Building A Large Thesaurus For Information Retrieval

Edward A. Fox, Jane Terry Nutter, Thomas Ahlswede, Martha W. Evens, Judith A. MarkowitzANLP 1988 · OpenReview

This paper tackles the challenge of improving information retrieval (IR) systems by automatically building a comprehensive lexical-semantic network (thesaurus). The authors propose a method to extract rich relational information from machine-readable dictionaries, aiming to bridge the gap between user queries and document content for more effective retrieval.

TL;DR

This paper presents a method for automatically constructing a large lexical-semantic network (thesaurus) from machine-readable dictionaries. By extracting various lexical and semantic relations, the authors aim to improve information retrieval systems, particularly for end-users who lack specialized search intermediary skills.

What Problem Does This Paper Address?

Traditional information retrieval systems often struggle with the "vocabulary mismatch" problem: users and document authors use different words to describe the same concepts. This leads to users missing relevant documents or retrieving irrelevant ones. While trained intermediaries can help bridge this gap, many users prefer to search directly. Automatic methods for expanding queries with related terms are desirable, but building comprehensive thesauri automatically has been a significant challenge. Manual construction is expensive and time-consuming, and existing techniques for automatic thesaurus construction based on collection statistics are unreliable. This paper addresses the need for a large, automatically built thesaurus that captures rich lexical and semantic relationships to aid end-user information retrieval.

Methods

The core of the paper's methodology involves extracting lexical and semantic relations from machine-readable dictionaries. This process is broken down into several key stages:

  • Dictionary Text Processing: The authors utilize two primary dictionaries: the Collins Dictionary of the English Language (CDEL) and Webster's Seventh New Collegiate Dictionary (W7). They employ text processing tools (like UNIX utilities, lex, and yacc) to convert these dictionaries into structured fact bases (Prolog relations). This involves cleaning, parsing, and extracting specific fields like headwords, definitions, sample usages, and part-of-speech information.
  • Definition Analysis: A significant portion of the work focuses on analyzing the definition texts themselves. They developed a parser (based on the Linguistic String Project's grammar) to process these definitions. This required extensive preprocessing, including handling punctuation, formatting, and ensuring all words in the definitions were present in the system's Word Dictionary. The goal is to identify and extract relational triples (word-relation-word) from the definitions.
  • Relation Extraction: The authors explore various techniques for relation extraction:
  • Explicitly Tagged Relations: For W7, they easily extracted synonymy relations marked by "small capitals."
    • Morphological Analysis: They leveraged prefixes and suffixes to infer relations between root words and their derived forms.
  • Defining Formulas: They identified common phrases within definitions that act as "defining formulas" to mark specific relations (e.g., "used for," "part of").
  • Hand Parsing: Simplified manual parsing was used to identify headwords and adjuncts, revealing stereotyped phrase structures indicative of relations.
  • Text Processing Approach: Recognizing the slowness of parsing, they developed a more efficient text processing approach based on identifying defining formulas, particularly for intransitive verb definitions. This method proved faster and yielded good quality triples.
  • Semantic Network Construction: The extracted relational triples are integrated into a large semantic network using the SNePS system. SNePS allows for a unified representation of associative information and predicate-logic-style inference, enabling efficient storage and retrieval of the lexical-semantic data.

Datasets

  • Collins Dictionary of the English Language (CDEL): Approximately 82,000 headwords.
  • Webster's Seventh New Collegiate Dictionary (W7): Approximately 69,000 headwords.
  • Microlytics Inc. synonym file: Used to augment the extracted information.
  • A collection of 82 documents: Used for initial testing of query expansion.
  • A larger collection: Used for subsequent experiments.

The paper does not introduce new benchmarks but rather uses existing collections for evaluating the utility of their constructed thesaurus.

Key Insights and Ideas

  • Dictionaries as Rich Sources: Machine-readable dictionaries are far more than simple word lists; they contain a wealth of structured information about lexical and semantic relationships that can be automatically extracted.
  • The Power of Definitions: Analyzing the definition texts themselves, not just explicit tags, is crucial for uncovering nuanced relationships like taxonomy, part-whole, and causal links.
  • Defining Formulas as Shortcuts: Identifying common "defining formulas" within definitions provides an efficient way to extract relational information, bypassing the need for full syntactic parsing in many cases.
  • Phrasal Lexicons are Key: The authors strongly advocate for including phrases in thesauri, citing evidence that they significantly improve retrieval effectiveness.
  • SNePS for Unified Representation: The SNePS semantic network system provides a robust framework for integrating diverse types of lexical-semantic information and supporting both rule-based and path-based inference.
  • Path-Based Inference for Efficiency: For tasks like query expansion, path-based inference within the semantic network is more efficient than rule-based inference, as it involves traversing predefined paths rather than complex rule matching.

Key Results

  • Successful Extraction of Relations: The authors successfully extracted thousands of relational triples from dictionaries, including taxonomies, modification relations, and various other lexical-semantic links.
  • Parsing Performance: The LSP parser achieved varying success rates across different parts of speech, with nouns having the highest parsing success (77.63%).
  • Efficiency Gains: The text processing approach for relation extraction (especially for intransitive verbs) was significantly faster (hours vs. days for parsing) and yielded comparable or better quality results.
  • Mild Improvement in IR: Initial experiments with the SMART system showed mild improvements in system performance when query expansion involved related terms extracted from their thesaurus.
  • Comparison with Other Expansion Methods: In one study, their lexical-semantic relation method performed comparably to or better than expansion using a thesaurus or dictionary definitions.

Limitations and Future Work

  • Parsing Limitations: The LSP parser did not achieve 100% success, and certain types of definitions (e.g., those involving biological taxonomy or complex adverbial structures) were more challenging to parse.
  • Manual Intervention: While aiming for automation, some manual checking and correction were necessary during the data extraction and refinement stages.
  • Future Work:
    • Expand Dictionary Coverage: Process additional dictionaries (e.g., Longman Dictionary of Contemporary English) and more data from the Oxford Text Archive.
  • Enhance Semantic Network: Extend SNePS to handle a much larger thesaurus and integrate it with IR systems like SMART and CODER.
  • User Involvement: Develop more flexible processing paradigms to allow end-users to be more directly involved in query expansion and term selection.
  • Intelligent Query Expansion: Leverage the semantic network for effective automatic or semi-automatic "intelligent" query expansion.
  • Further Testing: Conduct more extensive testing of the thesaurus's utility in interactive information retrieval sessions.

Summary updated

BibTeX
@inproceedings{fox1988building,
  author = {Edward A. Fox and Jane Terry Nutter and Thomas Ahlswede and Martha W. Evens and Judith A. Markowitz},
  title = {{Building A Large Thesaurus For Information Retrieval}},
  booktitle = {ANLP},
  year = {1988},
  url = {https://openreview.net/forum?id=ZrpfXZkiQj},
}
Figure