Probabilistic Labeling for Efficient Referential Grounding based on Collaborative Discourse

When humans and artificial agents (e.g. robots) have mismatched perceptions of the shared environment, referential communication between them becomes difficult. To mediate perceptual differences, this paper presents a new approach using probabilistic labeling for referential grounding. This approach aims to integrate different types of evidence from the collaborative referential discourse into a unified scheme. Its probabilistic labeling procedure can generate multiple grounding hypotheses to facilitate follow-up dialogue. Our empirical results have shown the probabilistic labeling approach significantly outperforms a previous graphmatching approach for referential grounding.