The emergence of online collaboration platforms has dramatically changed the dynamics of human teamwork, creating a veritable army of virtual teams composed of workers in different physical locations. The global world requires a tremendous amount of collaborative problem solving, primarily virtual, making it an excellent domain for computer scientists and team cognition researchers who seek to understand the dynamics involved in collaborative tasks to
develop virtual agents that support collaboration. This dissertation presents several technical innovations in the usage of machine learning towards analyzing, monitoring, and predicting collaboration success from multiparty dialogue by successfully handling the problems of resource scarcity and natural distribution shifts. 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. 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 the features predictive of successful collaboration, we propose multi-feature embedding (MFeEmb) to improve the generalizability of collaborative task success prediction models under natural distribution shifts and resource scarcity. MFeEmb leverages the strengths of semantic, structural, and textual features of the dialogues by incorporating the most meaningful information from dialogue acts (DAs), sentiment polarities, and vocabulary of the dialogues. To further enhance the performance of MFeEmb under a resource-scarce scenario, we employ synthetic data generation and few-shot learning. Results show that the proposed multi-feature embedding is an excellent choice for the meta-pretraining stage of the few-shot learning. Our proposed data augmentation method showed significant performance improvement. Our research has potential ramifications for the development of conversational agents that facilitate teaming.
Major: Computer Science
Educational Career:
Bachelor's of Computer Science, BS, 2011, University of Balochistan
Master's of Computer Science, MS, 2016, Balochistan University of IT, Engineering, and Management Sciences.
Committee in Charge:
Gita Sukthankar, Chair, Computer Science
Fei Liu, University of Central Florida.
Ivan Garibay , University of Central Florida
Shawn Burke , University of Central Florida
Approved for distribution by Gita Sukthankar, Committee Chair, on April 5, 2023.
The public is welcome to attend.