Abstract
This paper describes a cascaded discourse parsing system developed for the CoNLL 2015 Shared Task. The system employs a pipeline of Maximum Entropy (MaxEnt) classifiers to identify discourse connectives, extract their argument spans, and determine their semantic senses. By decomposing the complex task of shallow discourse parsing into discrete classification steps and utilizing syntactic tree features, the system achieved a competitive 10th-place ranking. Evaluation on the Penn Discourse Treebank (PDTB) 2.0 shows high performance in connective identification, though overall end-to-end parsing remains a significant challenge due to error propagation.
Introduction
Discourse parsing aims to uncover the logical relationships between segments of text, a critical component for high-level natural language understanding. For instance, in the sentence "I stayed home because it was raining," a parser must identify "because" as the connective, "I stayed home" as Argument 1, "it was raining" as Argument 2, and the sense as "Causal."
This work focuses on "shallow" discourse parsing, which involves identifying individual relations without constructing a full hierarchical document tree. The task is divided into identifying explicit relations, signaled by specific words like "but" or "however," and non-explicit relations, where the link is implicit (e.g., "I'm tired. I'm going to bed.").
System Architecture
The authors implemented a cascaded system using the Maximum Entropy (MaxEnt) framework. The architecture is split into two primary components: an Explicit Relation Parser and a Non-Explicit Relation Parser.
Explicit Relation Parser
The explicit parser follows a four-step pipeline:
- Connective Classifier: Determines if a potential connective word is actually serving a discourse function in its context.
- Argument Position Classifier: Identifies whether Argument 1 resides in the same sentence as the connective or a preceding one.
- Argument Extractor: Utilizes constituent parse trees to label specific nodes as part of the argument spans. A "tree subtraction" method is used to refine boundaries when nodes overlap.
- Explicit Sense Classifier: Assigns one of 15 semantic senses to the identified relation.
Non-Explicit Relation Parser
For relations lacking a clear connective (Implicit and AltLex), the system analyzes adjacent sentence pairs. It relies on word-pair features and Mutual Information to select the most informative lexical cues, reducing noise in the classification process.
Data and Evaluation
The system was trained and evaluated using the Penn Discourse Treebank (PDTB) 2.0 corpus. The training set comprised sections 02-21 (32,535 relations), while section 22 served as the development set (1,436 relations). Testing was performed on section 23 and a separate "blind" test set provided by the CoNLL organizers to evaluate generalization to new domains.
Results
The system's performance highlights the varying difficulty of discourse parsing sub-tasks. While connective identification reached high F1 scores (above 0.90), the end-to-end parser performance was significantly lower, reflecting the impact of error propagation through the pipeline.
As shown in the results table, the system achieved an end-to-end F1 score of 0.1814 on the development set and 0.1505 on the test set. The team ranked 10th out of 17 participants. The sharp drop in performance from connective identification to full parsing underscores the difficulty of accurately extracting argument spans and senses.
Discussion and Conclusion
The study demonstrates that a pipeline of MaxEnt classifiers is a robust baseline for shallow discourse parsing. Key insights include the effectiveness of syntactic path features for argument extraction and the utility of probability thresholding to improve classification confidence.
However, several limitations remain. Implicit relations continue to be the most challenging aspect of the task, as word-pair features can be noisy. Furthermore, the system is highly dependent on the quality of external syntactic parsers; errors in the initial parse tree directly impede argument extraction. Future work will focus on more sophisticated feature engineering for implicit relations and exploring methods to mitigate error propagation within the cascaded architecture.
Summary updated