Linguistically Motivated Descriptive Term Selection

K. Sparck Jones, J.i. TaitCOLING 1984

TL;DR

This paper introduces a novel approach to automatic document indexing that leverages sophisticated natural language processing (NLP) to extract semantically rich descriptive terms. By moving beyond simple keyword matching and focusing on the underlying meaning representations of texts, the authors aim to improve the accuracy and flexibility of information retrieval systems, particularly for complex technical documents.

What Problem Does This Paper Address?

Traditional automatic indexing often relies on shallow text processing, leading to terms that are too simplistic or tied to a single textual expression. This can hinder effective information retrieval, as user queries might not precisely match the limited index terms assigned to documents. The authors identify a need for more sophisticated, linguistically motivated indexing that captures the deeper semantic concepts within texts, enabling more robust matching between user information needs (expressed as requests) and document content. They propose that this can be achieved without requiring full natural language understanding of every document, focusing instead on generating flexible textual representations of concepts.

Methods

The core of the approach involves a powerful "syntactico-semantic apparatus" – specifically, Boguraev's analyser – to process texts. This analyser builds meaning representations in the form of dependency trees, capturing semantic relationships between words. The key steps are:

  1. Request Analysis: The analyser processes user requests to generate a meaning representation.
  2. Term Source Identification: Elements within this meaning representation are identified as potential "term sources." This involves defining "scale counts" based on the complexity of the linguistic structures (e.g., verb-noun relationships) to determine the scope of concepts to be indexed.
  3. Semantic Variant Generation: For each term source, the system generates semantic variants. This includes handling compound nouns by inferring relationships between their constituent words and generating alternative prepositional links.
  4. Syntactic Variant Expression: Finally, the system generates a range of textual expressions (syntactic variants) for these semantic concepts. This allows for flexible matching at the text level, even if the underlying conceptual representation is the same. The output is a Boolean expression of these variants for searching.

Datasets

Not discussed in the paper. The authors mention using a small-scale test with 10 requests against 11,429 abstracts in the area of electronics for initial evaluation.

Key Insights and Ideas

  • Indexing as Shallow Text Processing: The paper argues that effective indexing can be achieved through "shallow text processing" that is nevertheless "semantically based," meaning it requires sophisticated language understanding but not necessarily full comprehension of every text.
  • Meaning Representation is Key: The use of a syntactico-semantic analyser to create meaning representations (dependency trees) is central. This allows for the identification of concepts beyond simple word occurrences.
  • Semantic Variants for Flexibility: Generating semantic variants (e.g., handling compound nouns, inferring relationships) is crucial for capturing the intended meaning and allowing for different textual expressions of the same concept.
  • Text-Level Matching is Still Necessary: While conceptual understanding is the goal, the authors acknowledge that practical information retrieval often requires matching at the text level. Therefore, generating diverse textual expressions of concepts is essential.
  • "Redundant" Indexing is Beneficial: The process naturally leads to overlapping term sources, which the authors suggest is beneficial for indexing.

Key Results

The paper reports on initial, very limited tests. They matched 10 requests against 11,429 abstracts in electronics. The results showed that different indexing strategies (varying term source scale and compound noun inference) performed identically. However, the overall results were "not impressive," which the authors attribute to the small request sample size and the ad hoc nature of forming search specifications. They emphasize that the primary goal of this initial work was to demonstrate the feasibility of processing real requests and generating query specifications.

Limitations and Future Work

The authors acknowledge several limitations:

  • Small-Scale Testing: The retrieval experiments were very limited, making it impossible to draw valid performance conclusions. A much larger sample is needed.
  • Ad Hoc Search Specification: The method for forming search specifications from the generated term material is currently "fairly ad hoc."
  • Limited Compound Noun Processing: The approach to compound nouns was a "crude form of linguistic inference" and could be extended.
  • Need for More Testing: Extensive testing is required to establish the relative merits of different indexing strategies (e.g., exclusive vs. inclusive indexing, different scale counts) and to explore applications in various environments.

Future work should focus on larger-scale experiments, refining the search specification generation process, and exploring the application of this linguistically motivated indexing approach in a wider range of contexts.

Summary updated

BibTeX
@inproceedings{jones1984linguistically,
  author = {K. Sparck Jones and J.I. Tait},
  title = {{Linguistically Motivated Descriptive Term Selection}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1984},
}
Figure