A Quasi-Dependency Model for the Structural Analysis of Chinese BaseNPs

Jun Zhao, Changning HuangCOLING 1998

TL;DR

This paper introduces a novel "quasi-dependency" model for analyzing the structure of Chinese noun phrases (baseNPs), which are crucial for information retrieval. They also propose an MDL-based algorithm for learning the strength of these quasi-dependencies, outperforming traditional ML approaches.

What Problem Does This Paper Address?

Information retrieval systems often rely on understanding the structure of noun phrases to accurately retrieve relevant documents. While English has a clear definition of "baseNPs" as simple, non-recursive noun phrases, this definition is insufficient for Chinese. Chinese noun phrases can be complex and recursive, making their structural analysis challenging. Existing models for English noun phrase analysis, like adjacency or dependency models, are not directly applicable or sufficient for Chinese. The paper aims to bridge this gap by proposing a new model and a method to learn the necessary parameters for analyzing Chinese baseNPs.

Methods

The core of the paper is the quasi-dependency model. Unlike traditional dependency models that require a definitive dependency relation, this model focuses on the possibility of a dependency relation between words. This "quasi-dependency relation" can be learned from corpora. The model then uses these quasi-dependency strengths to infer the most likely syntactic structure of a baseNP.

For learning these strengths, the authors propose an MDL (Minimum Description Length) based algorithm. This approach aims to find a model that minimizes the combined description length of the data and the model itself. This is achieved through an iterative process that alternately optimizes the model's structure (which dependencies exist) and its parameters (the strength of those dependencies). This MDL approach is presented as an improvement over traditional Maximum Likelihood (ML) methods, which can overfit the data.

Datasets

The paper uses a corpus of Chinese baseNPs for training and testing. Specifically, they mention:

  • A training set of 7,500 N2+N2+N2-type baseNPs (where N2 represents bi-syllable nouns).
  • A close testing set of 500 baseNPs from the training set.
  • An open testing set of 500 baseNPs not included in the training set.

The paper does not explicitly state the creation of a new benchmark dataset but rather uses an existing corpus for their experiments.

Key Insights and Ideas

  • "Quasi-dependency" is key for Chinese: The concept of a "quasi-dependency" relation, which signifies the possibility of a dependency rather than a strict one, is crucial for handling the nuances of Chinese noun phrase structure.
  • MDL for robust learning: The MDL principle offers a principled way to balance model complexity and data fit, leading to more robust quasi-dependency strength estimation compared to traditional ML methods, mitigating overfitting.
  • Iterative optimization is effective: The proposed two-step iterative optimization (fixing structure, optimizing parameters, then fixing parameters, optimizing structure) is shown to be an effective strategy for learning the model.
  • Linguistic knowledge can pre-process: Incorporating linguistic knowledge (e.g., identifying independent word pairs like Noun+Adjective) can improve the efficiency and accuracy of the baseNP structure analysis.

Key Results

The proposed quasi-dependency model and MDL-based acquisition algorithm demonstrate superior performance:

  • Improved Analysis Precision: The quasi-dependency model achieves higher analysis precision for N2+N2+N2-type baseNPs (e.g., 91.5% on close test) compared to an adjacency model (84.6% on close test).
  • MDL Outperforms ML: The MDL-based algorithm for quasi-dependency-strength acquisition leads to better baseNP analysis precision (e.g., 91.5% on close test) than a traditional ML-based algorithm (89.0% on close test).
  • Left-binding Dominance: In N2+N2+N2-type baseNPs, left-binding structures are observed to be about twice as common as right-binding structures.

Limitations and Future Work

  • Data Sparseness: The paper acknowledges that a large lexicon size can lead to data sparseness issues in word-based acquisition. They propose a "complex-feature-based modeling" approach to address this by mapping words to feature sets.
  • Incorporating More Linguistic Knowledge: The authors suggest that future research should focus on integrating more linguistic knowledge into their statistical model to further enhance performance.
  • Generalizability: While the paper focuses on N2+N2+N2-type baseNPs, the generalizability of the model to other types of Chinese baseNPs could be a point for further investigation.

Summary updated

BibTeX
@inproceedings{zhao1998quasi,
  author = {Jun Zhao and Changning Huang},
  title = {{A Quasi-Dependency Model for the Structural Analysis of Chinese BaseNPs}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1998},
}
Figure