A Statistical Parser for Czech

Michael Collins, Jan Hajic, Lance A. Ramshaw, Christoph TillmannACL 1999 · OpenReview

Abstract

This paper considers statistical parsing of Czech, which differs radically from English in at least two respects: (1) it is a highly inflected language, and (2) it has relatively free word order. These differences are likely to pose new problems for techniques that have been developed on English. We describe our experience in building on the parsing model of (Collins 97). Our final results—80% dependency accuracy—represent good progress towards the 91% accuracy of the parser on English (Wall Street Journal) text.

Introduction

In the late 1990s, the field of statistical parsing was largely dominated by models optimized for English, such as those developed using the Penn Treebank. However, English is morphologically simple with a relatively rigid word order. This study explores the portability of these models to Czech, a language characterized by high inflection (HI) and free word order (FWO).

Czech presents two primary challenges:

  1. Data Sparsity: A single word can take thousands of forms, making it rare for a parser to encounter the same exact form twice.
  2. Structural Flexibility: Subject-verb-object triples can appear in any of the six possible permutations, making traditional fixed-position phrase-structure grammars less effective.

The Lexicalized Parsing Framework

The researchers utilized Michael Collins’ Model 1 as their foundation. This model relies on lexicalized Probabilistic Context-Free Grammars (PCFGs), where each non-terminal in the parse tree is associated with a headword and its part-of-speech (POS) tag. To apply this to Czech, the authors had to convert the native dependency format of the Prague Dependency Treebank (PDT) into lexicalized trees.

Lexicalized parse tree structure showing headword propagation and associated grammar rules.

Adapting to Czech: Tree Re-engineering

A critical step in the adaptation was "Tree Re-engineering." The authors moved away from flat baseline trees toward more linguistically motivated structures. They experimented with different ways to represent dependencies as constituents, comparing flat, right-branching, and binary structures to determine which best captured the relationships in a free-word-order context.

Comparison of (a) flat, (b) right-branching, and (c) binary tree structures used to convert dependency data into phrase-structure trees.

Key technical modifications included:

  • Bigram Dependencies: Shifting to a first-order Markov model where the probability of a modifier depends on both the head and the previously generated modifier.
  • POS Tagset Optimization: To combat sparsity, the authors reduced the 3,000+ possible Czech morphological tags to a 58-tag set focusing on "Main POS" and "Grammatical Case."

Key Insights and Results

The study revealed that more data is not always better; adding gender and number features to POS tags actually degraded performance by thinning the data too much. Interestingly, the parser performed better when trained on machine-generated tags rather than perfect human-annotated ones, as this better reflected the "real-world" errors it would encounter during testing.

Performance Summary

  • Baseline Accuracy: 72.3%
  • Final Accuracy: 80.0%
  • Relative Error Reduction: 26%

While the 80% accuracy is a significant milestone for a free-word-order language, it still trails the 91% accuracy achieved on English Wall Street Journal text. The parser performed best on news and business text (81.4%) but struggled with the longer, more complex sentences found in scientific articles (76.0%).

Conclusion and Future Work

The adaptation of the Collins model proves that lexicalized statistical parsers can be successfully tuned for highly inflected languages. Future improvements are expected to come from more sophisticated morphological treatments, such as using word lemmas to separate syntactic dependencies from semantic ones and employing automatic clustering to find the most informative POS groupings.

Summary updated

Abstract

This paper considers statistical parsing of Czech, which differs radically from English in at least two respects: (1) it is a highly inflected language, and (2) it has relatively free word order . These differences are likely to pose new problems for techniques that have been developed on English. We describe our experience in building on the parsing model of (Collins 97). Our final results- 80% dependency accuracy - represent good progress towards the 91% accuracy of the parser on English (Wall Street Journal) text.
BibTeX
@inproceedings{collins1999statistical,
  author = {Michael Collins and Jan Hajic and Lance A. Ramshaw and Christoph Tillmann},
  title = {{A Statistical Parser for Czech}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1999},
  url = {https://openreview.net/forum?id=rJbNoseuWB},
}
Figure