PoliSpace : Representing Political Questions and Opinions in a Shared Embedding Space

Sammanfattning

While AI has transformed many areas of society, its application in modeling political opinions remains relatively underexplored. Existing strategies often rely on large language models (LLMs), either through prompt-based personalization, where the individual’s information is included in the prompt, or through fine-tuning on user-specific data. However, such techniques face several challenges, including reliance on prompt engineering, vulnerability to hallucinations, and the costly requirement of training separate models for each individual. Most critically, they provide little transparency in how political opinions are represented or how predictions about an individual’s views are generated. To address these limitations, we propose POLISPACE, a simple and interpretable framework for political opinion modeling that embeds both individuals and political questions in a shared vector space. POLISPACE leverages the recent advances and improved semantic capabilities of text encoders to generate representations of political questions. The model simultaneously learns an embedding for each individual that captures its political view, such that the alignment between an individual’s embedding and a question embedding indicates agreement. We further extend this approach with POLISPACE+, which introduces individual-specific transformations of the question space to capture nuanced personal interpretations of questions. This embedding-based framework allows us not only to predict political opinions but also to perform confidence-aware reasoning such as opinion retrieval and thresholding to flag questions where the model lacks sufficient information to provide a reliable answer. We evaluate our proposed models using real-world data from the Swiss voting advice application Smartvote, testing across four increasingly difficult settings. POLISPACE+ outperforms most baselines by a wide margin and achieves performance comparable to ChatGPT prompting on unseen opinion prediction for indirectly seen questions, achieving an F1 score of 89.81%. In the most challenging unseen-question setting, POLISPACE and POLISPACE+ fail to outperform the baselines, yielding at best marginal gains. Despite this, we see POLISPACE as a foundation for building more robust and interpretable models for political opinion modeling.