Abstract
Discourse parsing in Chinese presents unique challenges due to the high frequency of non-explicit relations and the flexible nature of connectives. This paper introduces an end-to-end discourse parser that utilizes a uniform framework for both explicit and non-explicit relations. A key innovation is the "seed-expansion" approach for argument extraction, which incrementally builds argument spans. Evaluated on the CoNLL-2016 Shared Task, the system achieved a competitive 26.90% F1 score on the blind test set, demonstrating robustness particularly in non-explicit relation recognition.
Introduction
Discourse parsing involves identifying logical relationships between text segments (arguments), such as cause, contrast, or conjunction. While English discourse parsing has matured through the Penn Discourse TreeBank (PDTB), Chinese discourse parsing faces distinct linguistic hurdles. Approximately 80% of Chinese discourse relations are "non-explicit," lacking clear connective words, compared to roughly 55% in English. Furthermore, Chinese connectives are often parallel or scattered, and punctuation plays a critical role in signaling boundaries. This research addresses these challenges by developing a parser specifically adapted to Chinese linguistic characteristics rather than porting English-centric models.
Methodology
The proposed system follows a pipeline architecture consisting of four main subtasks, unified to handle both explicit and non-explicit relations within a single framework.
1. Anchor Identification
Instead of searching solely for connective words, the system identifies "anchors." For explicit relations, these are standard connectives. For non-explicit relations, specific punctuation marks are treated as anchors, allowing for a consistent processing flow across both relation types.
2. Argument Extraction: The Seed-Expansion Approach
The core novelty of this work is the "seed-expansion" method for argument extraction. Rather than predicting the entire span of an argument at once, the system proceeds incrementally:
- Seed Identification: It identifies the "seed" (the clause containing or adjacent to the anchor).
- Directional Classification: It determines the direction of expansion.
- Expansion: The argument span is expanded clause-by-clause until a classifier identifies a boundary.
3. Sense Classification and Argument Relabeling
A multi-class classifier predicts the semantic relationship (e.g., Conjunction, Expansion) between arguments. Finally, an "Argument Relabeling" step ensures that "Arg1" and "Arg2" labels align with the semantic definitions of the Chinese Discourse TreeBank (CDTB), as linear order in Chinese does not always dictate semantic roles. The system employs 17 Maximum Entropy classifiers utilizing lexical, syntactic, and word embedding features.
Experimental Setup and Results
The system was trained and evaluated on the CDTB as provided for the CoNLL-2016 Shared Task, using standard splits for training, development, and testing, along with a blind test set for robustness evaluation.
| Category | Dev F1 | Test F1 | Blind F1 |
|---|---|---|---|
| Explicit Parser | 45.21% | 48.59% | 24.46% |
| Non-Explicit Parser | 39.82% | 38.69% | 27.12% |
| Overall Parser | 41.20% | 40.89% | 26.90% |
The results indicate that the punctuation-anchor approach is particularly effective for non-explicit relations, where the system actually outperformed its explicit counterpart on the blind test set. This suggests that the unified framework is robust against the lack of explicit connectives.
Discussion and Key Insights
The research highlights several critical insights for Chinese NLP:
- Punctuation as Anchors: Treating punctuation as a first-class signal for non-explicit relations successfully unifies the parsing pipeline.
- Incremental Expansion: The seed-expansion method is more suited to the multi-clause nature of Chinese arguments than traditional tree-subtraction methods used in English.
- Semantic Flexibility: The necessity of a relabeling step underscores the importance of semantic over positional labeling in Chinese discourse.
Limitations and Future Work
Despite its successes, the system faces challenges with error propagation; failures in anchor identification can derail the entire pipeline. Additionally, the current model primarily assumes arguments are within adjacent sentences, struggling with long-range dependencies. Future work will focus on improving "span cohesion" features to better define argument boundaries and addressing multi-sentence argument spans.
Summary updated