Towards a Theory of Comprehension of Declarative Contexts

Fernando GomezACL 1982 · DOI

TL;DR

This paper proposes a novel theory for computer comprehension of declarative texts, like scientific or technical documentation. It distinguishes between concepts as "rules" (conceptual specialists) and concepts as "abstract representations" (schemata/frames), arguing that high-level conceptual knowledge should override lower-level linguistic processing for effective understanding.

What Problem Does This Paper Address?

Traditional approaches to natural language understanding (NLU) in AI have often focused on understanding human actions, plans, and intentions, particularly in dialogue systems. These systems typically rely on understanding verbs that describe actions and inferring goals. However, this paper identifies a gap in how computers comprehend declarative contexts – texts that describe states, properties, and relationships rather than actions. Examples include scientific books, textbooks, and technical manuals. The authors argue that these contexts are characterized by descriptive predicates and verbs (e.g., "contain," "refer," "consist of") and that meaning representation centered on action verbs is insufficient. They aim to develop a theory that specifically addresses the unique challenges of understanding this type of text.

Methods

The core of the proposed theory is based on a Kantian distinction between two aspects of concepts:

  • Conceptual Specialists (Concepts as Rules): These are seen as active agents containing inferential knowledge. They are essentially clusters of production rules that allow the system to interpret information, anticipate incoming data, and guide the comprehension process.
  • Schemata/Frames (Concepts as Abstract Representations): These hold declarative knowledge and structural information about a given topic or situation. They represent what the system "remembers" about a theme.

Comprehension is modeled as a process where a parser segments sentences into a case frame structure. However, unlike traditional parsers, this parser's primary role is not to produce an output for semantic routines. Instead, it initiates the parsing process and continues until a concept relevant to the text's theme is recognized. At this point, the corresponding conceptual specialist takes control, overriding the lower-level linguistic processing. This specialist then guides the parsing until the sentence's meaning is mapped into the internal knowledge structures (active frames). This "top-down" conceptual control is a key differentiator.

The system described, LLULL, uses this approach to understand programming problems. It employs specialized routines for parsing noun groups (DESCRIPTION), prepositions (e.g., FOR-SP, ON-SP), and different classes of descriptive verbs (D-VERBS) and action verbs.

Datasets

The paper does not explicitly mention specific datasets used for training or evaluation. Instead, it refers to understanding "programming problems taken verbatim from introductory books on programming" and uses examples from "Conway and Gries (1975)" to illustrate its concepts.

Key Insights and Ideas

  • Conceptual Dominance: High-level conceptual knowledge (via specialists) should override low-level linguistic processing for effective comprehension of declarative texts. This is a significant departure from purely bottom-up linguistic analysis.
  • Dual Nature of Concepts: Distinguishing between concepts as inferential rules (specialists) and concepts as structural representations (frames) is crucial for separating memory aspects from comprehension processes.
  • Parser as a Trigger: The parser's role is to facilitate the recognition of relevant concepts, which then take over the comprehension task. It's not an end in itself but a means to activate higher-level understanding.
  • Thematic Frames: These frames store structural knowledge about a topic and contain lists of relevant concepts, guiding the specialists.
  • Specialist-Driven Parsing: Once a specialist is activated, it controls the interpretation of the sentence, mapping constituents directly into internal knowledge structures.

Key Results

The paper describes the LLULL system, which has been used to understand "ten programming problems taken verbatim from textbooks." While specific quantitative results (e.g., accuracy scores) are not provided, the qualitative outcome is that the system can successfully parse and represent the meaning of these descriptive programming problem statements. The examples demonstrate how the system can identify relevant concepts, handle complex nominals referring to the same concept, and map sentence structures into a structured output representing the problem's requirements.

Limitations and Future Work

  • Recognition of Frames: The paper acknowledges that recognizing the correct thematic frame associated with a text can be a "very hard problem," although it notes it's simpler for the programming problems studied.
  • Parser's Role in Prepositional Phrases: The paper notes that if a relevant concept is in a prepositional phrase that starts a sentence, it might not immediately take control.
  • Future Work: The authors plan to extensively experiment with the system. They anticipate that adding new problem areas (e.g., understanding programming problems about Roman numerals) will involve integrating new verbs and prepositions into the parser and building new conceptual specialists and thematic frames for the new domain. They aim to fully evaluate their ideas in the coming months.

Summary updated

BibTeX
@inproceedings{gomez1982towards,
  author = {Fernando Gomez},
  title = {{Towards a Theory of Comprehension of Declarative Contexts}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1982},
  doi = {10.3115/981251.981259},
}

Loading related content...

Figure