A Framework Based on Graphical Models with Logic for Chinese Named Entity Recognition

Xiaofeng Yu, Wai Lam, Shing-Kit ChanIJCNLP 2008 · OpenReview

Abstract

Chinese Named Entity Recognition (NER) faces unique challenges due to the lack of explicit word boundaries and the presence of complex, nested entity structures. This paper presents a hybrid framework that integrates the statistical sequence labeling capabilities of Conditional Random Fields (CRFs) with the logical reasoning power of Markov Logic Networks (MLNs). By encoding linguistic domain knowledge into first-order logic rules, the proposed system effectively corrects common errors made by purely statistical models. Experimental results on the People’s Daily corpus demonstrate a 23.75% relative error reduction over state-of-the-art CRF baselines, achieving an F-measure of 92.39%.

Introduction

Named Entity Recognition (NER)—the identification of people (PER), locations (LOC), and organizations (ORG)—is a fundamental task in natural language processing. In the context of Chinese text, NER is particularly difficult because the language lacks spaces between words, and entities often contain nested components. For instance, an organization name like "Haidian District, Beijing City School of Computer Science" contains two distinct locations within its structure.

Traditional systems often treat NER as a character-level sequence labeling problem. However, this approach frequently ignores the rich linguistic relationships and structural rules that human readers use to identify entities. Purely statistical models are prone to "silly" mistakes, such as misidentifying common nouns as names or failing to respect punctuation boundaries. This research addresses these shortcomings by combining probabilistic graphical models with formal logic.

Methodology

The proposed framework utilizes a two-stage approach to entity extraction, leveraging the strengths of both statistical and relational learning.

The two-stage framework: CRFs provide initial candidates, which are then refined by MLNs using domain knowledge.

1. The Base System (CRFs)

The first stage employs Conditional Random Fields (CRFs), the standard for sequence labeling. This model performs the initial identification by analyzing character-level features, part-of-speech tags, and local context. It produces both highly confident entities and "weakly confident" candidates that require further validation.

2. The Logical Refiner (MLNs)

To refine the CRF output, the authors introduce Markov Logic Networks (MLNs). MLNs allow for the integration of first-order logic with probabilistic graphical models by attaching weights to logical rules. Unlike rigid rule-based systems, MLNs are "soft"; if a rule is occasionally violated in the data, the model learns a lower weight for it rather than failing entirely.

This stage treats NER as a relational problem rather than a simple sequence. By viewing entities as related objects, the MLN can enforce global consistency across a sentence or document.

Logic and Knowledge Representation

The core of the MLN component is a set of 15 manually crafted first-order logic formulas. These rules capture domain knowledge that statistical models often miss. For example, a rule might state that if a candidate entity ends with "University," it is highly likely to be an Organization.

Examples of First-Order Logic (KB) rules and the features they generate for specific entity instances.

As shown in the logic examples, the system can define relationships such as employment or colleagues to help disambiguate entities. If the model knows that "Peter" and "Smith" are colleagues, and "Peter" works for "IBM," it can logically infer a higher probability that "Smith" also works for "IBM," thereby reinforcing the identification of "IBM" as a Company/Organization.

Experimental Results

The system was evaluated using the People’s Daily corpus, a massive dataset containing over 357,000 sentences. To support the logic rules, the researchers also utilized external knowledge from Wikipedia and LDC lists to identify "salient words" for locations and organizations.

Key findings include:

  • Performance Gain: The hybrid CRF+MLN approach achieved an F-measure of 92.39%.
  • Error Reduction: This represents a 23.75% relative error reduction (RER) over the standalone CRF baseline.
  • Inference Efficiency: While MCMC-based methods like MC-SAT provided the highest accuracy, MAP/MPE inference offered a competitive 20.54% RER with significantly lower computational costs.
  • Organization Tagging: The logical rules were most effective at improving the detection of Organizations, which are historically the most difficult entities to identify in Chinese NER due to their length and internal complexity.

Conclusion

By combining the statistical robustness of CRFs with the relational reasoning of MLNs, this framework provides a powerful solution for Chinese NER. The ability to treat entities as interconnected objects rather than independent labels allows the model to overcome the limitations of local sequence labeling. Future work may involve "Structure Learning" to automatically discover logical rules from data and extending the framework to full Entity Relation Extraction.

Summary updated

BibTeX
@inproceedings{yu2008framework,
  author = {Xiaofeng Yu and Wai Lam and Shing-Kit Chan},
  title = {{A Framework Based on Graphical Models with Logic for Chinese Named Entity Recognition}},
  booktitle = {IJCNLP},
  year = {2008},
  url = {https://openreview.net/forum?id=49vgQfAjvd},
}
Figure