Detecting Student Idea Attribution: A Computational Approach to Characterizing Equitable Teaching Practice

• In Event: Straight Talk: Natural Language Processing Approaches to Characterizing Student Opportunities to Learn in Mathematics

Abstract

Scholars interested in disrupting traditional race, gender, and socio-economic status-based achievement patterns in mathematics identified acknowledging student competence (also known as positioning students as competent and assigning competence) as a potentially important instructional practice that leads to more equitable student outcomes (Boaler & Staples, 2008; Johnson et al., 2022; Author et al., 2019). One measurable signal of this practice is uptake, when a teacher takes up a student contribution by repeating, revoicing, or reformulating it. Recent efforts to automate the measurement of uptake offer timely teacher feedback and may increase student satisfaction (Author et al., 2021; 2024a).

Building upon this research, we focus on an explicit instructional move: student idea attribution (SIA), when a teacher references a specific student and their mathematical idea, strategy, or practice. A teacher’s explicit attribution of students’ sociomathematical contributions may encourage their identities as thinkers and doers of mathematics (Staples, 2007; Author et al., 2024b), while signaling the value and legitimacy of diverse contributions to the collective learning environment (Author et al., 2019). However, there is no automated, scalable measure of SIA beyond resource-intensive manual annotation and classroom observation. We extend frontier work on automated instructional feedback by developing a computational measure of SIA. Our goals are to: (1) assess the prevalence and change in attribution practices across an academic year; (2) analyze the patterns of attribution, including which students are positioned as competent, how this group changes over multiple lessons, and what characterizes frequently or infrequently attributed students; (3) examine how SIA correlates with positive math outcomes.

We analyze transcript data from a study that collects classroom data from 300 grade 4-8 classrooms using five microphones and voice enrollment to achieve near-universal, traceable recordings of teacher and student speech. Using data collected in the 2024-25 school year, a team of experts in mathematics instruction and computational text analysis collaboratively developed a coding scheme for SIA and labeled more than 2,200 teacher–student exchanges from the classroom transcripts of 19 teachers across 12 schools in a large district in North Carolina. We train supervised natural language processing models, including a fine-tuned BERT-base (Devlin et al., 2019), to detect teacher utterances that explicitly attribute mathematical contributions to students.

Our work adds to the collection of empirically grounded instructional practices that “make potentially productive routines of action visible, and thus learnable by others” (Author et al., 2019, p. 368). By identifying patterns in how teachers position competence over time, and to whom, we can support more intentional and equitable instructional decision-making. Like prior work on automated uptake (Author et al., 2021; 2024a), this measure can generalize to new teacher utterances and be integrated into coaching conversations and reflection tools, providing timely and personalized feedback at scale, e.g. through high-leverage coaching that is individualized, focused, sustained, and context-specific (Kraft et al., 2018). Ultimately, we contribute to the field a novel automated measure of student idea attribution, offering practitioners and researchers an empirical understanding of—and opportunity to act upon—how often teachers position students as competent and for whom.

Authors

• Jackie Ramos Draper, HARVARD GRAD SCHOOL OF ED

• Jeannette Garcia Coppersmith, Harvard University

• Bailey Buchanan, Harvard University

• Heather C. Hill, Harvard University