An NLP-assisted and explainable rule-based platform for formative feedback from academic similarity reports

Abstract

Academic similarity reports identify textual overlap but do not, by themselves, determine plagiarism or explain how students should revise their writing. This study describes an implemented, explainable, teacher-mediated workflow and reports a preliminary feasibility-oriented assessment of teacher-perceived usefulness. The platform processed information from 200 real, deidentified undergraduate similarity-report cases from a single institution. Fifteen instructors reviewed editable feedback drafts through a web interface and selected Accept, Review, or Redo, mapped analytically to Correct, Partial, or Adjust. Case-level exploratory analyses examined associations between teacher ratings and recorded severity, operational issue category, source type, similarity percentage, and number of sources. Correct, Partial, and Adjust ratings were recorded for 65 (32.5%), 75 (37.5%), and 60 (30.0%) cases, respectively. Correct plus Partial represented 140 cases (70.0%) as an aggregate indicator of initial-draft usability, not system accuracy. Exploratory analyses found no strong association between teacher ratings and recorded severity, operational issue category, source type, similarity percentage, or number of sources. Because no persistent teacher identifier was retained, instructor clustering and inter-rater agreement could not be estimated. No students participated, and no learning, feedback-uptake, writing-improvement, fairness, or academic-integrity outcomes were measured. PDF extraction and experimental machine-learning components were not independently evaluated. The contribution is therefore a transparent workflow that converts similarity-report information into rule-based, teacher-reviewable feedback drafts rather than a validated plagiarism detector, predictive model, or autonomous feedback system.