Analyzing Dependencies of Japanese Subordinate Clauses based on Statistics of Scope Embedding Preference

Takehito Utsuro, Shigeyuki Nishiokayama, Masakazu Fujio, Yuji MatsumotoNAACL 2000

TL;DR

This paper tackles the ambiguity in analyzing Japanese sentence structure, specifically how subordinate clauses relate to each other. They propose a novel statistical method that leverages the "scope embedding preference" of these clauses to improve dependency analysis accuracy, outperforming previous rule-based and statistical approaches.

What Problem Does This Paper Address?

Japanese, with its flexible word order, presents significant challenges for syntactic dependency analysis. A major source of ambiguity arises from subordinate clauses, where determining which clause modifies which is difficult. This ambiguity cascades into other syntactic ambiguities, making accurate parsing challenging. Existing rule-based methods for resolving these ambiguities have limited coverage, and purely statistical methods haven't fully exploited the linguistic insight of how subordinate clauses naturally embed within each other based on their scope.

Methods

The authors propose a corpus-based statistical approach to learn the dependency preference of Japanese subordinate clauses. They frame this as a classification problem, encoding various linguistic features of subordinate clauses to predict their scope embedding preference. They then use this preference to disambiguate dependencies. Specifically, they employ the decision list learning method (Yarowsky, 1994). This involves:

  1. Feature Engineering: Extracting features from subordinate clauses, including punctuation, grammatical information (part-of-speech, conjugation), and lexicalized forms.
  2. Evidence-Decision Pairs: Transforming sentence structures into pairs of evidence (feature sets of two subordinate clauses) and decisions (whether the first modifies the second or "beyond").
  3. Decision List Construction: Learning a sorted list of rules based on the statistical correlation between evidence and decisions, prioritizing rules with higher confidence.

This statistical learning approach aims to overcome the coverage limitations of previous rule-based methods.

Datasets

  • EDR bracketed corpus: Used for both training and evaluation. The corpus contains 210,000 Japanese sentences.
  • 95% of the EDR corpus was used for training (199,500 sentences), and 5% for testing.
  • From the training data, 162,443 pairs of subordinate clauses were extracted.

Key Insights and Ideas

  • Scope Embedding Preference is Key: The core insight is that the inherent "scope" of a subordinate clause (how much of the sentence it "covers") dictates how it can be embedded within or modify other clauses. Clauses with broader scopes tend to contain or be modified by clauses with narrower scopes.
  • Statistical Learning for Linguistic Rules: The paper successfully translates a linguistic theory (Minami's scope embedding preference) into a statistical learning problem, enabling broader coverage than manual rule creation.
  • Decision Lists for Disambiguation: The decision list learning method proves effective in capturing complex feature interactions for dependency disambiguation.
  • Improved Dependency Analysis: By accurately predicting subordinate clause dependencies, the overall accuracy of a statistical dependency analyzer is significantly improved.

Key Results

  • The proposed model ("Our Model") outperforms other models (a decision tree model and a simpler decision list model) in terms of coverage and precision for deciding dependencies between two subordinate clauses.
  • Specifically, the model achieves 0.8-1.8% improvement in chunk-level precision and 1.6-4.7% improvement in sentence-level precision when integrated into an existing statistical dependency analyzer (Fujio and Matsumoto, 1998).
  • The baseline precision for deciding dependency between two subordinate clauses is 53.78%.

Limitations and Future Work

  • Coverage vs. Precision Trade-off: The paper presents plots showing a trade-off between coverage and precision, indicating that higher precision is achieved at the cost of covering fewer dependency pairs.
  • Focus on Specific Dependency: The modeling ignores cases where the head VP chunk of Clause₁ modifies a subordinate clause that precedes Clause₂, assuming this is less related to scope embedding preference. This is a deliberate simplification.
  • Future Work: The paper mentions an upcoming publication detailing experimental evaluation (Utsuro, 2000), suggesting further analysis and refinement of the proposed method. The authors also implicitly suggest that further exploration of feature sets and learning algorithms could yield even better results.

Summary updated

BibTeX
@inproceedings{utsuro2000analyzing,
  author = {Takehito Utsuro and Shigeyuki Nishiokayama and Masakazu Fujio and Yuji Matsumoto},
  title = {{Analyzing Dependencies of Japanese Subordinate Clauses based on Statistics of Scope Embedding Preference}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2000},
}
Figure