TL;DR
This paper tackles the challenge of improving HPSG supertagging accuracy by incorporating long-distance dependency information, which is often missed by traditional sequential models. By integrating dependency relations into a point-wise averaged perceptron model, they achieve a significant boost in supertagging accuracy, leading to improved performance in the HPSG parser.
What Problem Does This Paper Address?
Supertagging is a crucial step in speeding up deep parsing for lexicalized grammars like HPSG. It assigns a "supertag" to each word, which is essentially a lexical template encoding its syntactic behavior. While sequence labeling models are commonly used for supertagging, they struggle to capture long-distance dependencies between words. This limitation hinders further improvements in accuracy, especially when dealing with the large number of fine-grained HPSG supertags. The authors aim to bridge this gap by leveraging dependency parsing, which excels at modeling these long-range relationships, to enhance supertagging performance.
Methods
The core of this paper is the integration of dependency information into a supertagging model. The authors first model dependencies using the projective MSTParser formalism, chosen for its accuracy in capturing long dependencies. They then extract dependency relations from HPSG treebanks, treating non-head daughters as modifiers of head-daughters.
The key innovation lies in how this dependency information is integrated into the supertagging process. Instead of relying solely on local context (surrounding words and POS tags), they augment a point-wise averaged perceptron (PW-AP) model with dependency-based features. Specifically, for the word being predicted, they extract its "modifiers" (MODIN) and its "head" (MODOUT) based on dependency arcs. These MODIN and MODOUT relations are then combined as features with the existing local context features. This creates a "dependency-informed" model (PW-DEP).
Datasets
The experiments were conducted on the WSJ-HPSG treebank. Sections 02-21 were used for training the dependency parser, the dependency-informed supertagger, and the HPSG parser. Section 23 served as the testing set.
Key Insights and Ideas
- Long-distance dependencies are crucial for supertagging: The significant accuracy gains observed when incorporating dependency information highlight that these long-range relationships are not just a theoretical concern but a practical bottleneck for current supertagging methods.
- Dependency parsing as an interface: The paper demonstrates that dependency parsing can serve as a valuable interface to inject richer syntactic information into sequence labeling tasks like supertagging, which are traditionally limited by local context.
- Simple integration yields strong results: The authors show that a straightforward integration of dependency features into a standard point-wise model can lead to substantial improvements, suggesting that complex architectural changes might not always be necessary.
- Improved supertagging benefits downstream parsing: The evaluation in the HPSG parser demonstrates that a more accurate supertagger directly translates to better parsing performance, particularly when the number of supertag candidates is limited.
Key Results
- Gold standard dependencies: With gold standard first-order dependencies, the dependency-informed supertagger (PW-DEP) achieved an accuracy of 97.43%, a significant improvement over the baseline PW-AP model's 92.48%. This represents an absolute gain of 4.95%.
- Automatically predicted dependencies: When using automatically predicted dependencies, PW-DEP still outperformed the baseline PW-AP, achieving 92.18% accuracy compared to 91.14%.
- HPSG Parser Performance: In the HPSG parser evaluation, using the dependency-informed supertagger resulted in a 5.06% higher F-score when only one supertag candidate was provided per word. This gain narrowed as more candidates were allowed, indicating the supertagger's effectiveness in reducing ambiguity early on.
- Comparison to prior work: The dependency-informed model showed competitive results compared to previous state-of-the-art methods, particularly when compared to BPM plus CFG-filter, with statistically insignificant differences in accuracy.
Limitations and Future Work
- Parsing Speed: The authors acknowledge that while supertagging accuracy improved, the overall parsing time was not optimized. The additional cost of MSTParser currently outweighs the parsing time saved by the improved supertagger. Future work could focus on optimizing the dependency parser for speed or reusing acquired dependencies for more effective pruning.
- Dependency Modeling: The paper suggests that further improvements in supertagging accuracy could be achieved with better dependency modeling, potentially through semi-supervised dependency parsers.
- Extensibility: The authors believe their approach is not limited to HPSG and can be extended to other lexicalized grammar supertaggers.
- Comparison with other dependency integration methods: The paper mentions that incorporating dependency constraints at the schema application stage of HPSG parsing has been explored. A direct comparison between their method and this approach is left as future work.
Summary updated