Interaction-Based Document Matching for Implicit Search Result Diversification

Xubo Qin, Zhicheng Dou, Yutao Zhu, Ji-Rong Wen

Abstract: To satisfy different intents behind the queries issued by users, the search engines need to re-rank the search result documents for diversification. Most of previous approaches of search result diversification use pre-trained embeddings to represent the candidate documents. These representation-based approaches lose fine-grained matching signals. In this paper, we propose a new supervised framework leveraging interaction-based neural matching signals for implicit search result diversification. Compared with previous works, our proposed framework can capture and aggregate fine-grained matching signals between each candidate document and selected document sequences, and improve the performance of implicit search result diversification. Experimental results show that our proposed framework can outperform previous state-of-the-art implicit and explicit diversification approaches significantly, and even slightly outperforms ensemble diversification approaches. Besides, with our proposed strategies the online ranking latency of our framework is moderate and affordable.