Work

Scaffolded large language model feedback builds durable translation competence whereas unrestricted use fosters dependency

Year: 2026
Type: preprint
Abstract: Large language models (LLMs) are rapidly entering translator education, yet it remains unclear whether they build students' own translation competence or merely improve the products students submit. We combined a validation study on public benchmark data with a 16-week quasi-experiment to test a scaffolded human--LLM pedagogy, the Draft--Diagnose--Revise--Reflect (D2R) cycle, for English--Chinese translation. In Study 1, a rubric-guided error-annotation prompt increased the agreement between LLM feedback and expert Multidimensional Quality Metrics judgements on 2,500 WMT segments for all four models tested; with this prompt GPT-4o reached a segment-level Kendall's $\tau$ of 0.30, on a par with the strongest neural metric, and the highest error-detection accuracy. In Study 2, 179 English majors received conventional instruction, unrestricted LLM access or the guardrailed D2R cycle. Unrestricted access maximised LLM-assisted performance but yielded no advantage in unassisted translation over conventional instruction at a delayed test 12 weeks after the course (Hedges' $g$ = $-$0.08) and produced the largest AI-dependency gap. Scaffolded use produced durable gains in unassisted translation ($g$ = 0.70), a 13\% reduction in error rate and higher metacognitive regulation. Process logs indicated that the benefit was transmitted mainly through lower verbatim adoption of LLM output. Pedagogical design, not LLM access per se, determines whether LLMs support or supplant translation learning.
Language: English
Open Access status: green
OpenAlex ID: W7233557264
Cites:
Cited by:
Related to: