Personalizing Query Auto-Completion for Multi-Session Tasks

摘要

Query auto-completion (QAC) displays a list of completions that start with input characters and is integrated into modern search engines. The goal is not only to reduce typing effort but also to help users formulate their search intent. Most prior QAC models focus on ranking completions on the basis of query log, whether considered on a whole or split into sessions based on time. However, a great amount of queries are issued to accomplish complex search tasks which straddle several sessions, and no previous work investigates QAC problem in this scenario. To tackle this challenge, we propose a supervised framework for QAC personalization, where three levels of task-related factors are considered separately and synthetically, including history-level, session-level, and query-level. Experimental results on a real-world search log confirm that our learning to rank model significantly outperforms the competitive baselines and enables a more comprehensive understanding of users' search history.