Translating Questions for Cross-Lingual QA

Proceedings of the 13th Annual Conference of the EAMT, pages 112–119, Barcelona, May 2009 Translating Questions for Cross-Lingual QA J¨org Tiedemann Information Science University of Groningen PO Box 716 9700 AS Groningen, The Netherlands j.tiedemann@rug.nl Abstract In this paper we investigate possibili- ties of the development of a task-specific translation component for cross-language question answering. We focus on the optimization of phrase-based SMT for models trained on very limited data re- sources. We also look at the combina- tion of such systems with another approach based on example-based MT with propor- tional analogies. In our experiments we could improve a strong baseline of a gen- eral purpose MT engine with more than 5 BLEU points. 1 Introduction Question answering (QA) is a popular task in the NLP research community. It combines various exciting sub-tasks coming from research in infor- mation extraction, retrieval and different kinds of linguistic processing in a real-world application. Cross-lingual QA adds yet another component to such systems, namely a translation component, in order to open QA systems for different languages. The motivation is to enable users to post questions in their favorite language and to make it possible to find answers in documents using other languages. So far, cross-lingual QA is mainly of academic in- terest because of the general shortcomings in the accuracy of QA systems and also the quality of current general-purpose machine translation (MT) engines. In this paper, we investigate the use of standard techniques in statistical MT (SMT) for the development of a task-specific translation com- ponent to be integrated in a cross-lingual QA ap- plication. However, we focus exclusively on this c⃝ 2009 European Association for Machine Translation. component without evaluating the effect of trans- lation quality on a particular QA engine. The predominant approach to cross-lingual QA is to translate incoming questions into the language the QA system understands and then to run the sys- tem as usual. The answers are usually not trans- lated, which is, especially for factoid questions, often also not necessary. Furthermore, users may well be able to browse through answers in another language (or may use on-line translation services to get a general understanding) but still feel more comfortable in asking in their own language. In our study, we will also follow this approach and, therefore, concentrate on the translation of ques- tions. In particular, we take the case of English-to- Dutch question answering, mainly because of our interest in Dutch QA. The simplest approach to cross-lingual QA is to use available on-line MT engines for the translation of questions (see for instance Larosa et al. (2005)). There are several problems with this approach: First of all, most of these en- gines are general-purpose MT systems which are not optimized for the specific task of translating questions. However, questions have a very spe- cific syntactic structure, often very different from other sentence types. They often show similar pat- terns, especially factoid questions, which makes them suitable for a data-driven approach modeling these specific patterns in …

Content truncated to 500 words