TL;DR
This paper introduces a novel "Lexical Knowledge Base (LKB) approach" to extract rich semantic information from the free-text definitions within machine-readable dictionaries. By treating dictionaries as a source of general knowledge and developing procedural techniques to analyze and structure definitions, the authors aim to build more comprehensive and semantically aware lexical resources.
What Problem Does This Paper Address?
Traditionally, computational linguistics has sometimes overlooked the importance of the lexicon, focusing more on core syntactic issues. However, there's a growing recognition of the lexicon's crucial role, leading to efforts to leverage existing machine-readable dictionaries (MRDs) for building extensive NLP-oriented lexicons. While MRDs contain a wealth of implicit semantic information, extracting it effectively has been a challenge. Previous approaches, like lexical databases (LDBs), require explicit structuring of all information upfront. This paper addresses the gap by proposing a method to procedurally acquire and organize semantic information that is implicitly embedded within the natural language definitions of dictionaries, moving beyond simple hierarchical classifications.
Methods
The core of the proposed approach is a "discovery procedure technique" that operates on natural language definitions. This process is designed to be recursive and refined.
- Morphological Analysis: The first step involves applying a morphological analyzer to the free-text definitions to break them down into their constituent word forms and identify their grammatical properties.
- Homograph Disambiguation: A disambiguation program is then used to resolve ambiguities arising from homographs, employing both general Italian rules and ad-hoc rules tailored for lexicographic definitions.
- Pattern-Matching Discovery Procedures: The system then employs a set of "discovery procedures" that act on the analyzed definitions. These procedures search for recurring patterns, co-occurrences of words, and syntactic structures.
- Labeling and Categorization: A "basic vocabulary" of labels is established to represent broad conceptual categories (e.g., 'INSTRUMENT', 'ANIMAL') and relations (e.g., 'STUDY', 'USED'). These labels help group semantically similar terms and identify relationships within definitions.
- Incremental Learning: The system is designed to be incremental. It starts with a generalized expertise and gradually learns and refines its capabilities by analyzing failures and incorporating new patterns and rules. This iterative process allows it to cope with the variations in how semantic information is expressed in natural language definitions.
- Formalized Representation: The acquired knowledge is represented using a formalism where each element is a Function characterized by a Type (e.g., Hypernym, Relation, Qualifier) and Arguments. This allows for a structured representation of word senses and their relationships.
Datasets
- Italian Machine Dictionary (DMI): This dictionary is already structured as a Lexical Database (LDB) and is primarily based on the Zingarelli Italian dictionary (1970). It contains approximately 1 million word forms from 120,000 lemmas and includes information on morphology, POS, usage codes, and specialized terminology. Semantic access paths for synonyms, hyponyms, and hypernyms are implemented.
- Garzanti Italian Dictionary (1984): This dictionary is being restructured using the described approach. The paper details the parsing of its photocomposition codes to identify the rough structure of lexical entries.
Key Insights and Ideas
- Dictionaries as General Knowledge Sources: The paper champions the idea that dictionaries are not just repositories of linguistic information but also valuable sources of basic, general world knowledge.
- Procedural Acquisition: The key innovation is the shift from static, explicitly defined lexical databases to a procedural approach that learns semantic information from unstructured text definitions.
- Leveraging Natural Language Structure: The authors recognize that lexicographic definitions, while natural language, are often constrained and somewhat regular, making them amenable to computational analysis.
- Incremental Refinement: The system's ability to learn and refine its patterns iteratively is crucial for handling the complexity and variability of natural language definitions.
- "Minimal Pairs" for Semantic Distinction: The idea of constructing "minimal pairs" of sense-definitions that differ by only one conceptual or relational feature is a powerful method for isolating the semantic distinctions between words.
Key Results
The paper focuses on the methodology and approach rather than presenting extensive quantitative results. However, it demonstrates the feasibility of the approach through:
- Successful Parsing and Disambiguation: Figures 4 and 5 show the output of the morphological analyzer and the disambiguation procedure on a sample definition, illustrating the initial processing steps.
- Pattern Identification: Figure 6 provides examples of queries to the Textual Data Base (DBT) that help identify patterns linking conceptual categories (like SCIENCE, BRANCA, SPECIALITA', RAMO, DISCIPLINA) with relations (like STUDIA, OCCUPA).
- Formalized Representation Examples: The paper provides concrete examples (page 4) of how dictionary definitions for "Batterio" and "Batteriologia" are transformed into a formal, structured representation using their proposed formalism.
Limitations and Future Work
- Initial Simplification: The authors acknowledge that the methodological approach was deliberately simplified at the beginning to introduce problems gradually.
- Need for Refinement: The system's initial runs are expected to break down at many points, necessitating evaluation and the addition of new capabilities and knowledge incrementally.
- Interactive Phase: An interactive phase is foreseen where definitions might be modified for normalization to acceptable analysis structures, suggesting a human-in-the-loop component.
- Future Work: The authors aim to develop a system capable of reorganizing free text into elaborately structured information, leading to a "Lexical Knowledge Base (LKB) approach." They also envision building hierarchies not only for hypernyms but also for more complex conceptual structures.
Summary updated