Abstract
Resolving ambiguous pronouns like "it" is a significant challenge for robots operating in physical environments. While statistical models excel at text-based coreference resolution, they often fail in imperative discourse where the correct referent depends on extra-linguistic context. This paper introduces a framework that combines logic-based reasoning via Answer Set Programming (ASP) with uncertainty modeling using Dempster-Shafer (DS) theory. By evaluating physical plausibility, social norms, and speaker intent, the system successfully identifies referents in situated tasks where traditional models struggle.
Introduction
In natural language processing, anaphora resolution involves determining the referent of a pronoun. In imperative discourse—commands given to a robot in a physical setting—this task becomes complex. For instance, in the sequence "Pick up the knife. Cut the tomato. Pass it to me," the referent of "it" depends on the current task state and social context. Standard statistical models typically rely on linguistic patterns and often fail to account for the physical state of the world or social expectations. The authors argue that effective resolution requires reasoning about physical capabilities, task constraints, and human intent.
Proposed Methodology
The authors propose a transparent, reasoning-based architecture that moves away from "black-box" neural networks. The system integrates two core components:
- Answer Set Programming (ASP): This declarative logic programming paradigm supports non-monotonic reasoning, allowing the robot to update its beliefs as new information arrives.
- Dempster-Shafer (DS) Theory: This framework represents and combines uncertain evidence, enabling the system to quantify ambiguity and ignorance across multiple potential referents.
The architecture employs three specialized Reasoning Consultants:
- Plausibility Reasoner ($\\mathcal{E}_p$): Evaluates physical feasibility (e.g., can an object be picked up?).
- Normativity Reasoner ($\\mathcal{E}_n$): Assesses adherence to social rules or task-specific constraints.
- Speaker Intent Reasoner ($\\mathcal{E}_i$): Determines relevance to the speaker's current activity.
Evaluation and Results
The framework was evaluated using situated challenge cases, including a "Block World" and various "Kitchen Scenarios." In the kitchen tasks, the correct resolution of "it" varied based on the human's activity (e.g., cooking vs. washing dishes).
The system's performance was compared against a state-of-the-art neural coreference system (Clark and Manning, 2016). While the neural baseline often provided incorrect links, the proposed reasoning-based system achieved high certainty in identifying the correct objects.
As shown in the results, the fusion of evidence from the three reasoners allows the system to resolve ambiguity effectively. For example, in a dish-washing context, the Normativity Reasoner correctly identifies the knife as the intended referent for "pass it," reasoning that tools should be passed to the person cleaning them.
Conclusion
This research demonstrates that treating anaphora resolution as a side effect of world-modeling and planning is a viable path for robotic language understanding. By filtering actions through physical, social, and intentional dimensions, robots can achieve more robust communication in dynamic environments. Future work will focus on automating the acquisition of "microtheories" and improving the scalability of the reasoning process for complex, real-world environments.
Summary updated