SBIR-STTR Award

In Phase I of PLATO (Plan Learning Across Textual Observations ), Language Computer Corporation will explore how best to migrate state-of-the-art plan recognition techniques to the more complex genre of textual data addressing such issues as varied textual inputs, diverse subject matter domains, changes in agent plans and goals, unclear or underspecified temporal relationships between actions, and action sequences which are simultaneously uncertain, noisy, and lossy. Specifically, we will demonstrate (1) how to extract plan actions from complex speech acts and task-oriented dialogues; (2) how to automatically produce plan and action resources for detecting plans and goals across multiple domains; (3) how to apply abductive inference to drive robust plan identification over textual data; and (4) how to identify agent progress towards or away from a goal, using state-of-the-art, multi-layered inference and a rich model of temporal reasoning.

In Phase II of PLATO, we will develop a prototype plan and goal identification system that employs a rich model of domain actions and events, extracted automatically from a massive amount of real-world domain-relevant information, to model agent-specific action costs and to process sequences of agent actions and events, extracted from text, as an agent plan in progress. The system will be able to (1) model the inferential knowledge of preconditions, effects, and quantities associated with domain actions and events, derived from web-scale textual resources; (2) extract and structure domain-relevant agent actions and states from text for incorporation into an explorable knowledge base; (3) model action costs for a particular agent in a variety of dimensions including financial cost, legal and moral constraints, and educational requirements; (4) perform goal identification over agent event sequences by estimating agent distance to a goal through state-of-the-art planning; and (5) provide visualization and exploratory methods for users to identify agent capabilities, likely next steps, and high-level steps within an agent plan.