Viewpoint detection on LGBT + reporting using contextual embeddings and qualitative thematic analysis: The use case on the word <i>deep</i>

This article presents an interdisciplinary study combining advanced natural language processing techniques by using contextual embeddings and manual thematic analysis. We analyse Slovenian news articles on LGBT + topics, focusing on the differences in the connotation of the word deep, the word whose usage differs the most between mainstream and conservative media groups, according to the system for automatic measuring of usage changes based on contextual embedding. At the content level, the study shows that, in mainstream media, the word deep is predominantly used in a conventional sense, with LGBT + topics being approached in a nonoppositional manner. On the other hand, conservative news portals predominantly refer to the construct of the ‘deep state’, its agents and practices, hence approaching LGBT + as one of the ‘forces in the shadows’ that push forward its agenda by subordinating main socio-political institutions with the help of its ‘allies’, for example, left-wing political parties. Additionally, we compare the results of the manual coding and the automatically defined clusters of usage and evaluate the potential of computational methods for qualitative analysis. The results show that computational analysis can act as a rough approximation to manual coding for main and emphasised differences, but it fails to detect fine-grained differences.