Modeling social causality and social judgment in multi-agent interactions.

摘要

Intelligent Agents are typically situated in a social environment and must reason about social cause and effect. Social causal reasoning is qualitatively different from physical causal reasoning that underlies most intelligent systems. Modeling the process and inference of social causality can enrich the capabilities of multi-agent and intelligent interactive systems. In this thesis, first we explore the underlying theory and process of how people evaluate social events, and present a domain-independent computational framework to reason about social cause and responsibility. The computational framework can be generally incorporated into an intelligent system to augment its cognitive and social functionality.; For the fidelity of the modeling, this work is based on psychological attribution theory. Attribution theory identifies several key factors people use in forming their judgments, such as physical cause, intentions, foreknowledge and coercion. Based on the theory, our work formalizes commonsense reasoning of deriving the beliefs about the key factors from natural language communication and task execution. In addition to developing the model, we design and conduct experiments to empirically validate the model, using real human data. The experimental results show that model predictions of the overall judgments, intermediate beliefs about the variables and inferential mechanism are consistent with people responses.; The computational framework has been applied to several applications, such as emotion modeling, natural language conversation strategies and performance assessment in group training. Other potential applications include interactive system design, adaptive user interfaces and coherent internal models for virtual humans. In the end of the dissertation, we summarize the research contributions and raise some issues for future considerations.