Query rewriting using statistical machine translation

Junwei Bao, De-Quan Zheng, Bing Xu, Tiejun Zhao

Abstract: In the area of Information Retrieval, user queries often mismatch the documents users exactly want. We regard this problem as a Query Rewriting task from user queries to document space. Using query logs containing query-keywords-CTR pairs, we trained a state-of-the-art statistical machine translation model to translate the user query to keywords of a web document. Using this method we successfully built the ¿lecical gap¿ between user queries and document keywords, and got the keywords as rewritings of the queries. We separately use BLUE and CTR-Recall as optimization target to complete eight comparable experiments. CTR-Recall is presented by us as an optimization target and evaluation indicator. It shows that if forcing the same word to be aligned in word alignment and using BLEU as optimization target we get both the best CTR-Recall and BLEU. At the same time using CTR-Recall as optimization target we get both the best CTR-Recall and BLEU too.