Chinese Temporal Tagging with HeidelTime

Hui Li, Jannik Strötgen, Julian Zell, Michael Gertz2014 · DOI

TL;DR

This paper introduces the first publicly available temporal tagger for Chinese, extending the existing HeidelTime system. By developing new Chinese-specific resources and re-annotating a key dataset, they achieve high accuracy in both extracting and normalizing temporal expressions, outperforming previous approaches.

What Problem Does This Paper Address?

Temporal information is crucial for understanding text, enabling tasks like event extraction and timeline generation. While significant research has focused on temporal tagging for English, other languages, including Chinese, have been less explored. Existing Chinese temporal taggers are not publicly available, leaving a gap in the NLP community. This paper aims to fill this gap by creating and releasing resources for Chinese temporal tagging, specifically addressing both the extraction of temporal expressions and their normalization into a standard format.

Methods

The authors leverage the existing HeidelTime system, a rule-based, multilingual temporal tagger known for its cross-domain capabilities. Instead of building a new system from scratch, they extend HeidelTime by developing language-specific resources for Chinese. This involves:

  • Linguistic Preprocessing: Utilizing the TreeTagger with its Chinese model for sentence segmentation, tokenization, and part-of-speech tagging, which are standard inputs for HeidelTime.
  • Resource Development: Translating and adapting English patterns, normalization rules, and extraction rules to the Chinese language. This includes creating new patterns for Chinese-specific expressions like numerals and adapting existing ones.
  • Rule Development: A systematic process of creating simple Chinese rules based on English ones, reviewing extracted expressions, improving rules, and adding new ones for undetected expressions. This iterative process also involved adapting pattern and normalization resources.
  • Re-annotation of Data: Critically, the authors re-annotated the TempEval-2 Chinese corpus to address its low quality and missing information, creating a more reliable benchmark for evaluation.

Datasets

  • TempEval-2 Chinese Corpus: This is the primary dataset used for training and evaluation. The paper notes that the original corpus had issues with missing normalized values and incorrect annotations.
  • Re-annotated TempEval-2 Corpus: The authors created an "improved" version of the TempEval-2 corpus by manually assigning normalized values and correcting errors, making it more suitable for robust evaluation. They also mention a "cleaned" version where expressions without value information were removed, as used by Angeli and Uszkoreit (2013).

Key Insights and Ideas

  • Extending Existing Systems is Viable: The success of extending HeidelTime for Chinese demonstrates that leveraging robust, existing multilingual systems can be an effective strategy for developing NLP tools for under-resourced languages, rather than always building from scratch.
  • Data Quality is Paramount: The significant effort in re-annotating the TempEval-2 corpus highlights how crucial high-quality annotated data is for training and evaluating NLP models, especially for complex tasks like temporal tagging. Even a well-established system can struggle with noisy data.
  • Chinese Temporal Tagging Presents Unique Challenges: The paper points out specific difficulties in Chinese, such as polysemous words (e.g., dynasty names also used as family names), ambiguous expressions, and the complexity of different calendar systems (lunar, solar terms, Taiwanese calendar).
  • Public Availability Matters: By releasing their Chinese HeidelTime resources and the re-annotated corpus, the authors contribute significantly to the research community, enabling further work and comparisons.

Key Results

  • High Accuracy: The re-annotated Chinese HeidelTime achieved high extraction and normalization quality, with value accuracies exceeding 90% on the improved and cleaned versions of the dataset.
  • Outperforming Previous Work: The developed Chinese HeidelTime system outperformed the multilingual parsing approach by Angeli and Uszkoreit (2013) in normalization accuracy, even though HeidelTime performed the full task (extraction and normalization) while the other approach assumed gold extents.
  • Improved Dataset: The improved TempEval-2 corpus contains 765 temporal expressions in the training set and 193 in the test set, with value information for almost all of them.

Limitations and Future Work

  • Taiwan-Specific Expressions: The current rules do not fully cover Taiwan-specific expressions, which were found in some test documents. This suggests a need for further expansion of the rule set.
  • Error Analysis: The authors plan to conduct a more detailed error analysis to gain further insights for improving HeidelTime's Chinese resources.
  • Ongoing Improvement: The iterative nature of rule development implies that continuous refinement is expected.

Summary updated

Abstract

Temporal information is important for many NLP tasks, and there has been extensive research on temporal tagging with a particular focus on English texts. Recently, other languages have also been addressed, e.g., HeidelTime was extended to process eight languages. Chinese temporal tagging has achieved less attention, and no Chinese temporal tagger is publicly available. In this paper, we address the full task of Chinese temporal tagging (extraction and normalization) by developing Chinese HeidelTime resources. Our evaluation on a publicly available corpus ‐ which we also partially re-annotated due to its rather low quality ‐ demonstrates the effectiveness of our approach, and we outperform a recent approach to normalize temporal expressions. The Chinese HeidelTime resource as well as the corrected corpus are made publicly available.
BibTeX
@misc{li2014chinese,
  author = {Hui Li and Jannik Strötgen and Julian Zell and Michael Gertz},
  title = {{Chinese Temporal Tagging with HeidelTime}},
  year = {2014},
  doi = {10.3115/v1/e14-4026},
}
Figure