Sequence Graph Network for Online Debate Analysis

TL;DR

This work tackles online debate analysis by jointly modeling temporal progression and inter-argument interactions. It introduces the Sequence Graph Network (SGN) and a Sequence Graph Attention (SGA) mechanism that propagate information along a debate graph constructed from sentence-level nodes connected by intra-turn, reinforcement, and counterargument edges, processed turn-by-turn. Utterances are encoded with SBERT and augmented with turn embeddings, then three GATs aggregate edge-type signals before a sequential update produces debater-specific representations read out for winner prediction via a pairwise cross-entropy objective. Across the Oxford-style debate dataset debate.org, the SGN outperforms sequence and graph baselines, with ablations demonstrating the critical role of counterarguments and edge construction choices, highlighting the model’s potential for understanding dynamic argumentative dynamics in discourse.

Abstract

Online debates involve a dynamic exchange of ideas over time, where participants need to actively consider their opponents' arguments, respond with counterarguments, reinforce their own points, and introduce more compelling arguments as the discussion unfolds. Modeling such a complex process is not a simple task, as it necessitates the incorporation of both sequential characteristics and the capability to capture interactions effectively. To address this challenge, we employ a sequence-graph approach. Building the conversation as a graph allows us to effectively model interactions between participants through directed edges. Simultaneously, the propagation of information along these edges in a sequential manner enables us to capture a more comprehensive representation of context. We also introduce a Sequence Graph Attention layer to illustrate the proposed information update scheme. The experimental results show that sequence graph networks achieve superior results to existing methods in online debates.

Figures (4)

• Figure 1: A “what-should-we-mention" information flow scheme that mimics the interaction process of a debater. At each time step $t$, the node features are updated by considering their peer nodes from the same turn and the connected nodes from previous turns, using Directed Graph Attention Network layers. Nodes associated with different debaters are colored differently. Each type of edge (colored arrows) contributes a corresponding representation, collectively forming $\mathbf{h}_i$. The node's utterance embedding $\mathbf{h}$ and the interaction representation $\mathbf{h}_i$ are used to update the node feature $\mathbf{h}'$.
• Figure 2: Graph Construction from Debate: Nodes establish connections through three distinct edge types, indicated by colored arrows. Intra-argument edges (blue) link nodes within the same turn, reinforcement edges (green) connect nodes from the same debater across different turns, while countering edges (orange) connect nodes from a debater to their opponent's, illustrating counter-argumentation. The sample debate is taken from data collected by durmus2019corpus.
• Figure 3: The proposed architecture consists of three key modules: (1) Information propagation is driven by the SGA layers, updating node features sequentially using a graph attention mechanism. (2) The readout layer identifies representative vectors associated with each debater, which are subsequently supplied as input to (3) an MLP classifier for predicting the debate winner.
• Figure 4: Impact of cross-argument construction values on network performance. Left: Edge construction using a threshold value. Right: Edge construction using top-k highest values.