This work is a step towards full automation of auto-mentoring processes in multi-player online environments such as virtual internships. We focus on automatically identifying speaker's intentions, i.e. the speech acts of chat utterances, in such virtual internships. Particularly, we explore several machine learning methods to categorize speech acts, with promising results. A novel approach based on pre-training a neural network on a large set of (and noisy) labeled data and then on expert-labeled data led to best results. The proposed methods can help understand patterns of conversations among players in virtual internships which in turn could inform refinements of the design of such learning environments and ultimately the development of virtual mentors that would be able to monitor and scaffold students' learning, i.e., the acquisition of specific professional skills in this case. [For the full proceedings, see ED593090.]