Grounding Language in Multi-Perspective Referential Communication

TL;DR

This paper presents a new task and dataset for referring expression generation and comprehension in multi-agent environments, emphasizing the importance of perspective-taking and demonstrating the challenges faced by automated models compared to humans.

Contribution

It introduces a novel multi-perspective referential communication task and dataset, and evaluates the performance gap between human and machine agents in this setting.

Findings

Models perform worse than humans in reference generation and comprehension.

Training with human-like success signals improves model performance.

Automated models can outperform proprietary baselines with targeted training.

Abstract

We introduce a task and dataset for referring expression generation and comprehension in multi-agent embodied environments. In this task, two agents in a shared scene must take into account one another's visual perspective, which may be different from their own, to both produce and understand references to objects in a scene and the spatial relations between them. We collect a dataset of 2,970 human-written referring expressions, each paired with human comprehension judgments, and evaluate the performance of automated models as speakers and listeners paired with human partners, finding that model performance in both reference generation and comprehension lags behind that of pairs of human agents. Finally, we experiment training an open-weight speaker model with evidence of communicative success when paired with a listener, resulting in an improvement from 58.9 to 69.3% in communicative success and even outperforming the strongest proprietary model.