Statistical Query Translation Models for Cross- Language Information Retrieval

Query translation is an important task in cross-language information retrieval (CLIR), which aims to determine the best translation words and weights for a query. This paper presents three statistical query translation models that focus on resolution of query translation ambiguities. All the models assume that the selection of the translation of a query term depends on the translations of other terms in the query. They differ in the way linguistic structures are detected and exploited. The co-occurrence model treats a query as a bag of words, and use all the other terms in the query as the context for translation disambiguation. The other two models exploit linguistic dependencies among terms. The noun phrase (NP) translation model detects NPs in a query, and translates each NP as a unit by assuming that the translation of a term only depends on other terms within the same NP. Similarly, the dependency translation model detects and translates dependency triples, such as verb-object, as units. The evaluations show that linguistic structures always lead to more precise translations. The experiments of CLIR on TREC Chinese collections show that all the three models have a positive impact on query translation, and lead to significant improvements of CLIR performance over the simple dictionary-based translation method. The best results are obtained by combining the three models.