Semantic and pragmatic precision in conversational AI systems

Abstract

For a conversational agent, to display intelligent interactive behavior implies the ability to respond to the user's intentions and expectations with correct, consistent and relevant actions with appropriate form and content in a timely fashion. In this paper, we present a data-driven analytical approach to embed intelligence into a conversational AI agent. The method requires a certain amount of (ideally) authentic conversational data, which is transformed in a meaningful way to support intelligent dialog modeling and the design of intelligent conversational agents. These transformations rely on the ISO 24617-2 dialog act annotation standard, and are specified in the Dialogue Act Markup Language (DiAML), extended with plug-ins for articulate representations of domain-specific semantic content and customized communicative functionality. ISO 24617-2 is shown to enable systematic in-depth interaction analysis and to facilitate the collection of conversational data of sufficient quality and quantity of instances of interaction phenomena. The paper provides the theoretical and methodological background of extending the ISO standard and DiAML specifications for use in interaction analysis and conversational AI agent design. The expert-assisted design methodology is introduced, with example applications in the healthcare domain, and is validated in human-agent conversational data collection experiments.

Keywords: conversational AI agents; dialog acts; dialog modeling; human-agent data collection; semantically and pragmatically motivated interaction analysis.

Figure 1

Probabilistic finite state automaton generated from ISO 24617-2 annotated VICO dialog data. SOM, Social Obligation Management; TSK, Task; AUF, Auto-Feedback; ALF, Allo Feedback; DST, Discourse Structuring; TRM, Turn Management; TMM, Time Management; OCM, Own Communication management; B, Bot; U, User.

Figure 2

Authoring content and setting a preference profile for a “hypertension” scenario. From top to bottom: (a) menu to author negotiation values, setting a preference profile and negotiation strategy selection either by an expert or user; (b) resulted authored or automatically generated preference profile; and (c) an action selection menu.

Figure 3

Example of a set participant's preference profile and action selection menu for “obesity” scenario.

Figure 4

ISO 24617-2 Extended Metamodel: (1) with communicative functions for task-specific DAs and dimension-specific functions for Task Management and IRM. (2) with additional qualifiers for emotions. (3) with additional dimension: Interpersonal Relations Management (IRM). (4) plug-in for articulate semantic content representation.