Reinterpreting the surprisal calculation of the Topic Context Model

In this paper, we propose and demonstrate an novel computation of wordwise surprisal within the information-theoretic Topic Context Model (TCM). Rather than calculating the average surprisal as in previous studies, the TCM is redefined to output a simple, non-averaged surprisal value. This adjustment facilitates a more straightforward comparison with other common surprisal models and measures, such as n-gram surprisal.