1. Introduction

Potential bias expressions in journalistic language emerge during the news production process—including the selection of facts, vocabulary, and tone—and can influence how the news is perceived (Park et al. 2009; Stroud 2011). In the initial stage of this research, combining computational linguistics (CL) and critical discourse analysis (CDA) has qualitatively demonstrated that specific linguistic features, drawn from subjectivity bias lexicons (Pryzant et al. 2020), reflect the distinct editorial policies and ideologies of major US and UK media outlets. While it provides insights into what constitutes media bias and how it manifests linguistically, it also raises a further question: to what extent do these features possess discriminatory power? If the identified subjective expressions are a robust signal of media origin, they should enable systematic differentiation among news sources. Building on this premise, this study tests whether subjectivity-bias markers in headlines can distinguish major US and UK newspapers, translating prior qualitative claims into a quantitative research question.

2. Data

The dataset consists of English news articles sourced from ProQuest Central (July 1, 2019–June 30, 2024), retrieved using the keyword “diplomacy,” with only headline text analyzed. To balance political perspectives, the dataset includes three US outlets (The Washington Post, The Wall Street Journal, The New York Times) and three UK outlets (The Daily Telegraph, The Guardian, Express), with headlines down-sampled to equalize counts across publications.

Figure 1. Data sources and counts by publication

3. Methodology

The analytical approach frames the research question as a classification task to determine if a machine learning model can distinguish between US and UK headlines based on linguistic characteristics. For feature extraction, lexicon-based features were computed for each headline using the subjectivity bias lexicons from Pryzant et al. (2020), introduced in Section 1. Each feature equals the proportion of tokens in the headline that appear in a given lexicon. Two models were employed to capture both linear and non-linear patterns: a Logistic Regression (LR) and an Explainable Boosting Machine (EBM) (Nori et al. 2019). Following an evaluation of model performance using the Area Under the Curve (AUC) metric, an error analysis was conducted to identify the linguistic patterns among different sources. Finally, Local Interpretable Model-agnostic Explanations (LIME) (Ribeiro et al. 2016) were utilized to provide granular, word-level interpretations of the EBM’s predictions, revealing how individual terms contribute to a headline’s predicted country of origin.

4. Result and discussion

Model comparison shows a non-linear approach is needed: LR scored an AUC of 0.601, while EBM reached 0.805, suggesting subjectivity-bias signals are complex and driven by interactions beyond simple word frequencies. While the AUC score confirms the EBM model is effective, a granular analysis of its confusion matrix (Figure 2) reveals how it works by uncovering distinct, country-specific “linguistic fingerprints” through its classification errors. Misclassifications primarily occurred within national borders—such as confusing The Guardian with other British outlets or The New York Times with The Washington Post. This indicates the existence of unique “British” and “American” styles of subjective expression, validating the bias lexicon’s ability to capture nationally-rooted journalistic styles. A notable exception, confusing the UK’s Daily Telegraph with the US’s Wall Street Journal, suggests a potential transnational “conservative business media” style, reinforcing that biased language is tied to national journalistic cultures.

Figure 2. Confusion matrix for EBM model

Figure 3 ranks global feature importance for the EBM model, identifying ratio_entailing, ratio_strong_subjectives, and ratio_negative_sentiment as the most influential predictors. The results show that the model prioritizes linguistic cues of logical commitment and strong subjective sentiment, while simple declarative statements like ratio_factives have minimal impact.

Figure 3. Global feature importance for the EBM Model

Leveraging LIME for local interpretations (Figure 4) reveals that, although the model confuses the Express with The Guardian, it correctly identifies the headline’s British origin. LIME analysis shows the emotive word “vow” drives the UK classification. Conversely, the direct conflict words “death,” “fight,” and “Battle” serve as the strongest evidence against an attribution to The New York Times. This demonstrates the model’s ability to separate national fingerprints of subjectivity, linking solemn declarations to a British style while finding visceral language counter-indicative of the American style in this context.

Figure 4. Local attribution for headline source (LIME)

5. Conclusion

This study quantitatively confirmed that subjectivity bias markers effectively differentiate US and UK news headlines. Its significance for digital humanities lies in offering a scalable and interpretable machine learning framework that reveals subtle ideological patterns imperceptible to close reading by making them quantitatively explicit. Future work will test and extend its generalizability on full news articles.