Motivational Computing: Transformer-Based Automation of Implicit Motive Coding

Implicit motives have a clear relation to observed behavior but are time-consuming to measure. Automating the content coding of implicit motives has been a pursuit since the 1960s, but has not been widely adopted due to limitations with the marker-word approach that underpinned those automation attempts. More recent attempts use a machine-learning-based approach to automating implicit motive coding. In this paper, we report on a new transformer model, in which machine-predicted scores of the Winter (1994) training manual benchmark dataset achieve over or near .85 correlation with hand-coded scores, as measured by both Pearson and intraclass correlation coefficients. We additionally provide evidence of the convergent, divergent, causal, and criterion validity of the machine-predicted motive scores from our model, and make the model freely available for research via a web-based API.¹ Finally, we compare performance between our model and two recent transformer-based models. We hope that triangulation of evidence from independent, parallel automation efforts will kindle greater advancement in this emerging field of "motivational computing", in which computers are used to detect and study motives in real time.