Source-linked AI summary
Blaming Across the Aisle: Political Contrasting and Blame Attribution in the Danish Parliament
Markus Lundsfryd Jensen, Rune Egeskov Trust, Kenneth Christian Enevoldsen, Sara Kolding
TL;DR
Parliamentary discourse remains comparatively underexamined despite its institutional weight, leaving robust evidence on perceived increases in hostility scarce. This study combines a blame classifier with multilevel modeling of Danish parliamentary discourse and finds blame declined before rising again, with political contrasting shaped by ideology.
Problem
Parliamentary discourse remains comparatively underexamined, despite its institutional weight and the role of blame attribution in political rhetoric.
Source: §1 · p. 1Method
The study combines a purpose-built blame classifier with multilevel statistical modeling of Danish parliamentary discourse from 1997 to 2026.
Source: §5 · p. 8Results
Blame declined before a sustained increase, while governing dampened blame less among right-wing parties and ideological extremity was more strongly associated with blame on the right.
Source: §4 · p. 7Takeaways & Limitations
The findings suggest the rise in harsh political language reflects ideologically asymmetric hardening concentrated on the right rather than a general rhetorical drift.
Source: §4 · p. 7Blame labels rely on a multi-step pipeline in which translating Danish political text into English may distort ironic, indirect, or idiomatic phrasing.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
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.
1 Introduction
The paper examines blame in comparatively underexamined parliamentary discourse and introduces Danish data and a model to study how blame varies over time and across political characteristics.
1 Introduction
- Parliamentary discourse remains comparatively underexamined despite its institutional weight, and blame attribution is a way politicians assign responsibility and criticize opponents.Parliamentary proceedings offer a view into policy positions and rhetorical strategies.Source: §1 · p. 1
1.1 Blame and Political Communication
- Opposition parties are expected to blame more than governing parties (“political contrasting”), while greater ideological distance from the center (“wingness”) is expected to increase blame.The account treats blame as combining causal attribution with negative sentiment.Source: §1.1 · pp. 1–2
1.2 Emotional Tone in Parliamentary Discourse
- Prior research suggests emotional arousal in Danish and Dutch intra-party speeches has risen while overall sentiment valence remained stable, but negative sentiment alone does not establish blame attribution.The study investigates blame’s relationship to political characteristics and contested claims of rising hostility, identifying blame through linguistic markers rather than assumptions about parties.Source: §1.2 · p. 2
1.3 Contributions and Hypotheses
- The paper annotates Danish parliamentary utterances from 1997–2026 with an annotation-efficient pipeline and presents the first model for blame detection in Danish politics.The annotations use computational assistance from DEBATE and are computationally validated against downstream performance on a manually constructed test set.Source: §1.3 · p. 2
- The analysis estimates temporal and political patterns in blame, including whether blame and the effects of extremity and opposition status have increased particularly in 2019–2026.These are the paper’s stated hypotheses, not reported findings.Source: §1.3 · p. 2
2 Methods
The methods combine a curated Danish parliamentary corpus, an annotation-efficient classifier pipeline, and multilevel models to study blame attribution over time and across political characteristics.
2.1 Dataset
- The merged parliamentary corpus spans 07/10-1997 to 26/02-2026; preprocessing reduced 6,553,133 sentences to 5,598,994, with 500,500 sampled for training, validation, and testing.The remaining 5,098,494 sentences were held out for inference.Source: §2.1 · pp. 2–3
- DEBATE labeled translated training sentences using five templates and blame, praise, and neutral labels; agreement-defined DIALs varied in blame prevalence from 1.71% to 0.50%.A manually annotated test set retained 424 sentences on which both annotators agreed, with 84.8% initial agreement (Cohen’s Kappa = .676).Source: §2.1 · pp. 3–4
2.2 Model Development
- The selected model was a full-precision LoRA fine-tune of mmBERT, trained with focal loss; hyperparameters were selected across DIALs, and validation performance was assessed by maximizing MCC.Training and validation used an 80-20 split of each DIAL subset.Source: §2.2 · p. 4
- BlameBERT, selected from DIAL-5 for its precision–recall trade-off, achieved average recall .81, average precision .80, and macro-averaged F1 .80, with no systematic party-specific error rate found.DIAL-4 achieved the same macro-F1, but DIAL-5 was selected; it used a learning rate of 1e−4 and square-root class weights.Source: §2.2 · p. 4
2.3 Analysis
- The analysis included 13 parties active in continental Denmark; positive standardized ideology scores defined the right wing and negative scores the left, and blame rates were modeled over 1997–2026.Negative-binomial mixed-effects models examined government status, ideological wing, and wingness, with party-level random intercepts and the temporal trend as a fixed-effect control.Source: §2.3 · pp. 4–5
- Sensitivity analyses tested increasingly conservative blame-probability thresholds and used sentence-level topic classifications to assess agenda differences and whether controlling for topics altered government-status effects.The models did not otherwise control for parliamentary topic or agenda.Source: §2.3 · p. 5
3 Results
Across the full period, blame followed a curved time pattern, while it increased linearly in recent years. Government status and its interactions with political wing and ideological extremity qualified these trends.
Results H1
- 0.0135 (b) was the significant positive quadratic time effect across the full period (p = .0347), while the linear time effect was negative but nonsignificant.Adding linear time improved fit over the intercept-only model, χ2(1) = 6.49, p = .0109; adding a quadratic term further improved fit, χ2(1) = 4.44, p = .0352.Source: §3 · p. 6, Figure 2 · p. 6
Results H2
- −0.478 (b) was the significant government-status effect across the full period (p < .001); positive interactions showed a weaker effect for right-wing parties and greater blame with right-wing ideological extremity.The government-status-by-right-wing interaction was 0.105 (b, p = .00236), and the right-wing-by-wingness interaction was 0.554 (b, p = .00265).Source: §3 · p. 6, Figure 3 · p. 6, Figure 4 · p. 7
Results H1.1
- 0.0921 (b) was the significant positive linear time effect on recent blame counts (p < .001), and adding a quadratic term did not improve model fit.The linear model fit better than the intercept-only model, χ2(1) = 107, p < .001; the quadratic addition yielded χ2(1) = 3.74, p > .05.Source: §3 · p. 6, Figure 2 · p. 6
Results H2.1
- 0.817 (b) was the significant positive right-wing-by-wingness interaction in recent years (p < .001), while government status had a negative effect (b = −0.515, p < .001).Adding the wing-by-wingness interaction improved model fit, χ2(3) = 15.5, p = .00143; adding a government-status-by-wing interaction did not further improve fit, χ2(1) = 0.857, p > .05.Source: Figure 3 · p. 6, §3 · p. 7, Figure 4 · p. 7
4 Discussion
Blame declined until about April 2016 and then rose, while government status and ideology structured who assigned blame. The later increase was especially pronounced among ideologically extreme right-wing parties.
Blame Attribution
- ∼April 2016 marked the global minimum in a banana-shaped 29-year trajectory, after which blame attribution increased at an accelerating rate.A significant positive linear trend was also found in recent years.Source: §4 · p. 7
Political Contrasting
- Governing left-wing parties blamed ∼62% as much as left-wing opposition parties, and government’s blame-dampening effect was ∼11% weaker among right-wing parties.The main effects of wing affiliation and wingness were not significant; significant interactions showed that ideological effects differed by political affiliation.Source: §4 · p. 7
- Each one-unit increase in wingness corresponded to a ∼74% stronger increase in blame attribution among right-wing parties than among left-wing parties; the full-period extremity effect was supported only on the right.This partially supports earlier findings linking ideological extremity to blame, but only for right-wing parties.Source: §4 · p. 7
- In recent years, governing parties blamed 60% as much as the opposition across both wings, while each unit increase in wingness brought a 126% greater increase in blame attribution among right-wing parties than left-wing parties.The recent-period main effect of wingness was not significant among left-wing parties; the wing-by-wingness interaction intensified the right-wing association.Source: §4 · pp. 7–8
- Blame became increasingly structured by ideology, intensifying most among ideologically extreme right-wing parties, which could potentially pull moderate parties toward generating more blame.The authors suggest this asymmetric hardening may have consequences beyond parliament.Source: §4 · p. 8
Sensitivity and Political Topic
- Findings held in direction and significance across classification thresholds, except the government-status-by-wing interaction was not significant at the most conservative threshold; controlling for topic did not change government’s blame-dampening effect.The test-set Cohen’s Kappa was .676, reflecting the difficulty of identifying blame.Source: §4 · p. 8
5 Conclusion
Blame in the Danish Parliament declined through 2016 before rising in a sustained, accelerating trend. Government status suppressed blame, but this effect and the recent intensification of blame varied by ideology.
5 Conclusion
- Blame declined through 2016 before entering a sustained, accelerating increase in recent years.The study examined nearly five million parliamentary sentences from 1997 to 2026; a standalone analysis corroborated the temporal drift from 2019 to 2026.Source: §5 · p. 8
- Government status suppressed blame attribution across both periods, while right-wing parties showed a weaker reduction in blame when governing and a stronger amplification as wingness increased.The authors describe this government-status effect as political contrasting.Source: §5 · p. 8
- In the recent period, the government effect applied across both wings, while the wing-by-wingness interaction intensified; core findings remained directionally and statistically robust across classification thresholds.Taken together, the findings offer partial empirical support for a perceived rise in political harshness in Denmark, while showing that the shift is nonuniform across the political landscape.Source: §5 · p. 8
Limitations
The findings are qualified by potential measurement noise, limited party-level information, Denmark’s support-party arrangements, and a smaller, disrupted recent-period sample.
Measurement limitations
- Blame labels may inherit distortions from translating Danish into English and from DEBATE, which was developed primarily for English political text.Translation artifacts may especially affect irony, indirect phrasing, and Danish idioms; BlameBERT may inherit systematic biases in DEBATE’s zero-shot labels.Source: Limitations · p. 9
Modeling limitations
- Wing affiliation and wingness vary across only 13 parties, and their shared basis in the same “lrgen” measure creates unavoidable information leakage.Their effective sample size is closer to 13 than to the full observation count, so standard errors warrant caution; mild zero-deflation is unlikely to materially affect conclusions. The measure is from the Chapel Hill Expert Survey (Rovny et al., 2025).Source: Limitations · p. 9
Support parties
- The binary government indicator treats formal support parties as opposition, although wing affiliation and its interaction with government status partially account for these arrangements.Support arrangements are typically within the same political wing, except for the cross-wing government formed in 2022.Source: Limitations · p. 9
Scope limitations
- Recent-period estimates use a smaller sample amid political disruption, while the classifier detects blame but not its target.COVID-19 and government transitions could confound observed trends; blame directed at the EU, pandemic, or global markets differs from blame directed at domestic rivals.Source: Limitations · p. 9
A Appendix · A.0.1 Data Processing Flowchart
Appendix A.0.1 points readers to a more detailed flowchart of the data-processing procedure described in Methods § 2.
A.0.1 Data Processing Flowchart
- Figure 5 provides a more in-depth flowchart of the data-processing procedure described in Methods § 2.Source: §A.0.1 · p. 13
A.1 Assessment of Dataset Compatibility
The section assesses whether merging the pre-2019 and post-2019 parliamentary speech datasets introduces systematic bias. It tests continuity in sentence volume and distributional differences at the January 2019 merge point to evaluate whether the datasets reflect a continuous data-generating process.
A.1 Assessment of Dataset Compatibility
- The assessment asks whether merging the pre-2019 and post-2019 speech datasets introduces systematic bias.Source: §A.1 · p. 13
- It examines continuity in sentence volume across the January 2019 merge point.Source: §A.1 · p. 13
- It also tests for distributional differences at the merge point to assess whether both datasets reflect a continuous data-generating process.Source: §A.1 · p. 13
A.1.1 Temporal Continuity, Distributional Comparisons and Discontinuity
Across checks of monthly counts, distributions, temporal-model residuals, and regression estimates, the study finds no evidence of a structural break at the 2019 dataset merge.
A.1.1 Temporal Continuity, Distributional Comparisons and Discontinuity
- Monthly sentence counts show no structural break at the 2019 merge, including after adjustment for active parliamentary days.The observed fluctuations do not align with the merge boundary and are consistent with natural variation in activity.Source: §A.1.1 · p. 13, Figure 6 · p. 13, Figure 7 · p. 13
- Monthly sentence-count distributions substantially overlap before and after 2019, with no systematic differences in density or boxplots.Source: §A.1.1 · p. 13, Figure 8 · p. 13, Figure 9 · p. 15
- D = 0.0847 and p = 0.810: a Kolmogorov–Smirnov test detects no difference between pre- and post-2019 residual distributions.Residual distributions came from a linear time model and remained highly similar across periods.Source: §A.1.1 · p. 15
- b = −58.2, SE = 2175, p > 0.05: the post-2019 coefficient is statistically insignificant in a regression with a linear time trend.The model provides no evidence of a discontinuity at the merge point.Source: §A.1.1 · p. 15
A.1.2 Overall Compatibility
The analyses found no evidence of a merge-related structural shift, supporting treatment of the combined dataset as a continuous time series.
A.1.2 Overall Compatibility
- The analyses found no evidence of an abrupt merge-related shift or structural artifacts, supporting treatment of the combined dataset as a continuous time series.Source: §A.1.2 · p. 15
A.2 Blame Examples … A.10 Controlling for Political Topics
Across these supplementary analyses, BlameBERT was selected for blame classification, and the government-status effect persisted after topic adjustment. Results were broadly robust to stricter classification thresholds, with one interaction losing significance at the strictest threshold.
A.2 Blame Examples
- Random examples showed lower classifier confidence for sentences labeled as containing blame than for those labeled as containing no blame.The examples were randomly sampled sentences classified by BlameBERT.Source: §A.2 · p. 15
A.3 Experiment Tracking; A.4 Performance Metrics
- DIAL-5 was selected as the final model because it shared the highest macro-averaged F1 and had a more balanced precision/recall distribution; model runs were tracked with Weights & Biases.Selected metrics for the best-performing model in each DIAL and training and validation metrics for all models were tracked there.Source: §A.4 · p. 17, §A.3 · p. 17
A.5 Comparing Against Other Models
- BlameBERT was identified as the best option for political blame classification; the generative Qwen 3.5 scored F1 .75 and precision .91 but severely underpredicted blame.Qwen 3.5 produced 68/424 (16%) null labels, and its abstain strategy had precision 1.00 and recall 0.42. Its inference took approximately 19 hours, versus a few minutes for BlameBERT after approximately 20 minutes of fine-tuning; the authors caution that the hardware differed.Source: §A.5 · pp. 17–18
A.6 Per Party Performance
- Confusion matrices showed no systematic difference in BlameBERT’s blame classification across parties represented by more than 10 test sentences.The per-party evaluation used gold-labeled test data.Source: §A.6 · p. 18, Figure 10 · p. 19
A.7 Wing and Wingness variables; A.8 Analysis Summaries
- CHES lrgen party left-right stance was z-scored across all parties; negative/positive values classify left/right, and absolute standardized values define wingness. Full H1-2.1 summaries report four LRT-best-fitting models.Source: §A.7 · p. 20, §A.8 · p. 20
A.9 Sensitivity Analysis: Classification Threshold Robustness
- Across stricter thresholds, reported effect directions were preserved and most significant effects remained significant, but the government-status-by-wing interaction lost significance at τ = 0.875.That interaction was significant at τ = 0.50, 0.625, and 0.75; at τ = 0.875 it was b = 0.076, SE = 0.079, p = .340. The authors note that increasing zero counts may reduce power at the strictest threshold.Source: §A.9 · pp. 22–23
A.10 Controlling for Political Topics
- Blame rates varied by topic—Fabric of Society exceeded Economy by more than threefold—while government and opposition discussed similar topic mixes.Adding topic significantly improved model fit, χ2(6) = 1082, p < .001; the authors conclude that agenda composition is unlikely to account for the government-status effect. The topic analysis sampled one sentence per speaker-day (84,581 observations) and grouped 56 classified topics into seven domains.Source: §A.10 · pp. 23–24
- After topic adjustment, government status remained negative (b = −0.410, SE = 0.0390, z = −10.5, p < .001), with governing-party relative blame probability at 66.3% versus 67.5% in the baseline model.The nearly unchanged estimates support the conclusion that topic composition does not explain the government-status effect.Source: §A.10 · pp. 24–25