Tracing Partisan Bias to Its Emotional Fingerprints: A Computational Approach to Mitigation

TL;DR

This work addresses the problem that media bias is embedded in emotional language and not just in topical polarity. It introduces NeutraSum, a neutralisation framework for multi-document summarisation that targets emotional fingerprints using a VAD-based analysis and two neutrality losses. The approach demonstrates that bias can be measurably reduced while preserving salient information, validated on the Allsides dataset with an auxiliary corpus to guide symmetry in representation. The proposed Chain-of-Thought based bias metric offers explainability, and the results suggest practical potential for producing emotionally neutral summaries in journalism and information synthesis.

Abstract

This study introduces a novel framework for analysing and mitigating media bias by tracing partisan stances to their linguistic roots in emotional language. We posit that partisan bias is not merely an abstract stance but materialises as quantifiable 'emotional fingerprints' within news texts. These fingerprints are systematically measured using the Valence-Arousal-Dominance (VAD) framework, allowing us to decode the affective strategies behind partisan framing. Our analysis of the Allsides dataset confirms this hypothesis, revealing distinct and statistically significant emotional fingerprints for left, centre, and right-leaning media. Based on this evidence-driven approach, we then propose a computational approach to mitigation through NeutraSum, a model designed to neutralise these identified emotional patterns. By explicitly targeting the VAD characteristics of biased language, NeutraSum generates summaries that are not only coherent but also demonstrably closer to an emotionally neutral baseline. Experimental results validate our framework: NeutraSum successfully erases the partisan emotional fingerprints from its summaries, achieving a demonstrably lower emotional bias score than other models. This work pioneers a new path for bias mitigation, shifting the focus from treating symptoms (political labels) to addressing the cause: the emotional encoding of partisan bias in language.

Figures (6)

• Figure 1: Emotional and Lexical Deviation from the Neutral Baseline. The chart visualizes the distinct emotional fingerprints of Left-leaning and Right-leaning media as deviations from the centre. The opposing shapes reveal their symmetrical but contrary emotional amplification strategies.
• Figure 2: Illustration of the training process of our model NeutraSum. The main task is to summarise a neutral summary $\hat{Y}^{D}$ from a Multi-document dataset $D$ using BART pre-trained language model. This dataset incorporates triplets of left, centre, and right-wing news articles $\left\{ X^{D}_{L}, X^{D}_{C}, X^{D}_{R} \right\}$, as well as expert-written summaries $Y^{D}$. Each triplet is reporting on the same issue. In the encoder framework, we introduced an Auxiliary Dataset $D_{\text{aux}}$ to propose polarised left-wing and right-wing news articles $X^{D_{\text{aux}}}_{L}$, $X^{D_{\text{aux}}}_{R}$. This dataset aims to join two neutrality losses (Equal-distance Loss and Contrastive Loss) by finetuning the whole semantic space. These two datasets jointly guide the neutral writing of a summary.
• Figure 3: An illustration of our Chain-of-Thought LLM-based Metric framework. Given media reports on the same issue and a generated summary, the framework employs a step-by-step reasoning process to assess media bias. It first identifies the main characters or topics and their associated attitudes, then analyses the underlying rationale behind these perspectives. Next, it extracts key words or phrases that reflect the detected attitudes and, ultimately, determines the political leaning of the text. After the chain-of-thought reasoning process, the framework leverages the VAD rating corpus to quantify the emotional tone of the extracted attitude-indicative words (from step 3) based on their Valence, Arousal, and Dominance scores.
• Figure 4: Scores of different models in political compass test. The horizontal axis represents the economic axis (ranging from left to right), and the vertical axis represents the social axis (ranging from liberal to conservative). It could be seen that the yellow points of our model NeutraSum and its variants have relatively lower bias scores in both axes. We also take the bias score of GPT4/4o and o1 for reference.
• Figure 5: Illustration of generated summary, the corresponding input, and expert-written summary Allsides:03. Yellow spans illustrates the different descriptions of the attitudes and reasons for cancelling the president's agenda. The generated summary for this issue extracts the salient and objective sentences from the centre and right articles, which are highlighted in green and red.