TLDR: This research explores how news media emotionally frames stories, particularly focusing on Bengali headlines. Using zero-shot inference with the Gemma-3 4B model, the study analyzed 300,000 Bengali news headlines and found a significant dominance of negative emotions like anger, fear, and disappointment. Based on these findings, the authors propose a “Bias-Sensitive News Interface” to visualize emotional cues, helping readers recognize affective bias and promoting more mindful news consumption and media literacy.
News media plays a significant role in shaping public sentiment, not just through the facts they report, but also by how they present those facts. The same event can be portrayed in vastly different emotional tones by various outlets, a phenomenon known as affective bias. This emotional framing can profoundly influence how audiences interpret issues, whom they empathize with, and what content they choose to share.
In today’s fast-paced digital world, emotionally charged news, especially content that evokes anger, fear, or outrage, tends to spread more rapidly and gain greater attention. This often encourages media outlets to frame stories in ways that provoke stronger reactions, contributing to what is known as the negativity bias in journalism. Over time, this can lead to affective polarization, where emotions, rather than evidence, become the primary drivers of news consumption.
Headlines are particularly powerful in this process, acting as emotional gateways that determine whether readers engage with a story at all. Studies suggest that a significant portion of shared articles on social media are never clicked, meaning headlines alone often shape much of the public’s initial emotional response to current events. The impact of such emotional exposure is not trivial; continuous consumption of negative news has been linked to increased worry, hopelessness, anxiety, and even the phenomenon of ‘doom-scrolling,’ which can worsen mood and heighten stress.
A recent study, titled How News Feels: Understanding Affective Bias in Multilingual Headlines for Human-Centered Media Design, delves into this critical area by examining Bengali news media, an ecosystem largely unexplored in affective computing research. The authors, Mohd Ruhul Ameen, Akif Islam, Abu Saleh Musa Miah, Ayesha Siddiqua, and Jungpil Shin, utilized zero-shot inference with the Gemma-3 4B model to analyze 300,000 Bengali news headlines and their content. This large-scale analysis aimed to identify the dominant emotion and overall tone of each headline without requiring any prior fine-tuning or manual annotation.
Key Findings: A Dominance of Negative Emotions
The research revealed a clear and significant dominance of negative emotions in Bengali news headlines. Emotions such as anger, sadness, disappointment, and fear collectively accounted for over half of all headlines analyzed. In contrast, neutral tones and positive emotions like joy were far less frequent. Notably, ‘disappointment’ emerged as a highly frequent category, suggesting that many headlines are framed through a lens of frustration or dissatisfaction.
This pattern highlights an inherent affective imbalance in Bengali news reporting, where emotionally negative framing appears to be the norm. The study also found that similar events are often framed differently by various outlets, with some amplifying fear or outrage while others report more calmly. This emotional intensity is a key driver of attention and virality in news, reinforcing the idea that ‘ill news runs a pace’ – negative stories spread faster and gain more engagement.
Methodology: Leveraging AI for Emotion Analysis
To conduct this extensive analysis, the researchers collected approximately 824,784 news items from major online portals in Bangladesh, including trusted sources like Prothom Alo, BDNews24, Ittefaq, and Samakal. A representative subset of 300,000 articles was then selected for detailed emotion analysis due to computational demands.
The Gemma-3 4B large language model was deployed locally through Ollama for emotion classification, utilizing the GoEmotions-28 taxonomy, which defines 28 fine-grained emotions. These were further grouped into seven coarse-level classes: joy, sadness, anger, fear, surprise, disgust, and neutral. To optimize efficiency, a load-balanced multi-instance setup was implemented, running three parallel Ollama endpoints on a single NVIDIA RTX 4090 GPU. This setup significantly reduced processing time, allowing the analysis of 300,000 headlines in approximately 40 hours.
Proposed Solution: The Bias-Sensitive News Interface
Building on these insights, the researchers propose a human-centered news aggregator called the ‘Bias-Sensitive News Interface.’ This system aims to transform large-scale emotion classification results into interactive, interpretable dashboards, promoting transparency, emotional balance, and reader well-being by revealing how emotional framing varies across outlets, topics, and time.
The proposed design includes three interconnected modules:
-
News Feed Analysis View: This central dashboard displays real-time emotional dynamics, including average valence (positivity/negativity), arousal (intensity), dominant emotion, and an Affective Polarization Index (API) to quantify emotional divergence across outlets. It allows for side-by-side comparison of how similar events are emotionally covered.
-
Bias-Sensitive News Interface: This module visualizes broader, cross-outlet emotional trends. It presents emotion distribution by outlet, traces emotional intensity over time, and introduces cross-outlet polarization metrics like the API and Jensen–Shannon Divergence (JSD) to quantify how differently outlets frame similar stories.
-
Detailed Emotion Analysis Panel: This panel offers a micro-level exploration of individual headlines, showing their emotion composition (e.g., sadness: 50%, fear: 30%), valence, and arousal scores. It highlights how the same event can be reported with different emotional tones across various outlets, helping users consciously and critically interpret news framing.
Also Read:
- Interpreting Digital Emotions: A Hybrid Framework for Nuanced Sentiment Analysis
- AI Uncovers Latent Political Narratives in Social Media Advertising
Significance for Human-Centered Media Design
The proposed system illustrates how emotion-aware AI can foster media literacy and promote emotional accountability in journalism. By visualizing how emotions drive engagement, it empowers readers to recognize affective manipulation in digital news. It helps users understand how sustained exposure to high-arousal, low-valence emotions like fear and anger can lead to cognitive overload and emotional fatigue.
Designed for real-time adaptability, the system can continuously ingest data, perform multilingual analysis, and maintain privacy by computing metrics at an aggregate level. Its modular architecture allows for scalable deployment, from local GPU setups to cloud-based infrastructures. Ultimately, this design envisions an emotionally transparent news ecosystem where readers can make informed choices about their media consumption, and newsrooms can self-assess the emotional balance of their coverage, advancing journalism toward data-driven emotional responsibility.
