Is it a good move? Mining effective tutoring strategies from human–human tutorial dialogues
Introduction
Dialogue-based Intelligent Tutoring Systems (ITS), similar to conventional ITS like SQL-Tutor [1], Algebra Tutor PAT [2], and eTeacher [3], aim at helping students construct knowledge and skills of different subjects by providing them with immediate and personalized instructions or feedback. Compared to conventional ITS, dialogue-based ITS deliver instructions or feedback by having natural and meaningful conversations with students [4], and are expected to act as competent as human tutors to engage students and provoke more in-depth thinking and learning. Given the promising potentials, both academic researchers and industrial practitioners have put great efforts in building various dialogue-based ITS, among which CIRCSIM-Tutor [5], AutoTutor [6], BEETLE II [7], and Why2 [8] are notable representatives. Noticeably, these systems have been deployed for use in practice and have assisted millions of students with their learning.
Despite being popular, most of the existing dialogue-based ITS are plagued by their inability in delivering personalized learning experiences to students [9]. The current dialogue-based ITS, as yet, fail to achieve their full potential and are unable to act as competently as human tutors [10]. One main reason is that these dialogue-based ITS, more often than not, lack sufficient pedagogical expertise as human tutors in guiding students [11], [12]. That is, these dialogue-based ITS typically have little knowledge about the tutoring strategies that can be of use to facilitate the tutoring process [13]. For instance, questioning a student’s progress of learning problems is a common strategy used to help tutors detect knowledge gaps of the student at the beginning of a tutorial session [14]. Then, in follow-up teaching activities, tutors can better direct their efforts, e.g., introducing relevant learning contents and designing appropriate teaching activities to enable students to develop mastery of those concepts. It should also be noted that successful applications of such a tutoring strategy often depends on (i) a tutor’s experience (and domain/contextual knowledge) in applying the strategy (e.g., when to ask questions and what type of question should be asked) and (ii) information about students (e.g., mastery level and learning progress) [15], [16], [17], [18].
Numerous studies have been conducted to investigate how dialogue-based ITS can be equipped with relevant pedagogical expertise to apply appropriate tutoring strategies [4], [11], [17], [19], [20], [21], [22]. In this strand of research, a recent trend is to mine large-scale data collected by existing dialogue-based ITS or generated between human tutors and students to discover effective tutoring strategies [11], [12], [19], [23]. However, existing data-intensive studies typically focused on the analysis of successful tutorial sessions (i.e., those in which students successfully solved problems or achieved meaningful learning) and the identification of effective tutoring strategies that tutors should take. We argue that, to provide students with necessary help, tutors should also learn from unsuccessful tutorial sessions and gain a better understanding of the factors contributing to such failures. Therefore, unsuccessful tutorial sessions should also be analysed to better guide the design and development of future dialogue-based ITS.
This study aimed to identify the frequent tutoring strategies used by tutors in not only successful but also unsuccessful tutorial sessions by mining a large-scale human–human tutorial dialogue dataset. The study also aimed to examine the extent to which these identified tutoring strategies are predictive of students’ problem-solving performance. Here, we described a tutoring strategy as the actions taken by a tutor in the tutorial process (e.g., asking thought-provoking questions and providing hints). Formally, our work was guided by three Research Questions:
- RQ1
What actions are commonly taken by tutors and students during tutorial sessions?
- RQ2
What patterns of actions, i.e., one or multiple consecutive actions, are associated with different levels of students’ performance in solving problems in tutorial sessions?
- RQ3
To what extent are the identified actions and action patterns predictive the problem-solving performance of students in tutorial sessions?
To answer the above questions, we first employed a widely-used dialogue act (DA) scheme (proposed by [24]) to characterize tutors’ (as well as students’) actions behind their utterances in a tutorial dialogue. Then, the derived actions were analysed by applying a sequence analysis to shed light on the frequent tutoring strategies employed by tutors, which were further used as input for a well-established machine learning method—Gradient Tree Boosting (GTB) [25]—to measure the contribution made by these strategies in predicting students’ problem-solving performance. To our knowledge, our study is the first to take students’ prior progress into account to reveal effective tutorial strategies in human–human online tutoring. By analysing a large corpus consisting of both successful and unsuccessful tutorial dialogues, our study contributed with an in-depth understanding of tutors’ as well as students’ behaviour in human–human online tutoring and offered empirical evidence to support existing good practices (e.g., providing timely feedback to students) for the development of dialogue-based ITS.
Section snippets
Related work
Tutoring strategies refer to principles and approaches employed by instructors to better assist students to learn in various educational settings [15], [26], e.g., raising a question to trigger in-depth thinking and acknowledging students’ achievement to motivate them to continue to learn. Effective tutoring strategies play essential roles in helping instructors better direct their teaching efforts and enabling students to construct meaningful knowledge, and thus have been investigated for
Dataset
With the ethics approval from Monash University for secondary data use (Project ID 26156), we used a deidentified tutorial dataset that was prepared by an educational technology company. The educational technology company provides a mobile phone application for tutors and students to work together to solve problems covering subjects like mathematics, chemistry, and physics. With the mobile application, a student could take a picture of an unsolved problem and initialize a request for help.
Results
With the method described in 3.3, we built a DA classifier which successfully assigned correct labels for 75% of the sentences in the labelled dataset. More specifically, the classifier achieved a performance of 0.742 and 0.828 in terms of F1-score and AUC, respectively. In particular, the classifier achieved a Cohen’s of 0.735, which demonstrated a sufficient performance level, especially given the large number of DA contained in our dataset (i.e., 31 second-level tags). This pre-trained DA
Discussion and conclusion
The construction of dialogue-based ITS with adequate pedagogic expertise is a longstanding task in the pathway towards delivering on-time, personalized, and meaningful learning experiences to students. Though quite some studies have been carried out, these studies often ignored the analysis of unsuccessful tutorial sessions and seldom paid attention to the reasons behind these unsuccessful tutorial sessions. This motivated us to analyse a large-scale dialogue corpus (over 14 K), which consisted
CRediT authorship contribution statement
Jionghao Lin: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Shaveen Singh: Conceptualization, Methodology, Validation, Formal analysis, Data curation. Lele Sha: Software, Validation, Data curation. Wei Tan: Software, Validation. David Lang: Resources. Dragan Gašević: Conceptualization, Validation, Supervision, Writing – original, Writing – review & editing, Project
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Jionghao Lin is a Ph.D. student in the Centre for Learning Analytics at Monash University, Melbourne, Australia. His primary research interests focus on the areas of learning analytics, natural language processing, and affective computing. Currently, Jionghao is mainly working on applying artificial intelligent technologies to understand and optimize the learning environment. He received his B.E. degree from Jianghan University, China, and Master degree in Data Science from Monash University,
References (73)
- et al.
eTeacher: Providing personalized assistance to e-learning students
Comput. Educ.
(2008) - et al.
Developing a coding scheme for analysing classroom dialogue across educational contexts
Learn. Cult. Soc. Interact.
(2016) - et al.
Learning from human tutoring
Cogn. Sci.
(2001) - et al.
Quality of educational dialogue and association with students’ academic performance
Learn. Instr.
(2018) - et al.
Study on student performance estimation, student progress analysis, and student potential prediction based on data mining
Comput. Educ.
(2018) - et al.
Integrating assessment, learning and instruction: Assessment of domain-specific and domaintranscending prior knowledge and progress
Studi. Educ. Eval.
(1996) - et al.
Exploring feedback and student characteristics relevant for personalizing feedback strategies
Comput. Educ.
(2014) Learning SQL with a computerized tutor
- et al.
Authoring content in the PAT algebra tutor
J. Interact. Media Educ.
(1998) - et al.
Recent advances in conversational intelligent tutoring systems
AI Mag.
(2013)
One-on-One Tutoring by Humans and Computers
AutoTutor: an intelligent tutoring system with mixed-initiative dialogue
IEEE Trans. Educ.
BEETLE II: Deep natural language understanding and automatic feedback generation for intelligent tutoring in basic electricity and electronics
IJAIED
When are tutorial dialogues more effective than reading?
Cogn. Sci.
Intelligent tutoring systems survey for the period 2000- 2018
Int. J. Acad. Eng. Res.
Intelligent tutoring systems: A comprehensive historical survey with recent developments
Int. J. Comput. Appl.
A tutorial Markov analysis of effective human tutorial sessions
Intelligent tutoring systems with conversational dialogue
AI Mag.
Classroom questioning
School Improv. Res. Ser.
Modelling human teaching tactics and strategies for tutoring systems: 14 Years on
IJAIED
SMART: Student modeling approach for responsive tutoring
User Model. User-Adapt. Inter.
A tutoring system that simulates the highly interactive nature of human tutoring.
J. Educ. Psychol.
A systematic literature review of intelligent tutoring systems with dialogue in natural language
IEEE Access
The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems
Educ. Psychol.
Designing engaging intelligent tutoring systems in an age of cognitive computing
IBM J. Res. Dev.
A study on two hint-level policies in conversational intelligent tutoring systems
Investigating the relationship between dialogue structure and tutoring effectiveness: a hidden Markov modeling approach
Int. J. Artif. Intell. Educ.
XGBoost: 2D-object recognition using shape descriptors and extreme gradient boosting classifier
Strategies for Effective Teaching
Teaching tactics in AutoTutor
Modell. Hum. Teach. Tact. Strateg. IJAIED
The Architecture of Cognition, Vol. 5
Cognitive tutors: Lessons learned
J. Learn. Sci.
Macroadapting animalwatch to gender and cognitive differences with respect to hint interactivity and symbolism
Cited by (0)
Jionghao Lin is a Ph.D. student in the Centre for Learning Analytics at Monash University, Melbourne, Australia. His primary research interests focus on the areas of learning analytics, natural language processing, and affective computing. Currently, Jionghao is mainly working on applying artificial intelligent technologies to understand and optimize the learning environment. He received his B.E. degree from Jianghan University, China, and Master degree in Data Science from Monash University, Australia.
Shaveen Singh is a Research Fellow at the Centre of Learning Analytics at Monash University. His research interests include the design and deployment of technology to increase the understanding and improve digital learning experiences. More specifically, his work examines the areas of learning analytics, personalized active learning, and building tools for teacher support. Shaveen is currently pursuing his Ph.D. at Monash University.
Lele Sha is a second-year Ph.D. student in the Centre for Learning Analytics at Monash University. His main research interest centres on applying Machine Learning and Natural Language Processing techniques to automatically processing educational forum posts. Specifically, he is focusing on improving model performance by applying extensive feature engineering and sentence embeddings. Before starting his Ph.D., Lele also worked in several software-as-a-service projects on learning management systems, which were successfully deployed to production and currently offering hundreds of online courses on its interactive training platform for Australian students.
Wei Tan is a Doctoral Researcher who studies the cutting-edge machine learning algorithm in Data Science. He specializes in Active Learning that optimize the labelling budget and time for the human annotator. His Ph.D. project is funded by Google Turning point. The aim is to develop the Surveillance System that will enable capture of a more complete set of coded ambulance data relating to SITB, mental health, and AOD attendances to inform policy, practice and intervention. He holds a master’s degree from Monash University, and has expertise in analytics design for the social media platform.
David Lang is a doctoral student in the Economics of Education program and an IES Fellow. He graduated from UCLA in 2008 with a B.A. in Economics & a B.S. in Actuarial Mathematics. Prior to his doctoral studies, David worked for five years as a research analyst at the Federal Reserve Bank of San Francisco. His research interests include higher education, online education, and quantitative methods in education research. At Stanford, David also obtained a master’s degree in Management Science and Engineering.
Dragan Gašević is Distinguished Professor of Learning Analytics in the Faculty of Information Technology and Director of the Centre for Learning Analytics at Monash University. As the past president (2015–2017) and a co-founder of the Society for Learning Analytics Research, he had the pleasure to serve as a founding program chair of the International Conference on Learning Analytics and Knowledge (LAK) and a founding editor of the Journal of Learning Analytics. His research centres on self-regulated and social learning, higher education policy, and data mining. He is a frequent keynote speaker and a (co-)author of numerous research papers and books.
Dr. Guanliang Chen is serving as a Lecturer in the Faculty of Information Technology, Monash University in Melbourne, Australia. Before joining Monash University, Guanliang obtained his Ph.D. degree at the Delft University of Technology in the Netherlands, where he focused on the research on large-scale learning analytics with a particular focus on the setting of Massive Open Online Courses. Currently, Guanliang is mainly working on applying novel language technologies to build intelligent educational applications. His research works have been published in international journals and conferences including AIED, EDM, LAK, L@S, EC-TEL, ICWSM, UMAP, Web Science, Computers & Education, and IEEE Transactions on Learning Technologies. Besides, he co-organized two international workshops and has been invited to serve as the program committee member for international conferences such as LAK, FAT, ICWL, etc.