Abstract
Text-level discourse parsing remains a challenge: most approaches employ features that fail to capture the intentional, semantic, and syntactic aspects that govern discourse coherence. In this paper, we propose a recursive model for discourse parsing that jointly models distributed representations for clauses, sentences, and entire discourses. The learned representations can to some extent learn the semantic and intentional import of words and larger discourse units automatically. The proposed framework obtains comparable performance regarding standard discoursing parsing evaluations when compared against current state-of-art systems.
Introduction
Discourse parsing is the task of uncovering the hierarchical structure of a text—identifying how small units (clauses) link together to form sentences, and how sentences link to form a complete, coherent document. Unlike sentence-level parsing, discourse lacks a formal grammar. Traditional systems rely heavily on surface features like cue words (e.g., "but," "however") or part-of-speech tags, which often fail to capture the deeper intentionality and semantics required to understand complex rhetorical relations.
Methodology
The proposed framework utilizes Recursive Neural Networks (RNNs) to model discourse structure bottom-up. The process involves four key stages:
- EDU Modeling: Text is segmented into Elementary Discourse Units (EDUs). Each EDU is represented as a fixed-length vector generated via recursive convolution over its syntactic parse tree.
- Binary Structure Classification: A three-layer neural network evaluates adjacent discourse units to determine if they should be merged into a larger subtree.
- Multi-class Relation Classification: For merged units, a second classifier identifies the rhetorical relation (e.g., Contrast, Attribution) connecting them.
- Recursive Composition: Merged units are represented by a new vector calculated from the children's vectors and the identified relation type. This allows the model to hierarchically represent the entire document.
For inference, a probabilistic CKY-like algorithm is employed to find the most likely global tree structure for a document.
Experimental Setup
The model was evaluated on the RST Discourse Treebank (RST-DT), consisting of 385 documents from the Wall Street Journal. Following standard practice, 110 fine-grained relations were grouped into 18 coarse-grained categories. The dataset was split into 347 documents for training and 38 for testing.
Results and Discussion
The recursive model achieves performance comparable to state-of-the-art systems of its time, such as HILDA and systems by Feng & Hirst.
Key insights from the results include:
- Meaning over Surface Cues: Distributed representations allow the model to capture relations without explicit cue words.
- Unified Treatment of Spans: The model treats within-sentence and cross-sentence spans uniformly, regardless of their level in the hierarchy.
- The Power of Composition: The semantic vector of a span is influenced by the specific rhetorical relation connecting its parts, capturing the "meaning" of the discourse structure.
While the model matches manual feature-engineered systems, it remains dependent on gold-standard segmentation and faces computational complexity challenges for long documents, requiring beam search (maintaining top-k options) during CKY inference.
Conclusion
This work demonstrates that deep learning can effectively model discourse structure by learning distributed representations. By jointly training classifiers and representation-learning parameters, the system captures the semantic and intentional "glue" of coherent text. Future work could apply these discourse-level embeddings to downstream tasks like sentiment analysis or summarization.
Summary updated