An Agent-based Approach to Chinese Named Entity Recognition

Shiren Ye, Tat-Seng Chua, Jimin LiuCOLING 2002 · OpenReview

TL;DR

This paper tackles the challenging problem of Chinese Named Entity Recognition (NER) by introducing a novel agent-based framework. It leverages a "rationality model" to identify potential entities and then uses a multi-agent negotiation process to resolve conflicts and select the most accurate entities, achieving over 92% F1 score on the MET-2 test corpus.

What Problem Does This Paper Address?

Named Entity Recognition (NER) is a crucial step in natural language processing, aiming to identify and categorize named entities like persons, locations, and organizations within text. However, Chinese NER presents unique difficulties. Unlike English, Chinese lacks explicit word delimiters (spaces) and capitalization, making word segmentation a significant hurdle. Furthermore, Chinese named entities often share characters with common words, leading to ambiguity. Existing methods, often relying on handcrafted rules or machine learning with extensive linguistic features, struggle with this inherent ambiguity and the lack of clear structural cues. This paper aims to overcome these challenges by proposing a more robust and flexible approach.

Methods

The paper introduces a two-pronged approach:

  1. Rationality Model for NE Detection: This model assigns a "rationality" score to token sequences, indicating the probability of them being a specific type of named entity. It treats NE recognition as a "coloring problem," where tokens are colored based on their likelihood of being a surname (red), first-name (blue), or other (white) for Person (PER) entities. This extends to more complex entities like Organizations (ORG) and Locations (LOC) by considering sequential patterns and grammatical rules. The rationality is computed using a probabilistic approach, akin to odds-likelihood, and then mapped to a specific range to handle potential overflow/underflow.
  2. Multi-Agent Framework for NE Confirmation: After identifying numerous potential NE candidates using the rationality model, the system employs a multi-agent framework to refine these candidates. Each potential NE is represented by an "agent" with its own "belief" (rationality value). These agents are grouped into teams based on their relationships (e.g., overlapping, supporting, competing). Agents then engage in a negotiation process:
  3. Competition: For overlapping NEs within a team, the agent with the highest belief "wins," and the beliefs of competing agents are decremented.
  4. Support: For supporting relationships, the winning agent's belief is used to reinforce other agents' beliefs. This negotiation process iteratively refines the beliefs of agents, ultimately leading to a more accurate selection of NEs.

Datasets

  • PKU-corpus: Used for training, it's a manually tagged corpus of Chinese news reports with over 30 POS tags, including specific tags for surnames and first-names. It contains approximately 37,000 sentences and over 10 million tokens.
  • Hownet: Used for training.
  • MET2 Chinese resources: Used for testing.
  • Two name lists (foreign and ORG names): Collected from the web via bootstrapping.
  • Common word dictionary: Built from the PKU corpus by excluding NEs.
  • Distribution information: Derived from the PKU corpus for token POS tags, NE types, and roles within NEs.
  • LOC list: Maintained from the MET-2 test corpus.

Key Insights and Ideas

  • Bridging Segmentation and Recognition: The paper cleverly bypasses the explicit need for perfect word segmentation by directly modeling token sequences and their likelihood of forming NEs, a significant advantage for Chinese.
  • Rationality as a Core Metric: The concept of "rationality" provides a principled way to quantify the confidence in a token sequence being a named entity, moving beyond simple statistical counts.
  • Agent Negotiation for Conflict Resolution: The multi-agent negotiation framework is a powerful mechanism to handle the inherent ambiguity and overlapping candidates generated by the initial detection phase, mimicking a collaborative decision-making process.
  • Handling Ambiguity through Competition and Support: The distinction between competitive and supportive relationships among NE candidates allows for nuanced refinement, where conflicting entities are resolved and reinforcing entities strengthen each other.

Key Results

The system achieves high performance on the MET-2 test corpus, with an F₁ value of over 92% across all NE types (PER, LOC, ORG) when using the agent-based modification. This represents a significant improvement over baseline methods and other reported systems on the same corpus. An informal test on general web-based articles also yielded F₁ measures of over 90%.

Limitations and Future Work

  • Limitations: The paper acknowledges that the initial greedy approach can lead to over-segmentation, necessitating the agent-based refinement. The rationality model's effectiveness is dependent on the quality and coverage of the training data and dictionaries.
  • Future Work: The authors plan to test their system on larger-scale corpora, refine techniques for various text types, and apply bootstrapping to address data sparseness. They also aim to extend their research to relation and information extraction in multilingual text.

Summary updated

BibTeX
@inproceedings{ye2002agent,
  author = {Shiren Ye and Tat-Seng Chua and Jimin Liu},
  title = {{An Agent-based Approach to Chinese Named Entity Recognition}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
  url = {https://openreview.net/forum?id=oM20LSY1Le},
}
Figure