A Sentiment-Based Approach to Measuring Multidimensional Party Positions using Transformer

Existing position scaling models from texts commonly rely on the conflict-driven assumption: that political subjects always exhibit divergent positions. Such scaling models search the linguistic space for differentiating features, often conflating rhetorical styles with political sentiments when the subjects are not inherently opposed. This article introduces ContextScale, a sentiment-based position scaling framework leveraging pre-trained Transformer models to measure multidimensional position scores at the sequence level, as well as predicting policy topics and political sentiments. Unlike conflict-driven models, ContextScale effectively distinguishes rhetorical styles from actual political sentiments. Further validity tests demonstrate its ability to capture nuances in party positioning, as well as its adaptability across languages, text domains, and theoretical models.