Strategize Before Teaching: A Conversational Tutoring System with Pedagogy Self-Distillation

Findings of the Association for Computational Linguistics: EACL 2023, pages 2268–2274 May 2-6, 2023 ©2023 Association for Computational Linguistics Strategize Before Teaching: A Conversational Tutoring System with Pedagogy Self-Distillation Lingzhi Wang1,2, Mrinmaya Sachan3, Xingshan Zeng4, Kam-Fai Wong1,2 1The Chinese University of Hong Kong, Hong Kong, China 2MoE Key Laboratory of High Confidence Software Technologies, China 3Department of Computer Science, ETH Zurich 1,2{lzwang,kfwong}@se.cuhk.edu.hk 3msachan@ethz.ch, 4zxshamson@gmail.com Abstract Conversational tutoring systems (CTSs) aim to help students master educational material with natural language interaction in the form of a dialog. CTSs have become a key pillar in educational data mining research. A key challenge in CTSs is to engage the student in the conversation while exposing them to a di- verse set of teaching strategies, akin to a hu- man teacher, thereby, helping them learn in the process. Different from previous work that gen- erates responses given the strategies as input, we propose to jointly predict teaching strate- gies and generate tutor responses accordingly, which fits a more realistic application scenario. We benchmark several competitive models on three dialog tutoring datasets and propose a unified framework that combines teaching re- sponse generation and pedagogical strategy pre- diction, where a self-distillation mechanism is adopted to guide the teaching strategy learning and facilitate tutor response generation. Our experiments and analyses shed light on how teaching strategies affect dialog tutoring. 1 Introduction Decades of research effort (Carbonell, 1970; Richardson, 1988; Brown, 2009) has been put into building intelligent tutoring systems (ITSs). An important feature of these systems is the ability to customize the instructional activities and strategies based on the learner’s characteristics and needs (Kele¸s et al., 2009). Conversational tutoring sys- tems (CTSs) that aim to offer automated tutoring through natural language dialog is a key pillar of ITS research. Earlier work in CTSs was based on conventional techniques such as Bayesian tech- niques with rule engines (Jeon and Su, 2010; Wer- agama and Reye, 2014) and hybrid neural networks (Kose and Arslan, 2017; Stasaski et al., 2020). While various advanced neural approaches have been applied to open-domain (Sordoni et al., 2015; Serban et al., 2016; Xing et al., 2017) and task- Teaching Strategy Tutor Response Restating Let me say back what I heard. Pressing for accuracy Can you tell us the steps you used to find the answer? (a) Two examples of teaching strategy and tutor response Tutor: Ok, now we have 'get cut off', 'put someone through' and 'get through’ scaffolding Tutor: I was talking to her, but suddenly we ___ ___ ___ (I couldn‘t hear her anymore) Can you choose one? eliciting Student: got cut off - Tutor: yes, good! Well done today! Have a lovely day! closing (b) An example of interactions between tutor and student Figure 1: Examples of teaching strategy and interac- tions between tutor and student. Teaching strategies in Figure 1(b) are in red. oriented dialogue systems (Zhao et al., 2017; Lei et al., 2018; Peng et al., 2020), conversational tu- toring systems have not benefited from the devel- opment …

Content truncated to 500 words