Statistics based hybrid approach to Chinese base phrase identification

Tiejun Zhao, Muyun Yang, Fang Liu, Jianmin Yao, Hao Yu2000 · DOI

TL;DR

This paper introduces a novel hybrid statistical approach for identifying seven types of Chinese base phrases, extending the concept of English Base Noun Phrases (BNPs). By simplifying sentence structure through this pre-processing step, the authors aim to significantly improve the efficiency and accuracy of Chinese parsers.

What Problem Does This Paper Address?

Syntactic parsing, the process of analyzing sentence structure, is a fundamental task in Natural Language Processing (NLP). However, parsing complex sentences can be computationally expensive and prone to errors. A common strategy to mitigate this is to decompose the parsing process into several phases, tackling simpler sub-problems first. One such sub-problem is identifying "base phrases" – fundamental syntactic units that, when identified, simplify the overall sentence structure for subsequent parsing stages. While Base Noun Phrase (BNP) identification has seen success in English, Chinese presents unique challenges. This paper addresses the need for a robust method to identify various types of base phrases in Chinese, thereby laying a crucial foundation for more accurate and efficient Chinese parsers.

Methods

The authors propose a "statistics based hybrid model" for Chinese base phrase identification. Their approach builds upon the concept of Base Noun Phrases (BNPs) and extends it to seven distinct types of Chinese base phrases, including base adjective phrases (BADJP), base adverbial phrases (BADVP), and base noun phrases (BNP), among others.

The core of their method involves a two-stage statistical modeling process:

  1. Duple Based HMM Parser: Initially, they employ a Hidden Markov Model (HMM) that considers "duples" of boundary tags (Left, Right, Inside, Outside) and phrase tags. This model is trained on a corpus of manually annotated sentences using bigram statistics. A Viterbi algorithm is used to find the most probable sequence of these duples.
  2. Triple Based MM Exploiting Linguistic Information: Recognizing that the duple-based model might be too coarse-grained, they introduce a "triple" model. This model incorporates POS (Part-of-Speech) information alongside the boundary and phrase tags. This "triple" (t, r, c) allows for a more nuanced representation of syntactic structures. This model is also trained using bigram statistics.
  3. Further Improvement Through TBED Learning: To further enhance accuracy and overcome potential ceilings reached by purely statistical models, they integrate a Transformation-Based Error Driven (TBED) learning strategy. This involves designing initial rules to correct statistical model errors. The system then iteratively learns new rules by comparing processed text with training data, generating rules based on pre-defined actions to correct mistakes and improve correct identifications.

Datasets

The paper mentions training their models on "10,000 manual annotated sentences." For evaluation, they use an "open test on additional 1000 sentences." The specific source or nature of these annotated sentences is not detailed beyond their quantity.

Key Insights and Ideas

  • Chinese Base Phrases are More Than Just NPs: The paper argues that Chinese sentences contain various structures, not just noun phrases, where components closely attach to central words to form basic phrases. This broadens the scope beyond traditional BNP identification.
  • Hierarchical Statistical Modeling: The progression from a duple-based HMM to a triple-based model (incorporating POS tags) demonstrates a clear strategy for increasing the granularity and accuracy of statistical models by leveraging richer linguistic information.
  • Hybrid Approach is Key: The combination of statistical modeling (HMMs) with rule-based learning (TBED) proves effective. Statistical models provide a strong baseline, while TBED learning fine-tunes the system by addressing specific error patterns, leading to significant accuracy improvements.
  • Pre-processing Simplifies Parsing: The core motivation is that successfully identifying these base phrases "simplifies the structure of the sentence," making the subsequent, more complex parsing task more manageable and accurate.

Key Results

The paper presents quantitative results in terms of precision for boundary tags (R), phrase tags (C), and both (R and C).

  • Duple Based HMM: Achieved a precision of 79.0% for both R and C in close testing and 74.7% in open testing.
  • Triple Based MM: Showed improvement, reaching 84.6% precision for both R and C in close testing and 83% in open testing.
  • TBED Module: This hybrid approach yielded the best results, with 89% precision for both R and C in close testing and 87.1% in open testing.

These results indicate a significant improvement in accuracy as the models become more sophisticated and incorporate linguistic information and error-driven learning.

Limitations and Future Work

  • Limited Coverage of Base Phrases: The authors acknowledge that the 7 types of Chinese base phrases they consider are "far from perfection" and that their theoretical definitions are still under development.
  • Need for Broader Coverage: Future work will focus on discussing and adding new base phrase types to achieve broader coverage of Chinese sentence structures.
  • Stricter Constraints: They also plan to define Chinese base phrases with "more strict constraints," either theoretically or empirically, to further refine their identification.
  • Continuous Accuracy Improvement: The pursuit of new techniques to improve the accuracy of statistical models remains a constant aim of their research.

Summary updated

Abstract

This paper extends the base noun phrase (BNP) identification into a research on Chinese base phrase identification. After briefly introducing some basic concepts on Chinese base phrase, this paper presents a statistics based hybrid model for identifying 7 types of Chinese base phrases in view. Experiments show the efficiency of the proposed method in simplifying sentence structure. Significance of the research lies in it provides a solid foundation for the Chinese parser.
BibTeX
@misc{zhao2000statistics,
  author = {Tiejun Zhao and Muyun Yang and Fang Liu and Jianmin Yao and Hao Yu},
  title = {{Statistics based hybrid approach to Chinese base phrase identification}},
  year = {2000},
  doi = {10.3115/1117769.1117781},
}
Figure