A Decision-Based Approach to Rhetorical Parsing

Daniel MarcuACL 1999 · DOI

Introduction

This paper presents the first data-driven rhetorical parser designed to uncover the hidden hierarchical structure of text using a shift-reduce framework. By framing discourse analysis as a sequence of supervised machine learning decisions, the system automates the partitioning of text into elementary units and their assembly into cohesive tree structures, representing a significant shift away from traditional, hand-coded rule-based systems.

Problem Statement

While computational linguistics has made significant strides in sentence-level syntactic parsing, understanding the "rhetorical structure" of entire documents remains a complex challenge. Rhetorical Structure Theory (RST) posits that text can be modeled as a tree where the leaves are Elementary Discourse Units (EDUs)—typically clauses or clause-like snippets. Internal nodes represent relationships such as Contrast, Elaboration, or Condition, distinguishing between a Nucleus (core information) and a Satellite (supporting information).

Prior to this work, rhetorical parsers relied heavily on manual "if-then" rules that were difficult to scale or adapt to diverse writing styles. This research seeks to automate this process by learning directly from human-annotated discourse trees.

Methodology

The parsing process is treated as a two-stage pipeline powered by the C4.5 decision tree algorithm:

  1. Discourse Segmenter: This component performs a lexeme-by-lexeme scan to identify EDU boundaries. It utilizes a combination of local features (e.g., Part-of-Speech tags, cue phrases like because) and global features (e.g., presence of punctuation or verbs).
  2. Shift-Reduce Action Identifier: Once segmented, a shift-reduce parser constructs the tree using a stack of partial trees and an input list of EDUs. The model chooses from 103 possible operations, including SHIFT (moving an EDU to the stack) and REDUCE (joining the top two trees on the stack).
The shift-reduce parsing process through several steps, illustrating how partial trees are built on a stack using specific reduction operations.

The parser employs a rich feature set, including structural data (tree height), lexical cues, and semantic similarity scores derived from WordNet to identify conceptual relationships between text spans.

Experimental Setup

The model was trained and evaluated on a diverse corpus of 90 manually annotated rhetorical structure trees:

  • 30 personal news stories (MUC7 coreference corpus).
  • 30 scientific texts (Brown corpus).
  • 30 editorials (Wall Street Journal).

To address data sparsity, the original taxonomy of 71 rhetorical relations was consolidated into 17 clusters based on rhetorical similarity.

Key Results

The system demonstrated robust performance, particularly in the segmentation phase. The segmenter achieved high accuracy across all three corpora, though it faced challenges with parenthetical units compared to standard sentence breaks.

Segmentation accuracy results across the MUC, WSJ, and Brown corpora, showing consistent performance above 96%.

In the action identification phase, the shift-reduce classifier correctly predicted parsing operations approximately 61% of the time, significantly outperforming the most-frequent-action baseline (50%). The study also highlighted that while the model is adept at identifying hierarchical spans and nuclearity—performing near human levels when provided with perfect segments—labeling specific rhetorical relations remains a more difficult task, with the machine trailing human performance by 15-20%.

Discussion and Insights

Several critical insights emerged from the study:

  • Segmentation as a Bottleneck: Errors in identifying EDUs have a cascading effect; a missed boundary prevents the parser from constructing a correct tree.
  • Learnability of Nuclearity: The model effectively learns to distinguish essential Nuclei from supporting Satellites, suggesting this aspect of discourse is highly systematic.
  • Structure vs. Labels: Identifying that a relationship exists (structure) is significantly easier for the model than determining the precise nature of that relationship (labeling).
  • Data Sensitivity: Learning curves indicated that performance had not yet plateaued, suggesting that increased training data would yield further improvements.

Conclusion

The primary limitation of this work is the relatively small dataset of 90 texts, which limits the model's exposure to the full variety of human rhetorical patterns. Future research will focus on scaling the training data and exploring more sophisticated semantic features. For applications like automated summarization, the current segmenter may already provide sufficient utility by focusing on sentence-level units.

Summary updated

BibTeX
@inproceedings{marcu1999decision,
  author = {Daniel Marcu},
  title = {{A Decision-Based Approach to Rhetorical Parsing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1999},
  doi = {10.3115/1034678.1034736},
}
Figure