Can Subcategorization Help a Statistical Dependency Parser?

Daniel ZemanCOLING 2002

TL;DR

This paper investigates whether verb subcategorization information, which describes a verb's argument requirements, can improve statistical dependency parsing. The authors demonstrate that by incorporating this lexical knowledge directly into a dependency parser, they achieve significantly better parsing accuracy compared to models that don't leverage subcategorization.

What Problem Does This Paper Address?

Statistical dependency parsing aims to determine the grammatical relationships between words in a sentence, often represented as a tree. While many parsers rely on word order and morphological features, they can struggle with languages like Czech, which have flexible word order. Verb subcategorization, a linguistic phenomenon describing which arguments (like subjects and objects) a verb "expects," offers a rich source of lexical information that could disambiguate these relationships. However, there's been limited empirical evidence on how effectively this information can be integrated into statistical dependency parsers, especially for morphologically rich, free-word-order languages. This paper seeks to bridge that gap by exploring different methods of incorporating subcategorization into a Czech dependency parser.

Methods

The authors build upon an existing statistical dependency parser for Czech. Their core contribution lies in exploring three levels of integrating subcategorization information:

  • Lexicalized Dependency Model (Naïve): This approach treats word-to-word dependencies directly, rather than just tag-to-tag dependencies. While this implicitly captures some lexical information, the authors found it had minimal impact on accuracy, suggesting that simply using word forms isn't enough.
  • Selective Lexicalization: This method selectively enriches the parser's lexicon. Instead of just using a verb's tag (e.g., "VB"), they might use a more specific tag like "Vgive" if the verb is "give." Similarly, prepositions and conjunctions are also "lexicalized" by replacing their generic tags with their actual lemmas. This allows the parser to distinguish between different types of verbs and their arguments.
  • Preferring Subcategorized Dependencies: This is the most sophisticated approach. They use an automatically acquired subcategorization dictionary, which lists verb-frame pairs (e.g., "verb X takes a direct object in the accusative case"). When parsing, the parser prioritizes dependencies that match these subcategorization frames. If a verb has multiple frames, it considers the frequency of each frame. They also incorporate a constraint that arguments should not "skip" closer potential arguments of the same verb.

Datasets

The experiments were conducted using the Prague Dependency Treebank (PDT). This treebank was automatically derived and contains approximately 25,900 sentences for training and 2,600 for testing. The authors also mention using a subset of the PDT that excludes elliptical constructions, coordinations, and appositions for some comparisons, to align with previous work.

Key Insights and Ideas

  • Subcategorization is Crucial for Free-Word-Order Languages: The paper strongly suggests that verb subcategorization is particularly valuable for languages like Czech, where word order is flexible and morphological case plays a key role in identifying arguments.
  • Direct Integration Outperforms Post-Ranking: Unlike some previous work that used subcategorization to re-rank parse trees after parsing, this paper shows that integrating subcategorization during the parsing process yields better results.
  • "Smart" Lexicalization is Key: Simply using word forms (naïve lexicalization) isn't sufficient. Selective lexicalization and prioritizing subcategorized dependencies are much more effective.
  • Frequency Matters: The frequency of subcategorization frames, as learned from corpora, is a vital component for the parser to make informed decisions.

Key Results

The paper presents two main tables summarizing the results:

  • The first table shows the impact of the "naïve" lexicalized model with different lambda values (a parameter controlling the influence of lexicalization). The accuracy remained largely unchanged (around 77-80%) across different subsets of the data, indicating minimal benefit from this approach.
  • The second table compares the baseline parser with the subcategorized parser. The subcategorized parser achieved significantly higher accuracy:
  • Baseline: 77.7% (all sentences), 78.7% (elliptical excluded), 79.9% (coordination/apposition excluded).
  • Subcategorized: 79.3% (all sentences), 77.7% (elliptical excluded), 82.1% (coordination/apposition excluded).

This shows a notable improvement, especially when dealing with more complex sentence structures.

Limitations and Future Work

  • Handling Coordinations and Appositions: The authors acknowledge that their current model is not well-adjusted to handle coordinations (e.g., "apples and oranges") and appositions (e.g., "my friend, John"). They suggest enriching the model to address these structures.
  • "Wild" Sentences: While they don't call non-projective dependencies "errors," they note that such constructions are rare in Czech and their parser's performance is still far from perfect in handling them.
  • Future Work: The authors explicitly mention investigating coordinated arguments and adjuncts and enriching the model to handle coordinations and appositions as future directions.

Summary updated

BibTeX
@inproceedings{zeman2002can,
  author = {Daniel Zeman},
  title = {{Can Subcategorization Help a Statistical Dependency Parser?}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure