Abstract: The objective of this work is to present a machine learning (ML) -based framework to identify evidence about collaborative problem solving (CPS) cognitive (teamwork) and social-emotional learning (SEL) skills from the dyadic (human-human-HH) interactions. This work extends our previous work (Chopade et al. IEEE HST 2018, LAK2019) , . Explicitly, we are interested in how teamwork skills and team dynamics are demonstrated as verbal and nonverbal behaviors, and how these behaviors can be captured and analyzed via passive data collection. For this work we use a two-player cooperative CPS game, Crisis in Space (CIS) from LRNG (Previously GlassLab Inc). During the summer of 2018, we implemented this CIS game for interns as a group study. A total of 34 participants played the game and provided study and survey data. During the study, we collected participants’ game play data, such as audio, video and eye tracking data streams. This research involves analyzing CIS multimodal game data, and developing skill models, and machine learning techniques for CPS skills measurement. In this paper, we present our ML framework for the analysis of audio data along with preliminary results from a pilot study. The analysis of audio data uses natural language processing (NLP) techniques, such as bag-of-words and sentence embedding. Our preliminary results show that various NLP features can be used to describe successful and unsuccessful CPS performances. The ML based framework supports the development of evidence centered design for teamwork skills-mapping and aims to help teams operate effectively in a complex situation. Potential applications of this work include support for the Department of Homeland Security (DHS), and the US Army for the development of learner and team centric training, cohort, and team behavioral skill-mapping.