Introduction
The detection of "hyperpartisan" news—content that is extremely one-sided or inflammatory—is a critical step in fighting misinformation. However, training machine learning models for this task faces two major hurdles: noisy text and noisy labels. Political news is filled with domain-specific slang and evolving jargon that often fall "out of vocabulary" (OOV) for standard models. Furthermore, large datasets often rely on "weak labeling," where every article from a specific publisher is automatically labeled as hyperpartisan, regardless of the individual article's actual content. This mismatch creates "label noise" that can significantly degrade model performance.
Methodology
Team yeon-zi addressed these challenges by proposing a two-step pipeline designed to prioritize data quality over raw quantity.
Step 1: Data De-noising via Pseudo-labeling
Instead of trusting the 750,000 weakly-labeled articles blindly, the team used a small but "clean" dataset of 645 human-labeled articles to train an initial classifier. This model was then used to predict labels for the large noisy dataset. Articles were only kept if the model's prediction matched the original publisher-level label with high confidence (a prediction score at least 0.2 away from the neutral 0.5 mark). This aggressive filtering shrunk the dataset from 750,000 articles down to a high-quality subset of roughly 32,000 articles.
Step 2: BERT-based Representation
To handle the "noisy text" problem, the team leveraged BERT (Bidirectional Encoder Representations from Transformers). They fine-tuned BERT on the news domain to better capture political nuances and used its byte-pair vocabulary to eliminate OOV issues. The final architecture added LSTM layers on top of the fine-tuned BERT to process both the article title and the body text.
Key Results
The effectiveness of the de-noising strategy and domain-specific fine-tuning is evident in the performance metrics on the final by-article test set.
The proposed approach (BERT_ft + De-noise) achieved 75.8% accuracy, ranking 11th out of 43 teams. Notably, training on just 32,000 "clean" articles (8.5% of the original data) yielded significantly better results than using the full noisy dataset. Furthermore, fine-tuning BERT specifically on the news dataset provided a ~10% boost in accuracy over vanilla BERT, proving that political discourse requires specialized language understanding.
Conclusion
Team yeon-zi's approach demonstrates that data quality and domain adaptation are more impactful than model complexity alone. By filtering out label noise and adapting language models to the specific nuances of political news, they achieved competitive results. Future work may involve explicitly leveraging "stance on sensitive topics" or topic modeling to further refine detection capabilities.
Summary updated