A Comparative Study for Query Translation using Linear Combination and Confidence Measure

A Comparative Study for Query Translation using Linear Combination and Confidence Measure Youssef Kadri Laboratoire RALI, DIRO Université de Montréal CP 6128, Montréal, Canada, H3C3J7 kadriyou@iro.umontreal.ca Jian-Yun Nie Laboratoire RALI, DIRO Université de Montréal CP 6128, Montréal, Canada, H3C3J7 nie@iro.umontreal.ca Abstract In Cross Language Information Retrieval (CLIR), query terms can be translated to the document language using Bilingual Dictionaries (BDs) or Statistical Transla- tion Models (STMs). Combining different translation resources can also be used to improve the performance. Unfortunately, the most studies on combining multiple re- sources use simple methods such as linear combination. In this paper, we drew up a comparative study between linear combina- tion and confidence measures to combine multiple translation resources for the pur- pose of CLIR. We show that the linear combination method is unable to combine correctly different types of resources such as BDs and STMs. While the confidence measure method is able to re-weight the translation candidate more radically than in linear combination. It reconsiders each translation candidate proposed by different resources with respect to additional fea- tures. We tested the two methods on differ- ent test CLIR collections and the results show that the confidence measure outper- forms the linear combination method. 1 Introduction Cross Language Information Retrieval (CLIR) tries to determine documents written in a language from a query written in another language. Query transla- tion is widely considered as the key problem in this task (Oard, 1998). In previous researches, various approaches have been proposed for query transla- tion: using a bilingual dictionary, using an off-the- shelf machine translation system or using a parallel corpus. It is also found that when multiple transla- tion resources are used, the translation quality can be improved, comparing to using only one transla- tion resource (Xu, 2005). Indeed, every translation tool or resource has its own limitations. For exam- ple, a bilingual dictionary can suggest common translations, but they remain ambiguous – transla- tions for different senses of the source word are mixed up. Machine translation systems usually employ sophisticated methods to determine the best translation sentence, for example, syntactic analysis and some semantic analysis. However, it usually output only one translation for a source word, while it is usually preferred that a source query word be translated by multiple words in or- der to produce a desired query expansion effect. In addition, the only word choice made by a machine translation system can be wrong. Finally, parallel corpora contain useful information about word translation in particular areas. One can use such a corpus to train a statistical translation model, which can then be used to translate a query. This approach has the advantage that few manual inter- ventions are required to produce the statistical translation model. In addition, each source word can be translated by several related target words and the latter being weighted. However, among the proposed translation words, there may be irrelevant ones. Therefore, one can take advantage of several translation resources and tools in order to produce better query translations. …

Content truncated to 500 words