A Contextual Bandit Approach To Dynamic Search

When users engage in complex search tasks, they encounter two main questions at each point in the search: how to re-formulate the query and whether to continue the search or stop. In this paper, we propose a contextual bandit algorithm to model the dynamic search process. The proposed algorithm uses the context surrounding the current state of the search to select how the search will continue through different query re-formulation tactics. Furthermore, the algorithm automatically decides stopping condition for a search process. Using data from the Text REtrieval Conference (TREC) 2016 Dynamic Domain Track, we evaluate our system's search effectiveness over time and compare to the official runs. Our results show that the use of context as well as an automated stopping condition is effective in a dynamic search system.