Language Specific Issue and Feature Exploration in Chinese Event Extraction

Zheng Chen, Heng JiNAACL 2009 · OpenReview

Abstract: Zheng Chen, Heng Ji. Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Short Papers.

2009.

The "Segmentation Trap" in Chinese Event Extraction

In the field of Event Extraction (EE), the primary objective is to identify triggers—words that signal an event, such as "fired" or "merged"—and their associated arguments, which are the entities involved (e.g., the "employer" or the "company"). While these tasks are well-established for English, Chinese presents a unique linguistic challenge: the absence of spaces between words.

Standard Natural Language Processing (NLP) pipelines typically begin with "word segmentation." However, this research highlights a critical "segmentation trap" for Chinese EE. Standard segmenters often fail to align with event triggers. For instance, the Chinese word jībì (击毙) means "shoot and kill." A standard segmenter treats this as a single unit, yet it contains two distinct triggers: "shoot" (an Attack event) and "kill" (a Die event). Conversely, some triggers may span across multiple segmented words. The authors found that approximately 13% of triggers in the ACE Chinese corpus do not align perfectly with word boundaries, leading to cascading errors in downstream extraction.

Methodology: Words vs. Characters

The authors explored two distinct strategies to overcome these segmentation issues for Trigger Labeling:

  1. Word-based Strategy: This utilizes a standard Maximum Entropy (MaxEnt) classifier. To mitigate segmentation errors, the authors implemented a "global errata table" derived from training data to help the model recognize triggers that deviate from standard word boundaries.
  2. Character-based Strategy: This approach treats trigger identification as a sequence labeling task using a Maximum Entropy Markov Model (MEMM). By assigning B-I-O (Beginning, Inside, Outside) tags to individual characters, the model can identify triggers regardless of how a standard segmenter might have split the text.

For Argument Labeling, a two-step MaxEnt approach was used (identification followed by classification). This stage relied on a mix of lexical, syntactic (parse tree paths), and semantic features (synonym dictionaries and PropBank).

Feature Engineering and Insights

The study revealed that while syntactic features like parse tree depth are often helpful in other NLP tasks, they can be a double-edged sword in Chinese EE. For trigger labeling, adding syntactic features actually decreased performance, as triggers are syntactically flexible and can appear as various parts of speech or within compounds.

However, for Argument Labeling, syntactic features proved essential. Specifically, the "Path" (the route through a parse tree from the trigger to the argument) and "Distance" features were crucial for correctly identifying and labeling roles.

Impact of different feature sets on argument identification and labeling.

Additionally, contextual "neighbors" were highly informative. For example, the presence of the word bèi (by) is a strong indicator that the following entity is the "Attacker" in a conflict event.

Key Results

The experimental results, conducted on the 2005 ACE Chinese corpus, demonstrate the clear superiority of the character-based approach. By bypassing traditional segmentation, the character-based model achieved a significant boost in precision for trigger identification, reaching 82.4%—a 14.3% improvement over the word-based baseline.

Overall performance comparison between baseline, word-based, and character-based systems.

As shown in the comparison table, the character-based system reached an F-measure of 59.9 for trigger labeling and 43.8 for argument labeling. While the character-based approach effectively addresses the segmentation problem and yields high precision, the authors note that overall recall remains a challenge.

Conclusion

This research confirms that character-level processing is a more robust foundation for Chinese Event Extraction than traditional word-based segmentation. Future work aims to address the recall gap by exploring cross-lingual event extraction, potentially using English extraction systems to bootstrap and improve performance in Chinese through consensus-based inference.

Summary updated

BibTeX
@inproceedings{chen2009language,
  author = {Zheng Chen and Heng Ji},
  title = {{Language Specific Issue and Feature Exploration in Chinese Event Extraction}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2009},
  url = {https://openreview.net/forum?id=r1E33G-uWS},
}
Figure