Conceptualizing Illiberalism Using Natural Language Processing

This paper's goal is to elaborate on how we can use a large corpus of text data together with machine learning to operationalise a multi-faceted and complex concept such as illiberalism. First, we elaborate on the different machine learning and Natural Language Processing methodologies available for such purpose. We present a review of methodologies that could be useful to conceptualise illiberalism using text data, and how we are planning to build a model using these methodologies. We then elaborate on potential sources of data and how to access them. We also report on how to link different resources and why manifestos, parliamentary speeches, and social media data are the best potential fit. Third, we elaborate on the conceptualisation of illiberalism and on the process we used to label the data. In the end we use a deep learning model to classify texts associated with illiberalism. We specifically will use a MdBERTa multilingual model trained on our whole corpus to classify texts characterised by illiberalism. We show the results of the model and we use the newly labelled text to build a ‘map of illiberalisms'. Our final goal is to understand the salient characteristics of texts containing illiberalism and to create a map that highlights the distinctive ideological features of illiberalism and neo-authoritarian values versus liberal democratic values. We do so by using Word Embeddings to place all the newly labelled texts on a relevant scale. This map built on our MdBERTa model gives us the options to explore new underlying dynamics and potential new associations of words and topics to illiberalism.