Representation of Texts for Information Retrieval

Nicholas J. Belkin, B. G. Michell, D. G. KuehnerACL 1980 · DOI

This research explores a fundamental shift in document retrieval (DR) by moving away from treating documents as simple "bags of words." Instead, the authors propose representing texts as associative networks—graph-like structures designed to map the "state of knowledge" contained within a document. This approach aims to better assist users who may be in an "anomalous state of knowledge" (ASK), where they recognize a gap in their understanding but lack the precise vocabulary to formulate a keyword-based query.

Methodology: The Word Association Algorithm

The foundation of this approach is a Word Association Algorithm that generates a graph where nodes represent words and edges represent the strength of their association. This strength is calculated based on proximity within the text. The algorithm assigns a rank ($r$) to word pairs based on their relative positions, which is then used to calculate a score.

Ranking values (r) based on word proximity used in the association algorithm.

To refine these statistical maps, the researchers proposed five discourse-based modifications to weight the graph more intelligently:

  1. Sentence Weighting: Giving more "gravity" to concepts in the first and last sentences.
  2. Paragraph Structure: Repeating the first sentence of a paragraph at its end to reinforce thematic links.
  3. Title Primacy: Overweighting words found in the title.
  4. Phrase Recognition: Treating multi-word phrases (e.g., "single-link method") as single conceptual units.
  5. Stop-word Filtering: Removing function words (e.g., "the", "and") to prevent them from distorting the conceptual map.

Evaluation and Key Results

The study utilized a corpus of 30 scientific abstracts to generate association maps. To evaluate the effectiveness of these representations, the researchers asked the original authors of the abstracts to assess the resulting graphs.

A sample association map (Figure 3) and the corresponding author evaluation results (Table 1).

As shown in the evaluation data (Table 1 in the figure above), the results were mixed:

  • Accuracy: 48% of authors felt the maps were an accurate reflection of their knowledge.
  • Connectivity: The system struggled with calibration; while 63% felt some concepts were too strongly connected, a significant 96.3% felt some were too weakly connected.
  • Omissions: 88.9% of respondents felt important concepts were missing from the maps.

Conclusion

While the authors acknowledged the connectivity issues and omissions, they viewed the 48% accuracy rate as a promising baseline. The study suggests that "rough equivalents" of linguistic analysis—such as simple discourse-based weighting—can effectively inject structural insights into statistical models. This provides a computationally efficient path toward representing the "aboutness" of a document for large-scale retrieval systems, laying the groundwork for more sophisticated conceptual matching in information retrieval.

Summary updated

BibTeX
@inproceedings{belkin1980representation,
  author = {Nicholas J. Belkin and B. G. Michell and D. G. Kuehner},
  title = {{Representation of Texts for Information Retrieval}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1980},
  doi = {10.3115/981436.981482},
}
Figure