A Block-Based Robust Dependency Parser for Unrestricted Chinese Text

Ming ZhouACL 2000

TL;DR

This paper introduces a novel "block-based dependency parsing" strategy for Chinese, integrating phrase structure and dependency parsing to overcome the limitations of traditional methods in handling unrestricted text. This hybrid approach achieves high robustness and performance, enabling practical application in machine translation.

What Problem Does This Paper Address?

Parsing Chinese text presents significant challenges compared to languages like English due to its unique grammatical structures and lack of explicit morphological markers. Traditional parsers, often relying on extensive rule bases and lexicons, struggle with the "unrestricted" nature of real-world text, which includes ungrammatical phenomena, complex sentence structures, and a vast vocabulary. Existing approaches often fall short in robustness, making them impractical for applications like machine translation. The paper aims to bridge this gap by developing a parser that can effectively handle the complexities and variations of Chinese text.

Methods

The core innovation is the "block-based dependency parsing" strategy. This approach first identifies "blocks" – fundamental information units within a sentence – using a partial parsing technique. This partial parsing focuses on identifying local structures and "meta-blocks" (like frame structures) rather than attempting to derive a full sentence structure from the root. This partial parsing is designed to be robust, allowing the parser to skip over ungrammatical portions.

Following block identification, a dependency parser is applied to build relationships between these blocks. This strategy leverages existing knowledge of phrase structure parsing for block identification and then applies dependency parsing to establish relationships, effectively combining the strengths of both paradigms. The system also incorporates a sophisticated part-of-speech tagging mechanism that goes beyond simple word categories to capture real grammatical features, addressing a key challenge in Chinese linguistics.

Datasets

The paper evaluates the system on two corpora:

  • "Primary school textbook of Singapore": A corpus of 1842 modern Chinese sentences, designed to cover various sentence types and morphological phenomena.
  • People's Daily (1998, 1999, 2000): A collection of real-world news articles, notable for containing unknown words, long and complex sentences, and ellipsis.

Key Insights and Ideas

  • Block-based parsing as a robustness enhancer: By breaking down sentences into blocks and parsing these units, the system can more effectively handle ungrammatical or complex segments without the entire parse failing.
  • Hybrid approach is key: Combining phrase structure parsing for block identification with dependency parsing for inter-block relations offers a powerful synergy, leveraging the strengths of both.
  • Part-of-speech tagging needs to be context-aware: The paper highlights that simply assigning a single part-of-speech tag to a word in Chinese is insufficient. Capturing the actual grammatical function within a context is crucial for accurate parsing.
  • Leveraging existing phrase structure knowledge: The authors recognize the wealth of phrase structure parsing knowledge for Chinese and propose to re-use it to compensate for the scarcity of word-based dependency knowledge.

Key Results

The system demonstrated good performance and high robustness on both test corpora.

  • On the "Primary school textbook of Singapore," it achieved an average precision of 90.4% with an average sentence length of 7.34 words.
  • On the more challenging People's Daily corpus, it achieved 67.7% precision with an average sentence length of 14.52 words. The paper notes that the parser has been incorporated into a successful commercial Chinese-Japanese machine translation product.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Part-of-speech identification precision: Currently relies on manually compiled linguistic rules, which limits precision. Future work will explore statistical or hybrid approaches.
  • Ambiguous word segmentation: The paper identifies cases where words can be segmented in multiple ways with different meanings (e.g., "将来").
  • Compound noun identification: The system struggles with exhaustively enumerating all compound nouns, leading to inevitable errors.
  • Proper noun and predicate identification: These areas require further improvement.
  • Syntactical ambiguity: The paper points out ambiguities, particularly with "V+A+N" patterns, and suggests exploring different reduction methods.
  • Handling more predicate candidates: The current parser can handle up to 7 predicate candidates; sentences with more will be partitioned. Future work may involve improving this capability.

Summary updated

BibTeX
@inproceedings{zhou2000block,
  author = {Ming Zhou},
  title = {{A Block-Based Robust Dependency Parser for Unrestricted Chinese Text}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2000},
}
Figure