A language-feature model predicted recorded English essay scores more accurately than a simple baseline.
Its researchers caution that predicting scores is not equivalent to giving trustworthy writing advice.
What researchers studied
Researchers at the University of Nottingham Malaysia analysed 3,911 essays from the released ELLIPSE English-language-learner corpus training partition.
Six existing score variables covered cohesion, syntax, vocabulary, phraseology, grammar and conventions.
They extracted 77 interpretable linguistic features and compared models using repeated nested cross-validation.
The team assessed both score prediction and the more demanding task of identifying an individual writer’s weaker areas.
The study did not measure improvements in students’ writing.
What they found
Ridge regression achieved a mean columnwise root-mean-square error of 0.5271, a 19.3% reduction compared with a constant-mean baseline.
Performance was not detectably different from a TF-IDF representation under the reported comparisons.
Dimension-specific models performed better than a holistic model, but selecting one or two weak areas was not reliably better than a simple base-rate strategy.
Some feature coefficients also reversed direction when correlated predictors were considered, undermining straightforward interpretations.
What it means
Automated scoring may help educators summarise patterns across large collections of writing.
However, an apparently precise prediction should not automatically be presented as an explanation of what a student needs to practise.
A responsible feedback system would distinguish between reproducing existing ratings and establishing a learner’s underlying needs.
Human assessment, uncertainty information and validation against educational outcomes remain important, especially when results could affect academic decisions.
Important limitation
The researchers describe the feedback implications as exploratory rather than diagnostic.
The corpus’s existing ratings are themselves an imperfect target, and model accuracy against those ratings does not establish accuracy against actual writing ability.
The work used one corpus partition and cross-validation, not a prospective classroom trial.
Inter-rater agreement was modest and supporting validation evidence for the original scores was limited.
No causal learning benefit was demonstrated.
Source Information
Study: Interpretable linguistic features for multi-trait prediction of english essay scores, and what they do not license a feedback tool to claim
Authors: Muhammad Syukran Shabaruddin; Doreen Ying Ying Sim
Journal / Source: Frontiers in Education
Published: 9 November 2026
