17)

IJCAI 1987

TL;DR

This paper introduces a novel approach to building natural language interfaces (NLIs) for rule-based expert systems, moving beyond tedious menu-driven interactions. It presents a semantic interpreter that translates user statements into facts for the expert system and a new inference engine, Director, designed to efficiently utilize these facts for focused and less query-intensive interactions.

What Problem Does This Paper Address?

Traditional expert systems often rely on menu-driven interfaces for data gathering, which can be cumbersome and limit the system's capabilities. While natural language interfaces (NLIs) offer a more intuitive way for users to interact with expert systems, translating natural language into the structured facts required by these systems is challenging. Specifically, expert systems lack the inherent regularity of databases, making it difficult to create domain-independent semantic parsers. Furthermore, existing inference engines may not efficiently handle the unconstrained flow of information from an NLI, leading to unnecessary questions and a loss of focus. This paper addresses these two core challenges: developing a robust semantic translation mechanism for expert systems and designing an inference engine that can effectively leverage this input.

Methods

The paper proposes a two-pronged approach:

  • Semantic Interpreter: This component focuses on translating natural language statements into facts for the expert system. Its core is a verb categorization system, where verbs are organized into hierarchical categories. Each hierarchy incorporates selectional restrictions based on noun features (derived from thesauri like Roget's) to guide the parsing process. This approach aims for modularity and some domain independence, unlike previous methods that relied on assumed data structures. The parser uses an ATN (Augmented Transition Network) to derive syntactic structure, which then influences semantic disambiguation and fact derivation.
  • Director: An Inference Engine: This new interpreter is designed to efficiently handle the unconstrained input from the NLI. It employs a heuristically controlled combination of forward and backward chaining.
  • Forward chaining ensures that all facts entered by the NLI are fully considered, promoting a focus based on user input.
  • Backward chaining is guided by heuristics that prioritize rules with recently entered user facts in their left-hand-sides, aiming for a focused and coherent interaction.
  • Director also utilizes rule-fact maps (RF map and FR map) for efficient access to rule dependencies, enabling faster self-description and optimized rule selection. It also incorporates a limited form of user control over backward chaining.

Datasets

Not discussed in the paper. The authors mention testing their ideas in the domain of tax and financial advising using a small expert system called "Taxpert," which deals with personal income tax matters.

Key Insights and Ideas

  • Verb Categorization for Semantics: The paper highlights that organizing verbs hierarchically, with selectional restrictions, provides a powerful and modular way to derive semantics for expert systems, even without a predefined underlying structure.
  • Balancing Forward and Backward Chaining: Director's key innovation is its heuristically controlled blend of forward and backward chaining. This allows it to fully incorporate user-provided facts (forward chaining) while intelligently pursuing goals and minimizing user queries (backward chaining).
  • Focusing on User Input: The heuristics for backward chaining prioritize rules that utilize recently entered user facts. This is crucial for making the interaction feel natural and efficient, as the system can leverage what the user has just said.
  • Rule-Fact Maps for Efficiency: The introduction of RF and FR maps significantly speeds up operations like self-description and rule selection, which are critical for responsive expert systems.
  • User Control for Expertise: Allowing users to provide hints for backward chaining acknowledges that users might have domain knowledge and can help guide the system more effectively.

Key Results

The paper presents a conceptual framework and an implemented system (Director) that demonstrates the feasibility of their approach. While specific quantitative results (e.g., accuracy improvements, query reduction percentages) are not detailed, the qualitative benefits are emphasized:

  • Reduced User Queries: Director aims to minimize the number of questions posed to the user by efficiently utilizing volunteered information.
  • Focused and Coherent Interaction: The heuristic control of chaining and the use of user-provided facts lead to a more natural and less confusing dialogue.
  • Improved Efficiency: By avoiding unnecessary searches and rule evaluations, Director can lead to faster problem-solving.
  • Domain Independence: The semantic approach, based on verb categorization, is designed to be somewhat domain-independent, making it adaptable to different expert systems.

Limitations and Future Work

The authors acknowledge several areas for future investigation:

  • Deriving Goals from User Queries: While Director supports setting goals, the semantic translation of user queries into system goals is still under consideration.
  • Partial Matching: The system needs to handle inputs that only partially match a fact.
  • Semantically Incomplete Input and Anaphora: The current system may need further development to robustly handle incomplete statements and references to previously mentioned entities.
  • Instantiation of Variables: The exact algorithm for instantiating variables within facts needs formalization.
  • Full Natural Language Interface: The ultimate goal is to replace menu systems entirely with a full natural language interface, allowing users to pose questions as well as provide facts.

Summary updated

BibTeX
@inproceedings{1987,
  title = {{17)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}
Figure