No abstract

RESUMEN

Conversational agents that engage children in collaborative storytelling need a collection of domain knowledge to formulate its responses. However, the current manual processes of collecting, annotating and extracting data from a corpus of unstructured text is time-consuming that often impedes the population of a knowledge base. An alternative is to design the agent as a teachable peer that can continuously learn from its human users. In this paper, we describe a storytelling agent that can expand its domain knowledge base by learning new assertions from its story-based conversation with children. We also use conversations as a means of validating the acquired knowledge. Results showed that the agent can achieve a recall value from 83% to 94% in extracting capabilities, property and instance relations while encountering difficulty in identifying assertions that describe the location of story events. Analysis of the conversation logs showed that the performance of the agent is affected by how children describe the characters and events in their stories.