Knowledge Discovery Through Spoken Dialog

People can acquire knowledge not only from media such as books, but also through interactions with other people. Automated dialog agents, however, typically limit their learning from labeled data, and programming. This thesis describes an approach that enables such agents to actively acquire new knowledge through spoken dialog interaction. To acquire knowledge in different situations, we propose several key dialog-driven strategies that include user-initiated, system-detected and system-initiated processes. Using these techniques, an agent can acquire domain-specific knowledge both from the public domain (through open domain knowledge bases) and in the informal space (through human users). Our approach incorporates two sets of techniques: First, we design techniques that allow a spoken dialog agent to detect when an interaction contains unknown entities. These unknowns can manifest as unseen words/phrases, or unseen references to known entities. We use various spoken dialog features to detect and handle unknown entities. We then use open-domain knowledge bases and appropriate semantic-relatedness measures to deal with unknown references to known entities. Second, we show that our novel conversational strategies allow our agent to acquire additional information about detected unknowns, which can improve the execution of appropriate dialog tasks. These strategies are designed to support two primary goals: (i) Validate the genuineness of detected unknowns, (ii) Elicit relevant information about these unknowns. In this thesis we show that the proposed strategies are effective, and useful. We demonstrate their effectiveness through a system that solicits situational information to augment its knowledge base from its users in a domain that provides information on events. We find that this knowledge is consistent and useful and that it provides reliable information to users.