XLING: Matching Query Sentences to a Parallel Corpus using Topic Models for WSD

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 167–170, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics XLING: Matching Query Sentences to a Parallel Corpus using Topic Models for Word Sense Disambiguation Liling Tan and Francis Bond Division of Linguistics and Multilingual Studies, Nanyang Technological University 14 Nanyang Drive, Singapore 637332 alvations@gmail.com, bond@ieee.org Abstract This paper describes the XLING system partici- pation in SemEval-2013 Crosslingual Word Sense Disambiguation task. The XLING system introduces a novel approach to skip the sense disambiguation step by matching query sentenc- es to sentences in a parallel corpus using topic models; it returns the word alignments as the translation for the target polysemous words. Although, the topic-model base matching under- performed, the matching approach showed po- tential in the simple cosine-based surface simi- larity matching. 1 Introduction This paper describes the XLING system, an un- supervised Cross-Lingual Word Sense Disam- biguation (CLWSD) system based on matching query sentence to parallel corpus using topic models. CLWSD is the task of disambiguating a word given a context by providing the most ap- propriate translation in different languages (Lefever and Hoste, 2013). 2 Background Topic models assume that latent topics exist in texts and each semantic topic can be represented with a multinomial distribution of words and each document can be classified into different semantic topics (Hofmann, 1999). Blei et al. (2003b) introduced a Bayesian version of topic modeling using Dirichlet hyper-parameters, La- tent Dirichlet Allocation (LDA). Using LDA, a set of topics can be generated to classify docu- ments within a corpus. Each topic will contain a list of all the words in the vocabulary of the cor- pus where each word is assigned a probability of occurring given a particular topic. 3 Approach We hypothesized that sentences with different senses of a polysemous word will be classified into different topics during the LDA process. By matching the query sentence to the training sen- tences by LDA induced topics, the most appro- priate translation for the polysemous word in the query sentence should be equivalent to transla- tion of word in the matched training sentence(s) from a parallel corpus. By pursuing this ap- proach, we escape the traditional mode of dis- ambiguating a sense using a sense inventory. 4 System Description The XLING_TnT system attempts the matching subtask in three steps (1) Topicalize: match- ing the query sentence to the training sentences by the most probable topic. (2) Rank: the matching sentences were ranked according to the cosine similarity between the query and matching sentences. (3) Translate: provides the translation of the polysemous word in the matched sentence(s) from the parallel corpus. 4.1 Preprocessing The Europarl version 7 corpus bitexts (English- German, English-Spanish, English-French, Eng- lish-Italian and English-Dutch) were aligned at word-level with GIZA++ (Och and Ney, 2003). The translation tables from the word-alignments were used to provide the translation of the poly- semous word in the Translate step. The English sentences from …

Content truncated to 500 words