TL;DR
This paper explores whether expert systems, rather than traditional database systems, offer a more suitable environment for researching natural language pragmatics, specifically how discourse context influences response generation. The authors argue that expert systems, with their inherent goal-driven nature and interactive information gathering, provide a richer testbed for developing more tailored and context-aware natural language interfaces.
What Problem Does This Paper Address?
The paper tackles the question of where to best conduct research on advanced natural language processing (NLP) capabilities, particularly those related to pragmatics. While database systems have historically been used for NLP research, the authors posit that they may not fully capture the complexities of natural language interaction, especially when it comes to understanding and generating responses that are sensitive to the ongoing conversation. The core challenge is to find an environment that allows for the study of how previous utterances in a dialogue influence the meaning and generation of subsequent responses, moving beyond simple query-response mechanisms.
Methods
The paper compares two primary environments for studying natural language pragmatics: database systems and expert systems.
- Database Systems: Research in this area has focused on two main approaches:
- Gricean Principles: Using conversational maxims to infer user intentions and avoid generating false implicatures. This involves analyzing what is known in the database and applying principles to generate appropriate responses.
- User Plans: Modeling possible user plans and speech acts as plans to derive the speaker's intended meaning from a query.
- Expert Systems: The authors propose that expert systems, particularly those with interactive, backward-chaining architectures, offer a more promising environment. In this setting, the system actively seeks information to achieve a goal. The natural language interface can leverage this goal-driven interaction to:
- Tailor Explanations: Generate explanations that are specific to the user's context, varying aspects like point of view, level of detail, discourse strategy, and surface choice.
- Utilize Extended Discourse: Handle longer conversational segments where the user's underlying goal is made explicit, allowing the system to direct the dialogue and incorporate user-provided information for more relevant justifications.
The authors' own work at Columbia focuses on developing natural language for expert systems, specifically aiming to generate explanations tailored to the user based on prior discourse.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Expert Systems as a Richer Pragmatic Testbed: The paper argues that the goal-oriented and interactive nature of expert systems provides a more natural setting for studying how discourse context influences language generation compared to traditional database systems.
- Tailored Explanations are Key: The ability to generate different explanations for the same advice, based on dimensions like point of view and level of detail, is crucial for effective user interaction in expert systems.
- Discourse as Goal-Directed: In expert systems, natural language dialogue can be structured around a clear user goal, allowing the system to actively guide the conversation and gather relevant information for justification.
- User Modeling from Discourse: Extended discourse in expert systems allows for richer user modeling, as volunteered information can be used to tailor system responses and justifications.
- Limitations of Database Discourse: While database systems can handle some pragmatic issues like ellipsis and anaphora resolution, they are less suited for studying the generation of contextually sensitive responses beyond basic query interpretation.
Key Results
The paper does not present quantitative results from experiments. Instead, it offers a qualitative argument for why expert systems are a better environment for studying extended discourse and tailored response generation. The authors illustrate their points with hypothetical natural language interactions within a student advisor expert system, demonstrating how different explanations can be generated and how user input can shape the system's advice.
Limitations and Future Work
- Limitations:
- The paper acknowledges that determining where discourse segments are divided can be a challenging task for natural language systems.
- It is unclear how natural language would be effectively used in non-interactive expert systems, except for generating explanations after advice has been produced.
- Future Work:
- The authors are exploring the use of natural language as an "overlay" on menu systems to allow users to directly state goals, skip questions, and provide information in a single utterance.
- They are working on expanding the capabilities and knowledge of their student advisor expert system to handle more general problem-solving scenarios.
- The paper suggests that further research is needed to explore how to effectively segment discourse and leverage previous conversational turns for response generation in expert systems.
Summary updated