Nearly five million Danish parliament sentences show blame falling until 2016 then climbing, with the sharpest rise among extreme right-wing parties

The study builds BlameBERT, a blame-detection model for the Danish political context, reaching a macro-averaged F1 of 0.80, average recall of 0.81, and average precision of 0.80 on a manually annotated test set. The authors describe it as the first model for blame detection in a Danish political context, released together with code, model, and dataset; training labels were generated by machine translation plus the zero-shot NLI model DEBATE, forming Datasets of Increasing Agreement Levels (DIAL-1 to DIAL-5), with the most conservative DIAL-5 chosen. The test set contains 424 manually annotated sentences (148 blame, 34.9%), with 84.8% inter-annotator agreement (Cohen's Kappa 0.676) and only agreed sentences retained; BlameBERT's macro-F1 exceeds a Qwen 3 embedding baseline (0.67) and a Qwen 3.5 generative baseline (0.75), and per-party inspection found no systematic error rate.

Across 1997 to 2026, parliamentary blame follows a banana-shaped trajectory: declining to a global minimum in April 2016, then rising at an accelerating rate. This supplies parliamentary-scene quantitative evidence for the contested claim that political discourse is growing more hostile, where much prior empirical work draws on social media; the authors note the timing coincides with Denmark's 2015 immigration crisis and reports of hardening rhetoric afterward. Likelihood ratio tests on negative binomial mixed-effects models show that adding linear and then quadratic time terms significantly improves fit, with a significant positive quadratic coefficient; a separate model for 2019 to 2026 shows a significant positive linear trend.

Government status consistently suppresses blame, the effect the authors call political contrasting: governing left-wing parties blamed only about half as much as left-wing opposition parties. The effect holds both over the full period and in the recent subsample, and remains nearly unchanged when topic is controlled, suggesting it reflects rhetorical framing rather than agenda composition. The government-status coefficient is significantly negative in the full-period model and again in the recent-period model. A supplementary analysis classified topics with ManifestoBERTa on 84,581 observations; governing and opposition parties had similar topic distributions, and controlling for topic left the government-status coefficient nearly identical.

Ideology moderates political contrasting, and this moderation intensified in recent years: the blame-dampening effect of governing is weaker among right-wing parties, and blame rises with ideological extremity more strongly on the right. The authors report partial support for prior findings that more ideologically extreme parties blame more, but only for right-wing parties; in recent years each unit increase in wingness produced a larger blame increase among right-wing than left-wing parties, suggesting an ideologically asymmetric hardening concentrated on the right. In the full-period model the government-status-by-right-wing interaction and the right-wing-by-wingness interaction are both significant; in the recent-period model the right-wing-by-wingness interaction is significant and stronger than over the full period. Sensitivity analysis preserved the direction of every effect across thresholds, with the sole exception that the government-status-by-wing interaction in the full-period analysis lost significance at the strictest threshold.

The study applies to Danish parliamentary speech from 1997 to 2026 and to the 13 parties active in continental Denmark at the time of analysis; its classifier and annotation pipeline target parliamentary text in low-to-mid resource languages, and the framework can be borrowed for other parliamentary corpora. For a reader, the value lies in moving the debate over hostile political discourse from social media into an institutional setting and providing reusable model and data resources, so later work can compare blame dynamics across countries, periods, or institutional settings on a shared measurement basis.

The classifier identifies whether blame occurs but not whom it targets, so a party with a high blame rate might mainly blame non-partisan targets such as the EU, the pandemic, or global markets, which is a different phenomenon from the rival-directed political contrasting the framework emphasizes. Training labels passed through Danish-to-English machine translation before DEBATE, which was developed and validated mainly on English political text, so irony, indirect phrasing, and language-specific idioms may be affected and BlameBERT may inherit systematic biases. Wing and wingness are time-invariant characteristics varying across only 13 parties, so the effective sample size is closer to 13 than to the full observation count, and both derive from the same expert-survey measure, leaving some information overlap. The recent subsample is smaller and spans the COVID-19 pandemic and multiple government transitions. In addition, some numeric values in the loaded text (such as blame prevalence for each DIAL and some coefficients and standard errors) are missing from the tables, so exact effect sizes would require consulting the original appendices.