When a party moves from government to opposition, its speeches change. Opposition members accuse and demand; governing members request and consolidate. The party's underlying commitments mostly stay put. Text classifiers that read ideology from speech cannot tell these two sources of style apart. A model trained while one party governed learns that consensual tone means that party. It is then misled when the party loses power and starts sounding combative.
This note proposes a model with two separate internal spaces, trained together. The first is a shared space that must predict ideology but is penalized for carrying information about whether the speaker currently governs. The second is a private space that is explicitly trained to predict governing-versus-opposition rhetoric, including fine-grained speech acts such as accusations, demands and promises. The test is a "role-flip" evaluation. We hold out the period after a party switches between government and opposition and ask whether the model still recognizes the party from its speeches. Time-matched splits in which nobody's role changes serve as the control.
Core Claim
Much of what looks like temporal drift in political text classifiers is really institutional-role drift. A multi-task model can remove it, given two ingredients. The first is a role adversary that is conditioned on party. The second is a private branch trained on role-specific speech acts, which gives role rhetoric somewhere to go other than the shared space.
The prediction is concrete. Take the extra accuracy drop that occurs when a test period includes a role flip, beyond the drop seen at a same-length boundary without a flip. That excess drop should shrink substantially for the proposed model. It should not shrink for single-task models, hard-sharing multi-task models, or models that are simply told the speaker's role.
Why This Gap Is Real
The motivating evidence comes from rhetorical-behavioral alignment work.
- How to Do Politics with Words built a 14-class speech-act taxonomy for two decades of Bundestag debates. It showed that parties entering government shift toward consolidating acts and parties entering opposition shift toward confrontational ones.
- Zhang et al. found a parallel shift in how parliamentary questions function in UK Hansard.
- More than Votes? fine-tuned BERT to predict Supreme Court justices' appointing party from oral arguments. Macro-F1 fell from 0.85 in-domain to 0.70 on a chronological split. That paper attributes the drop to temporal and topical drift; we hypothesize that in legislatures a large share of it is role change.
The closest multi-task prior work constrains the design.
- Policy-focused Stance Detection in Parliamentary Debate Speeches jointly trained policy-category and stance heads on ParlVote+ with hard parameter sharing. It failed to beat a single-task bag-of-words model conditioned on whether the motion came from government or opposition. BERT also lost there because of 512-token truncation of speeches averaging over 700 words. So naive sharing is a known negative result, and any encoder we use must read whole speeches.
- TWISTED showed that auxiliary speech-act and toxicity heads help stance detection on tweets. Its auxiliary labels are dictionary-derived and generic. It never separates role from stance; it just adds heads. We instead use institutional speech acts as the target for a separated private space.
The method source is Adversarial Multi-task Learning for Text Classification. Its ASP-MTL model, called the adversarial shared-private model below, keeps shared and task-private representations apart. It does this with an adversary that tries to guess the task from the shared features, plus an orthogonality penalty between shared and private features. We change what the adversary targets: it predicts governing role, not task identity.
Research Program
The encoder has to see a whole speech. It splits each speech into 510-token chunks, encodes each chunk with a pretrained transformer, and pools the chunk vectors with attention. This avoids the truncation failure documented on ParlVote+ and PTPARL-V.
From the pooled vector, two projections produce:
- a shared representation $s$, which feeds the ideology heads: party identity, plus per-motion stance where available;
- a private representation $p$, which feeds role heads: governing versus opposition, and, on Bundestag data, the 14 speech-act classes predicted at the chunk level.
A subtlety decides whether the idea works at all. Within any single legislative period, party and role are perfectly collinear for the major parties. A plain adversary that strips role from $s$ would therefore strip party too. We avoid this in two ways.
First, the adversary is given the gold party label $y$ alongside $s$. It is then rewarded only for recovering role information that $s$ holds beyond party.
Second, training data must contain role flips for some parties. Only then does "role beyond party" become learnable.
The objective is
$$ \mathcal{L} = \mathcal{L}_{\text{ideo}}(s) + \mathcal{L}_{\text{role}}(p) + \mathcal{L}_{\text{act}}(p) - \lambda \, \mathcal{L}_{\text{adv}}(r \mid s, y) + \mu \, \lVert S^\top P \rVert_F^2 . $$
The terms are:
- $\mathcal{L}_{\text{ideo}}$ is the ideology loss on the shared representation;
- $\mathcal{L}_{\text{role}}$ and $\mathcal{L}_{\text{act}}$ are the role and speech-act losses on the private representation;
- $\mathcal{L}_{\text{adv}}$ is the cross-entropy of an adversary predicting role $r$ from $s$ and party $y$, and the encoder is trained to increase it through gradient reversal;
- $S$ and $P$ are batch matrices of shared and private vectors, so the last term is the orthogonality penalty from ASP-MTL;
- $\lambda$ and $\mu$ are weights tuned on validation data.
Speech-act labels are sparse and skewed; rare classes such as Promise score poorly in the original taxonomy paper. We therefore test gradient magnitude balancing in the style of RotoGrad as a secondary variant, not a core ingredient.
The central evaluation is leave-one-flip-out. Hold out one party's period after a role change and train on everything else, including other parties' flips. Then measure party-identification macro-F1 on the held-out period.
The data are:
- Bundestag. The SPD, Greens, FDP and CDU/CSU each change role during the two decades covered, and speech-act labels exist.
- ParlVote+ (UK). The 2010 change of government and the Liberal Democrats' 2010–2015 coalition are the flips.
For each flip we build a control. It is a split of the same length at a boundary where no party changes role (for example, a UK election the same party won). This gives a difference-in-differences quantity: the flip-split drop minus the control-split drop. We call it the role-attributable drop. It is the primary metric.
Secondary measurements are:
- a linear probe for role on frozen $s$, with party given as input;
- stability of a continuous party-lean score across a flip. This follows the language-partisanship score of More than Votes?. A flip should not move a party's mean score when its voting record does not.
- ParlVote+ stance macro-F1, to confirm that separating role costs nothing on the standard task.
Baselines
Do nothing. A majority-class predictor and a single-task bag-of-words model with TF-IDF features and a small neural classifier, trained on the same splits. This establishes how much of any drop is visible without deep models.
Strongest existing method on ParlVote+. The motion-dependent single-task bag-of-words model from Policy-focused Stance Detection in Parliamentary Debate Speeches. It reports 64.1% macro-F1 on policy preference and 70.8% macro-F1 on stance. It is the strongest competitor for two reasons: it already beat both BERT and multi-task learning on this data, and it is given role information at input.
- We reimplement it on our splits with the same tuning budget, and we give it the speaker's role as an extra input.
- Beating it means a statistically significant reduction in the role-attributable drop, via bootstrap over speeches, on a majority of held-out flips.
- On the standard ParlVote+ stance split, the proposed model should come within noise of, or exceed, 70.8% stance macro-F1.
Single-task long-context encoder. The same chunked encoder trained only on the ideology head. This isolates the gain from simply reading whole speeches.
Hard-sharing multi-task model. The same encoder with ideology, role and speech-act heads on one shared vector. This replicates the design that failed before, now with a better encoder.
Key-ingredient ablations. Four variants isolate what matters:
- single-task plus a role adversary, with no private branch and no speech-act supervision;
- the full model with the adversary not given party;
- the full model without the orthogonality term;
- the full model without speech-act labels, using role labels only.
All models share the encoder, chunking, optimizer, and a fixed number of hyperparameter trials selected on a validation flip distinct from the test flip.
Frontier language model as a reference. A zero-shot party-identification prompt to a large language model. The only related figure in the file is GPT-4o's roughly 84% vote-simulation accuracy on EuroParlVote, reported by Demographics and Democracy. That is a different task and dataset. No reported result was found on Lacuna for party identification on ParlVote+ or Bundestag role-flip splits.
Why not just…?
Why not just give the classifier the speaker's role as an input feature? This is what motion-dependent modeling does, and at test time role is known. The flaw appears under a flip. A model that learns an interaction between party and role has never seen the held-out party in its new role, so it cannot know how that party sounds there. It can only fall back on whatever confrontational style it associated with other parties. The role-input baseline above runs exactly this test. If it matches the proposed model's role-attributable drop, the disentanglement is unnecessary, and we will report that.
Why not just remove role with an adversary on a single-task model? This is the most serious threat, because it is the proposal minus the multi-task parts. Our hypothesis is that, without a private branch trained on speech acts, the encoder has nowhere to put role rhetoric. The adversary then either fails to remove it or also erodes party information that correlates with it. The first ablation above settles this directly. If single-task-plus-adversary matches the full model, the honest conclusion is that multi-task learning adds nothing here, and the contribution reduces to the conditional adversary and the role-flip evaluation.
Why not just use bag-of-words, which already won on ParlVote+? Bag-of-words avoids truncation. But function words and speech-act markers such as "we demand" or "the government has failed" are exactly what shifts with role, so a bag-of-words model should be highly exposed to flips. Its accuracy on PTPARL-V vote prediction was only 59–60.75% against a 53% majority baseline. The difference-in-differences measurement will show whether its role-attributable drop is large. It could turn out small; the test is symmetric.
Why not just prompt a frontier language model? Such models bring their own heuristics to political speech. Demographics and Democracy documents a centrist skew, worse performance at ideological fringes, and a male-by-default bias on EuroParlVote. None of that suggests they separate role from ideology. Their pretraining also likely contains the historical record of which party governed when, so a zero-shot score on past flips does not cleanly test generalization. We include one only as a reference and flag this contamination risk.
Decisive Checks
The idea succeeds if three things hold together:
- the role-attributable drop falls clearly below that of every baseline on most held-out flips;
- the role probe on the shared representation, given party, is near chance;
- standard ParlVote+ stance performance is preserved.
It fails if one of these holds instead:
- the control splits show drops as large as the flip splits, meaning role was never the cause and More than Votes?-style drift is purely temporal;
- the single-task adversary ablation matches the full model, in which case the multi-task claim is withdrawn.
Both outcomes are informative for anyone who reads ideology from legislative speech.
References
- How to Do Politics with Words: Investigating Speech Acts in Parliamentary Debates
- Asking too much? The rhetorical role of questions in political discourse
- More than Votes? Voting and Language based Partisanship in the US Supreme Court
- Policy-focused Stance Detection in Parliamentary Debate Speeches
- Toxicity, Morality, and Speech Act Guided Stance Detection
- Adversarial Multi-task Learning for Text Classification
- PTPARL-V: Portuguese Parliamentary Debates for Voting Behaviour Study
- RotoGrad: Gradient Homogenization in Multi-Task Learning
- Demographics and Democracy: Benchmarking LLMs' Gender Bias and Political Leaning in European Parliament