Mapping Democracy Towards AI-Strengthened Political Discourse

In this work, we set out to investigate the potential of Artificial Intelligence (AI), to aid democratic practice. To that end, we developed an AI pipeline leveraging advancements in Natural Language Processing (NLP) to perform machine-based analysis of political texts. We present two proof-of-concept implementations of our method: (1) alignment analysis of party manifestos to policy texts like the EU Green Deal and NATO's 2022 Strategic Concept, and (2) identification of convergence trends with far-right discourse. Our method reveals how AI-driven text analysis can facilitate transparent, evidence-based policy discussions and enhance democratic accountability. While our findings primarily validate existing political science observations, they demonstrate the methodological potential for systematic, data-driven political discourse analysis.

Key Results and Take-Aways

1. Policy Alignment
   • EU Green Deal: The Greens and SPD exhibit the strongest alignment, while the AfD and the Left show more skepticism or alternative frameworks.
   • NATO's 2022 Strategic Concept: Greens and SPD again rank highest in defense-related alignment, while AfD and the Left remain most distant.
   • Policy Gaps: Certain themes—such as stricter legislative enforcement or deeper EU and NATO integration—are underrepresented, pointing to potential opportunities in lobbying or coalition building with selected parties.
2. Party Convergence with Far-Right Narratives
   • General Trend (2013–2025): All mainstream parties moved closer to AfD positions, especially in topics linked to national identity and law & order—likely influenced by increased coverage of immigration and security post-2015.
   • The Left remains consistently furthest from AfD, showing a counter-trend to other parties by diverging again after a brief dip in 2017.
   • Normalization of Far-Right Frames: Most parties show convergence on far-right narratives. Shared language on immigration or cultural identity can legitimize far-right discourse and shape broader political debate.
3. AI for Democracy
   • Methods: A modular framework combining semantic embeddings, sentiment, stance classification, and topic modeling.
   • Potential: The approach can be extended beyond manifestos, enabling voter guidance tools, strategic voting analyses, and real-time discourse monitoring.Introduction

Introduction

Artificial Intelligence (AI) increasingly impacts the political landscape. While there has been research directed towards the potential negative implications of AI on democratic processes— such as AI-generated content, mis- and disinformation, or manipulative campaign tactics—AI also presents an opportunity to strengthen democratic processes. In recent years, far-right and populist movements have effectively leveraged AI-driven methods, underscoring the urgency for a counterpoint to mitigate the negative perception of AI in political discourse. AI can enable transparent, evidence-based political analysis, policymaking, and more.

To that end, we set out to demonstrate how AI can benefit democracy by developing an end-to- end AI pipeline for analyzing political text similarity. Although our work emphasizes the tool, numerous research opportunities could arise from the effective use of text analysis and data- driven policy analysis. We illustrate practical applications through two prototypical use cases intended as inspiration for future research directions:

• Policy Alignment: Comparing German party manifestos from the 2025 elections to NATO's 2022 Strategic Concept and the EU Green Deal, revealing areas of high alignment, potential policy gaps, lobbying opportunities, and possible coalition pathways.
• Party Convergence: Examining convergence trends with the far-right party Alternative für Deutschland (AfD) starting from 2013, providing insights for early warning systems, strategic voting, and discourse monitoring.

Our analysis focuses on manifestos from parties elected to the German Parliament in 2025 and those represented in the preceding election period: Die LINKE (The Left), Sozialdemokratische Partei Deutschlands (SPD), Bündnis 90/Die Grünen (The Greens), Freie Demokratische Partei (FDP), the CDU/CSU (Christian Democrats), and Alternative für Deutschland (AfD). For policy alignment, only the most recent 2025 manifestos were considered. To analyze political convergence over time, manifestos from 2013 to 2025 were incorporated. While we exclusively focus on German parties, we designed our method in a way that allows the use across countries and regulatory frameworks providing opportunities for further analysis.

As we see our method as a prototype and the potential for misuse, we do not provide our data, code, or models. However, if you are interested in trying it out or want to reach out, please contact Julian Schön 

Methodology

Data

The analysis utilizes political manifesto data from the Manifesto Project [1], encompassing German political parties (AfD, CDU/CSU, FDP, SPD, Greens, and Die LINKE) from 2013 to 2021. Manifestos from the 2025 German parliamentary election were additionally collected and processed. Each manifesto underwent sentence segmentation and translation into English using DeepL [2]. The Manifesto Project categorizes each sentence into one of 72 predefined topics. To streamline further analysis we condense these into 18 broader topics and one for non-thematic sentences.

Methods

We developed a modular framework for analyzing sentence-level political text similarity. First, we obtained sentence embeddings in vector space using a semantic embedding model. Given that semantic embedding is a well-established field, we relied on pre-trained models and the SentenceBert library [3]. Our choice of embedding model was based on a mixture of the MTEB Leaderboard [4] and visual inspection of the resulting similarity scores. The final model selected was intfloat/multilingual-e5-large-instruct [5].

Semantic embedding models, as implied by the name, position semantically similar sentences closely together and dissimilar ones farther apart in their embedding space. However, a limitation is that political nuances, such as opposing stances on the same topic, can yield high similarity scores. To address this, we enhanced our similarity calculation by incorporating sentence sentiment and political stance. We employed the pre-trained sentiment classification model distilbert-base-uncased-finetuned-sst-2- english [6] to classify sentences as positive or negative. For political stance classification into left or right, we utilized a zero-shot classification based on mlburnham/Political_DEBATE_large_v1.0 [7]. We experiment with fine-tuning the models as the performance of the sentiment and political stance models could be improved. However, given that our data is not labelled for stance and sentiment improvements were not significant.

Finally, to classify new documents into our sentence classification scheme, we trained an ensemble classifier. The ensemble leveraged ManifestoBerta [8] and Roberta models, specifically fine-tuned on our customized simplified category system.

We integrated this modular approach into a weighted similarity scoring method, combining cosine similarity for embeddings, with a weight of 0.5, and categorical distance measures for sentiment and stance, each weighted at 0.25. For the categorical distance, we consider sentences labelled as having opposite sentiment (positive vs. negative) receive a lower similarity of 0 and 1 otherwise. This helps distinguish opposing viewpoints that might otherwise share overlapping keywords. For the political stance classification, we obtain labels corresponding to left–center-right on the political spectrum. If both are labelled as having the same political stance, the similarity is 1; if they are not equal but adjacent, their similarity is 0.5. Otherwise, the similarity is 0. The weights are chosen to balance the trade-off of noisy labels obtained from stance and sentiment classification, while still adding them to the analysis to strengthen the handling of opposing views. Thus, the cosine similarity between the two sentences indicates how closely the sentences align in semantics, while the addition of stance and sentiment aim to improve similarity with regards to political nuances of the text, while the weights aim to balance the noisy labels generated by sub-optimal models.

Analysis

In this report, we primarily use two methods of analysis. The first form of analysis is based on a sentence-level comparison of a certain topic based on the above-described methods. The second method is based on simple topic modelling to perform subtopic analysis.

Sentence Level Similarity

The goal of this analysis is to give a broad overview of the similarity of two given documents on a specific topic, e.g. foreign-policy. To achieve this, we use the embedding along with sentiment and stance labels. We calculate the weighted similarity score as described above between any two sentences of the given documents that fall in the topic of interest as classified by our classification ensemble. Finally, to perform the topic-wide analysis, we aggregated the 15 sentences most similar within a topic using the median to weaken the influence of outliers.

By combining these factors, we aim to move beyond only considering lexical or semantic overlap to capture differences in ideological and tone. For our purpose, sentence-level similarity is especially useful for:

• Granular Comparisons: Analysing where two manifestos—e.g., from different parties—use nearly identical language or framing within a certain topic.
• Quantitative Aggregations: Summarizing how close or far each party is from a reference on a sentence-by-sentence basis.
• Trend Analyses: Observing how distances evolve over time for a particular theme.

Sentence Clustering for Sub-Topic Identification

To understand why certain sets of sentences within a classified topic appear similar, we cluster sentences from a reference text, e.g. far-right manifesto or EU policy document, or any other source within a topic of interest. Each cluster represents a sub-topic, following a minimal topic-modelling approach:

1. Reference Text Clustering
   • We gather all sentences from the chosen reference corpus within a topic.
   • A clustering algorithm groups them based on semantic embeddings, forming sub-topics.
2. Cluster Summaries
   • For each cluster, we sample the top 10 sentences closest to the centroid and summarize them to get a description of the subtopic using ChatGPT.
3. Matching Other Parties
   • We then map each sentence from other parties' manifestos to the closest cluster centroid in the reference.
   • This identifies where a party overlaps with specific sub-themes, highlighting the policy frames or rhetorical patterns they share with the reference text.

Through clustering, we capture how parties distribute their statements across sub-topics rather than relying on a single distance measure for an entire theme, giving us more nuanced insights.

Policy Alignment with Party Stand-points

In this policy project we aim to show how our tool can be used to compare different political text. To that end, we consider comparing the 2025 manifestos to key policy documents. We compare the manifestos to both the EU's Green Deal and NATO's 2022 Strategic Concept. This analysis gives insight into how our method can be used to identify how parties align with certain policy.

Additionally, it allows the identification of policy gaps in a political landscape if none of the parties cover certain parts of the policy in question, identify particular lobbying or coalition opportunities for certain policies, and more. Notice, that the policies chosen for comparison are arbitrary and the approach can easily be adapted to any other policy paper.

Case 1: EU’s Green Deal

Since the EU launched its Green Deal, member states' parties have increasingly incorporated (or resisted) key ideas like emission reduction targets, sustainable investment, and stronger climate laws. This analysis compares the six chosen major German parties—AfD, CDU, FDP, SPD, Greens, and the Left—against core Green Deal pillars, highlighting both aggregate rankings and cluster-specific thematic overlaps.

We compare the manifestos to the EU Green Deal using a sentence wise relative ranking based on our similarity measure. The ranking indicates closeness to the policy, i.e., lower is better. We compare topics if a topic has more than 10 sentences dedicated to it in the policy paper, as otherwise it is not considered a central theme of the policy. Further, as we calculate the similarity based on topics, we calculate a weighted overall similarity for each party. The weighted overall similarity is calculated as the weighted sum of all topic similarities, where the weights correspond to the percentage of overall sentences dedicated to that specific topic in the policy document (i.e., EU’s Green Deal). By following this approach, more weight will be given to the similarities than cover the central theme of the policy project resulting in a more robust overall ranking. The relative ranking can be seen in Table 1. Further, we perform a cluster analysis of the main Category 'Environment & Sustainability' of the Green Deal. The results can be seen in Figure 1. In the cluster analysis we consider the average distance to the cluster center, i.e., lower means closer textual/semantic similarity to Green Deal language and size, i.e., number of sentences matched, showing how extensively each party discusses that sub- theme.

Table 1. Relative Sentence Based Similarity of Parties Compared to EU’s Green Deal

Image

Figure 2. Clusters over EU’s Texts in Category 16 and their Corresponding matches by the parties. Size of the circle indicates the number of matched sentences and distance indicates the mean similarity.

Interpretation & Implications

Alignment Spectrum: The Greens and SPD demonstrate the strongest commitment to EU-aligned climate action, evidenced by both high rankings and robust coverage in environment clusters.

CDU and FDP do engage with Green Deal themes (notably in clusters emphasizing technology, investments, or market instruments), but often with narrower or more market-centric language. Left invests substantial text on environment (135 sentences in Cluster 1) but sometimes diverges from EU-oriented growth frameworks, likely focusing more on social justice and state intervention.

AfD generally lags, though it still references climate and environment themes (size=60 in Cluster 1, 14 in Cluster 4), indicating partial engagement—albeit with higher distances from EU proposals indicating negative references of the discussed topics.

Potential Policy Outcomes: Most parties now discuss emission reductions, sustainable investments, and legislative enforcement. This suggests some normalization of EU climate discourse in the German political mainstream.

Variation in average distances (and small cluster sizes for some parties) hints that not all are equally committed to binding targets or robust EU governance. FDP, for instance, might favor lighter regulatory frameworks, while the Left might insists on more radical change. However, we clearly see the strong alignment of the Greens with EU environmental endeavors.

Policy Gap Identification

1. Ambition Levels & Enforcement
   • Cluster 6 ("Focus on strengthening climate laws...enforcement by the Commission") sees limited coverage for CDU (size=3), almost none for AfD (size=1) and none for FDP and SPD. This indicates a gap in endorsing strict EU oversight.
2. Global Climate Action vs. Domestic Priorities
   • Clusters 2–3 stress EU-led global leadership. Parties with fewer sentences matched there might prioritize national measures over EU-level or global frameworks.
3. Socioeconomic Implications
   • While the Left invests heavily in environmental text, its lower alignment with certain EU macro- economic or market-based solutions shows a gap in bridging social aims with EU-style green growth strategies.

Overall as the Greens show good coverage there is no party wide policy gap regarding EUs Green Deal, however there are clear party specific gaps.

Key Takeaways

• Greens & SPD as EU Green Deal Front-Runners: High overall rankings and expansive cluster coverage mark them as the most consistently aligned with EU climate proposals.
• CDU & FDP: Moderate Overlap with Specific Emphases: They discuss environmental topics in a more selective, market-friendly way, often matching narrower sub-themes like sustainable investment (Cluster 4).
• Left's Broad Engagement, Divergent Economic Angle: Large text volume suggests strong environmental interest, but average distances in certain clusters reflect tension with the EU's growth-based approach.
• AfD's Skeptical Engagement: Although it references environmental policy, higher distances, and lower rankings confirm less alignment with the Green Deal's core tenets.
• Normalization of EU Climate Discourse: Across parties, multiple references to emissions reductions and regulatory frameworks signal that EU-led environmental narratives are increasingly mainstream, even if full consensus on enforcement or market instruments remains elusive.

Case 2: NATO’s 2022 Strategic Concept

NATO's 2022 Strategic Concept lays out key commitments to defense readiness, collective security, and responses to emerging threats. In this section, We assess how closely the chosen six German parties—AfD, CDU, FDP, SPD, Greens, and the Left—align with these priorities. We compare the documents using the same methods as above. The ranking and clustering results are provided in Table 2 and Figure 2, respectively.

Table 2. Relative Sentence Based Similarity of Parties Compared to NATO’s 2022 Strategic Concept.

Interpretation & Implications

• Greens & SPD: Their ranking suggests a robust acceptance of NATO's strategic outline. Cluster data confirms emphasis on deterrence (Cluster 2), crisis prevention (Cluster 5), and addressing emergent threats (Cluster 3).
• CDU & FDP: While generally pro-defense, textual distances suggest a narrower alignment

—likely focusing on conventional defense readiness and cooperative security but diverging on nuclear deterrence or deeper global commitments.

• AfD: Scores near the bottom of overall ranking, indicating skepticism toward NATO's broad scope. The moderate cluster sizes show it still addresses NATO concerns, albeit from a more distant rhetorical stance (higher average distance).
• Left: Reinforces its position as the most NATO-critical party. Although it discusses multiple clusters, the consistently higher distances signify reluctance about defense spending and nuclear weapons.

Policy Implications:

Image

Figure 3. Clusters over NATO’s Texts in Category 2 and their Corresponding matches by the parties. Size of the circle indicates the number of matched sentences and distance indicates the mean similarity.

• A potential cross-party consensus emerges around crisis prevention (Cluster 5) and threats from non- state actors (Cluster 6), though the Left and AfD remain more resistant to robust collective defense measures.
• Variation in nuclear deterrence stances (Cluster 4 ) underscores persisting divides on nuclear policy, possibly affecting Germany's role in NATO's extended deterrence frameworks.

Gap Analysis

• Nuclear Posture (Cluster 4): Mixed engagement from CDU/FDP, while Greens/Left show relatively higher distances, highlighting policy friction over nuclear deterrence.
• Defense Readiness vs. Domestic Emphasis: SPD and Greens adopt more language on readiness (Clusters 2 and 5), whereas AfD and Left remain more distant, revealing a gap in prioritizing NATO-led military initiatives.

Key Take-Aways

• Greens & SPD: Most closely aligned with NATO's 2022 Strategic Concept across defense and foreign policy.
• CDU & FDP: Show moderate alignment, focusing on conventional deterrence but revealing some distance in nuclear or global threat sub-topics.
• AfD & Left: Far less aligned, each for different reasons—AfD leaning toward skepticism about collective NATO policies, and the Left historically anti-militarist.
• Cluster Insights: All parties address multiple aspects of NATO's strategic priorities, but average distances indicate substantial variation in how fully they embrace shared defense, nuclear postures, and crisis prevention.

Political Convergence

Over recent years, traditional far-right narratives have increasingly assimilated into center- democratic parties. This convergence legitimizes far-right topics and seldom weakens those parties. To provide insights and validate these trends, we evaluated political convergence. We specifically considered categories 12, 13, and 18 as primary drivers of far-right politics. This analysis can be expanded to include EU-positive or negative stances and more.

Using manifestos from 2013-2025, we compared mainstream parties' manifestos to AfD manifestos, focusing on specified topics. For overall similarity, we calculated average similarity over topics rather than sentence-level similarity for entire manifestos, given the limitations of embedding models detailed later.

Figure 3. Showcases each party's distance to the AfD manifestos over time. As we consider a distance metric, lower means a closer standpoint to the AfD. Because each distance

Image

Figure 3. Political Convergence on selected topics of all party manifestos to the AfD.

measurement is made against the same year's AfD manifesto, the sharp distance decrease we see from 2013 to 2017 is likely because of the expanded manifestos of the AfD between those years. An additional explanation for the sharp shift between 2013 and 2017 could be attributed to the 2015 refugee crisis increasing the amount of text each party dedicates to these topics and consequently also potentially increasing the similarity of all manifestos to that of the AfD.

Topic 12: National Way of Life

• CDU: Starts at a higher distance (~0.37 in 2013) but drops notably by 2017 and remains around ~0.19–0.20 by 2025
• -FDP/Greens/SPD: Show moderate but steady decreases. In all cases, their distance in 2025 is significantly lower than in 2013.
• Left: Begins with the highest distance (~0.44 in 2013), then dips but consistently remains above 0.28–0.30 by 2025. Despite some convergence, it still stands out as the farthest from AfD positions on nationalistic narratives.

Key Observation:

Topic 12 shows the most marked convergence for nearly all parties. Nationalism/immigration/ patriotism often becomes mainstream election talk, leading to rhetorical overlap—even if that was not the intention.

Topic 13: Culture & National Identity

Again, we observe a general shift in similarity closer to the AfD.

• CDU starts out lowest. Given that they are a conservative Christian party, this is not surprising.
• Interestingly, the SPD showcases a significant drop, resulting in their standpoint being closest to that of the AfD by 2025
• Interestingly, both the Greens and the SPD will converge with the AfD on the topic in 2025, while all other parties will diverge.
• The other parties remain relatively stable with yearly fluctuations.

While the mainstream parties may have always referenced cultural identity in some fashion, leading to the overall closer distance to the AfD, especially the change in the trend of both SPD and Greens is noteworthy.

Topic 15: Law & Order

Data availability was not given for some parties in 2013, leading us to consider only the manifestos from 2017 onwards.: Many parties only show values from 2017 onward for Topic 15. This already bypasses the initial big shift from 2013 to 2017, so we see less of a "sharp drop" effect.

Convergence: Distances are in the 0.18–0.25 range. CDU and SPD often have the smallest distances, possibly reflecting commonly accepted positions on policing or public safety. The Left remains higher than the other parties but still trends downward slightly.

Interpretation:

Conservative-leaning positions or mainstream approaches to criminal justice (e.g., stronger police presence, security expansions) might show more textual overlap with the AfD's 2017+ manifestos, reducing measured distance.

Interpretation & Implications

Mainstreaming and Legitimization of Right-Wing Narratives

The data indicates that all mainstream parties increasingly adopt language once viewed as far- right, especially on Topics 12 (National Way of Life) and 15 (Law & Order). When traditional parties co-opt these narratives—whether to address voter concerns or to neutralize the far-right—they often legitimize the framing and policy positions in public discourse. Empirical cases in other European contexts (e.g., Austria, Netherlands, or France) illustrate how normalizing far-right themes can increase far-right parties' credibility and electoral appeal.

Increased Text Similarity Driven by Refugee Crisis Coverage

One reason for the substantial drop in distances between 2013 and 2017 is the expansion of content about immigration and security in all-party manifestos. This is partly a response to the 2015 refugee crisis, which heightened public attention to issues of immigration, cultural

identity, and law enforcement. As mainstream parties devoted more sentences to these topics, the measured overlap in language with the AfD, which was already emphasizing these topics, naturally rose. This is also due to the inherent nature of text similarity measurement, as texts are generally more similar if they cover the same talking points, even if they are from opposing views.

The Left’s Diverging Trend and 2017 Dip

While the Left is consistently the furthest from AfD positions, it shows a temporary dip in 2017, reducing distance more than in other years. However, from 2017 onward, its distance generally increases - or at least remains higher relative to other parties. This pattern suggests the 2017 dip may be largely artifactual, driven by an overall increase in manifesto text addressing immigration and cultural topics. The Left's year-over-year rebound could indicate a stronger opposition to far-right narratives than raw similarity scores might initially suggest. For example, in 2017, many left-wing parties across Europe found themselves compelled to address immigration concerns at length (thus raising textual overlap), but they maintained a critical or humanitarian tone, distinguishing them ideologically even if some language overlapped. This gives us an insight into the general trend between 2013 and 2017, indicating that the significant changes in similarity are likely influenced by the general increase in text similarity due to the refugee crisis.

Subtopic Analysis

We perform the same subtopic analysis as before on the topic of 'National Way of Life (Nationalism, Immigration, Patriotism), as it covers the main topics of the far-right. The result are visualised in Figure 4.

Interpretation & Implications

Multiple Sub-Theme Overlaps: CDU strongly aligns with AfD on restrictive asylum measures (Cluster 1), showing both high volume (15 sentences) and low average distance (~0.157). This suggests they frequently discuss topics like external asylum processing and stricter deportations in terms that resemble AfD's framing.

Greens & FDP focus heavily on welfare or economic critiques of immigration (Cluster 2), while SPD exhibits moderate coverage across several clusters but stands out due to an extremely close match (~0.122) in Cluster 5. Even though it's only a few sentences, the extremely close match indicates that they are likely to be very close to AfD's sentences.

The Left typically has fewer matches, reinforcing its known distance from far-right positions. Yet in Clusters 1 and 6 (strict residency separation), a few sentences show relatively low distance. This highlights occasional rhetorical similarities—perhaps referencing the same issues (e.g., residency regulations) in parallel language.

Finally, it needs to be noted that, though we have tried to address this, low average distance likely means similar wording or topic; it does not perfectly correlate with ideological support. Some parties might address (and critique) the same issues, which still boosts textual overlap. However, when we see both high volume and low distance, it frequently implies a more thorough thematic convergence—discussing the same sub-theme in a manner reminiscent of AfD.

Overall Key Take-Aways

• All mainstream parties (CDU, SPD, FDP, Greens, Left) show a decline in distance to AfD positions from 2013 to 2025 on far-right themes (Topics 12, 13, and 15).
• Largest Drops (2013–2017): Much of the convergence occurs in this period, likely due to mainstream parties discussing traditionally right-wing topics more extensively, increasing textual overlap with the AfD.
• CDU Leads Convergence: It exhibits the strongest shift toward AfD rhetoric on nationalistic/ immigration-related content, as reflected both in overall distance and targeted cluster matches.
• FDP & SPD: Both parties show moderate but steady downward trends, converging around distance scores of ~0.20 by 2025.
• Greens: Also move closer on immigration/cultural themes—though less sharply than the CDU— and frequently engage AfD's economic/welfare arguments.
• Left as the Furthest: It remains most distant from AfD across all three topics, showing only partial convergence between 2013 and 2017. Notably, since 2017, it diverges again—making it the only party maintaining a more consistent gap.

• Shrinking Gaps by 2025: Across parties, differences in distance to AfD narrow, indicating a collective shift in rhetorical stance on immigration, cultural identity, and law enforcement.

  Image

  
  

  
  

  In the more nuanced cluster view, all parties partially adopt or at least address AfD's core frames—  potentially normalizing far-right narratives in mainstream debate.

• Normalization of Right-Wing Language: As mainstream parties increasingly address or echo far- right frames, these topics become more normalized in public debate—potentially legitimizing the AfD's core positions and shifting the overall political discourse.

Limitations

Our working still has key limitations that we aim to address in further in future work. One of the main limitations is the data. While the manifesto project classifies the data based on human experts, the classification scheme is quite complex, and the sentence splitting they perform sometimes results in sentences that are non-thematic if stripped from their context. This leads to several limitations for our work. First, it limits our ability to train downstream classifiers as the data quality is insufficient; second, the quality of our classifier significantly impacts the results of the 2025 manifesto classification as it is purely based on the classification model. Therefore, the quality of the labeled data indirectly influences the quality of labelling of the 2025 manifesto and, thereby, also the quality of the results.

Another limitation of our work are the underlying models. The topic classification model would further benefit from additional training as the classification accuracy of 64% can likely be improved further. This however plays hand in hand with the data quality that would need to be improved.

Additionally both the sentiment and stance model are not task specific. Therefore, both models would benefit greatly from fine tuning on task-specific data.

Addressing these key limitations would strengthen the work significantly and allow for a more founded analysis.

Given the time constraints of this project, we see this as an outlook of what is possible in the field of AI based political text analysis and look forward to improving our approach further in future work.

Outlook

Looking forward, there are multiple ways to expand and refine the AI-based political text analysis demonstrated here:

• Additional Use Cases:
  • Voter Guidance Tools: Offer citizens an interactive interface to compare personal preferences with detailed party stances.
  • Strategic Voting: Help voters identify which party's policies best serve specific objectives, encouraging more informed electoral choices.
  • Accountability Checks: Compare parliamentary voting records with manifesto pledges or coalition agreements.
  • Regional Politics & Parliament Data: Integrate local election manifestos and legislative transcripts for a broader, multi-level political landscape.
  • Question Answering: Build a system enabling direct queries about party positions, track records, or proposed legislation based on manifestos and MPs' voting behaviour.
  • Given the insights into political convergence, a future research project could take our initial insights further and develop a more comprehensive analysis and potential early warning system for convergence on specific topics. Additionally, a comparison between political stance convergence including an investigation on how linguistic characteristics such as populist language or key phrases propagate through the center-democratic parties would be an insightful analysis.
• Technological Enhancements: Fine-tune the sentiment and stance models on domain-specific data and incorporate advanced language models to better capture political nuances.

By adopting these expansions, future research could harness AI's full potential as a democratic resource, bridging gaps between political representation, policy evaluation, and citizen engagement.

Conclusion

We set out to investigate how AI can benefit democratic processes. To achieve this we developed an end-to-end NLP pipeline allowing us to analysis political text with AI revealing granular overlap and macro thematic trends within party manifestos. Our approach brings together sentence-level similarity and sub-topic clustering in an effort to understand the fine-grained nature of rhetorical alignment as well as the overarching frameworks of operation for parties. Through these approaches to comparing German parties to the EU Green Deal and NATO's 2022 Strategic Concept, we see clear differences in policy commitment and find convergence with far-right narratives on immigration, national identity, and law and order. While our approach is subject to limitations regarding data labelling, classification accuracy, and the nature of political speech, these initial findings show AI's potential to enhance transparency in democratic processes and policy. There are numerous applications: voter education applications, strategic voting apps, tighter accountability tools, and more policy gap analysis. To realize these opportunities fully, follow-up studies will need to improve data annotation methods, broaden domain-specific stance and sentiment models, and expand the pipeline to other policy topics and contextual datasets, such as legislative histories. Through these constant endeavors, text analysis based on AI can turn out to be a valuable democratic tool, aiding sharper communication among groups, empowering better- informed citizenship, and enabling a more transparent political system to flourish.

References

1. Lehmann, L.. (2024). The Manifesto Data Collection. Manifesto Project (MRG/CMP/MARPOR). Version 2024a.
2. https://www.deepl.com/en/translator
3. Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.
4. Muennighoff, N., Tazi, N., Magne, L., & Reimers, N. (2022). MTEB: Massive Text Embedding Benchmark. arXiv preprint arXiv:2210.07316.
5. Wang, L., Yang, N., Huang, X., Yang, L., Majumder, R., & Wei, F. (2024). Multilingual E5 Text Embeddings: A Technical Report. arXiv preprint arXiv:2402.05672.
6. Victor Sanh, Lysandre Debut, Julien Chaumond, & Thomas Wolf (2019). DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. CoRR, abs/1910.01108.
7. Burnham, M. (2024). Stance detection: a practical guide to classifying political beliefs in text. Political Science Research and Methods.
8. Burst, L.. (2024). manifestoberta. Version 56topics.context.2024.1.1.