SEISD: An environment for extraction of Semantic Information from on-line dictionaries

Alicia Ageno, Irene Castellón, Maria Antònia Martí, German Rigau, Francesc Ribas Framis, Horacio Rodríguez, Mariona Taulé, Felisa VerdejoANLP 1992 · DOI · OpenReview

TL;DR

This paper introduces SEISD, an interactive environment designed to automatically extract semantic information, specifically taxonomies and other relations, from Machine Readable Dictionaries (MRDs). It tackles the challenge of building knowledge bases by leveraging existing lexical resources, offering a semi-automatic approach that guides human experts through complex decision-making processes.

What Problem Does This Paper Address?

Building robust Knowledge-Based Systems (KBS) is a significant challenge, largely due to the difficulty of acquiring comprehensive knowledge. Traditional methods often involve manual knowledge engineering, which is time-consuming and expensive. This paper addresses the problem of efficiently extracting structured semantic information from readily available on-line dictionaries (MRDs) to populate these KBS. The core challenge lies in transforming the unstructured or semi-structured text within dictionary definitions into a formal, machine-understandable conceptual structure, like a taxonomy. This process is complicated by the inherent ambiguity and variability in natural language definitions.

Methods

SEISD employs a multi-stage approach:

  • Taxonomy Extraction: This module uses a syntactic-semantic analyzer (FPar) and a morphological analyzer (Seg-Word) to identify hypernyms and other semantic relations within dictionary entries. The process starts with a "top entry" and works downwards to build a hierarchical structure.
  • Semantic Extraction: Once a taxonomy is established, this batch process uses a more comprehensive grammar to extract "differentia" (defining characteristics) from definitions, enriching the conceptual nodes.
  • Heuristic Validation: A key innovation is the use of weighted heuristics to guide decision-making in ambiguous situations where algorithmic solutions are insufficient. The system allows for the evaluation and updating of these heuristics to improve its autonomy.
  • Knowledge Integration: The extracted information is integrated into a Lexical Knowledge Base (LKB) using a typed Feature Structure (FS) system with default inheritance.

The system is designed to be interactive, guiding human experts through crucial decision points and providing access to relevant information, rather than being fully automatic.

Datasets

The system was evaluated using the Vox Spanish dictionary, a monolingual dictionary containing approximately 90,000 entries and around 150,000 senses. The authors focused on specific narrow domains within the dictionary, including:

  • Noun domains: "substance," "food," "drink," "person," "place," and "instrument" (approx. 3000 senses).
  • Verb domains: "movement," "ingestion," and "cooking" (approx. 300 senses).

Key Insights and Ideas

  • Leveraging MRDs for Knowledge Acquisition: The paper demonstrates that existing dictionaries, often overlooked for deep semantic extraction, can be a valuable source for building knowledge bases.
  • Interactive Extraction is Crucial: Fully automatic extraction from natural language definitions is difficult. An interactive environment that guides human experts through decision points is essential for accuracy and validation.
  • Heuristics as a Decision-Making Aid: In situations with linguistic ambiguity or underspecification, a system of parameterized heuristics can effectively guide the extraction process and improve performance.
  • Modular Design for Reusability: The environment is designed with reusability in mind, both for its own components and for integrating existing lexical software.

Key Results

The paper reports on the successful construction of several taxonomies from narrow but significant domains within the Vox Spanish dictionary. The system's performance is controlled by heuristics, and the authors have developed a mechanism for evaluating and modifying these heuristics to enhance the system's autonomy. The linguistic knowledge sources (grammars, rules) were developed concurrently with the environment.

Limitations and Future Work

  • Limitations: The extraction process is not fully automatic and requires human expert validation. The need for different grammars for parsing entries from distant semantic fields highlights the complexity of handling diverse linguistic phenomena. The paper also notes the challenge of dealing with the vast number of possible words in dictionary definitions, which would require a very broad morphological analyzer.
  • Future Work: The authors suggest further refinement of the heuristic sets to improve the system's autonomy. The design also considers the possibility of further use of their own software pieces, implying ongoing development and potential for broader application.

Summary updated

BibTeX
@inproceedings{ageno1992seisd,
  author = {Alicia Ageno and Irene Castellón and Maria Antònia Martí and German Rigau and Francesc Ribas Framis and Horacio Rodríguez and Mariona Taulé and Felisa Verdejo},
  title = {{SEISD: An environment for extraction of Semantic Information from on-line dictionaries}},
  booktitle = {ANLP},
  year = {1992},
  doi = {10.3115/974499.974553},
}
Figure