---
title: 'EuroParlVote: Voting Networks & Political NLP'
url: https://www.emergentmind.com/topics/europarlvote
type: topic
---

# EuroParlVote: Voting Networks & Political NLP

EuroParlVote denotes a set of research artifacts centered on voting in the European Parliament, but the term is not monosemous in the literature represented here. In one line of work, it refers to a densely weighted signed-network dataset derived from roll-call votes in the 7th European Parliament, designed to study the informational value of negative links for graph partitioning [1507.04215]. In a later line of work, it refers to a benchmark for evaluating large language models on politically sensitive tasks such as gender classification and vote prediction from parliamentary debate speeches and demographic metadata, accompanied by the ParlAI Vote demonstration platform [2509.06164; 2509.16264]. Related 2024 studies also use the European Parliament electoral setting to examine LLMs as voting assistants and as generators of synthetic vote predictions, extending the EuroParlVote label into broader political-NLP and computational-social-science contexts [2407.08495; 2409.09045].

## 1. Terminological scope and research lineage

The earliest EuroParlVote usage in the material considered here is a network-science construction based on raw roll-call records made publicly available by VoteWatch Europe for the entirety of the 7th European Parliament, from June 2009 to June 2014. Its central object is a complete signed graph on Members of the European Parliament (MEPs), where an edge encodes average voting agreement or disagreement over a selected time window and policy area [1507.04215].

A later usage reframes EuroParlVote as an NLP benchmark linking debate speeches to roll-call outcomes and demographic metadata such as gender, age, country, and political group. This version is introduced explicitly as “a novel benchmark for evaluating large language models (LLMs) in politically sensitive contexts,” with downstream tasks including gender classification and vote prediction [2509.06164]. The associated ParlAI Vote system presents the same benchmark as a basis for interactive exploration, bias auditing, and model comparison over debates and votes [2509.16264].

Beyond those two core meanings, the European Parliament electoral setting is used in 2024 for two adjacent purposes. One study investigates Mistral and Mixtral as Voting Advice Applications on the “EU and I” questionnaire and discusses recommendations “for Production EuroParlVote” [2407.08495]. Another formulates a public-opinion prediction task labeled “EuroParlVote” over approximately 26,000 Eurobarometer voter profiles, asking GPT-4-Turbo to predict turnout and party choice in the 2024 European Parliament elections [2409.09045]. A plausible implication is that EuroParlVote has developed into a family of benchmarks and application settings rather than a single fixed dataset.

## 2. The 2015 signed-network dataset

The 2015 EuroParlVote construction starts from roll-call records in which each of roughly 736 MEPs casts one of three votes—For, Against, or Abstain—or is absent, with absences recorded as Absent, Did not Vote, or Documented Absence. Dataset extraction proceeds in two steps: first, filtering by policy area and time period to select documents \(d_1,\dots,d_\ell\); second, computing pairwise “document-wise agreement” scores \(m_{uv}(d_i)\in[-1,+1]\) for each unordered pair of distinct MEPs \((u,v)\) and averaging them across documents [1507.04215].

In the preferred abstention variant, agreement is defined as follows:
\[
m_{uv}(d_i)=
\begin{cases}
+1, & \text{if }u,v\in\{\text{For, Against}\}\text{ and agree}\\
-1, & \text{if }u,v\in\{\text{For, Against}\}\text{ and disagree}\\
0, & \text{if exactly one of }(u(d_i),v(d_i))\text{ is Abstain}\\
+1, & \text{if }u(d_i)=v(d_i)=\text{Abstain}\\
0, & \text{if at least one is absent.}
\end{cases}
\]
The average similarity is then
\[
m_{uv}=\frac{1}{\ell}\sum_{i=1}^{\ell}m_{uv}(d_i).
\]

From these similarities, the dataset builds a complete graph \(G=(V,E,s,w)\), where each edge \(e=\{u,v\}\) has sign
\[
s(e)=\mathrm{sign}\bigl(m_{uv}\bigr)
\]
and weight
\[
w(e)=|m_{uv}|\in[0,1].
\]
Positive links therefore encode overall agreement, negative links systematic disagreement, and magnitudes the strength of agreement or disagreement. No thresholding or sparsification is applied. The resulting network has \(|V|\approx736\), \(|E|=\binom{|V|}{2}\approx270{,}000\), and density approximately \(1.0\) [1507.04215].

Several empirical regularities are reported. Across policies and time windows, the distribution of \(m_{uv}\) is “consistently bimodal on the positive side,” with peaks near \(0\) and \(0.6\)–\(0.7\), while the negative side is relatively flat with minimum values around \(-0.5\). Approximately 20–30 percent of edges are negative, and the mean absolute agreement \(E[|m_{uv}|]\) typically lies around \(0.4\). Since the graph is complete, every node has identical degree in the unweighted sense, but weighted strength
\[
\mathrm{str}(u)=\sum_{v\neq u}w(\{u,v\})
\]
varies from about 200 to 450 depending on how centrally an MEP votes relative to the plenary [1507.04215].

## 3. Structural balance, correlation clustering, and the role of negative links

The analytical core of the 2015 study is a comparison between two families of partitioning methods. The first family treats the signed graph through structural balance, equivalently the Correlation Clustering (CC) problem, and seeks a partition \(P=\{V_1,\dots,V_k\}\) minimizing the total weight of imbalanced edges. Writing
\[
E^+=\{e\in E: s(e)=+\},\qquad E^-=\{e\in E: s(e)=-\},
\]
the imbalance of \(P\) is defined by
\[
I(P)=\sum_{i=1}^k\sum_{e\in E^-\cap E[V_i]}w(e)
+\sum_{1\le i<j\le k}\sum_{e\in E^+\cap E[V_i:V_j]}w(e).
\]
The study solves this objective using Parallel Iterative Local Search (Parallel ILS), initialized from a greedy seed partition and configured with a reduced perturbation level of 15 and subsampling of 70 percent of possible moves per iteration [1507.04215].

The second family consists of four community-detection algorithms that do not natively handle negative links: InfoMap, EdgeBetweenness, FastGreedy, and WalkTrap, all as implemented in the igraph library. To apply them, the signed graph is projected into two unsigned variants: the positive subgraph
\[
G^+=(V,E^+)
\]
and the complement of the negative-edge graph,
\[
\overline{G^-}=(V,\mathcal{P}_2(V)\setminus E^-).
\]
These projections discard part of the original signed information by construction [1507.04215].

The comparative results are the main substantive finding. On the positive-only graph \(G^+\), InfoMap yields partitions with \(I(P)\) typically below 5 percent of \(|E|\), but with \(k\) ranging from 4 to 14. EdgeBetweenness, FastGreedy, and WalkTrap often produce imbalances in the 20–60 percent range and fragment the network into dozens of small communities. On \(\overline{G^-}\), InfoMap degenerates to a single community, while the other methods return moderate imbalance but few clusters because the graph is nearly complete. By contrast, Parallel ILS on the full signed graph attains \(I(P)\approx0\%\), frequently under 0.5 percent of \(|E|\), with \(k\in\{2,3\}\), corresponding to the major left–right split plus occasional centrist or swing groups [1507.04215].

A further comparison uses Normalized Mutual Information between the InfoMap partition and the Parallel ILS partition, reported at approximately 0.01–0.04. The study interprets this as confirmation that even the best-performing unsigned community-detection method returns a clustering structurally quite different from the balance-optimal signed partition. The broader conclusion is that, in legislative voting networks, negative links are “indispensable” for recovering polarized structure, and that the relevance of negative ties for graph partitioning remains an open question requiring further domain-specific study [1507.04215].

## 4. The 2025 benchmark linking debates, speeches, and votes

In 2025, EuroParlVote is introduced as a benchmark for LLM evaluation on parliamentary data. The benchmark links debate speeches to roll-call vote outcomes and attaches demographic metadata for each MEP. The description in “Benchmarking Gender and Political Bias in Large Language Models” reports 21,996 debate utterances, 956 distinct motions, and 1,232 unique individuals who were present and voting; metadata include gender, age, country, and political group, with age ranging from 25 to 83 years and mean \(52.3\) with \(\sigma=8.7\) [2509.06164]. The later ParlAI Vote overview describes the benchmark as containing 969 roll-call votes, approximately 24,200 speeches, and approximately 683,000 MEP–vote instances, with four key demographic attributes—gender, age, country of election, and political group—and with debates split chronologically into 60 percent training, 20 percent validation, and 20 percent test [2509.16264].

| Aspect | [2509.06164] | [2509.16264] |
|---|---|---|
| Speeches / utterances | 21,996 | \(\approx 24{,}200\) |
| Votes / motions / debates | 956 distinct motions | 969 debates / roll-call votes |
| Individuals | 1,232 unique individuals | 705 sitting MEPs, \(\approx 683{,}000\) MEP–vote instances |

The 2025 benchmark construction emphasizes alignment between debate records and vote records. In one description, roll-call votes are aligned from HowTheyVote.eu to EuroParl debates via document IDs, retaining only speeches by MEPs who cast “For” or “Against” votes, excluding “Abstain” votes and instances with missing text or topic, preserving the original 24 EU languages, and applying only minimal normalization such as unchanged casing and punctuation [2509.06164]. The ParlAI Vote overview, by contrast, treats vote prediction as a three-way problem over yes, no, and abstain, and reports a corpus-wide vote distribution of about 75 percent “yes,” 20 percent “no,” and 5 percent “abstain,” together with a gender distribution of roughly 60 percent male and 40 percent female [2509.16264]. This suggests that the platform overview presents either an extended benchmark definition or a broader system-level view than the binary benchmark formulation.

The metadata schema is central in both descriptions. Country is represented either as one of 28 states, including the UK, or as country of election over the 27 EU member states, while political group is represented with categories such as GUE/NGL, GREEN_EFA, SD, RENEW, EPP, ECR, ID, and NI, organized from far-left to far-right in one benchmark description [2509.06164]. These attributes are not merely descriptive; they are explicitly available as model inputs, fairness dimensions, and interactive filters in the accompanying demonstration system [2509.16264].

## 5. Tasks, metrics, and observed bias patterns

EuroParlVote’s two principal tasks are gender classification and vote prediction. In the benchmark formulation, gender classification takes a single MEP debate speech \(x\) and predicts \(y\in\{\text{Male},\text{Female}\}\), optionally returning a confidence score \(s\in\{1,\dots,5\}\). Vote prediction takes debate topic \(t\), speech \(x\), and optional hints such as gender \(g\) or political group \(p\), and predicts \(v\in\{\text{For},\text{Against}\}\) [2509.06164]. In the ParlAI Vote overview, gender classification is given as a supervised mapping over speech text and optional metadata, while vote prediction is formulated as a three-way classification over structured topic representation \(T\), speech text \(s\), and demographic profile vector \(d\), with target \(y_{\text{vote}}\in\{\text{yes},\text{no},\text{abstain}\}\) and categorical cross-entropy loss [2509.16264].

The metric suite includes standard classification measures and fairness disparities. The benchmark reports accuracy, precision, recall, F1, and AUC-ROC for calibration, together with Demographic Parity Difference
\[
\Delta_{\rm DP}=P(\hat Y=1\mid G=A)-P(\hat Y=1\mid G=B)
\]
and Equality of Opportunity Difference
\[
\Delta_{\rm EO}=P(\hat Y=1\mid Y=1,G=A)-P(\hat Y=1\mid Y=1,G=B),
\]
computed across gender and political groups [2509.06164]. The ParlAI Vote overview adds a group-wise error disparity
\[
\Delta_{\text{error}}=|err(\text{female})-err(\text{male})|
\]
and an Equalized Odds Gap defined as \(\max\{|TPR_A-TPR_B|,|FPR_A-FPR_B|\}\) [2509.16264].

The reported empirical results emphasize systematic gender and ideological bias. For gender classification on the test set, the benchmark reports the following excerpted accuracies: LLaMA-3.2 at 60.01 percent, Mistral-large at 63.60 percent, Claude-3.5 at 64.25 percent, Gemini-2.5 at 65.23 percent, and GPT-4o at 61.02 percent. The female F1 versus male F1 gap is substantial for several models; for LLaMA-3.2 it is 33.52 percentage points, and the misclassification rate includes 71.13 percent of Female predicted as Male versus 18.38 percent of Male predicted as Female [2509.06164]. The ParlAI Vote overview reports a different task realization in which GPT-4o reaches 95.4 percent overall accuracy on gender classification and 77.3 percent on vote prediction, while LLaMA-3.2 reaches 85.1 percent and 64.9 percent respectively, with error disparities between female and male groups [2509.16264].

Vote prediction results also reveal sensitivity to demographic framing. In the benchmark’s “Without Gender” setting, LLaMA-3.2 reaches 67.10 percent accuracy; under “All-Female” this falls to 63.33 percent, while “All-Male” yields 67.64 percent. Proprietary models such as GPT-4o vary by less than 1 percentage point across gender manipulations, which the study interprets as robustness relative to open models [2509.06164]. Across political groups, centrist groups are the easiest: for GPT-4o, accuracy reaches 88.49 percent on RENEW and 88.09 percent on SD, while far-left groups are lower, such as 75.19 percent on GUE/NGL. The ParlAI Vote overview similarly reports a centrist bias, with error around 22 percent for centrist groups, around 32 percent for the extreme-left, and around 25 percent for extreme-right groups; including explicit political-group identifiers in the demographic vector reduces \(\Delta_{\text{error}}\) across ideology by 20 percent [2509.16264].

Methodologically, the benchmark’s discussion attributes these biases to training-data skew, model priors that formal political rhetoric is male-dominated, and prompt framing effects when demographic context is absent. Proposed mitigations include demographic hints, counterfactual gender swapping, constrained decoding, post-hoc calibration, and future incorporation of discourse structure or fairness-aware losses [2509.06164]. The system-oriented paper extends this agenda toward multilingual transfer, graph-based modeling of MEP interaction networks, additional covariates such as tenure and committee assignments, and debiasing that preserves accuracy while reducing demographic and ideological disparities [2509.16264].

## 6. Related EuroParlVote applications in voting assistance and electoral prediction

A separate, application-oriented strand studies LLMs in European Parliament election settings rather than parliamentary debate-to-vote modeling. One paper audits Mistral-7B and Mixtral-8×7B as Voting Advice Applications using the “EU and I” 2024 questionnaire of 30 political statements, party answers, and “gold-standard” justifications. The evaluation binarizes the five-point Likert scale into agreement versus disagreement and reports zero-shot baseline accuracy of 76 percent for Mistral and 82 percent for Mixtral. Web-based RAG improves these to 84 percent and 85 percent, self-reflection to 77 percent and 84 percent, and expert-augmented context to 91 percent for both models [2407.08495].

The same study reports substantial heterogeneity across political groups, with performance disparity from 50 to 95 percent in the abstract and an approximate range from 80 percent to 85 percent across countries or euro-party groups in Setting 0. It also notes that custom-curated manifesto RAG improves Mixtral by approximately 1.5 percentage points over web RAG, especially for underperforming euro-groups such as EPP and ECR, while self-reflection gives only modest gains without gold context [2407.08495]. For “Production EuroParlVote,” the paper recommends hybrid RAG over a vetted manifestos corpus supplemented by timely newswire snippets, source display for transparency, drill-down into party summaries and extracted opinions, toggles between concise and guided modes, auditing per-party and per-group misrepresentation, exposure of confidence scores, and fact-checking or expert review on contentious statements [2407.08495].

Another study investigates GPT-4-Turbo as a generator of “synthetic samples” for predicting the 2024 European Parliament elections from approximately 26,000 anonymized voter profiles drawn from Eurobarometer 99.4. Prompts encode socio-demographic features in short English or native-language sentences and, in the full version, four underlined attitudinal variables: political interest, left–right ideology, trust in the EU, and attitudes toward further European integration. The output space includes “No” for non-voters and the country-specific set of parties, randomized to avoid positional bias [2409.09045].

The reported findings are markedly pessimistic. Under the English full prompt, predicted turnout averages 83 percent versus actual approximately 49 percent, an overestimation of 34 percentage points. The party winner is identified in only 11 of 27 countries, mean party-rank accuracy is 8 percent with median 0 percent, and mean absolute error in party vote shares per country lies between 7 and 15 percentage points. Performance is context-dependent: Western Europe shows smaller turnout errors, while Eastern and Southern Europe reach overestimation up to 50 percentage points; Romance and Germanic language countries show lower vote-share MAE than Slavic and Baltic countries. Native-language prompting worsens turnout predictions across the five focus countries, while attitudinal information consistently improves turnout and vote-share metrics relative to demographics-only prompting [2409.09045].

Taken together, these adjacent studies broaden the EuroParlVote landscape beyond parliamentary roll-call analysis. They examine whether LLMs can infer party stances from manifestos and retrieved context, and whether they can predict individual electoral behavior across heterogeneous national and linguistic contexts. Their conclusions are more cautious than the benchmark papers: automated RAG remains limited by noisy retrieval, and synthetic-sample electoral prediction “largely fail[s],” especially outside Western and English-dominant contexts [2407.08495; 2409.09045].

## 7. Significance and open research questions

Across its variants, EuroParlVote occupies an intersection of network science, legislative studies, and political NLP. The 2015 signed-network dataset demonstrates that dense, continuous signed relations extracted from roll-call agreement are analytically consequential: negative links materially alter graph partitions and are not safely discarded in the European Parliament setting [1507.04215]. The 2025 benchmark shifts the emphasis from inter-MEP similarity to text-conditioned prediction and fairness auditing, using speeches, topics, and metadata to probe model behavior under demographic and ideological variation [2509.06164; 2509.16264].

Several open questions recur across the literature. One is representational: whether signed community detection and structural-balance objectives can be unified with link-density principles in a genuinely signed community-detection framework for legislative data [1507.04215]. Another is multimodal and multilingual generalization: the 2025 platform explicitly notes that debates occur in 24 EU languages and that generalization to new topics, languages, or future parliamentary terms is untested [2509.16264]. A third concerns fairness and accountability: benchmark results indicate that model accuracy and group disparities vary materially by gender, ideology, country, and language, and that explicit context can reduce some disparities but not eliminate them [2509.06164; 2409.09045].

A broader synthesis is that EuroParlVote has become a testbed for two distinct but connected scientific questions. The first concerns the structure of political conflict, coalition formation, and polarization as revealed by signed voting relations. The second concerns the extent to which contemporary LLMs can model, predict, or assist with political decision processes without reproducing systematic demographic, ideological, or contextual biases. The literature considered here does not present these questions as resolved. Instead, it frames EuroParlVote as an evolving research infrastructure for studying them under increasingly technical, comparative, and fairness-sensitive conditions [1507.04215; 2509.06164; 2509.16264].

Source: https://www.emergentmind.com/topics/europarlvote