Instructable and Adaptive Web Agents that Learn to Retrieve and Extract Information

We present a system for rapidly and easily building instructable and selfadaptive Web agents for information-retrieval and information-extraction tasks. Our Wisconsin Adaptive Web Assistant (Wawa) constructs a Web agent by accepting user preferences in form of instructions and adapting the agent’s behavior as it encounters new information. Wawa has two neural networks that are responsible for the adaptive capabilities of its agents. User-provided instructions are compiled into these neural networks and are modified via training examples. Users can create these training examples by rating pages that are retrieved by Wawa, but more importantly our system uses techniques from reinforcement learning to internally create its own examples. Users can also provide additional instruction throughout the life of an agent. We evaluate Wawa on a “home-page finder” agent and a “seminarannouncement extractor” agent.