Research Mentor(s): Lorenzo Garcia-Amaya, Assistant Professor
Research Mentor School/College/Department: Romance Languages and Literatures, College of Literature, Science, and the Arts
Presentation Date: Thursday, April 22, 2021
Session: Session 2 (11am – 11:50am)
Breakout Room: Room 20
One of the key elements underlying successful second language (L2) learning is the interaction with other language users in the target language (Gass & Varonis, 1985). While the expectation exists that learners will use their linguistic skills in theStudy Abroad (SA) context (e.g., Fraser 2002; Freed, 1990; 2000; Hernández, 2010; Martinsen, 2011; Mendelson 2004; Vande Berg et al., 2009, Whitworth, 2006), it is also known that many SA learners interact substantially less with native speakers (NSs) than they initially anticipate (DeKeyser, 1986; Dewey et al., 2014; Rivers, 1998; Wilkinson, 1998a, 1998b; García-Amaya, 2017). Recent research further shows that SA learners’ target-language use decreases over a six-week SA experience (García-Amaya, 2017; forthcoming). While interaction in the L2 is essential for achieving L2 gains, research has found that learners must negotiate meaning to notice gaps in their acquisition and benefit from their interlocutor(s)’ efforts to adjust interaction and facilitate comprehension (Long 1981, 1983, 1986). One could thus argue that deeper conversations, involving such negotiating of meaning, may be necessary for facilitating L2 learning. Along with this hypothesis, several questions come to mind. How can we distinguish deep from shallow interactions? How can we understand interaction depth without necessarily observing learners’ interactions in real time? In this presentation, we will discuss modeling depth of interaction through neural networks (cf. Frank, Monaghan, Tsoukala, 2019). Instead of asking participants to self-rate the depth of their interactions via online questionnaires, we have designed a predictive model that takes as input several structural parameters of interaction, including length, interactiveness, activeness, and number of participants involved. In terms of precision and recall, our model’s preliminary results are 76% and 82%, respectively. These results will be discussed in light of our modest dataset and future steps.