ARTIMIS: Natural Dialogue Meets Rational Agency

IJCAI 1997

TL;DR

This paper introduces ARTIMIS, a novel system for cooperative spoken dialogue that grounds communication in a formal theory of rational agency. By integrating a generic logical framework for rational behavior with specialized natural language processing components, ARTIMIS aims to achieve robust and human-like dialogue capabilities.

What Problem Does This Paper Address?

The paper tackles the significant challenge of creating truly intelligent systems capable of natural human-computer dialogue. Traditional systems often struggle with the nuances of cooperative behavior, such as providing suggestive answers, which is crucial for user acceptance. The authors argue that a generic, intelligent-agent-based approach, built upon formal theories of reasoning and interaction, is necessary to overcome these limitations and move beyond simple task-oriented dialogues.

Methods

ARTIMIS is built around a "rational unit" which is the kernel of an intelligent agent. This kernel is an implementation of a formal theory of interaction expressed in a first-order (multi)modal logic of mental attitudes (belief, uncertainty, intention) and actions. This logic provides a homogeneous framework to model principles of rational behavior, communication, and cooperation.

Key aspects of their approach include:

  • Formal Theory of Interaction: A set of generic axioms models rational behavior, communication, and cooperation. This theory is interpreted by a specifically designed reasoning engine.
  • Mental Attitudes and Intentions: The logic formalizes beliefs (B) and intentions (I), treating intention as a macro-attitude. Agents are assumed to be fully introspective and have consistent beliefs.
  • Communicative Acts (CAs): CAs are treated as regular actions, characterized by their rational effect (RE) and feasibility preconditions (FPs).
  • Logical Rationality Principles: Axioms govern how agents plan actions based on their intentions and beliefs about feasibility. For instance, Axiom A2 states that an agent will plan an act if it intends to achieve its RE and has no reason not to carry it out.
  • Cooperative Behavior Axioms: Specific axioms (A7, A8, A9) enable mental attitude transfer (e.g., believing what another agent wants to communicate) and the adoption of intentions from other agents, forming the basis for cooperative dialogue.
  • Specialized Natural Language Components: While the core is logical, ARTIMIS includes subsystems for natural language input (island-driven parsing, semantic completion) and output (linguistic act planner, linguistic realizer). These are intended to be eventually integrated into the logical framework.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and a sample dialogue rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Dialogue as a Derivable Activity: The paper posits that dialogue is not a separate module but a dynamic outcome of explicit reasoning processes motivated by the observation and planning of communicative acts.
  • Genericity and Robustness: By using a generic logical framework for rational behavior, ARTIMIS aims for robustness, allowing it to handle complex and possibly underspecified situations.
  • Planning as Deduction: The authors claim that planning communicative acts can be a fully deductive process, directly derived from rationality principles and act models, without needing external planning algorithms.
  • Unified Logical Framework: The core contribution is the integration of rational agency principles and communicative acts within a single, homogeneous logical framework, enabling a unified approach to intelligent dialogue.

Key Results

The paper presents a sample dialogue demonstrating ARTIMIS's ability to handle cooperative spoken dialogue, specifically in the context of the AGS domain (spoken query of information services). The dialogue shows the system understanding user requests, providing relevant information, and offering alternatives cooperatively. A reasoning example illustrates how the system plans a communicative act (informing the user of a phone number) based on its logical framework. The authors state that the system is implemented and has been evaluated, with results found in a separate publication.

Limitations and Future Work

  • Algorithmic Bottlenecks: The authors acknowledge that algorithmic bottlenecks, particularly in the input side's search space, currently prevent the direct logical specification of natural language processing components within their theorem-proving framework.
  • Integration of NL Components: A key future direction is the eventual integration of the natural language input and output subsystems as logical theories within the rational unit.
  • Experimentation in Real Services: The authors expect the dialogue technology to be experimented with in real services in the short to middle term.

Summary updated

BibTeX
@inproceedings{1997artimis,
  title = {{ARTIMIS: Natural Dialogue Meets Rational Agency}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure