TL;DR
This paper introduces a novel generalization of Categorial Grammar (CG) that uses recursive constraints to define lexical categories. This approach elegantly handles complex linguistic phenomena like Dutch cross-serial dependencies and adjunct scope, offering a more computationally tractable and linguistically powerful framework than previous methods.
What Problem Does This Paper Address?
Traditional Categorial Grammar, while powerful, has struggled to capture certain linguistic phenomena without resorting to complex, often non-recursive, feature structures or cumbersome lexical rules. Specifically, phenomena requiring recursive definitions of grammatical functions (like subject/object roles) or flexible handling of optional and iterative elements (like adjuncts) have been challenging. Previous attempts to introduce polymorphism in CG were limited by the non-recursive nature of feature descriptions. This paper aims to bridge this gap by proposing a system where lexical categories are defined by recursive constraints, allowing for a more "lexicalist" grammar where the lexicon carries significant grammatical information, simplifying the syntactic component.
Methods
The core innovation is the introduction of recursive constraints to define lexical categories in Categorial Grammar. Instead of relying solely on simple feature structures, the authors propose using relational constraints that can refer to themselves. This allows for the definition of a richer set of polymorphic categories and the direct modeling of recursive lexical rules.
Key techniques include:
- Relational Constraints: These allow for more expressive definitions of categories, going beyond simple feature matching.
- Recursive Constraints: The central idea, enabling categories to be defined in terms of themselves, mirroring recursive linguistic processes.
- Delayed Evaluation: To handle the potential non-termination introduced by recursive constraints, a processing strategy is employed where constraints are only evaluated when their arguments are sufficiently instantiated. This is implemented using a "block" facility in Prolog.
- Lexical Rules as Constraints: The paper demonstrates how previously proposed lexical rules (e.g., for passive or adjunct modification) can be elegantly reformulated as recursive constraints.
Datasets
Not discussed in the paper. The paper focuses on linguistic analysis and theoretical framework development, using example sentences from Dutch.
Key Insights and Ideas
- Recursive Constraints for Linguistic Power: The paper argues that recursive constraints are essential for capturing the full expressive power needed for certain grammatical phenomena, particularly those involving recursive definitions or optional, iterative elements.
- "Lexicalist" Grammars: By moving complexity to the lexicon via recursive constraints, the syntactic component can be simplified to just application, aligning with a "lexicalist" approach to grammar.
- Unified Treatment of Adjuncts: Adjuncts are re-analyzed as arguments of the verb, rather than functors, which, when combined with recursive constraints, provides a more flexible and accurate account of their position and scope, especially in complex verb clusters.
- Computational Tractability via Delayed Evaluation: The non-termination problem arising from recursive constraints is effectively managed by delaying their evaluation until sufficient information is available, making the system computationally viable.
- Cross-Serial Dependencies Explained: The framework offers a principled account of Dutch cross-serial dependencies by implementing a "generalized disharmonic division" as a recursive constraint.
Key Results
The paper demonstrates the linguistic merits of its approach through analyses of:
- Dutch Cross-Serial Dependencies: The proposed framework successfully accounts for the obligatory ordering of verbs and their arguments in Dutch verb clusters, a phenomenon that has been challenging for previous CG models.
- Adjunct Position and Scope: The analysis of adjuncts as arguments, coupled with recursive constraints, provides a unified account for their distribution and scope, including ambiguous cases, which is a significant improvement over functor-based approaches.
- Subject-Verb Agreement: The paper shows how agreement can be naturally incorporated by reducing it to a form of subcategorization, managed by recursive constraints.
While no quantitative results (e.g., parsing speed, accuracy on a corpus) are presented, the paper provides detailed linguistic analyses and derivations to support its claims.
Limitations and Future Work
- Processing Complexity: While delayed evaluation addresses non-termination, the paper acknowledges that efficient parsing with complex constraints remains a challenge. Chart-based parsing strategies are discussed as potential solutions, but practical implementation with highly complex constraints is noted as difficult.
- Semantic Integration: The paper briefly touches upon semantics, showing how it can be integrated with the recursive constraints for adjuncts. However, a full semantic treatment for all aspects of the grammar is not elaborated.
- "Hand-Fabricated" Context-Free Grammar: For practical parsing, the authors found success using a hand-crafted context-free grammar as a first phase. The relationship between this CFG and the full CUG grammar, and ensuring the CFG produces a superset of derivations, is noted as a difficult problem.
- Future Work: The authors suggest that their fragment is part of a larger system for Dutch verb clusters and that additional recursive constraints can be introduced. The general approach of locating grammatical complexity in the lexicon via recursive constraints is presented as a promising direction for future research in grammar development.
Summary updated