TL;DR
This paper proposes a novel approach to language generation by integrating it into a general problem-solving framework, viewing language as an action planned to achieve specific goals. This contrasts with traditional methods that treat generation as a separate linguistic module, aiming to create a unified system capable of producing natural language utterances as part of a broader planning process.
What Problem Does This Paper Address?
Traditional approaches to language generation often treat it as a specialized, independent module within a larger system. This "generation module" receives a pre-defined message in an internal representation and translates it into natural language. The problem with this separation is that it disconnects the decision of what to say from how to say it, failing to capture the nuanced interplay between communicative goals and the planning of linguistic actions. This paper addresses this by proposing an integrated approach where language generation is viewed as a form of problem-solving, akin to planning physical actions, and is directly incorporated into a general planning system. This allows for a more holistic approach to communication, where the content and form of an utterance are jointly determined by the speaker's goals and the context.
Methods
The paper introduces KAMP (Knowledge and Modalities Planner), a planning system that builds upon existing AI planning techniques (specifically, Sacerdoti's NOAH system). KAMP represents actions and goals in a procedural network, allowing for actions to be described at various levels of abstraction. A key innovation is the integration of linguistic actions within this general planning framework. KAMP employs a "critic" system that examines the plan globally, identifying interactions, resolving conflicts, and optimizing the plan. This iterative process of planning, expanding, and criticizing aims to produce executable actions, including the generation of natural language utterances. The system also incorporates a knowledge representation capable of handling beliefs and a deduction system for efficient reasoning.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Language as Action: The core idea is that linguistic actions are not fundamentally different from physical actions; both are planned to achieve specific goals. This perspective blurs the line between deciding what to say and how to say it.
- Integrated Planning: Language generation is not a separate module but an integral part of a general problem-solving and planning system. This allows for a more coherent and contextually aware generation process.
- Goal-Driven Generation: Utterances are generated to satisfy various types of goals, including knowledge state goals (changing beliefs), discourse goals (managing the conversation flow), and social goals (adhering to politeness norms).
- Levels of Abstraction: KAMP can represent actions at different levels of detail, allowing high-level communicative goals (like informing) to be expanded into more concrete linguistic and non-linguistic actions.
- The Role of Critics: The critic system is crucial for refining plans, identifying interactions between actions, and ensuring the generated plan is efficient and coherent. This is particularly important for handling the complexities of natural language.
- Contextual Nuance: The paper highlights how discourse goals and social considerations can influence the specific wording and structure of an utterance, demonstrating that the "how" of communication is deeply intertwined with the "what."
Key Results
The paper describes the conceptual framework and the development of KAMP, rather than presenting quantitative results from a fully implemented and tested system. It demonstrates how KAMP can plan for complex scenarios involving multiple agents and knowledge acquisition. An example is provided of planning to remove a bolt, where the system plans to inform the apprentice about the action and the location of a tool. The paper illustrates how the system can expand high-level "INFORM" actions into more specific linguistic constructs, considering factors like the hearer's knowledge and social goals. The system's ability to combine informing actions and handle concept activation is also discussed.
Limitations and Future Work
The authors acknowledge that KAMP is in the early stages of implementation and has not yet been directly applied to language generation. Future work will involve formally describing linguistic actions (focusing, concept activation) and implementing the critic system precisely. The goal is to develop a system capable of producing utterances that exhibit the richness of human communication. The paper also notes that while social goals are important, there is no simple formula for generating polite utterances, suggesting this remains a complex area for research.
Summary updated