Language Comprehension in a Problem Solver

Douglas Wong

Abstract

Historically, artificial intelligence research has treated natural language processing and automated problem solving as distinct disciplines with separate representational requirements. This paper introduces BRUIN, a unified system that employs a single frame-based knowledge representation to perform both tasks. By framing language comprehension as a process of "plan recognition," BRUIN demonstrates how a system can infer high-level context from specific actions, effectively bridging the gap between linguistic input and practical reasoning.

Introduction

In the late 1970s and early 1980s, AI research faced a significant divide. Problem solvers typically relied on formal logic and predicate calculus to navigate state spaces, while language comprehension systems utilized "frames" or "scripts" to model stereotypical human situations. This dual-representation approach created a barrier: systems could not easily apply their practical knowledge to understand narratives, nor could they use linguistic context to inform problem-solving strategies. Douglas Wong's work at Brown University sought to unify these domains through the BRUIN system.

The BRUIN Architecture

The BRUIN system is built upon three integrated components:

  1. FRAIL (Frame-based AI Language): A foundational knowledge representation where all concepts—from physical objects to complex procedures—are defined as frames.
  2. NASL: The problem-solving engine that executes plans by navigating FRAIL frames to achieve specific goals.
  3. PRAGMATICS: The context-recognition module that interprets linguistic input by mapping actions to potential plans within the FRAIL database.

Methodology: Plan Recognition as Comprehension

The core innovation of BRUIN is treating comprehension as "bottom-up context recognition." When the system encounters an action (e.g., "sawing a board"), the PRAGMATICS component searches for plans that include that action as a sub-step. By identifying a "build-house" plan, the system can disambiguate the roles of objects mentioned in the text, such as identifying the "board" as the "lumber" slot in the construction frame.

Evaluation Domains

Wong evaluated BRUIN across three distinct symbolic domains:

  • Inventory Control: Solving manufacturing problems involving "economic lot size" calculations.
  • Restaurant Dining: Understanding script-based narratives (e.g., interpreting "picking up a menu" as a precursor to ordering).
  • Blocks World: A standard AI environment for manipulating geometric objects.

Key Results

The integration of problem-solving and language processing allowed BRUIN to perform complex inferences. In the inventory domain, the system demonstrated the ability to select appropriate mathematical models based on textual context. For instance, it could distinguish between two equations (EQ1 and EQ2) by determining if the "production rate" variable was present in the problem description.

Equation EQ2 used by BRUIN in the inventory control domain.

Furthermore, the system utilized slot constraints to disambiguate contexts. If an action like "sawing" involved a "person" rather than a "board," the system would correctly pivot from a "construction" frame to a "magician's act" frame, as the latter's "victim" slot accepts human entities.

Limitations and Conclusion

While pioneering, BRUIN was characterized by Wong as a "first impression" system. Its primary limitations included:

  • Action-Centricity: The PRAGMATICS module only triggered on actions, failing to recognize context from states (e.g., "being hungry").
  • Lack of Revision: The system struggled with "garden path" scenarios, where initial ambiguous information led to a context that later proved incorrect.
  • Bottom-Up Bias: The system was less effective at using high-level expectations to resolve low-level ambiguities.

Despite these constraints, BRUIN laid essential groundwork for modern cognitive architectures by demonstrating the power of a unified representational framework for reasoning and language.

Summary updated

BibTeX
@inproceedings{wong1981language,
  author = {Douglas Wong},
  title = {{Language Comprehension in a Problem Solver}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure