Single Character Chinese Named Entity Recognition

Xiaodan Zhu, Mu Li, Jianfeng Gao, Changning HuangSIGHAN 2003 · DOI

TL;DR

This paper tackles the challenge of recognizing single-character named entities (SCNEs) in Chinese text, a common but error-prone aspect of Named Entity Recognition (NER). The authors propose a novel source-channel model and compare it to classification-based approaches, demonstrating superior performance, particularly for location and person names.

What Problem Does This Paper Address?

Named Entity Recognition (NER) is crucial for understanding text, but Chinese NER faces a unique challenge: Single Character Named Entities (SCNEs). These are named entities composed of just one Chinese character, like "中" (China) or "俄" (Russia). While common (making up over 8% of named entity tokens), SCNEs are a significant source of errors in existing NER systems. This is partly due to their distinct generation and contextual properties compared to multi-character entities, and a general lack of focused research on them. The paper aims to improve SCNE recognition by formulating it within a source-channel framework and comparing it to classification-based methods.

Methods

The paper proposes two main approaches for SCNE recognition:

  • Improved Source-Channel Model: This is the core contribution. It adapts the standard source-channel model, commonly used for Chinese word segmentation, to specifically handle SCNEs. This involves:
  • Defining SCNEs (Single Character Location - SCL, Single Character Person - SCP) as distinct word classes.
  • Developing specialized channel models for SCL and SCP, calculating their probabilities based on character occurrences in training data.
  • Introducing weighting coefficients (CW) to balance the probabilities of different word classes.
  • Retraining the source model on a corpus where SCNEs are explicitly tagged (e.g., "周" in "周总理" as SCP, not just part of a larger person name).
  • Character-based Classifiers: As an alternative, the authors frame SCNE recognition as a binary classification problem. They build two classifiers:
  • Maximum Entropy (ME) Model: Uses a log-linear model with features derived from a window of six surrounding characters.
  • Vector Space Model (VSM): Also uses a six-character context window, employing a tf-idf weighting scheme with an adaptation for character positions.

Datasets

  • Training Corpus: A 10MB corpus, manually annotated for SCNEs.
  • Test Corpus: A 1MB corpus, also manually annotated.
  • Corpus Source: Texts are randomly selected from "People's Daily," covering 10 subjects and 5 writing styles. This dataset is noted as being significantly larger than typical test sets used in prior SCNE research (e.g., MET-2).
  • Benchmark: The paper establishes its own standard for NE/SCNE based on MET-2 and IEER-99, aiming for better comparability.

Key Insights and Ideas

  • SCNEs are a distinct challenge: The paper highlights that SCNEs are not just smaller versions of multi-character entities; they have unique contextual and generative properties that require specialized handling.
  • Source-channel model excels: The adapted source-channel model significantly outperforms classification-based methods (ME and VSM) and even state-of-the-art general NER systems for SCNE recognition. This suggests that the richer, global and local information captured by the source-channel framework is more effective for this specific task.
  • Context is important, but not everything: While context features are useful for ME and VSM, the paper implies that simply increasing the context window size might not yield substantial gains, reinforcing the need for a more robust model like the source-channel approach.
  • Ambiguity is a major error source: Human error analysis reveals that ambiguity between SCNEs and multi-character NEs, and between SCNEs and regular lexicon words, are the primary causes of errors.

Key Results

  • Source-Channel Model Performance:
    • SCL: Achieved an F-score of 81.01%.
  • SCP: Achieved an F-score of 68.02%.
  • Improvement over baseline: The adapted source-channel model (M2) showed significant improvements over the original source-channel model (M1) for SCL and SCP.
  • Comparison with other systems: The paper's system (M2) outperformed other state-of-the-art systems (MSWS, LCWS, PBWS) in both SCL and SCP recognition, achieving the best results. For example, it achieved an F-score of 81.01% for SCL compared to PBWS's 43.63%.
  • Classification models: ME and VSM classifiers performed reasonably well but were consistently outperformed by the source-channel model.

Limitations and Future Work

  • Ambiguity Resolution: Errors stemming from ambiguity between SCNEs and multi-character NEs, and between SCNEs and regular words, are difficult to resolve with current features. This is a significant area for future work.
  • Data Sparsity: The scarcity of SCNEs in training data, especially for less common SCNEs, makes training data sparser and can lead to more errors.
  • Expanding SCNE Lexicon: The authors suggest addressing errors from missing SCNEs (like "卡" for Qatar) by acquiring additional SCNE entries from resources like abbreviation dictionaries. However, they caution that this needs careful selection to avoid introducing rare or irrelevant entries.
  • Unsupervised Methods: The paper suggests that unsupervised methods could be explored to mitigate the reliance on large, annotated training data.

Summary updated

BibTeX
@inproceedings{zhu2003single,
  author = {Xiaodan Zhu and Mu Li and Jianfeng Gao and Changning Huang},
  title = {{Single Character Chinese Named Entity Recognition}},
  booktitle = {SIGHAN},
  year = {2003},
  doi = {10.3115/1119250.1119268},
}

Loading related content...

Figure