ParlaSent: Mapping Sentiment in Political Discourse with Large Language Models - OSCARS-funded project ParlaCAP

This is a public repository hosting the preprint of the article titled ParlaSent: Mapping Sentiment in Political Discourse with Large Language Models. Abstract This paper introduces ParlaSent, a deep learning model for multilingual sentiment analysis, specifically designed to study sentiment in political discourse. Built upon the XLM-R-parla model—a domain-specific multilingual transformer based on XLM-RoBERTa—ParlaSent benefits from additional pre-training on 1.72 billion words from the parliamentary proceedings of 26 European parliaments. ParlaSent outperforms sentiment dictionaries and achieves performance comparable to GPT-4o in a zero-shot sentiment classification setting, while being significantly faster and more cost-effective. By enabling high-quality sentiment analysis across multiple languages, ParlaSent enhances the study of political sentiment within European democracies and supports efficient, large-scale sentiment annotation in parliamentary discourse.