Textual Expertise in Word Experts: An Approach to Text Parsing Based on Topic/Comment Monitoring

Udo HahnACL 1984

TL;DR

This paper introduces "word experts" as a novel approach to text parsing, focusing on understanding text cohesion and coherence by analyzing topic and comment structures. The system builds "text graphs" that represent the thematic organization of texts, enabling more abstractive summarization and deeper text analysis.

What Problem Does This Paper Address?

Understanding the deeper meaning and structure of text goes beyond simply processing individual words. Traditional text analysis often struggles to capture how sentences and paragraphs connect to form a coherent whole, especially at the level of thematic progression and topic development. This paper addresses the challenge of parsing text not just for its grammatical structure, but for its semantic and thematic organization, aiming to bridge the gap between micro-level textuality (cohesion) and macro-level textuality (coherence). The authors propose a method to represent and analyze these higher-level structures, which is crucial for tasks like automatic summarization and information retrieval.

Methods

The core of the proposed approach is the "word expert" system. Each word expert is a procedural agent that encapsulates linguistic and world knowledge about a specific word. This knowledge is represented declaratively in a decision net. Word experts operate by testing conditions against two primary knowledge sources:

  • Context: A frame knowledge base storing conceptual world knowledge. Conditions here check for active frames, hierarchical relationships between frames, and slot assignments.
  • Co-text: A data repository tracking the linear sequence of text analysis. Conditions here check for temporal relationships between text elements (e.g., "before," "after," "in phrase") and whether elements have been processed or modified.

When a word expert is activated, it evaluates these conditions. If a condition is met, it can trigger actions like updating the frame knowledge base (e.g., assigning a slot value) or generating a "reading." These readings, such as "SHIFT" or "SPLITTING_THEMES," are passed to other system components. The ultimate goal is to generate a "text graph" that visually represents the thematic progression and coherence of the text. The system also includes a word expert editor that uses a specialized representation language.

Datasets

The paper uses a German text segment about microcomputers as an example for demonstrating the word expert system. No specific large-scale datasets are mentioned for training or evaluation; the focus is on illustrating the methodology with a carefully chosen example.

Key Insights and Ideas

  • Word Experts as Building Blocks: The idea of using specialized "word experts" that encode knowledge about individual words and their roles in text structure is a powerful modular approach.
  • Bridging Cohesion and Coherence: The system effectively links low-level cohesive devices (like reference and substitution) to high-level coherence phenomena (like thematic progression), demonstrating a unified view of textuality.
  • Text Graphs for Thematic Structure: The generation of text graphs provides a clear, graphical representation of how topics are introduced, developed, and related, offering a valuable output for understanding text structure.
  • Topic/Comment Monitoring: The parsing process is framed as a "topic/comment monitoring device," suggesting that the system actively tracks the flow of information and focus within the text.
  • Functional Description: The paper emphasizes a functional description of the parsing device, focusing on what it achieves rather than the intricate procedural details, which is useful for understanding its capabilities.

Key Results

The paper demonstrates the successful application of word experts to identify and represent textual cohesion and coherence phenomena in a sample text. Specifically, it shows how word experts can:

  • Identify lexical cohesion through relationships between terms.
  • Recognize thematic progression patterns like "SPLITTING_THEMES" and "CASCADING_THEMES."
  • Generate "readings" that describe these phenomena (e.g., SVAL_ASSIGN(Mikrocomputer, Mikroprozessor, Z-80)).
  • Construct text graphs that visually depict the thematic structure, such as the relationships between "Mikrocomputer," its components, and its manufacturers.

The results are presented through detailed trace examples of word expert execution and the resulting text graph structures, illustrating the system's ability to parse text beyond simple syntax.

Limitations and Future Work

  • Simplified Examples: The paper acknowledges that the word experts presented are simplified prototypes, restricting their scope to specific text structures. A complete system would require a much larger and more complex set of word experts.
  • Intermediary Steps Skipped: The representation of the textual analysis skips many intermediary steps, particularly concerning phrasal word experts, indicating that the presented analysis is not exhaustive.
  • Completeness of Representation: The authors note that the representation of the textual analysis lacks completeness, suggesting that a more comprehensive system would need to handle more types of word experts and their interactions.
  • Future Work: The authors suggest that further effort will be devoted to graphical modes of system interaction and that the text graph structures are a natural medium for accessing TOPIC as a text information system. They also mention that word experts are available for other cohesive phenomena like pronominal anaphora and conjunctions, implying future extensions.

Summary updated

BibTeX
@inproceedings{hahn1984textual,
  author = {Udo Hahn},
  title = {{Textual Expertise in Word Experts: An Approach to Text Parsing Based on Topic/Comment Monitoring}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}

Loading related content...

Figure