An experimental Critical Multimodal Discourse study to the AI-driven sentiment analysis of online crisis communication

In response to the challenges of crisis management and communication in business digital scenarios (Umar et al. 2022; Catenaccio 2021), this case study presents an example of the evolution and implications of online crisis communication and discursive practices that combine human input and AI for the management of crisis events and trust repair. Reliability of communication in web-based business scenarios cannot be easily achieved because of the huge amount of data involved. In addition, reputation management and trust are constantly put under threat by misinformation and misunderstanding (Garzone, Giordano 2018). With a view to countering these risks, companies are increasingly outsourcing digital marketing services (Palttala et al. 2012) based on AI methods to analyse consumers’ online needs and behaviour (Schwaiger et al. 2021), even though research has recently raised some concerns about over-reliance on AI-based tools (Tam, Kim 2019). We intend to show how multimodal critical discourse studies (Djonov, Zhao 2014) can help identify and understand the potentials and limits of social media listening tools that are designed for crisis prediction and management (van Zoonen, van der Meer 2015). To do so, we present a case study that we engaged with during our research traineeship at Digital Trails, a B2B company dealing with online visibility, digital marketing and reputation analysis. In reporting the outcomes of online reputation analysis to gauge possible damage after a crisis event, we present a comparison between AI-driven sentiment analysis (conducted via Meltwater) and the human-based revision and fine-tuning of AI-driven sentiment analysis. Our aim is to discuss the potentials and the criticalities of AI-driven sentiment detection. Among the latter, we highlight the unrecognizability of languages other than English, and flaws in interpreting pragmatic aspects, as well as multimodal digital artefacts. Consistent with our findings, we argue that AI models based on a unimodal and decontextualized architecture still require human validation. We conclude by indicating research directions for the detection of sentiment polarity, which include higher collaboration between IT developers and multimodal discourse analysts so that multimodally-informed models can assist crisis communication and management more efficiently.