Learning to Translate: A Query-Specific Combination Approach for Cross-Lingual Information Retrieval

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 589–599, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Learning to Translate: A Query-Specific Combination Approach for Cross-Lingual Information Retrieval Ferhan Ture Raytheon BBN Technologies 10 Moulton St Cambridge, MA, 02138 USA fture@bbn.com Elizabeth Boschee Raytheon BBN Technologies 10 Moulton St Cambridge, MA, 02138 USA eboschee@bbn.com Abstract When documents and queries are pre- sented in different languages, the com- mon approach is to translate the query into the document language. While there are a variety of query translation approaches, recent research suggests that combining multiple methods into a single ”structured query” is the most effective. In this pa- per, we introduce a novel approach for producing a unique combination recipe for each query, as it has also been shown that the optimal combination weights dif- fer substantially across queries and other task specifics. Our query-specific combi- nation method generates statistically sig- nificant improvements over other combi- nation strategies presented in the litera- ture, such as uniform and task-specific weighting. An in-depth empirical anal- ysis presents insights about the effect of data size, domain differences, labeling and tuning on the end performance of our ap- proach. 1 Introduction Cross-lingual information retrieval (CLIR) is a special case of information retrieval (IR) in which documents and queries are presented in different languages. In order to overcome the language barrier, the most commonly adopted method is to translate queries into the document language. Many methods have been introduced for translat- ing queries for CLIR, ranging from word-by-word dictionary lookups (Xu and Weischedel, 2005; Darwish and Oard, 2003) to sophisticated use of machine translation (MT) systems (Magdy and Jones, 2011; Ma et al., 2012). Previous research has shown that combining evidence from differ- ent translation approaches is superior to any sin- gle query translation method (Braschler, 2004; Herbert et al., 2011). While there are numer- ous combination-of-evidence techniques for both mono-lingual and cross-lingual IR, recent work suggests that there is no one-size-fits-all solution. In fact, the optimal combination weights (i.e., weights assigned to each piece of evidence in a linear combination) differ greatly across queries, tasks, languages, and other variants (Ture et al., 2012; Berger and Savoy, 2007). In this paper, we introduce a novel method for learning optimal combination weights when build- ing a linear combination of existing query transla- tion approaches. From standard query-document relevance judgments we train a set of classifiers, which produce a unique combination recipe for each query, based on a large set of features ex- tracted from the query and collection. Experi- mental results show that the effectiveness of our method is significantly higher than state-of-the-art query translation methods and other combination strategies. 2 Related Work The earliest approaches to query translation for CLIR used machine-readable bilingual dictio- naries (Hull and Grefenstette, 1996; Balles- teros and Croft, 1996), achieving around up to 60% of monolingual IR effectiveness. Xu and Weischedel (2005) showed that effectiveness can be increased to around 80% by weighting each …

Content truncated to 500 words