Predicting the topic of your next query for just-in-time IR

Proactive search technologies aim at modeling the users' information seeking behaviors for a just-in-time information retrieval (JITIR) and to address the information needs of users even before they ask. Modern virtual personal assistants, such as Microsoft Cortana and Google Now, are moving towards utilizing various signals from users' search history to model the users and to identify their short-term as well as long-term future searches. As a result, they are able to recommend relevant pieces of information to the users at just the right time and even before they explicitly ask (e.g., before submitting a query). In this paper, we propose a novel neural model for JITIR which tracks the users' search behavior over time in order to anticipate the future search topics. Such technology can be employed as part of a personal assistant for enabling the proactive retrieval of information. Our experimental results on real-world data from a commercial search engine indicate that our model outperforms several important baselines in terms of predictive power, measuring those topics that will be of interest in the near-future. Moreover, our proposed model is capable of not only predicting the near-future topics of interest but also predicting an approximate time of the day when a user would be interested in a given search topic.