Abstract
This paper tackles the challenge of "non-tree" argument mining—where arguments form complex graphs rather than simple hierarchies—by introducing a model that jointly predicts proposition types and their relationships. By combining Task-Specific Parameterization (TSP) to separate feature encoding and Proposition-Level Biaffine Attention (PLBA) to score potential connections, the researchers achieved state-of-the-art results on the Cornell eRulemaking Corpus (CDCP).
Introduction
Most traditional argument mining systems assume that arguments follow a "tree" structure: a single main claim (the root) supported by several premises (the leaves), where each premise connects to exactly one parent. However, real-world discourse is rarely that tidy. In many contexts, a single premise might support multiple claims, or some statements might be entirely isolated.
The authors identify a gap in capturing these non-tree structures. Existing models often struggle with "flexible" arguments where a proposition can have several parents or none at all. The challenge lies in effectively encoding the propositions so the model knows which features matter for what a statement is (its type) versus how it relates to others (its edges).
Methodology
The authors propose a neural architecture consisting of two primary innovations designed to handle the complexities of non-tree discourse:
- Task-Specific Parameterization (TSP): Instead of using a single shared representation for all tasks, the model uses a shared BiLSTM encoder followed by two distinct "attention-to-encoder" layers. One branch focuses on identifying the proposition type (e.g., Fact vs. Value), while the other focuses on edge prediction (e.g., Statement A supports Statement B). This separation allows the model to capture task-specific nuances—like discourse markers for edges and pronouns for types—that might otherwise get blurred together.
- Proposition-Level Biaffine Attention (PLBA): Borrowing from token-level dependency parsing, the authors adapted biaffine attention to work at the proposition level. This mechanism computes a score for every possible pair of propositions in a text, allowing the model to predict multiple outgoing edges for a single node, effectively building a graph rather than a tree.
The model is trained using a joint objective function, meaning it learns to classify types and predict edges simultaneously, allowing the two tasks to inform one another.
Key Insights and Results
The primary dataset used is the Cornell eRulemaking Corpus (CDCP). Unlike the more common "Essay" datasets which are strictly tree-structured, CDCP contains 731 arguments with flexible, non-tree edges.
Performance Metrics
The proposed model significantly outperformed previous state-of-the-art (SOTA) models on the CDCP dataset:
- Edge Prediction: Achieved an F1 score of 34.04, a substantial jump over the previous best of 29.29.
- Proposition Type: Achieved an F1 score of 78.91, surpassing the previous SOTA of 73.5.
Feature Divergence
A critical finding was that the features required to identify a proposition's type differ from those needed to find its relationships. Edge prediction relies heavily on discourse markers like "because" or "so," whereas proposition types (specifically "Testimony") are often signaled by first-person pronouns like "I" or "me." Ablation studies confirmed that the task-specific layers (TSP) are crucial for stable performance in edge prediction.
Conclusion
By moving away from rigid tree constraints and employing task-specific parameterization, this research provides a more robust framework for mining arguments in unstructured, real-world text. While challenges remain regarding label imbalance—particularly for rare edge types like "Evidence"—the biaffine approach sets a new benchmark for non-tree argument mining.
Summary updated