An Inference-based Approach to Dialogue System Design

Johan Bos, Tetsushi OkaCOLING 2002

TL;DR

This paper proposes a novel approach to dialogue system design that leverages first-order logic inference for robust understanding and action. By translating dialogue meaning into first-order logic and using theorem provers and model builders, the system can handle complex instructions, resolve ambiguities, and manage dialogue flow more effectively, leading to increased portability across domains.

What Problem Does This Paper Address?

Designing effective spoken dialogue systems, particularly for complex tasks like instruction following, is challenging. A key hurdle is accurately interpreting user utterances, which often involve context-sensitive phenomena like anaphora and presuppositions, and then determining the appropriate system response or action. Traditional systems often rely on hand-crafted rules or statistical models that can be brittle and difficult to adapt to new domains. The authors aim to address this by developing a more principled, inference-driven approach that can handle a wider range of linguistic phenomena and be more easily ported to new applications.

Methods

The core of the proposed approach is the use of Discourse Representation Theory (DRT) to represent the meaning of dialogue utterances. DRT is extended to handle dialogue-specific phenomena, including actions, modal expressions, and questions. Crucially, these extended DRT structures are then translated into first-order logic (FOL) formulas. This translation allows the system to leverage powerful, off-the-shelf inference tools:

  • Theorem Provers (TP): Used to check for inconsistencies in the dialogue interpretation. If a contradiction is found, it signals a misunderstanding or disagreement.
  • Model Builders (MB): Used to generate finite, minimal first-order models of the dialogue state. These models are not just abstract representations but are used to directly drive system actions and generate responses.

The system is implemented as a society of OAA (Open Agent Architecture) agents, with specialized agents for speech recognition (ASR), dialogue move engine (DME), resolution (RES), generation (GEN), and inference (INF). The INF agent acts as a mediator, coordinating the TP and MB agents.

Datasets

Not discussed in the paper. The paper focuses on a specific application (home automation) and describes the setup and evaluation within that context, rather than using pre-existing datasets for training or evaluation.

Key Insights and Ideas

  • Inference as a Foundation for Dialogue Management: The paper argues that inference, beyond simple consistency checking, is crucial for sophisticated dialogue behavior, including ambiguity resolution and action execution.
  • Model Building for Action: A key innovation is the use of first-order model builders not just for theoretical analysis but for generating concrete representations that can directly trigger system actions. This makes the system's understanding more actionable.
  • DRT to FOL Translation for Portability: By translating dialogue meaning into FOL, the system can leverage existing, well-developed inference engines. This translation mechanism is key to making the system more portable to new domains, as only the lexicon and domain-specific knowledge base need to be adapted.
  • Hybrid DRS-conditions for Richer Representation: The introduction of hybrid DRS-conditions, particularly those linking discourse referents to possible worlds, allows for a more nuanced representation of context and knowledge.

Key Results

The paper demonstrates the feasibility of the inference-based approach through an example application in home automation. The system successfully handles commands like "turn on a light," resolving the ambiguity of "a light" by using model building to identify specific lights that can be turned on. It can also signal inconsistencies when commands are contradictory. While specific quantitative results are not the primary focus, the evaluation sections highlight:

  • Adaptability: The system can be adapted to new domains by modifying the lexicon and knowledge base, with the core dialogue system remaining largely unchanged.
  • Portability: The domain-specific axioms and semantic descriptions are the main components requiring adaptation for new domains.
  • Scalability: For small dialogues (a few utterances) in a small domain, the system achieves reasonable real-time performance, though inference can take several seconds.

Limitations and Future Work

  • Non-Incremental Approach: The current approach is not incremental; the entire dialogue history is translated to FOL and checked for consistency after each utterance. The authors suggest that an incremental approach, particularly for model building, would be beneficial.
  • Performance on Large Dialogues: Unrestricted theorem proving and model building fail for large dialogues.
  • Sorted Logics and Discourse Structure: Future work could explore using sorted logics and incorporating discourse structure to limit the complexity of the inference tasks.
  • Experimenting with Modal Formulations: Further investigation into different modal formulations for representing dialogue phenomena is suggested.

Summary updated

BibTeX
@inproceedings{bos2002inference,
  author = {Johan Bos and Tetsushi Oka},
  title = {{An Inference-based Approach to Dialogue System Design}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure