Abstract
Discourse Argument Identification (DAI) is a critical component of discourse parsing, tasked with identifying the text spans (Arg1 and Arg2) associated with a discourse connective. Traditional approaches often treat DAI as a sequence labeling task, which ignores the underlying syntactic structure of language. This paper presents a unified framework that treats DAI as a shallow semantic parsing task, focusing on constituents within a parse tree. By shifting the focus from individual tokens to structural units, the proposed system achieves significant performance gains on the Penn Discourse Treebank (PDTB) for both intra-sentence and inter-sentence relations, outperforming previous state-of-the-art models in robustness and accuracy.
Introduction
In Natural Language Processing, discourse parsing aims to uncover how different parts of a text relate to one another. A central challenge in this field is Discourse Argument Identification (DAI). When a "connective" word like because or however is identified, DAI must determine the exact boundaries of the two arguments it connects.
Historically, most approaches have treated this as a "chunking" or sequence labeling problem, asking whether individual words are part of an argument. The authors argue this approach is fundamentally flawed because words belong to syntactic structures. By ignoring the "tree" structure of a sentence, older models miss the rich grammatical context that defines argument boundaries.
Methodology
The authors propose a unified framework that handles two distinct scenarios: intra-sentence and inter-sentence arguments.
Intra-sentence Identification
For arguments within the same sentence, the task is recast as a Shallow Semantic Parsing problem. The discourse connective acts as a predicate, and its scope is mapped to constituents (phrases) in a parse tree.
- Pruning: To manage complexity, a tree-traversal strategy filters out unlikely candidates by examining the "siblings" and "parents" of the connective. The authors found that a "Level" threshold of 4 (searching four steps up the tree) provides the optimal balance between coverage and noise reduction.
- Classification: A two-stage classifier is used. A binary classifier first determines if a constituent is a valid argument, followed by a multiclass classifier that labels it as Arg1, Arg2, or Null.
- Features: The model utilizes lexical, syntactic, and connective-driven features. Notably, adding features for words immediately following a connective (next1, next2) significantly improves Arg1 localization.
Inter-sentence Identification
For arguments spanning different sentences, the system uses a lightweight, rule-based solution. Based on statistics from the Penn Discourse Treebank (PDTB), the system typically assigns the immediately preceding sentence as Arg1 and the current sentence (minus the connective) as Arg2.
Structural Representation
A key insight of this framework is that discourse arguments almost always align with syntactic constituents. By treating a whole phrase (e.g., a Noun Phrase) as a single unit, the model avoids the "independence assumption" errors common in token-level models.
Results
The framework was evaluated on the PDTB using both "Gold Standard" and "Automatic" parse trees. Key findings include:
- Arg1 Identification: On gold-standard trees, the system achieved an F-measure of 62.98%, surpassing the previous state-of-the-art of 59.39% (Ghosh et al., 2012).
- Robustness: Using automatic (imperfect) parsers, the system maintained a significant lead with an F-measure of 48.11%, compared to 39.00% in previous models.
- Position Accuracy: The improved feature set reached a 97.55% F-measure in identifying whether Arg1 was intra- or inter-sentence.
Conclusion
Treating DAI as shallow semantic parsing provides a more accurate and robust framework for discourse analysis. While the approach is dependent on the quality of the initial parse tree, it significantly reduces errors by respecting syntactic boundaries. Future work will explore tree kernel-based methods to learn complex structural patterns directly and implement post-processing sanity checks to ensure argument coherence.
Summary updated