Conversational Agents Improve Peer Learning through Building on Prior Knowledge

Introduction Conversational agents In the field of technology-enhanced learning, pedagogical agents have been developed to serve a wide variety of instructional roles, such as expert, motivator, or mentor (Baylor & Kim, 2005). Conversational agents are typically regarded as a subgroup of pedagogical agents involving learners in natural language interactions (Kerly, Ellis, & Bull, 2009). Research has shown using conversational agents to engage learners in one-to-one (student-agent) tutorial dialogues to improve students' comprehension and foster students' engagement and motivation (Veletsianos & Russell, 2014). During the past decade, researchers also focused on developing conversational agents for collaborative learning support (e.g., Kumar & Rose, 2011). Despite the established cognitive and social benefits of computer-supported collaborative learning (CSCL), collaborative knowledge construction is not a given but depends on the quality of interactions taking place among learners (Dillenbourg & Tchounikine, 2007; Kreijns, Kirschner, & Jochems, 2002). Under this prism, well-targeted supportive interventions can be used as a method to increase the probability of constructive peer interactions occurring by means of stimulating cognitive processes, such as conflict resolution, mutual regulation or explicit explanation (Tchounikine, Rummel, & McLaren, 2010). Evidence suggests that conversational agents with social interaction capabilities can enhance learning and idea generation productivity by providing dynamic support for learners working together (Kumar & Rose, 2011; Kumar, Beuth, & Rose, 2011). Chaudhuri et al. (2008) reveal that agents guiding peers through prescribed lines of reasoning on specific topics can improve learning performance. A study by Walker, Rummel, and Koedinger (2011) indicates that an agent displaying reflective prompts in a scripted peer tutoring activity can help students produce conceptually richer statements. Academically productive talk Another research direction has recently emerged focusing on an agile form of conversational agent support, which emphasizes the key role of social interaction in student engagement and learning (e.g., Adamson, Dyke, Jang, & Rose, 2014). This approach draws on the academically productive talk (APT) framework, itself originating from a substantial body of work on useful classroom discussion practices and norms (Michaels, O'Connor, & Resnick, 2008). According to APT, a peer dialogue in class should be accountable to the learning community, accurate knowledge and rigorous thinking, irrespective of the subject area (Sohmer, Michaels, O'Connor, & Resnick, 2009). In view of the above, peers should paraphrase and expand on each other's ideas (i.e., being accountable to the learning community), support the validity of their claims making explicit references to a pool of knowledge accessible to the group (i.e., being accountable to accurate knowledge), and logically connect their statements through rigorous argumentation (i.e., being accountable to rigorous thinking). Unlike other well-known discourse frameworks such as the IRE (Initiation, Response and Evaluation), the APT framework does not entail closing down a conversation after successfully eliciting a correct learner's response; instead, APT aims to promote and scaffold open-ended discussions where learners explicate their reasoning, compare their contributions with their partners' and construct logical arguments based on accurate evidence (Michaels, O'Connor, Hall, & Resnick 2010). Indeed, APT does not expect the teacher to maintain full control over learners' discussions, and prioritizes reasoning over correctness. The importance of the explicit articulation of reasoning is universally acknowledged by researchers, despite the different conceptualization of studies exploring the key features of a productive peer dialogue (for example, "transactivity," "group cognition," and "productive agency") (Stahl & Rose, 2011). …