ValuesML: A new multilingual dataset for values detection in news and political manifestos

Values are important building blocks of political ideologies and are frequently invoked in political debates. Yet, because values are abstract, understanding how they are expressed in real-world discourse, and how this varies across contexts, remains a challenge. In this paper, we introduce a large-scale, expert-annotated dataset of value expression in political text, comprising news articles and political manifestos across nine languages. The dataset is grounded in the refined theory of human values (Schwartz et al., 2012) and was developed through an iterative process of annotation guideline development, annotator training, and expert curation. Value annotations capture both the type of value expressed and its evaluative framing, distinguishing whether values are expressed as attained or constrained. The final dataset comprises 2648 texts and 74,231 sentences across nine languages. This resource enables systematic, cross-linguistic analysis of value expression in political communication and provides a benchmark for developing and evaluating computational models of value detection.