TL;DR
This paper introduces a novel "generalized categorial grammar" that aims to improve the recovery of unbounded dependencies in natural language processing. By combining the transparent predicate-argument structure of categorial grammars with the flexible category system of HPSG, the authors achieve state-of-the-art performance on an unbounded dependency recovery task.
What Problem Does This Paper Address?
Accurately identifying predicate-argument dependencies is crucial for tasks like information extraction and question answering. A significant challenge in this area is the recovery of "unbounded dependencies," which are relationships between words that are separated by other words or phrases. Traditional parsing methods, often based on PCFGs and trained on resources like the Penn Treebank, struggle with these long-distance dependencies. While existing categorial grammars (like CCG) are good at representing these dependencies, they often require a large number of categories, leading to data sparsity issues. This paper seeks to bridge this gap by developing a categorial grammar that is both effective at recovering unbounded dependencies and more learnable from data.
Methods
The core of this paper lies in two main contributions:
- Generalized Categorial Grammar: The authors propose a categorial grammar that generalizes categories to limit their number while maintaining grammatical constraints. This is achieved by allowing arbitrary type-constructing operators and grouping syntactically interchangeable signs into equivalent categories. This approach aims to strike a balance between the small, uniform composition operations of traditional categorial grammars and the richer, more flexible category system found in HPSG.
- Reannotation System: To train their generalized categorial grammar, the authors developed a novel reannotation system. This system transforms existing resources, like the Penn Treebank, which are based on Government and Binding Theory, into the new categorial representation. This process involves a single top-down pass that extracts arguments and modifiers, making the trace and function labels in the Treebank accessible for training.
Datasets
The primary dataset used is the Penn Treebank. The paper describes a process of reannotating sections 02-21 of the WSJ portion of the Penn Treebank into their generalized categorial grammar format. For evaluation, the authors utilize development and test sets from Rimell et al. (2009) and Nivre et al. (2010), which are specifically designed for unbounded dependency recovery tasks.
Key Insights and Ideas
- Bridging the Gap: The paper successfully demonstrates that a generalized categorial grammar can achieve performance comparable to state-of-the-art systems on unbounded dependency recovery, suggesting that this approach offers a promising alternative to existing methods.
- Reannotation as a Key Enabler: The development of a flexible reannotation system is crucial. It allows leveraging existing, widely-used treebanks for training novel grammar formalisms, overcoming the bottleneck of creating new annotated corpora from scratch.
- Category Generalization for Learnability: By generalizing categories and limiting their number, the proposed grammar aims to mitigate data sparsity issues often encountered with highly specific category sets, leading to more robust learning.
- Transparent Dependencies: The generalized categorial grammar retains the desirable property of transparent predicate-argument dependencies, which is beneficial for downstream interpretation tasks.
Key Results
The proposed system achieved unbounded dependency parsing accuracy favorably comparable to all the systems recently studied by Rimell et al. (2009) and Nivre et al. (2010). While specific percentage scores are presented in Table 1, the key takeaway is that their system performs competitively with established parsers on a challenging task. The reannotation process achieved approximately 94% coverage of the training set.
Limitations and Future Work
- Limited Scope of Generalization: The paper notes that the conversion to the generalized categorial grammar replaces some Treebank category labels with more general specifications (e.g., "A-aN" for adjective phrases). However, it also acknowledges that certain contexts, like those involving "become" or "start," cannot be reliably identified by the current reannotation.
- Dependency Label Simplification: For evaluation, dependencies were mapped to numerical relations (e.g., '1' for 'nsubj' and 'nsubjpass'). While this simplification is noted as being losslessly reversible, it does represent a departure from more detailed dependency labels.
- Future Work: The authors suggest that the modularity of their reannotation rules allows for experimentation with different syntactic analyses, implying future work could explore variations in grammar formalisms or reannotation strategies. The availability of the system and rules online suggests a path for further research and development by the community.
Summary updated