Language Technologies Thesis Proposal
- Gates Hillman Centers
- Traffic21 Classroom 6501
- SREECHARAN SANKARANARAYANAN
- Ph.D. Student
- Language Technologies Institute
- Carnegie Mellon University
Collaborative Conversational Support Across Contexts
Collaboration is a fundamentally human activity spanning almost every sphere of interaction including the school, the workplace, the home, and the social spheres. In order to support collaboration in the contexts associated with these spheres, it becomes important to understand what effective collaboration means in them, and what idiosyncrasies could come in the way of offering this support. We start with a construct called transactivity, the notion of original reasoning expressed in an utterance operating on the reasoning expressed in another. Transactivity is known to be associated with improved learning, knowledge integration, knowledge transfer, and effective collaborative problem-solving. However, prior work providing conversational support for transactive exchange in a classroom context provides a clue that factors unique to that context such as difficulty of the materials, and the skill level of the learner interacts with these positive effects. Related concepts in the workplace and organizational contexts also produce positive effects but point to context-specific factors interacting with them. In this thesis, we stress-test transactivity across more strikingly different contexts in the four different spheres of human collaboration identified above. In doing so, we expect to identify context-specific factors that might prevent its benefits from being realized, and infer a more nuanced set of best practices for generalizing the insights from one context to a broader set of contexts.
First, in an online classroom context, we use a transactivity-based team formation strategy to provide team recommendations to students. We find that student preferences encourage them to undermine these recommendations undercutting the ability of transactivity to have a positive impact. Second, we investigate an instructional scaffold for collaboration programming called Online Mob Programming (OMP), that involves learners being supported by a conversational agent in taking up interdependent roles. We see that the instructional materials not being well adapted to learner prior knowledge potentially interferes with the benefits of transactivity. Further, some functions of the scaffold mirroring the expected benefits from transactivity could wash out benefits from it.
In proposed work, we intend to use the data collected in the online classroom, to learn reinforcement learning-informed dialogue strategies for better agent-support, and for generating code explanations that can provide additional shared context in support of knowledge co-construction. Further, we propose to extend this work to the community college context, allowing us to test the generality of injecting workplace-relevant collaborative experiences into university learning environments in another context where employment is the salient goal. Finally, in the home sphere, we propose the use of the conversational agent to engage patients and their caregivers in discussions about living healthfully together post a life-altering medical procedure and in the social sphere, we propose to use the conversational agent to support transactive exchange between individuals with opposing political inclinations in order to achieve the related notion of civil political discourse.
Carolyn Penstein Rosé, (chair)
Alan W. Black
Stian Håklev (École polytechnique fédérale de Lausanne)
Pierre Dillenbourg (École polytechnique fédérale de Lausanne)