Cross-lingual question answering

Abstract

Question Answering has become an intensively researched area in the last decade, being seen as the next step beyond Information Retrieval in the attempt to provide more concise and better access to large volumes of available information. Question Answering builds on Information Retrieval technology for a first touch of possible relevant data and uses further natural language processing techniques to search for candidate answers and to look for clues that accept or invalidate the candidates as right answers to the question. Though most of the research has been carried out in monolingual settings, where the question and the answer-bearing documents share the same natural language, current approaches concentrate on cross-language scenarios, where the question and the documents are in different languages. Known in this context and common with the Information Retrieval research are three methods of crossing the language barrier: by translating the question, by translating the documents or by aligning both the question and the documents to a common inter-lingual representation. We present a cross-lingual English to German Question Answering system, for both factoid and definition questions, using a German monolingual system and translating the questions from English to German. Two different techniques of translation are evaluated: \n•\tdirect translation of the English input question into German and\n•\ttransfer-based translation, by using an intermediate representation that captures the “meaning” of the original question and is translated into the target language. \nFor both translation techniques two types of translation tools are used: bilingual dictionaries and machine translation. The intermediate representation captures the semantic meaning of the question in terms of Question Type (QType), Expected Answer Type (EAType) and Focus, information that steers the workflow of the question answering process. \nThe German monolingual Question Answering system can answer both factoid and definition questions and is based on several premises: \n•\tfacts and definitions are usually expressed locally at the level of a sentence and its surroundings;\n•\tproximity of concepts within a sentence can be related to their semantic dependency;\n•\tfor factoid questions, redundancy of candidate answers is a good indicator of their suitability;\n•\tdefinitions of concepts are expressed using fixed linguistic structures such as appositions, modifiers, and abbreviation extensions.\nExtensive evaluations of the monolingual system have shown that the above mentioned hypothesis holds true in most of the cases when dealing with a fairly large collection of documents, like the one used in the CLEF evaluation forum.