Abstract
The second version of the Teacher-Student Chatroom Corpus (TSCC) significantly expands upon its predecessor, offering a rich dataset for analyzing one-to-one English language tutoring. TSCC v2 includes 260 chatroom lessons, totaling 41.4K conversational turns and 363K word tokens across 13 students of varying proficiency levels (B1–C2). Beyond data expansion, this release introduces pedagogical annotations using the Self-Evaluation of Teacher Talk (SETT) framework and explores the use of transformer-based models to automatically detect discourse sequence shifts. The corpus is freely available for research, providing a foundation for developing pedagogically aware AI tutors and enhancing teacher training.
Introduction: The Challenge of One-to-One Tutoring Data
While massive datasets exist for generic web-based chat, high-quality data specifically capturing the nuances of language tutoring remains scarce. In a pedagogical setting, dialogue is structured into "sequences"—functional blocks moving from openings and warm-ups to grammar exercises or "repair" sequences following student errors. Understanding these transitions is crucial for helping teachers reflect on their interactional competence and for building AI dialogue managers that can steer lessons naturally rather than just responding to prompts in isolation.
The TSCC v2 Dataset
The primary contribution of this work is the significant expansion of the TSCC. Version 2 increases the number of lessons from 102 to 260, involving two teachers and 13 students with seven different first languages (Italian, Japanese, Mandarin, Russian, Spanish, Thai, and Ukrainian).
The corpus now covers a broader range of learner proficiencies, from intermediate (B1) to proficient (C2) on the CEFR scale. This volume and diversity make it a robust resource for studying how teachers manage one-to-one interactions in a digital environment.
Pedagogical Annotation and Sequence Detection
To enrich the data, the researchers applied the Self-Evaluation of Teacher Talk (SETT) framework. This categorizes turns into "modes" (e.g., Managerial, Materials/Skills) and "interactures" (e.g., Scaffolding, Direct Repair), providing a granular view of how a lesson is actually constructed turn-by-turn.
A key finding from this annotation is the teacher's role as the primary architect of the lesson. While conversational turns are split nearly 50/50, teachers set three times as many "modes" as students, focusing heavily on managerial tasks.
The researchers also framed lesson management as a machine learning task: New Sequence Detection. Using DistilBERT, they experimented with predicting whether the next turn would initiate a new discourse sequence. They found that providing the model with conversational context (the previous turn) significantly improved performance, suggesting that topic shifts are often foreshadowed in the preceding exchange.
Key Results and Future Directions
The experiments demonstrated that detecting sequence shifts is a solvable task. The best-performing configuration—DistilBERT with one turn of context—achieved an F1-score of 0.489, significantly outperforming probabilistic baselines. Notably, the model achieved a high recall (0.636), which is particularly promising for real-time applications where missing a topic shift is more disruptive than a false alarm.
Future work will address the subjectivity of sequence boundaries and move toward multi-class classification to predict the specific type of upcoming sequence (e.g., "Grammar Repair" vs. "Homework Setting"). The ultimate goal is to integrate these models into real-time feedback tools for teachers and more sophisticated, pedagogically aware language learning chatbots.
Summary updated