Machine Learning Approaches to Facial and Text Analysis

The advent of machine learning (ML) tools presents researchers with the possibility of using large and new datasets related to text and image repositories. In this paper, we make a methodological contribution to strategy research by documenting a novel synthesis of two machine learning methods--the unsupervised topic modeling of textual data and the supervised ML coding of facial images with a neural network algorithm. We employ these novel methods to study CEO oral communication, using videos and corresponding transcripts of emerging market CEO interviews to conduct our analysis. Building on Helfat and Peteraf (2015) who document the importance of "oral language" as an important managerial cognitive capability, we code the topics and sentiments expressed in the text of what the CEOs say (verbal language) and separately code the facial expressions of the CEOs (non-verbal communication). Using the interview text sentiment scores as well as our video-based facial expression sentiment variables, we conducted factor analysis to construct four distinct CEO oral communication "styles," which we label Expressive, Stern, Dour, and Contented. We also reveal that CEOs who communicate with certain styles also tend to focus on specific topics, even controlling for their country of origin and gender. For example, CEOs who tend to be more expressive devote more attention to topics related to society at large and avoid topics related to the government. By contrast, dour CEOs are more likely to dwell on topics related to both the government and society. These results suggest that a CEO's communication style reveals a substantial amount of information about their attention to certain aspects of their businesses