Abstract
This paper presents a standalone, end-to-end Chinese discourse parser designed to transform raw text into hierarchical Rhetorical Structure Theory (RST) trees. By integrating BERT-based encoders with a revised dynamic oracle training strategy and unsupervised data augmentation (UDA), the proposed model overcomes common hurdles such as pipeline error propagation and data sparsity. Experimental results on the Chinese Discourse Treebank (CDTB-14) demonstrate that the model achieves state-of-the-art performance across all discourse parsing sub-tasks.
Introduction
Discourse parsing involves analyzing the macro-level structure of a document by segmenting text into Elementary Discourse Units (EDUs) and identifying their hierarchical relationships. In the RST framework, this requires determining nuclearity (centrality) and rhetorical relations (e.g., Causality, Contrast). For Chinese discourse parsing, two primary challenges exist: the complexity of traditional pipelines that rely on error-prone pre-processing steps like word segmentation, and the sparsity of labeled data in datasets like CDTB-14.
Methodology
The authors propose a transition-based Shift-Reduce parser that processes text linearly using a stack and a queue. This architecture is "complete" as it handles the entire pipeline from raw character input to the final discourse tree.
Key technical components include:
- BERT-based Encoder: A
BERT-base-Chinesemodel serves as a character-level encoder for both EDU segmentation (utilizing a CRF layer) and feature extraction for the transition-based action classifier. - Revised Dynamic Oracle: To improve robustness against inference errors, the authors implemented a two-pass training strategy. In "teacher mode," the model follows the gold path; in "student mode," it explores its own predictions to learn error recovery.
- Unsupervised Data Augmentation (UDA): Back-translation was employed to generate paraphrased Chinese sentence pairs from unlabeled data. A consistency loss ensures the model predicts identical rhetorical labels for both original and paraphrased versions, aiding generalization.
- Modified Beam Search: The decoding process was adjusted to average probabilities only for steps involving active choices, preventing bias toward sequences with many forced moves.
Experimental Setup
The model was primarily evaluated on the CDTB-14 (Chinese Discourse Treebank), which contains 500 articles annotated in the RST style. Additionally, the CDTB-12 dataset (PDTB-style) provided a source of unlabeled text for the UDA back-translation process.
Results and Discussion
The proposed model established a new state-of-the-art (SOTA) for Chinese discourse parsing. When evaluated on the comprehensive "All" metric, it achieved an F1-score of 48.9, significantly outperforming previous benchmarks.
Key findings from the evaluation include:
- Segmentation Success: EDU segmentation is nearly solved, with an F1-score of 97.4.
- UDA Efficacy: Data augmentation significantly improved performance for rare rhetorical relations. For instance, the F1-score for the "Explanation" relation increased from 51.1 to 55.2.
- Decoding Philosophy: Surprisingly, increasing beam width (k > 1) led to a slight decrease in performance. This suggests that the sequential nature of human discourse may be better captured by a greedy approach within the shift-reduce paradigm.
Conclusion
By adopting an end-to-end character-level approach and a robust training oracle, this research effectively mitigates error propagation and data scarcity in Chinese discourse parsing. While the model achieves SOTA results, the mid-50s F1-scores for relation and nuclearity labeling indicate that deep semantic understanding remains a challenge. Future work may explore cross-lingual transfer learning to leverage more abundant English discourse datasets.
Summary updated