A Comparative Analysis of Approaches for Automatic Frame Detection

Framing analysis in news is essential in computational linguistics to detect specific ways of highlighting certain points of view that help to shape public opinion. In recent years, transformer-based models have significantly improved this type of analysis, although some questions remain regarding their adaptability and ability to perform well on different datasets. In this thesis, I focus on comparing different methods of automatic frame detection, with particular attention to the improvement and evaluation of FrameFinder, an open-source framework. I tested the system on the Gun Violence Frame Corpus (GVFC) and introduced some structural optimisations, primarily to handle external configuration files that define frame labels, dimensions and polarities in a more flexible manner. In addition to comparing contextual and non-contextual approaches to polarity detection, the study analyses the informational impact of using titles alone versus a combination of titles and article abstracts. To ensure reproducible results, all experiments were run on the CAPRI cluster using Singularity containers.