From Annotated Multimodal Corpora to Simulated Human-Like Behaviors

摘要

Multimodal corpora prove useful at different stages of the development process of embodied conversational agents. Insights into human-human communicative behaviors can be drawn from such corpora. Rules for planning and generating such behavior in agents can be derived from this information. And even the evaluation of human-agent interactions can rely on corpus data from human-human communication. In this paper, we exemplify how corpora can be exploited at the different development steps, starting with the question of how corpora are annotated and on what level of granularity. The corpus data can be used either directly for imitating the human behavior recorded in the corpus or rules can be derived from the data which govern the behavior planning process. Corpora can even play a vital role in the evaluation of agent systems. Several studies are presented that make use of corpora for the evaluation task.