BiasNet: A Contrastive GNN-Based Framework for Classifying Political Stance in News

01 Introduction

News shapes public opinion and underpins accountable, democratic decision-making — which is exactly why the rise of hyperpartisan content is a problem. Hyperpartisan articles display strong ideological lean toward the left or right, distort perception, and widen political divides. Detecting this lean automatically is a prerequisite for balanced information consumption, yet conventional text classifiers process each article in isolation, in a vacuum, missing the relational signals — shared sources, overlapping topics, similar phrasing — that connect biased narratives to one another.

BiasNet addresses this gap directly: instead of classifying articles one at a time, it models the corpus as a graph and lets a Graph Neural Network propagate context between related articles, while a contrastive objective sharpens the separation between ideological clusters in embedding space.

02 Related Work

Hyperpartisan detection research has moved through three broad waves, summarized below.

Stylometric & lexical baselines

Potthast et al. showed that readability, vocabulary richness and syntactic patterns alone give strong discrimination with classical ML. Srivastava et al. combined a bias lexicon with GloVe / Doc2Vec embeddings and logistic regression, reaching 82.0% accuracy — the top F1 score at SemEval 2019.

Contextual & transformer-based models

Jiang et al.'s ELMo Sentence Representation CNN won SemEval 2019 outright at 82.2% accuracy. Huang & Lee combined BERT and ELMo; Isbister & Johansson applied ULMFiT transfer learning (80%); Hanawa et al. ensembled 400 models to reach 80.9%. Several of these works also surfaced a recurring failure mode: transformers fine-tuned naively tend to overfit to publisher-specific signals rather than genuine ideological cues.

Robustness, disentanglement & graph methods

Baly et al. used adversarial training (FGM) for better out-of-domain stability. Raza et al.'s Dbias disentangles bias-specific from bias-neutral features, but — like most prior work — does not model relationships between articles. Ding et al.'s HyperGAT applies hypergraphs to text classification generally, without a contrastive or bias-aware objective. GraphCL (You et al.) introduced contrastive pretraining for graphs but not for this task. Most recently, Mohan & Pengyu compared n-grams, ELMo and LLM embeddings (Gemma3, LLaMA), finding that LLaMA embeddings with a downstream classifier reached 92.6% — the strongest prior baseline BiasNet is compared against.

03 Methodology

BiasNet combines a graph representation of the news corpus with a message-passing GNN encoder and a dual training objective: standard classification plus supervised contrastive alignment.

3.1 Dataset

BiasNet uses the SemEval-2019 Task 4 dataset, which labels political news either byarticle (crowdsourced, article-level, high annotator agreement) or bypublisher (inferred from the publisher's overall lean, per BuzzFeed / MediaBiasFactCheck.com). For tractability, the study uses a subset of 3,832 articles split 80 / 10 / 10 into train / validation / test, moderately imbalanced across classes.

A second, independent set of 150 recently published articles — spanning publishers across the political spectrum (Counterpunch, FoxBusiness, DailyWire, Texas Tribune) — was manually labeled with the AllSides bias-checker tool, purely for out-of-distribution evaluation.

3.2 Graph construction

The corpus is modeled as an undirected graph G = (V, E). Each node is an article, initialized with its 768-dimensional LaBSE embedding — chosen over English-only encoders like vanilla BERT for its multilingual, cross-source robustness. Two articles are connected by an edge when their embedding cosine similarity exceeds a tuned threshold of 0.8: low enough to link genuinely related articles, high enough to avoid a dense, uninformative graph.

3.3 Graph Neural Network model

A K-layer message-passing GNN learns node and graph-level representations. At each layer, every node aggregates its neighbors' representations and combines them with its own:

After K layers, a READOUT function pools node embeddings into a graph-level summary, which an MLP maps to classification logits:

Two GNN variants are compared:

Both variants share two task heads: a classification head producing stance logits, and a projection head (2-layer MLP) mapping embeddings into the contrastive space.

3.4 Training objectives

Classification uses standard cross-entropy over the three stance classes:

To sharpen the embedding space itself, a supervised NT-Xent contrastive loss pulls together projections of same-label articles and pushes apart the rest. For anchor i with positive set P(i) = { j ≠ i : label(j) = label(i) }:

The two objectives are combined with weight λ:

3.5 Graph augmentation for contrastive views

Each training step generates two stochastic views of the article graph:

• Edge dropout — randomly removes a fraction of edges, reducing reliance on any single neighbor and mitigating spurious publisher-specific links.
• Feature masking — randomly masks a fraction of node features, simulating partial or noisy text representations.

Both views pass through the same GNN encoder and projection head; NT-Xent is then applied across the two views so same-label nodes are pulled together regardless of which perturbation they saw.

3.6 Optimization & hyperparameters

Models train for 100 epochs with Adam, weight decay, and early stopping on validation accuracy. Depth is kept shallow (K = 2) with dropout to counter over-smoothing.

A similarity threshold of 0.7 produced an over-dense, less discriminative graph; 0.9 was too sparse for effective aggregation. 0.8 offered the best trade-off. Lower NT-Xent temperature (τ = 0.5) sharpened positive/negative contrast and improved cluster separation.

3.7 Application to unseen articles

For real-time use, a new article is embedded with the same LaBSE model, added as a node connected by cosine similarity to the existing graph, and classified by the trained GNN using its own features and its neighbors' influence — enabling continuous, incremental classification.

04 Results

Four architectures were evaluated: GCN and GAT, each with and without the contrastive objective. All metrics are averaged over five runs with 95% confidence intervals.

4.1 Comparison with prior SemEval systems

On the 150-article, out-of-distribution evaluation set, the best-performing configuration held up at 86.0% accuracy — evidence of meaningful generalization beyond the training distribution, across publishers it had never seen.

4.2 Embedding visualization (UMAP)

4.3 Statistical significance

Contrastive learning delivers a statistically significant lift for GAT, but not for GCN — while both contrastive variants converge to comparable, strong performance overall.

05 Limitations

06 Discussion

The gap between training and test accuracy — e.g. GCN + Contrastive reaches 99.8% train but 95.02% test — signals residual overfitting: models are still partly learning dataset- or source-specific artifacts rather than pure ideological structure. Contrastive learning narrows this gap but doesn't close it. The authors point to harder regularization, more aggressive graph augmentation, and more diverse training data as the likely next steps toward consistent performance on unseen or temporally shifting news content.

07 Conclusion

BiasNet reframes hyperpartisan news detection as a relational problem rather than an isolated text-classification one. By representing articles as nodes in a similarity graph and training a GNN with a joint classification + contrastive objective, it captures both content and structure — outperforming prior SemEval-2019 systems substantially and holding up on genuinely unseen, recent articles. Future work aims at heterogeneous graphs that jointly model articles, authors and sources, multilingual and cross-platform extension, and temporal modeling to track how bias evolves over time.

— References