TL;DR
This paper introduces COMPLEX, a "meta-lexicon" designed to store shared lexical information for Natural Language Processing (NLP) systems, aiming to reduce redundancy and improve coverage. The authors highlight the challenges of polysemy and mapping between different dictionary sources, proposing a unified approach to build a broad-coverage, theory-neutral lexicon.
What Problem Does This Paper Address?
Natural Language Processing (NLP) systems rely heavily on computational lexicons, which are essentially dictionaries for computers. However, each NLP system has unique needs, leading to different types and amounts of information being stored in their individual lexicons. This results in significant redundancy, as many systems require similar basic lexical information (e.g., part of speech, subcategorization for verbs, gender for nouns). Furthermore, existing lexicons often suffer from limited coverage, as they are frequently hand-built and tailored to specific systems. The paper addresses the need for a centralized repository of shared lexical information that can be accessed and adapted by various NLP systems, thereby promoting efficiency and broader coverage. A major challenge identified is the "polysemy problem" – how to represent words with multiple meanings – and the "mapping problem" – how to reconcile differing sense distinctions across various machine-readable dictionaries.
Methods
The authors propose building COMPLEX, a "computational meta-lexicon," by drawing primarily on explicit and implicit information from machine-readable dictionaries (MRDs). The core idea is to create a broad-coverage lexicon by extracting and unifying shared lexical data. They aim for a neutral representation and uniform data structure to ensure compatibility with a range of NLP system requirements.
Key methodological aspects include:
- Leveraging Existing MRDs: The primary sources for COMPLEX are MRDs, with plans to expand to text corpora. They specifically mention using enhanced UDICT (LexSys), the Brandeis Verb Lexicon, and LDOCE (Longman Dictionary of Contemporary English).
- Addressing Polysemy: The paper acknowledges the difficulty of representing multiple word senses. While some systems (like Boguraev's) attempt to make sense distinctions, this often forces adherence to the sense distinctions of a particular dictionary (e.g., LDOCE). COMPLEX aims to represent information in a way that can accommodate different senses, though the exact mechanism for this is still under development.
- The Mapping Problem: Recognizing that sense distinctions vary across dictionaries, the paper identifies the need for methods to map senses between different sources. This is crucial for integrating information from multiple MRDs.
- Structured Data Representation: The LexSys data structure, for instance, organizes information into a feature hierarchy with categories like MORPHological, PHONological, SYNTACTIC, SEMANTIC, STYLISTIC, and GRAPHEMIC. This structured approach allows for efficient querying and avoids the unstructured list format of older systems like UDICT.
- Evaluation of Semantic Codes: A significant portion of the paper details an evaluation of LDOCE's "Box Codes" for selectional restrictions on verbs. This involved constructing tests to determine if verbs truly required human subjects in all senses, revealing limitations in the LDOCE data.
Datasets
The paper primarily discusses the use of existing machine-readable dictionaries as datasets:
- UDICT (enhanced LexSys): An existing computational lexicon used as a starting point.
- Brandeis Verb Lexicon: A lexicon providing subcategorization information for verbs.
- LDOCE (Longman Dictionary of Contemporary English, 1978): A key source for definitions, grammatical codes, and semantic selectional restrictions.
- Webster's Seventh New Collegiate Dictionary (Merriam, 1963): Planned for use, along with taxonomy files derived from it.
- Collins Thesaurus (1984): Planned for use.
- Collins bilingual dictionaries: Planned for use for English/Italian, English/French, English/German, and English/Spanish.
- Text Corpora: Planned for future use to confirm intuitions and extract implicit information.
No new benchmarks or datasets were introduced in this paper.
Key Insights and Ideas
- The "Meta-Lexicon" Concept: The idea of a central repository for shared lexical information is a powerful concept for reducing redundancy and promoting interoperability among NLP systems.
- Polysemy is a Fundamental Hurdle: The paper strongly emphasizes that any useful computational lexicon must grapple with word sense disambiguation. Simply listing features that apply to all senses or any sense is insufficient; a more nuanced approach is needed.
- MRDs are Rich but Flawed: Machine-readable dictionaries are invaluable resources, but their data (like LDOCE's Box Codes) is not always precise, complete, or consistent, requiring careful evaluation and potential augmentation.
- Structured Hierarchies Improve Usability: Organizing lexical data into structured hierarchies (e.g., by feature type) makes it more efficient to query and manage compared to flat lists.
- Hypothesis: Genus Terms Reflect Properties: The study on LDOCE's selectional restrictions suggests that genus terms in dictionary definitions (e.g., in a taxonomy) can reflect properties of their headwords, hinting at potential inheritance of semantic features.
Key Results
- LDOCE Selectional Restriction Evaluation: The study on LDOCE's "human subject" selectional restrictions found that only 59% of the candidate verbs truly required human subjects in all senses. This highlights the need for careful re-evaluation of such data.
- Filtering and Morphological Methods:
- Filtering using Merriam's taxonomy yielded 312 candidate verbs, with nearly half proving acceptable in the tests.
- The morphological method (using "-er" suffixes) was less successful, with only 20% of candidate verbs passing the tests.
- Overall: The paper demonstrates that while MRDs are essential, their data requires significant processing, evaluation, and integration with other sources to build a robust computational lexicon.
Limitations and Future Work
Limitations:
- Polysemy Representation: The paper acknowledges that a definitive solution to representing multiple word senses is still an open problem.
- Mapping Problem: Developing practical methods for mapping senses between different dictionaries remains a significant challenge.
- LDOCE Data Quality: The evaluation revealed limitations in the accuracy and completeness of LDOCE's selectional restriction codes.
- Preliminary Results: The authors state that their results are still too preliminary to be fully reported, particularly regarding the proposed approach to data structure and integration.
Future Work:
- Extracting Data from Text Corpora: To confirm intuitions and extract implicit information.
- Using Example Sentences: To verify hypotheses about lexical features and identify contradictions in existing dictionary data.
- Developing Sense Division Mechanisms: Creating methods to automatically divide senses based on semantic or syntactic cues.
- Exploring Multiple Dictionaries: Developing strategies to use several dictionaries for maximum coverage.
- Practical Solution to Mapping: Continuing work on a practical solution for the mapping problem.
- Expanding Data Structures: Incorporating more information and refining data structures as the lexicon grows.
- Building a Rich Computational Lexicon: The ultimate goal is to create a comprehensive lexicon usable by NLP systems, actively addressing both the polysemy and mapping problems.
Summary updated