Learning to Translate for Multilingual Question Answering

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 573–584, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Learning to Translate for Multilingual Question Answering Ferhan Ture Comcast Labs∗ 1110 Vermont Ave NW Ste 600 Washington, DC, 20005 USA ferhan ture@cable.comcast.com Elizabeth Boschee Raytheon BBN Technologies 10 Moulton St Cambridge, MA, 02138 USA eboschee@bbn.com Abstract In multilingual question answering, either the question needs to be translated into the docu- ment language, or vice versa. In addition to direction, there are multiple methods to per- form the translation, four of which we explore in this paper: word-based, 10-best, context- based, and grammar-based. We build a fea- ture for each combination of translation direc- tion and method, and train a model that learns optimal feature weights. On a large forum dataset consisting of posts in English, Arabic, and Chinese, our novel learn-to-translate ap- proach was more effective than a strong base- line (p < 0.05): translating all text into En- glish, then training a classifier based only on English (original or translated) text. 1 Introduction Question answering (QA) is a specific form of the information retrieval (IR) task, where the goal is to find relevant well-formed answers to a posed question. Most QA pipelines consist of three main stages: (a) preprocessing the question and collec- tion, (b) retrieval of candidate answers in the col- lection, and (c) ranking answers with respect to their relevance to the question and return the top N an- swers. The types of questions can range from fac- toid (e.g., “What is the capital of France?”) to causal (e.g., “Why are trees green?”), and opinion ques- tions (e.g., “Should USA lower the drinking age?”). The most common approach to multilingual QA (MLQA) has been to translate all content into its ∗This work was completed while author was an employee of Raytheon BBN Technologies. most probable English translation via machine trans- lation (MT) systems. This strong baseline, which we refer to as one-best MT (1MT), has been successful in prior work (Hartrumpf et al., 2009; Lin and Kuo, 2010; Shima and Mitamura, 2010). However, re- cent advances in cross-lingual IR (CLIR) show that one can do better by representing the translation space as a probability distribution (Ture and Lin, 2014). In addition, MT systems perform substan- tially worse with user-generated text, such as web forums (Van der Wees et al., 2015), which provide extra motivation to consider alternative translation approaches for higher recall. To our knowledge, it has yet to be shown whether these recent advance- ments in CLIR transfer to MLQA. We introduce a novel answer ranking approach for MLQA (i.e., Learning to Translate or L2T), a model that learns the optimal translation of question and/or candidate answer, based on how well it discrimi- nates between good and bad answers. We achieve this by introducing a set of features that encapsulate lexical and semantic similarities between a question and a candidate answer through various translation strategies (Section 3.1). The model then learns …

Content truncated to 500 words