Simulating User Querying Behavior Using Embedding Space Alignment.

Simulation is used as a cost-efficient and repeatable means of experimentation to support Information Retrieval (IR) systems and digital libraries with more realistic directives when user interaction data is lacking. While simulation has been mainly used for the study of user search behaviours, we argue that the potential of using it for generating querying behavior is in its infancy. In this paper, we describe a pilot study to simulate query sessions via embedding alignment techniques and assess its capability of transferring knowledge from one source to the other. Besides, we validate the simulated query sessions using a cost and benefit evaluation task and investigate whether they can reproduce properties of real query sessions. We found that Skip-thought vectors are well suited for query embedding. In particular, using Skip-thought with vocabulary expansion as a support for embedding alignment techniques is effective in transferring contextualized embeddings that represent a user querying behavior. While the results of the performance evaluation are encouraging and comparable to related research, it is still challenging to simulate interactive query reformulations.