A process‐oriented perspective on pre‐service teachers' self‐efficacy and their motivational messages: Using large language models to classify teachers' speech

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Background: Recent studies have examined the relation between teacher motivation, motivational messages and student learning but are limited to an achievement‐related context, primarily using survey data. Moreover, our understanding of the relation between various teacher characteristics, such as teacher self‐efficacy (TSE), and their motivational message use remains limited. Aims: Our study tested whether teacher speech can be classified into self‐determination (SDT)‐based motivational messages and reliably assessed with a large language model (LLM). Additionally, we analysed the relation between pre‐service TSE and their motivational message use. Sample: For our first aim, we used human‐rater annotations from 119 pre‐service teachers' classroom recordings. For our second aim, we used data from 103 pre‐service teachers (52.69% female; M age = 22.98, SD age = 3.26, Min age = 19, Max age = 34) who participated in a survey and were video‐recorded while teaching. Methods: First, we manually classified pre‐service teachers' motivational messages based on transcripts and used human‐rater annotations to fine‐tune an LLM. Second, we analysed the relation between pre‐service TSE and motivational message use. Results and Conclusions: The fine‐tuned LLM demonstrated promising performance in assessing SDT‐based motivational messages but needs further refining to assess thwarting messages. The analysis with human annotation showed that pre‐service TSE for classroom management positively affected the frequency of relatedness‐supportive messages. Pre‐service TSE for student engagement increased the likelihood of never using a competence‐ or relatedness‐thwarting message. Pre‐service TSE for instructional strategies reduced the frequency of autonomy‐supportive messages. LLM‐based analyses showed slightly different results but did not contradict human annotation‐based analyses.

This publication has 48 references indexed in Scilit: