This study presents several technical innovations on the usage of machine learning towards analyzing, monitoring, and predicting collaboration success from multiparty dialogue for the development of a virtual agent to support effective teamwork. First, we examine the problem of predicting team performance from embeddings learned from multiparty dialogues such that teams with similar conflict scores lie close to one another in vector space. We extract the embeddings from three types of features: 1) dialogue acts, 2) sentiment polarity 3) syntactic entrainment. Although all of these features can be used to predict team performance effectively, their utility varies by the teamwork phase. We separate the dialogues of players playing a cooperative game into stages: 1) early (knowledge building), 2) middle (problem-solving), and 3) late (culmination). Unlike syntactic entrainment, both dialogue act and sentiment embeddings effectively classify team performance, even during the initial phase.Second, we address the problem of learning generalizable models of collaboration. Machine learning models often suffer domain shifts; one advantage of encoding the semantic features is their adaptability across multiple domains. We evaluate the generalizability of different embeddings to other goal-oriented teamwork dialogues. Finally, in addition to identifying features predictive of successful collaboration, we aim to improve the generalizability of collaborative task success prediction models under natural distribution shifts and resource scarcity using a combination of domain invariant and domain specific features. Research is conducted on several multiparty dialogue datasets, including Teams, ASIST, and GitHub.