GraphNLI: A Graph-based Natural Language Inference Model for Polarity Prediction in Online Debates

• Predicting whether a reply supports or attacks its parent can help measure debate controversy and formally identify attacked, rebutted, or supported arguments.Source: §1 · p. 1
• Text-only models that consider a reply and its parent in isolation risk losing crucial information from the wider debate.The paper asks whether incorporating the whole discussion thread can improve polarity prediction.Source: §1 · pp. 1–2
• GraphNLI uses graph walks to sample discussion-tree context and adds the sampled nodes to the local reply-and-parent inputs.A weighted root-seeking walk performed best at capturing wider Kialo debate context.Source: §1 · p. 2
• 83% accuracy, outperforming baselines including S-BERT, is the reported overall result for GraphNLI.Source: §1 · p. 2
• Parent and other ancestor nodes help more than siblings and children, and their importance decreases with distance.Source: §1 · p. 2

• An abstract argumentation framework represents arguments and their attack relations, while bipolar frameworks add support relations and assign each edge a polarity.Polarity denotes whether an edge is attacking or supporting.Source: §2 · p. 2
• Representing online debates as bipolar argumentation frameworks enables argument-theoretic analyses of support and attack relations.Such analyses include identifying justified and rebutted arguments.Source: §2 · pp. 2–3
• Prior polarity-prediction approaches generally use only the replying argument and the argument it replies to as model inputs.Source: §2 · p. 3
• Whether adding ancestor texts, nearby comments, or graph features improves polarity accuracy had not yet been investigated in the argument-mining literature, to the authors’ knowledge.Source: §2 · p. 3

• Kialo users submit claims in discussions that develop into directed trees because each non-thesis claim replies to exactly one other claim.Source: §3 · p. 3
• The Kialo snapshot contains 1,560 discussion threads, each with a root thesis and replies labeled as attacking or supporting their parent.Source: §3 · p. 3
• Kialo’s strict moderation means each post is treated as a self-contained argument with a clear claim backed by reasoning.Source: §3 · p. 3

4.1 Representing Debates as Discussion Trees

• Each debate is represented as a tree whose thesis is the root, with directed edges from each reply to the post it answers and labels for support or attack.Source: §4.1 · p. 3

4.2 GraphNLI Architecture

• GraphNLI captures local and global debate context through graph-based walks to predict the polarity of replies.Source: §4.2 · p. 3, §4 · p. 3

4.2.1 Capturing Global Context through Graph Walks.

• Kialo data collection stopped after a January 2020 backend change made the crawler obsolete.The authors considered the collected data sufficient to test GraphNLI.Source: §4.2.1 · p. 3
• Walks ignore edge direction to reach neighboring context, including siblings; root-seeking walks weight nearer ancestors more heavily, while biased random walks favor the parent.The weighted root-seeking walk uses L=5, and the biased random walk uses L=4; empirically, p=0.75 gives the best random-walk results.Source: §4.2.1 · pp. 4–5

4.2.2 Model Overview.

• For each point-of-interest post, GraphNLI encodes the post and sampled neighbors with RoBERTa, pools neighbor embeddings, then classifies their concatenation with the absolute embedding difference.The classifier is a softmax layer; the model is end-to-end trainable during backpropagation.Source: §4.2.2 · p. 5

5.1 Dataset Preprocessing

• The randomly selected 80% training split contained 259,499 arguments, while the test split contained 64,874; the debates were generally balanced between support and attack.Source: §5.1 · p. 5

5.2 Training Details

• GraphNLI used a root-seeking walk with ancestor weights decreasing by distance, then fine-tuned with cross-entropy for four epochs.The discount parameter was γ = 0.75; training used batch size 16 and Adam with a 2 × 10−5 learning rate.Source: §5.2 · pp. 5–6

5.3 Baselines

• Baselines included unigram bag-of-words, prompt embeddings, Sentence-BERT, and a two-layer MLP over fixed BERT embeddings.Source: §5.3 · p. 6

5.4 Model Evaluation

• 82.87% accuracy with weighted-average root-seeking GraphNLI exceeded all reported baselines; the unweighted variant reached 81.96%.Source: Table 1 · p. 6, §5.4 · p. 6

5.5 Ablation Study

• Root-seeking walks outperformed biased root-seeking random walks by 1.52% accuracy, consistent with upstream text contributing more than sibling and child context.Source: §5.5 · pp. 6–7
• Weighted-average aggregation performed better than summation and simple averaging, supporting progressively lower weights for more distant ancestors.Source: §5.5 · p. 7
• Concatenating the node embedding, its neighborhood embedding, and their absolute difference performed best among the tested concatenation methods.The authors report that adding element-wise multiplication decreased performance, as also reported by [38].Source: §5.5 · p. 7, Table 2 · p. 7
• A model with fixed BERT embeddings performed worse than Sentence-BERT, while end-to-end training produced embeddings suited to polarity prediction.Source: §5.5 · p. 7

6 CONCLUSIONS AND FUTURE WORK

• Weighted root-seeking walks with weighted ancestor aggregation gave the best classification accuracy, and ablations found parents and other ancestors more relevant than siblings or children.Ancestor importance decreased with distance from the reply node.Source: §6 · p. 7
• Polarity predictions could support conversation-health measures, help flag possible hate or misogyny, and help identify filter bubbles or represent viewpoints in debates.Source: §6 · p. 7
• The authors identify applying GraphNLI to the described applications as future work.Source: §6 · p. 7
• The approach was demonstrated and evaluated on Kialo, a tightly moderated platform; validity on noisier, weakly moderated discourse such as Reddit remains to be tested.Some forums also lack explicit reply structure, so reply links would need to be inferred before applying the graph walks.Source: §6 · p. 8