Performance Evaluation and Error Analysis for Multimodal Reference Resolution in a Conversation System

Multimodal reference resolution is a process that automatically identifies what users refer to during multimodal human-machine conversation. Given the substantial work on multimodal reference resolution; it is important to evaluate the current state of the art, understand the limitations, and identify directions for future improvement. We conducted a series of user studies to evaluate the capability of reference resolution in a multimodal conversation system. This paper analyzes the main error sources during real-time human-machine interaction and presents key strategies for designing robust multimodal reference resolution algorithms.