Revisiting Open Domain Query Facet Extraction and Generation

Abstract

Web search queries can often be characterized by various facets. Extracting and generating query facets has various real-world applications, such as displaying facets to users in a search interface, search result diversification, clarifying question generation, and enabling exploratory search. In this work, we revisit the task of query facet extraction and generation and study various formulations of this task, including facet extraction as sequence labeling, facet generation as autoregressive text generation or extreme multi-label classification. We conduct extensive experiments and demonstrate that these approaches lead to complementary sets of facets. We also explored various aggregation approaches based on relevance and diversity to combine the facet sets produced by different formulations of the task. The approaches presented in this paper outperform state-of-the-art baselines in terms of both precision and recall. We confirm the quality of the proposed methods through manual annotation. Since there is no open-source software for facet extraction and generation, we release a toolkit named Faspect, that includes various model implementations for this task.