Artificial Chatting – Pretty close to the real thing, minus the awkward pauses

The best training involves interacting with a team. But this approach is expensive and dependent on schedules, so not efficient enough for all of the training offered to today’s military force. Enter “virtual agents,” artificially intelligent reps. By integrating these agents into learning, military personnel can participate in team training regardless of others’ availability at any given time. TIER1 developed a toolkit to rapidly develop virtual agents to help the Air Force be better prepared for the challenges they face in protecting our freedom.

Training agent development acceleration

Project Details

Proposal Title:
SynChat: Synthetic Language Interface Toolkit for Chat
United States Air Force
Contract Numbers:
FA8650-13-M-6443, FA8650-14-C-6576
Start Dates:

SynChat supports training system developers in simplifying and accelerating the integration of advanced language capable virtual agents into training systems. It includes a suite of tools to support rapid development of agent models and a flexible communications interface built upon pervasive open protocols and deployed as a set of web services. Because of this work, agents can also be developed faster, which means they can be integrated into other learning systems less expensively.

How we did it

This natural language interface toolkit allows rapid development of domain-specific lexicon items, as well as the extension of the general reasoning capabilities of virtual agents into new domains. The solution includes a well-defined interface for text chat between training environments and virtual agents. The chat interface uses open text-based chat standards to enable additional implementation options on either side of the interface. During the development of the toolkit, we leveraged prior Air Force investments associated with synthetic teammate research.

This work advanced the integration and application of text-based communications for command and control training environments through the inclusion of:

  • A web-based capability to upload and mine text documents for new linguistic information
  • An intuitive user interface that allows domain analysts to rapidly adopt the tool without requiring the analyst to be a software developer or cognitive modeler
  • A flexible tag-based system to accommodate future linguistic ontology revisions
  • A standards-based interface definition between language-capable software agents and training systems.

In order to allow implementation into other training scenarios, SynChat includes the capability to mine and integrate multiple data sources which accelerates knowledge elicitation and management to populate the domain knowledge repositories for the virtual agents. This allows the solution to be replicated in a variety of training environments.