Enhancement of a Chinese Discourse Marker Tagger with C4.5

Benjamin K. T’sou, Tom B.y Lai, Samuel W.k. Chan, Weijun Gao, Xuegang ZhanACL 2000

TL;DR

This paper tackles the challenge of automatically identifying and classifying discourse markers in Chinese text, crucial for tasks like text summarization. They enhance an existing system by applying the C4.5 machine learning algorithm to disambiguate these markers, significantly improving accuracy.

What Problem Does This Paper Address?

Discourse markers are linguistic cues that signal the structure and relationships between parts of a text. Understanding these markers is vital for Natural Language Processing (NLP) tasks like automatic text summarization, where they help identify important sentences and rhetorical structures. However, automatically tagging these markers in Chinese is difficult due to several factors:

  • Ambiguity: Many discourse markers have multiple meanings or can function as discourse markers in one context and not in another (apparent vs. real discourse markers).
  • Discontinuity: Discourse markers can be split into multiple parts, with one part potentially missing (represented by a NULL marker).
  • Lack of Delimiters: Chinese text lacks explicit word delimiters, requiring a robust segmentation step.
  • Context Dependency: The correct interpretation of a discourse marker often depends heavily on its surrounding context and the rhetorical relation it signifies.

The authors aim to improve an existing system (SIFAS) by using machine learning to better handle the disambiguation of these complex markers, particularly the challenging NULL marker cases.

Methods

The core of the paper's contribution lies in applying the C4.5 decision tree algorithm to improve discourse marker tagging. The process involves:

  1. Data Preparation and Segmentation: Input text is prepared, and words are segmented.
  2. Candidate Discourse Marker (CDM) Identification: Words identified as potential discourse markers are flagged.
  3. Real Discourse Marker (RDM) and Apparent Discourse Marker (ADM) Tagging: Initially, a manual or rule-based approach identifies RDMs and ADMs.
  4. Feature Extraction: For untagged CDMs, relevant features are extracted. These include:
  5. The discourse marker itself.
  6. Two words immediately to its left and right (syntactic information).
  7. The number of delimiters (commas, periods) before and after the marker (for context and NULL marker location).
  8. A Boolean indicating if the marker is a Real Discourse Marker (IsRDM).
  9. C4.5 Application: The C4.5 algorithm is trained on these features to:
  10. Classify CDMs: Distinguish between RDMs and ADMs.
  11. Locate NULL Markers: Determine the position of missing marker components.

The paper also outlines several principles used by the tagging algorithm to resolve ambiguity, such as greediness (prioritizing the first match), locality (prioritizing shorter distances between marker parts), explicitness, superiority (prioritizing inter-sentence relations), and back-marker preference.

Datasets

  • SIFAS Corpus: The system is developed and evaluated on a corpus of Chinese newspaper editorials.
  • Training Data: 80 tagged editorials from Ming Pao (Hong Kong) were used for training.
  • Testing Data: 20 randomly selected editorials from Mainland China and Hong Kong newspapers were used for automatic tagging and manual checking.

Key Insights and Ideas

  • Machine Learning for Disambiguation: Applying ML, specifically C4.5, is a powerful way to handle the inherent ambiguity of discourse markers, which is difficult for rule-based systems alone.
  • The NULL Marker Problem: The paper highlights that a significant portion of discourse marker pairs involve a NULL marker, making their identification a key challenge. The C4.5 approach shows promise in locating these missing components.
  • Feature Engineering is Crucial: The effectiveness of C4.5 relies on carefully selected features, including syntactic information (part-of-speech of neighboring words) and positional information (delimiter counts), which capture crucial contextual clues.
  • Trade-offs in Tagging: The paper acknowledges that while ML improves accuracy, some errors (like misclassifying RDMs as ADMs) are difficult to avoid without deeper semantic understanding.

Key Results

  • The enhanced system achieved a Gross Accuracy (GA) of 95.38%, a significant improvement over the previous GA of 68.89%.
  • The paper reports that most errors are caused by the misclassification of CDMs (distinguishing RDMs from ADMs).
  • The system can correctly locate NULL markers with reasonable accuracy, with statistics showing about 60% of NULL markers being correctly located within one comma/period.
  • The distribution of relations in the test data shows a strong prevalence of INTRA-sentence relations (around 94.58%), with a substantial percentage of these involving NULL markers (around 67.22%).

Limitations and Future Work

  • Syntactic/Semantic Information: The authors acknowledge that their current approach relies heavily on syntactic information. They suggest that incorporating more comprehensive semantic information would further improve performance, though this is challenging due to the lack of readily available resources like thesauri.
  • ADM/RDM Classification Errors: Distinguishing between apparent and real discourse markers remains a difficult problem, especially without deeper semantic understanding.
  • Overgeneration: The system sometimes overgenerates by mismatching CDMs as RDM pairs.
  • Future Work: The authors suggest that more training data and potentially more sophisticated ML algorithms could further refine the system. They also mention the potential for interactive selection of low-precision rules to speed up semi-automatic tagging.

Summary updated

BibTeX
@inproceedings{tsou2000enhancement,
  author = {Benjamin K. T’sou and Tom B.Y Lai and Samuel W.K. Chan and Weijun Gao and Xuegang Zhan},
  title = {{Enhancement of a Chinese Discourse Marker Tagger with C4.5}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2000},
}
Figure