2019. The relationship between influential actors’ language and violence: A Kenyan case study using artificial intelligence
Scholarly work addressing the drivers of violent conflict predominantly focus on macro-level factors, often surrounding social group-specific grievances relating to access to power, justice, security, services, land, and resources. Recent work identifies these factors of risk and their heightened risk during shocks, such as a natural disaster or significant economic adjustment. What we know little about is the role played by influential actors in mobilising people towards or away from violence during such episodes. We hypothesise that influential actors’ language indicates their intent towards or away from violence. Much work has been done to identify what constitutes hostile vernacular in political systems prone to violence, however, it has not considered the language of specific influential actors. Our methodology targeting this knowledge gap employs a suite of third party software tools to collect and analyse 6,100 Kenyan social media (Twitter) utterances from January 2012 to December 2017. This software reads and understands words’ meaning in multiple languages to allocate sentiment scores using a technology called Natural Language Processing (NLP). The proprietary NLP software, which incorporates the latest artificial intelligence advances, including deep learning, transforms unstructured textual data (i.e. a tweet or blog post) into structured data (i.e. a number) to gauge the authors’ changing emotional tone over time.
Our model predicts both increases and decreases in average fatalities 50 to 150 days in advance, with overall accuracy approaching 85%. This finding suggests a role for influential actors in determining increases or decreases in violence and the method’s potential for advancing understandings of violence and language. Further, the findings demonstrate the utility of local political and sociological theoretical knowledge for calibrating algorithmic analysis. This approach may enable identification of specific speech configurations associated with an increased or decreased risk of violence. We propose further exploration of this methodology.