Abstract
Fang Kong, Hwee Tou Ng, Guodong Zhou. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP).
2014.
Introduction
Discourse parsing is a fundamental task in Natural Language Processing (NLP) that involves identifying the semantic relationships between text segments. A critical component of this process is discourse argument labeling: identifying the specific text spans (Arg1 and Arg2) that are linked by a discourse connective such as "but", "because", or "so".
This 2014 EMNLP paper proposed a novel framework that shifts the focus from word-level tagging to syntactic constituents. While the paper initially claimed significant performance gains through a joint inference mechanism, it has since become a notable case study in the importance of reproducibility following a 2015 addendum.
The Challenge of Argument Labeling
Identifying the full spans of discourse arguments is significantly more complex than identifying the connectives themselves. For instance, in the sentence:
"But its competitors have much broader business interests and so are better cushioned."
The connective "so" links an Arg1 ("its competitors have much broader business interests") to an Arg2 ("are better cushioned").
Historically, researchers used two main approaches:
- Linear Tagging: Treating the task as sequence labeling (word-by-word). While effective for local patterns, it often ignores the hierarchical structure of the sentence.
- Subtree Extraction: Attempting to map an argument to a single node in a parse tree. This is often too restrictive, as arguments frequently span multiple nodes or only parts of subtrees.
The Constituent-Based Approach
To address these limitations, Kong et al. proposed a "middle ground" that leverages syntactic structure while maintaining flexibility. Instead of individual words, the system treats every node (constituent) in a syntactic parse tree as a potential candidate for an argument label (Arg1, Arg2, or NULL).
System Workflow
- Pruning: To manage the high number of nodes in a parse tree, a recursive algorithm filters out constituents unlikely to be arguments based on their structural distance from the connective.
- Unified Handling: The model treats the immediately preceding sentence as a "special constituent," allowing it to handle inter-sentential arguments using the same logic as intra-sentential ones.
- Parser Diversity: By combining candidates from different parsers (Charniak and Berkeley), the authors increased the recall, ensuring the correct argument was more likely to be present in the candidate pool.
Joint Inference via Integer Linear Programming (ILP)
The core theoretical novelty of the paper was the use of Joint Inference to enforce global consistency. A local classifier might independently label two overlapping phrases as Arg1 and Arg2, which is linguistically impossible. The authors used Integer Linear Programming (ILP) to apply global constraints, such as ensuring that Arg1 and Arg2 do not overlap and that every connective has at least one Arg2.
Results and the 2015 Addendum
The paper initially reported state-of-the-art results on the Penn Discourse Treebank (PDTB). Using gold-standard trees, the authors claimed an F1-score of 58.18% for exact matches of both Arg1 and Arg2, a substantial improvement over previous benchmarks.
However, a 2015 addendum revealed that these gains could not be replicated. Due to errors in the experimental setup and evaluation code, the conclusion that joint inference (ILP) significantly improves performance was retracted. While the constituent-based candidate selection remains a valid architectural contribution, the specific benefits of the ILP layer were found to be unsupported.
Conclusion
The work by Kong et al. highlights the importance of syntactic structure in discourse parsing. While the joint inference claims were later corrected, the shift toward constituent-based analysis provided a robust framework for future research. The paper serves as a vital reminder of the necessity for rigorous code validation and the value of reproducibility in the NLP community.
Summary updated