Detecting Political Bias in News Articles Using Headline Attention

Rama Rohit Reddy Gangula, Suma Reddy Duggenpudi, Radhika MamidiACL 2019

TL;DR

This paper introduces a novel "Headline Attention Network" to automatically detect political bias in news articles, particularly for the Telugu language. By focusing on how headlines influence the interpretation of article content, their model achieves significantly better bias detection accuracy than existing methods, offering a valuable tool for media verification.

What Problem Does This Paper Address?

Identifying political bias in news media is a crucial but challenging task. Traditional methods often rely on manual analysis, which is time-consuming and labor-intensive. While automated methods exist, they often treat headlines and articles as separate entities or simply concatenate them. This paper argues that headlines play a disproportionately important role in shaping a reader's perception of an article's bias, and proposes a model that leverages this headline-article interaction for more effective bias detection. The authors specifically address this challenge in the context of Telugu news, for which a suitable dataset was lacking.

Methods

The core of the paper is the Headline Attention Network. This architecture consists of three main components:

  • Headline Encoder: A bidirectional LSTM processes the headline words to create a contextual representation.
  • Article Encoder: Similarly, a bidirectional LSTM encodes the article's words, capturing contextual information for each word.
  • Headline Attention Layer: This is the novel component. It learns to weigh the importance of different words in the article based on their relevance to the headline's representation. This is achieved by calculating an attention score for each article word, which is essentially a measure of similarity between the word's representation and the headline's representation. These scores are then used to create a weighted sum of article word representations, forming a final article representation that is heavily influenced by the headline.

This final representation is then fed into a classifier to predict the political party towards which the article is biased. The model is trained using a negative log-likelihood loss function.

Datasets

The authors created a new dataset specifically for this research. It comprises 1329 Telugu news articles collected from various Telugu newspapers. Each article was manually annotated by four native Telugu speakers for bias towards one of five political parties (BJP, TDP, Congress, TRS, YCP) or marked as "None" if unbiased. A high Kappa score of 0.9 was achieved, indicating good inter-annotator agreement.

Key Insights and Ideas

  • Headline Primacy: The paper strongly supports the intuition that headlines are critical for detecting political bias. Readers often form initial impressions based on headlines, and this influences how they interpret the subsequent article content.
  • Attention is Key: Simply concatenating headlines and articles is not as effective as using an attention mechanism. The Headline Attention Network's ability to dynamically weigh article words based on their relevance to the headline is crucial for its superior performance.
  • Contextual Understanding: The use of bidirectional LSTMs for both headline and article encoding allows the model to capture rich contextual information, which is essential for nuanced bias detection.
  • Domain Specificity: The creation of a Telugu-specific dataset highlights the need for language-specific resources in NLP research, especially for under-resourced languages.

Key Results

The Headline Attention Network achieved the best performance among all tested methods, outperforming previous best baselines by 4.22%. Specifically, when comparing neural network methods, the Headline Attention Network significantly improved accuracy compared to simply concatenating headline and article content. The paper also shows that considering only the headline can yield good results, but the full attention mechanism integrating headline and article provides the highest accuracy.

Limitations and Future Work

The authors acknowledge that their model can be extended to other sentiment-based classification tasks involving texts with a title/headline and a body, such as blogs or online trending articles. They also suggest that visualizing the attention weights can effectively highlight words that contribute to bias, offering interpretability. Further research could explore applying this model to other languages and different types of bias beyond political affiliation.

Summary updated

BibTeX
@inproceedings{gangula2019detecting,
  author = {Rama Rohit Reddy Gangula and Suma Reddy Duggenpudi and Radhika Mamidi},
  title = {{Detecting Political Bias in News Articles Using Headline Attention}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure