TL;DR
This paper introduces a novel approach to rhetorical parsing that efficiently constructs underspecified representations of discourse structure using parse forests. By separating high-confidence information from hypothetical reasoning, their method allows for robust parsing even without explicit discourse markers, paving the way for more scalable and flexible discourse analysis.
What Problem Does This Paper Address?
Discourse parsing, the task of identifying the rhetorical structure of a text, is crucial for applications like automatic summarization. However, accurately determining this structure is notoriously difficult. Previous methods often rely heavily on heuristics or probabilistic information to find the "best" rhetorical tree. This paper tackles the challenge of building these structures more efficiently and robustly, especially in cases where explicit linguistic cues (discourse markers like "but" or "even though") are absent or ambiguous. The core problem is how to represent and compute possible discourse structures without getting bogged down in combinatorial explosion or requiring perfect input.
Methods
The authors propose a surface-based approach that combines an explicit rhetorical grammar with the construction of parse forests. Instead of generating a single, definitive parse tree, their system builds a forest – essentially an attributed And-Or-graph – that compactly encodes multiple possible analyses. This forest representation has two key properties:
- Subtree Sharing: If a particular substructure is used in multiple different analyses, it's represented only once, saving computational resources.
- Containment of Ambiguities: Disagreements or alternative interpretations are explicitly represented within the forest structure, allowing for later resolution or continued processing with multiple hypotheses.
The grammar itself is feature-based and designed to build these rhetorical trees from smaller units (chunks, phrases, sentences) up to larger "macro segments." They define specific rules for combining these units, distinguishing between rules that rely on discourse particles and those that can infer relations like SEQUENCE or ELABORATION more freely.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Parse Forests as a Solution to Ambiguity: The central idea is that representing all possible valid structures in a compact forest is more efficient and robust than trying to pick a single "best" tree upfront, especially when cues are weak.
- Separating Confidence Levels: The approach explicitly separates "high-confidence" information (e.g., from explicit markers) from "hypothetical reasoning" (inferred relations), allowing for a more principled way to handle underspecification.
- Grammar Design for Robustness: The grammar rules are designed to handle various combinations of text segments and discourse particles, including rules that can infer relations like SEQUENCE and ELABORATION even without overt markers. This makes the parsing more robust to variations in text.
- Cubic Complexity: By using chart parsing combined with parse forest construction, the authors claim their method achieves cubic complexity, which is a significant improvement over potentially exponential complexity in other approaches.
Key Results
The paper demonstrates the concept with a small example text, illustrating how the parse forest can represent ambiguities, such as the span of a relation or the choice between different relation types (e.g., CONTRAST vs. ELABORATION). They show a simplified parse forest and a sample parse tree derived from it. While no quantitative evaluation results are presented, the paper argues for the theoretical advantages of their approach in terms of efficiency and robustness.
Limitations and Future Work
- Underspecified Representation: The primary limitation acknowledged is that the output is an underspecified representation (a forest), not a single definitive tree. This means further processing is needed depending on the application, which could involve selecting the "best" tree or working with the set of hypotheses.
- Grammar Complexity: While the grammar is designed for efficiency, the paper notes that the inherent relational nature of RST trees can still lead to rule multiplication. They propose an inheritance hierarchy as a way to manage this, but it's an ongoing area of grammar design.
- Future Work: The authors suggest that depending on the application, further steps can either calculate the "best" tree out of the forest or continue working with a set of structured hypotheses. They also mention the work by Reitter and Stede on an XML-DTD scheme for representing these parse forests, indicating a direction for practical implementation and exchange.
Summary updated