Automatic Indexing and Government-Binding Theory

Robert J. KuhnsCOLING 1990

TL;DR

This paper introduces a novel news indexing system (NAS) that leverages Government-Binding (GB) Theory to understand the underlying syntactic and semantic structure of news articles. By moving beyond simple keyword matching, NAS can derive deeper conceptual relationships, enabling more accurate and comprehensive indexing for improved information retrieval.

What Problem Does This Paper Address?

The ever-increasing volume of text generated daily poses a significant challenge for efficient and accurate information retrieval and routing. Traditional indexing methods rely heavily on automatic keyword identification, which often miss the nuances of natural language. These methods fail to account for the inherent syntactic and semantic structure of text, leading to a superficial understanding of content. This gap means that even if articles discuss the same underlying concepts (e.g., "foreign trade"), they might be missed by retrieval systems if the exact keywords aren't present. The authors aim to bridge this gap by developing a system that can understand the meaning and relationships within sentences, not just the words themselves.

Methods

The core of the NAS system is a parser based on Government-Binding (GB) Theory, a linguistic framework that models sentence structure and relationships between words. This parser analyzes sentences to derive predicate-argument structures, identifying empty categories (like pronouns) and resolving thematic and binding relationships. This syntactic information is then fed into a semantic interpreter, which maps these structures onto conceptual representations stored in a "concept base." These conceptual representations are linked to a controlled indexing vocabulary.

Key novelties include:

  • GB Theory Integration: Applying a sophisticated linguistic theory (GB) for deep syntactic analysis, which is a departure from typical keyword-based approaches.
  • Semantic Interpretation: Translating grammatical structures into conceptual representations, allowing for a more abstract understanding of content.
  • Conceptual Indexing: Linking these conceptual representations to an indexing vocabulary to assign terms that denote the article's content, even if those specific terms don't appear in the text.

The system also includes preprocessors to handle different news feed formats and a template filter for rigidly formatted, non-textual articles.

Datasets

The paper mentions that the system's lexicon was developed by analyzing strings (words) from a newswire, containing over 17,000 items. Formal benchmarks were established using news from Reuters. No new datasets were introduced.

Key Insights and Ideas

  • Beyond Keywords: The paper strongly advocates that true understanding and effective retrieval require moving beyond simple keyword matching to analyze the underlying structure and meaning of language.
  • Linguistic Theory for Practical Applications: Demonstrates that theoretical linguistic frameworks like GB Theory can be successfully applied to solve real-world problems like information indexing.
  • Conceptual Abstraction: The ability to represent concepts like "foreign trade" without relying on explicit keywords is crucial for robust information retrieval.
  • The Power of Structure: Understanding how words relate to each other syntactically and semantically unlocks a deeper level of content analysis.

Key Results

  • Indexing Coverage: NAS can assign indexes to approximately 75% of news stories. The authors state a planned goal of at least 85% coverage is achievable.
  • Processing Speed: On a Symbolics 3640, NAS can process entire days of news (500-600 stories/day) in 35-40 minutes.
  • Accuracy: Accuracy was judged "extremely high" by independent indexers and editors. However, quantitative quality assessment was difficult due to the complexity and subjectivity of indexing terms.
  • Qualitative Success: The system successfully indexes numerical stories (e.g., earnings reports) and textual stories, deriving relationships and concepts beyond explicit mentions.

Limitations and Future Work

Limitations: * Quantitative Quality Assessment: The authors acknowledge that the complexity and subjectivity of indexing terms made quantitative quality judgment difficult. * Coverage Goal: While 75% coverage is good, the goal of 85% is still aspirational.

Future Work: * Deductive Databases: Development of an interface for NAS to deductive databases to store indexes as logical relations, enabling retrieval based on explicit, implicit, or inferred information. * Extensions to Other Domains: The natural language component is not restricted to news processing and has been used for insurance applications with appropriate preprocessors and expert systems. * Refining Linguistic Models: The application of a GB-based parser to real-world text helps identify the boundaries between core and peripheral aspects of linguistic theory, potentially leading to more robust and psychologically realistic models of human sentence processing.

Summary updated

BibTeX
@inproceedings{kuhns1990automatic,
  author = {Robert J. Kuhns},
  title = {{Automatic Indexing and Government-Binding Theory}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1990},
}
Figure