Finding frames with BERT: A transformer-based approach to generic news frame detection

Defined by Entmann (1993) as a process of selecting and making more salient the specific aspects of social reality, framing is among the most extensively used concepts in the field of communication science (Olsson & Ihlen, 2018). The abundant body of research utilising the concept of framing highlights the versatility of the concept: it has been used for examining the representation of armed conflict (Tschirky & Makhortykh, 2024), climate change (Vu et al., 2021), politics (Ogan et al., 2018), and racial injustice (Lane et al., 2020). The diversity of areas in which the concept of framing is applied and the vagueness of its operationalisation are, however, occasionally viewed as the concept's weakness: Cacciatore et al. (2016) note that it results in the unnecessarily broad understanding of framing that overlaps with other concepts, such as agenda-setting, and diminishes its explanatory potential. Despite the above-mentioned criticism, we suggest that framing remains an essential tool for understanding how certain interpretations of important societal issues become more visible and in which ways individuals are exposed to these interpretations. The importance of such an understanding increases under the conditions of the high-choice media environment (van Aelst et al., 2017) in which we are consuming information. With more available information sources and, consequently, more possibilities for being exposed to them -- both selectively (Messing & Westwood, 2014) and incidentally (Lee & Kim, 2014) -- it is crucial to be able to distinguish between frames coming from these sources, especially regarding the salience of epistemically contested issues which can easily amplify polarisation in the society. The ability to detect the presence or absence of specific frames in this context also becomes paramount for detecting attempts to manipulate public opinion. Another reason why frame detection is highly relevant is the growing reliance on artificial intelligence (AI)-powered systems for organising and generating information regarding societally relevant issues. The adoption of systems such as search engines and recommendations systems and, recently, generative AI-powered chatbots has profound implications for how individuals are exposed to information as these systems decide what information sources and interpretations are prioritised in response to the user input (e.g.