Augmenting speech transcripts of VR recordings with gaze, pointing, and visual context for multimodal coreference resolution

Abstract: Understanding transcripts of immersive multimodal conversations is challenging because speakers frequently rely on visual context and non-verbal cues, such as gestures and visual attention, which are not captured in speech alone. This lack of information makes coreferences resolution-the task of linking ambiguous expressions like it'' orthere'' to their intended referents-particularly challenging. In this paper we present a system that augments VR speech transcript with eye-tracking laser pointing data, and scene metadata to generate textual descriptions of non-verbal communication and the corresponding objects of interest. To evaluate the system, we collected gaze, gesture, and voice data from 12 participants (6 pairs) engaged in an open-ended design critique of a 3D model of an apartment. Our results show a 26.5\% improvement in coreference resolution accuracy by a GPT model when using our multimodal transcript compared to a speech-only baseline.