Measuring the accuracy of social network ideological embeddings using language models

The opinions of people on different issues are traditionally studied through polls. Recently, ideological embedding methods have proposed to position social network users in spaces where positions are informative of their opinions. These methods have been shown to be effective in the US and in European settings. However, validating their results is challenging, as required data must not rely on the social network structure used in the embedding (panel data, for example). In this article, we propose a validation method based on language models for classification of users in ideological spaces. We illustrate our methodology using political manifestos, political surveys on party positions, and text utterances produced by Twitter users in Chile and in France, two countries with different party structure. Using text classification related to issues relevant to the ideological spatialization, we show that positions can be shown to be highly accurate, thus allowing for the robust inference of opinions of users at large scales.