilingual approach is better suited for this task, and leads to improvements of up to 47% with respect to the monolingual baseline. 1 Introduction Semantic relatedness is the task of quantifying the strength of the semantic connection between tex- tual units, be they words, sentences, or documents. For instance, one may want to determine how se- mantically related are two words such as car and automobile, or two pieces of text such as I love an- imals and I own a pet. It is one of the main tasks explored in the field of natural language processing, as it lies at the core of a large number of applica- tions such as information retrieval (Ponte and Croft, 1998), query reformulation (Metzler et al., 2007; Yih and Meek, 2007; Sahami and Heilman, 2006; Broder et al., 2008), image retrieval (Leong and Mi- halcea, 2009; Goodrum, 2000), plagiarism detection (Hoad and Zobel, 2003; Shivakumar and Garcia- Molina, 1995; Broder et al., 1997; Heintze, 1996; Brin et al., 1995; Manber, 1994), information flow (Metzler et al., 2005), sponsored search (Broder et al., 2008), short answer grading (Mohler and Mihal- cea, 2009a; Pulman and Sukkarieh, 2005; Mitchell et al., 2002), and textual entailment (Dagan et al., 2005). The typical approach to semantic relatedness is to either measure the distance between the constituent words by using a knowledge base such as Word- Net or Roget (e.g., (Leacock and Chodorow, 1998; Lesk, 1986; Jarmasz and Szpakowicz, 2003; Peder- sen et al., 2004)), or to calculate the similarity be- tween the word distributions in very large corpora (e.g., (Landauer et al., 1991; Lin, 1998; Gabrilovich and Markovitch, 2007)). With almost no exception, these methods have been applied on one language at a time – English, most of the time, although mea- sures of relatedness have also been explored on lan- guages such as German (Zesch et al., 2007), Chinese (Li et al., 2005), Japanese (Kazama et al., 2010), and others. In this paper, we take a step further and ex- plore a joint multilingual semantic relatedness met- ric, which aggregates semantic relatedness scores measured on several different languages. Specifi- cally, in our method, in order to measure the re- latedness of two textual units, we first determine their relatedness in multiple languages, and conse- quently infer a final relatedness score by averaging the scores calculated in the individual languages. Our hypothesis is that a multilingual …

Content truncated to 500 words