Knowing Before Seeing: Incorporating Post-retrieval Information into Pre-retrieval Query Intention Classification

Xueqing Ma, Xiaochi Wei, Yixing Gao, Runyang Feng, Dawei Yin, Yi Chang

Abstract: Query intention classification is crucial for modern search engine, which explicitly limits the search range to the suggested leaf categories where related search results can be retrieved accurately. Recent studies in the field of query intention classification mainly rely on various post-retrieval information, e.g., click-graphs, document categories and click-through logs. However, despite the promising results of such methods, they are not always valid when applied to real search engines. This mainly can be attributed to the unavailability of the post-retrieval information especially for 1) the pre-retrieval scenario and 2) processing the massive long-tail data that have never appeared. To address the problems , we introduce a unique Post-retrieval Information Fusion (PostIF) framework to incorporate post-retrieval information into a pre-retrieval model during training, instead of directly leveraging a straightforward post-retrieval model in both training and inference. The PostIF framework consists of two parts: an imitation module and an integration module. The imitation module learns to transform existing pre-retrieval information into pseudo post-retrieval information under the guidance of real post-retrieval information, and the integration module predicts the final result by integrating the pre-retrieval and the pseudo post-retrieval representation. We further design two specific methods (i.e., detective method, generative method) to implement the imitation module. Extensive experiments on real-world search logs have validated the proposed method in the pre-retrieval query intention classification task.