Blaming Across the Aisle: Political Contrasting and Blame Attribution in the Danish Parliament

AI summary

Overview

• Research area: Computational social science / natural language processing applied to legislative speech, specifically blame attribution in the Danish Parliament (Folketinget) from 1997 to 2026.
• Technical level: Intermediate — it combines transformer fine-tuning (LoRA, focal loss), zero-shot NLI annotation, and negative binomial mixed-effects regression, but presents each step in accessible terms.
• Scope (1 sentence): The paper builds the first Danish political blame classifier, BlameBERT (macro-averaged F1 of .80), and uses it to model how blame attribution in roughly 4.9 million parliamentary sentences varies with time, government status, political wing, and ideological extremity.

What This Paper Is About

Media and society widely perceive political communication as growing harsher, but the empirical evidence is disputed and comes mostly from social media rather than parliaments, which carry greater institutional weight. This paper asks whether blame attribution — an evaluative act combining causal attribution with negative sentiment — has actually increased in the Danish Parliament, and how it is structured by whether a party governs or sits in opposition, which political wing it belongs to, and how ideologically extreme it is. The authors build a purpose-made Danish blame classifier and then apply multilevel statistical models to separate genuine trends from party-level structural differences.

Key Contributions

1. An annotation-efficient pipeline and dataset. The authors annotate direct utterances from Danish parliamentarians (1997–2026) using a pipeline computationally assisted by the NLI model DEBATE and computationally validated against a manually constructed gold test set.
2. The first blame-detection model for a Danish political context. BlameBERT was trained on the Danish parliamentary data and compared against two zero-shot Qwen baselines.
3. Quantified temporal and political structure of blame. Using negative binomial mixed-effects models over nearly five million sentences, the authors estimate how blame varies over time, with government status, political wing, and ideological extremity, and contrast the full period with recent years (2019–2026).
4. Public release of artifacts. Code is available on GitHub, and the model and dataset are available on Hugging Face.

Main Findings

• Banana-shaped temporal trajectory: Across 1997–2026, blame declined initially, reached a global minimum around April 2016, and then increased at an accelerating rate. Adding time as a linear predictor significantly improved fit over an intercept-only model (χ²(1) = 6.49, p = .0109), and adding a quadratic term improved fit further (χ²(1) = 4.44, p = .0352). The positive quadratic effect was significant (b = 0.0135, SE = 0.00640, z = 2.11, p = .0347), while the linear term was negative but not significant (b = −0.00980, SE = 0.00600, z = −1.63, p > .05).
• Sustained recent increase (2019–2026): For the recent period, the linear model fit significantly better than an intercept-only model (χ²(1) = 107, p < .001), and a quadratic term added nothing (χ²(1) = 3.74, p > .05). The linear time effect was significant and positive (b = 0.0921, SE = 0.00861, z = 10.7, p < .001).
• Political contrasting — government status dampens blame: Including government status significantly improved model fit over the full period (χ²(1) = 668, p < .001) and in recent years (χ²(1) = 119, p < .001). Over the full period, governing left-wing parties blamed about 62% as much as left-wing opposition parties; in recent years, governing parties blamed 60% relative to the opposition.
• Ideology moderates the governing effect: The interaction between government status and right wing was positive and significant over the full period (b = 0.105, SE = 0.0346, z = 3.04, p = .00236), meaning the blame-dampening effect of governing was about 11% weaker among right-wing parties. In the recent period, this interaction did not improve fit (χ²(1) = 0.857, p > .05), so the government effect applied across both wings.
• Ideological extremity amplifies blame asymmetrically: Over the full period, the right wing × wingness interaction was significant (b = 0.554, SE = 0.184, z = 3.01, p = .00265), with each one-unit increase in wingness associated with about a 74% increase in blame attribution for right-wing relative to left-wing parties. The main effect of wingness (which reflects left-wing parties) was not significant.
• The asymmetry intensified recently: In 2019–2026, the wing × wingness interaction was significant (b = 0.817, SE = 0.162, z = 5.05, p < .001), with each unit increase in wingness associated with a 126% increase in blame attribution among right-wing compared to left-wing parties. The main effect of right wing was negative and significant in this period (b = −0.522, SE = 0.151, z = −3.45, p < .001), and the wing × wingness interaction improved fit compared to the government-only model (χ²(3) = 15.5, p = .00143).
• Model performance and comparison: BlameBERT obtained an average recall of .81, average precision of .80, and a macro-averaged F1 of .80. On class-1 (blame) metrics, the Qwen embedding baseline scored 0.52 precision, 0.87 recall, 0.65 F1, and 0.67 macro F1; the generative Qwen 3.5:9B baseline scored 1.00 precision, 0.42 recall, 0.60 F1, and 0.75 macro F1. No systematic error rate was found across parties.
• Annotation quality: Inter-annotator agreement on the 500 upsampled sentences was 84.8% (Cohen's Kappa = .676); only sentences where both annotators agreed were kept, giving a test set of 424 sentences (148 blame, 34.9%).
• Sensitivity analysis: The direction and significance of findings held at all classification thresholds, with one exception — the government status × ideological wing interaction in the full-period analysis remained the same direction but was not statistically significant at the most conservative threshold.
• Agenda does not explain the effect: Governing and opposition parties did not notably differ in their topical agendas, and controlling for topic did not alter the blame-dampening effect of government status.

Methodology in Plain English

The authors combined two parliamentary speech sources: ParlSpeechV2, covering 07/10-1997 to 20/12-2018, and newer transcripts fetched from the Danish Parliament's SFTP server (earliest transcribed debate 06/10-2009, most recent 26/02-2026). They merged the two from 20/12-2018 with the next available date, 09/01-2019, covering 07/10-1997 to 26/02-2026, and excluded paragraphs spoken by a chairman.

Sentence segmentation used DaCy ("da_dacy_large_trf", version 0.2.0). Sentences shorter than five characters or containing parentheses were dropped, reducing the count from 6,553,133 to 5,598,994 (85%). From the cleaned data they randomly sampled 500,500 sentences for training, validation, and testing, holding out the remaining 5,098,494 for inference. Because DEBATE is a fine-tuned DeBERTa-V3 zero-shot classifier that is not multilingual and accepts only 512 tokens, only the training data was machine translated with Opus-MT-da-en and then labeled by DEBATE through five hypothesis templates, with candidate labels "blame", "praise", and "neutral". A sentence counted as blame when the blame probability was ≥ .80 and greater than both other labels; labels were mapped back onto the original Danish sentences.

From the five templates they built "Datasets of Increasing Agreement Levels" (DIALs). DIAL-n required at least n templates to agree, from DIAL-5 (most conservative) to DIAL-1 (least). Blame prevalence was 1.71%, 1.18%, 0.91%, 0.70%, and 0.50% for DIAL-1 through DIAL-5. For the gold test set they upsampled 250 blame and 250 non-blame sentences from DIAL-1; two of the authors (males, Danish, age 24-25) annotated them following Bilotta et al.'s definition of blame, and only agreed sentences were kept.

Model training used a full-precision LoRA fine-tune (rank 64, alpha scaling 128) with focal loss, a grid search over three learning rates (1e-5, 1e-4, 5e-4) with linear decay, and three alpha scaling constants (raw class weights, class weights to the power of two-thirds, square root of class weights), with the gamma focusing parameter fixed at 2.0. mmBERT was trained and validated on an 80-20 split of each DIAL subset, with model selection by maximizing the Matthews Correlation Coefficient on validation. DIAL-5 and DIAL-4 tied on macro-F1, but the DIAL-5 model (learning rate 1e-4, alpha = square root of class weights) was chosen for its better precision-recall trade-off and named BlameBERT.

For the analysis, only parties still existing and active in continental Denmark were kept, and non-attached members were excluded, leaving 4,938,119 unique sentences (96.9%) across 13 parties. Wing was assigned from Chapel Hill Expert Survey ideological placements (positive standardized scores = right wing), and wingness was the absolute standardized distance from the mean. BlameBERT classified all remaining sentences, aggregated by month, year, and party, giving 2,529 observations overall and 810 for 2019–2026. Hypotheses were tested with negative binomial mixed-effects models (glmmTMB) with party-level random intercepts and a sentence-count offset, compared via likelihood ratio tests. A sensitivity analysis applied increasingly conservative classification thresholds, and a supplementary analysis used ManifestoBERTa to check whether topic agendas differed between governing and opposition parties.

Why This Matters

The paper argues that the perceived rise in harsh political language is not a uniform rhetorical drift but an ideologically asymmetric hardening concentrated on the right of the political spectrum. Because parliamentary debate sits at the core of democratic institutions and helps set the stage for how political conflict is handled throughout society, this asymmetry may carry consequences beyond the chamber.

• Research impact: It provides a rare parliamentary (rather than social media) test of the "hostile politics" claim, an annotation-efficient pipeline for a low-to-mid resource language, and a released model and dataset others can reuse or challenge.
• Real-world applications:
  • Parliamentary monitoring and transparency tools that track tone and blame over time.
  • Media and political-communication analysis, giving journalists a measurable baseline instead of impressionistic claims about "sharper tones".
  • Comparative political science, by supplying a replicable template for measuring blame in other parliaments.
  • Danish-language NLP, by contributing a political-domain dataset and classifier for an under-resourced language.
• Industry relevance: Comparable classifiers and pipelines are directly applicable to media monitoring, political risk analysis, and content-moderation or discourse-quality systems operating on political text in smaller languages.

Future Directions

• Target identification: The classifier detects that blame occurs but not whom it targets. High blame rates could reflect blame directed at external or non-partisan targets (the EU, the pandemic, global markets) rather than domestic rivals — a conceptually different phenomenon from rival-directed political contrasting that the framework emphasizes.
• Cross-national and cross-lingual extension: The pipeline was designed for low-to-mid resource languages, so applying it to other parliaments and testing whether the ideologically asymmetric pattern generalizes is a natural next step.
• Reducing translation and label noise: Training labels passed through machine translation before DEBATE labeling, and the authors note that translation artifacts may be especially pronounced for irony, indirect phrasing, and idioms; they also note BlameBERT can inherit DEBATE's systematic biases rather than correct them.
• Better modeling of parliamentary structure and confounds: The binary government indicator treats support parties as opposition despite Denmark's minority-government tradition, the recent period (2019–2026) spans the COVID-19 pandemic and multiple government transitions, and wing and wingness are both derived from the same Chapel Hill "lrgen" measure, creating unavoidable information leakage.

Target Audience

This paper benefits computational social scientists and political communication researchers studying legislative discourse, NLP practitioners working on Danish or other low-to-mid resource languages, and political scientists interested in polarization, opposition-government dynamics, and the empirical measurement of political hostility. It is also relevant to analysts and journalists who want quantitative grounding for claims about rising political harshness.

Authors’ abstract

Political discourse is widely perceived to be growing more hostile, yet robust evidence remains scarce. This study examines blame attribution in the Danish Parliament from 1997 to 2026, combining a purpose-built classifier, BlameBERT (F1: 0.80), with multilevel statistical modeling. The classifier is constructed using an annotation-efficient pipeline for blame attribution in low-to-mid resource languages. The results reveal a banana-shaped trajectory, with blame declining until around 2016 before entering a significant and sustained increase in recent years (2019-2026). Government status consistently influenced blame attribution - an effect we term political contrasting - with opposition parties blaming substantially more than governing parties. This effect was moderated by ideology: The blame-dampening effect of governing was less pronounced among right-wing parties, and ideological extremity amplified blame more strongly on the right. In recent years, the interaction between political wing and ideological extremity intensified, suggesting an ideological hardening of the blame rhetoric concentrated on the right of the political spectrum. Taken together, these patterns suggest that the perceived rise in harsh political language reflects not merely a general rhetorical drift, but an ideologically asymmetric hardening of political discourse. A sensitivity analysis showed that the conclusions were robust to varying classification thresholds.