TL;DR
This paper explores two contrasting methods for automatically generating thesauri from dictionaries: a traditional, intuitive approach and a more systematic, computational method based on analyzing dictionary definitions. The authors propose that by leveraging the rich, formalized information within dictionaries like the Longman Dictionary of Contemporary English (LDOCE), we can move beyond manual thesaurus creation towards more robust and scalable automated systems.
What Problem Does This Paper Address?
Creating comprehensive thesauri, which group words with similar meanings, has traditionally been a labor-intensive process relying on human lexicographers' intuition and prior knowledge. This manual approach is time-consuming, expensive, and can lead to inconsistencies. The paper addresses the challenge of automating thesaurus production by exploring how to extract semantic relationships from dictionary entries, specifically focusing on the "genus term" (the superordinate term in a definition) and other formalized linguistic information present in machine-readable dictionaries. The authors highlight the limitations of existing automated approaches and propose a more refined method for leveraging dictionary data.
Methods
The paper contrasts two main approaches:
- Traditional/Intuitive Approach: Exemplified by the Longman Lexicon of Contemporary English (LOLEX), this method relies on lexicographers' intuition to select thesauric categories and assign lexical items.
- Amsler-type Procedure: This computational approach uses a computerized dictionary as its data base. It involves:
- Manual Selection and Disambiguation of Genus Terms: Identifying the first word in a definition that shares the same part of speech as the defined word and acts as its superordinate.
- Tree-Growing Algorithm: This algorithm builds filiations (hierarchies) between lexical entries and genus terms. It recursively finds items that use a given term as their genus term, creating a hierarchical structure.
- Leveraging Formalized Information: The authors emphasize the importance of using structured data within dictionaries, such as part-of-speech (POS) tags, grammatical fields (which describe syntactic environments), and semantic codes (inherent features for nouns, selectional restrictions for adjectives/verbs). They propose using these codes to disambiguate genus terms and even to infer thesauric links based on similarity.
Datasets
- Longman Dictionary of Contemporary English (LDOCE): The primary data source used by the authors. They specifically mention using the computer tape of LDOCE, which contains approximately 60,000 entries with formalized information.
- Merriam-Webster Pocket Dictionary (MPD): Used as the data base for Amsler's original approach.
Key Insights and Ideas
- Formalized Dictionary Data is Key: The rich, structured information within machine-readable dictionaries (like POS tags, grammatical fields, and semantic codes) is crucial for automated thesaurus production, offering more than just raw definitions.
- Genus Term as a Starting Point: The "genus term" in a definition is a valuable starting point for building hierarchical relationships, but its selection and disambiguation require careful handling.
- Controlled Defining Vocabularies Impact Hierarchy Depth: Dictionaries with controlled defining vocabularies (like LDOCE) tend to produce shallower thesauric hierarchies compared to those without (like MPD). This is because fewer words are available to serve as genus terms.
- Grammatical Codes for Disambiguation: Grammatical codes, which describe the syntactic environment of a word sense, can be used to disambiguate genus terms and infer thesauric links by matching codes between definiens and genus term.
- Beyond Genus Terms: While genus terms are important, analyzing the entire definition string is necessary to capture a broader range of thesauric relationships, especially when the genus term itself is a general pro-form.
Key Results
The paper doesn't present quantitative results in terms of accuracy metrics for thesaurus generation. Instead, it focuses on:
- Demonstrating the Amsler-type procedure: The authors illustrate how the tree-growing algorithm can construct hierarchical taxonomies using examples like "VEHICLE."
- Highlighting the impact of controlled vocabularies: They show how LDOCE's controlled vocabulary leads to shallower hierarchies compared to MPD, using the example of "LIMOUSINE."
- Proposing a method for disambiguating genus terms: They suggest using similarity in grammatical codes between a definition and its genus term to select the appropriate word sense for the genus term.
- Identifying limitations of genus-term-only analysis: They argue that analyzing the full definition string is essential for retrieving more nuanced thesauric links.
Limitations and Future Work
- Shallow Hierarchies: The use of a controlled defining vocabulary in LDOCE results in shallower thesauric hierarchies, which might be less revealing for organizational purposes.
- Genus Term Limitations: When genus terms become general pro-forms (like "SOMETHING"), they offer little specific information, necessitating the analysis of the entire definition.
- Need for Full Definition Analysis: The authors suggest that processing and disambiguating entire definition strings is necessary to retrieve information that a less controlled dictionary might embed in its genus terms.
- Future Work: The paper implies future work in fully exploiting the formalized information (grammatical codes, semantic codes) in LDOCE for high-power disambiguation and thesauric class setting. They also suggest that analyzing whole definition strings can reveal thesauric links beyond the simple hyponym-superordinate relation.
Summary updated