Automatic candidates identification for Robotic Process Automation based on textual process descriptions

Abstract Robotic Process Automation has been widely adopted across industries over the past years, evolving from a niche automation tool into a mature, commercially established technology. Nonetheless, the initial identification of suitable process candidates remains a critical and challenging step. The study concentrates on the automatic candidate identification from textual process descriptions based on the approach proposed by Leopold et al. We extend the original work through three principal contributions: the expansion of the training dataset with a larger, multi-domain corpus; the generation of several new features derived from consolidated process characteristics; and the comparison of several classifiers. Our experiments reveal that the contribution of the engineered features is limited, and identify the reason: two of the four features encode the same distinction, and that distinction carries no positive signal. Text representation has a modest effect, and the modern representations do not outperform plain word-level TF-IDF; among the evaluated configurations, the simple word-level TF-IDF with LinearSVC attains the best overall balance. Crucially, cross-dataset evaluation shows that models trained on synthetically generated text do not fully transfer to human-authored text. Overall, the study provides an empirically grounded assessment of the current limits and opportunities in text-based RPA candidate identification.