TY - GEN

T1 - Towards Modeling and Evaluating Instructional Explanations in Teacher-Student Dialogues

AU - Feldhus, Nils

AU - Anagnostopoulou, Aliki

AU - Wang, Qianli

AU - Alshomary, Milad

AU - Wachsmuth, Henning

AU - Sonntag, Daniel

AU - Möller, Sebastian

PY - 2024/9/4

Y1 - 2024/9/4

N2 - For dialogues in which teachers explain difficult concepts to students, didactics research often debates which teaching strategies lead to the best learning outcome. In this paper, we test if LLMs can reliably annotate such explanation dialogues, s.t. they could assist in lesson planning and tutoring systems. We first create a new annotation scheme of teaching acts aligned with contemporary teaching models and re-annotate a dataset of conversational explanations about communicating scientific understanding in teacher-student settings on five levels of the explainee’s expertise: ReWIRED contains three layers of acts (Teaching, Explanation, Dialogue) with increased granularity (span-level). We then evaluate language models on the labeling of such acts and find that the broad range and structure of the proposed labels is hard to model for LLMs such as GPT-3.5/-4 via prompting, but a fine-tuned BERT can perform both act classification and span labeling well. Finally, we operationalize a series of quality metrics for instructional explanations in the form of a test suite, finding that they match the five expertise levels well.1

AB - For dialogues in which teachers explain difficult concepts to students, didactics research often debates which teaching strategies lead to the best learning outcome. In this paper, we test if LLMs can reliably annotate such explanation dialogues, s.t. they could assist in lesson planning and tutoring systems. We first create a new annotation scheme of teaching acts aligned with contemporary teaching models and re-annotate a dataset of conversational explanations about communicating scientific understanding in teacher-student settings on five levels of the explainee’s expertise: ReWIRED contains three layers of acts (Teaching, Explanation, Dialogue) with increased granularity (span-level). We then evaluate language models on the labeling of such acts and find that the broad range and structure of the proposed labels is hard to model for LLMs such as GPT-3.5/-4 via prompting, but a fine-tuned BERT can perform both act classification and span labeling well. Finally, we operationalize a series of quality metrics for instructional explanations in the form of a test suite, finding that they match the five expertise levels well.1

KW - Dialogue

KW - Discourse Analysis

KW - Evaluation

KW - Explanations

U2 - 10.1145/3677525.3678665

DO - 10.1145/3677525.3678665

M3 - Conference contribution

SN - 9798400710940

SP - 225

EP - 230

BT - Proceedings of the 2024 International Conference on Information Technology for Social Good

CY - New York, NY, USA

ER -