CANTO: The Constructive Agent NLI Toolkit

ABSTRACT: SIFT will dramatically improve the Air Force's ability to train using synthetic teammates by leveraging prior synthetic teammate research to develop a generalized and reusable natural language interaction (NLI) development toolkit for constructive agents. We will simplify the creation of constructive agents by making it easy to develop and test the required situation and linguistic knowledge. We will (1) extend the existing model of the situation, (2) improve the cognitive fidelity of its procedures for language generation, and (3) encode knowledge abstractly in conjunction with a compiler for converting it to ACT-R. We will incorporate concepts from situation semantics and utilize a psycholinguistically-motivated model of word meaning based on packets representing bundles of semantic and pragmatic linguistic knowledge. Situation models compiled from this knowledge will interpret chat text in terms of observable situation elements and relevant activities. Synthetic agents built with CANTO will utilize expectations and default assumptions about other agents'goals and behaviors to prepare suitable responses and incrementally generate language from the situation model. BENEFIT: The proposed CANTO toolkit will increase the cognitive fidelity of interactive constructive agents in training simulations in order to provide more realistic communication with the human trainee. Ultimately, this will more effectively prepare trainees for mission-time coordination with their human teammates. CANTO will achieve this by leveraging generalizable elements in the existing Synthetic Teammate Project implementation and closing technological gaps using advances in knowledge representation, psycholinguistics, situation modeling, and natural language generation. In addition, CANTO will reduce the time and difficulty of creating constructive agents in new simulation domains. This is because (1) CANTO's foundational ontology supports domain-general aspects of situated natural language interaction, (2) the CANTO ontology is readily extendable for domain-specific situations and tasks, (3) CANTO will include UIs and APIs for specifying this knowledge, and (4) CANTO will utilize an extended Herbal compiler to support constructive agents in ACT-R and other cognitive architectures.