Auto-marking for multiple-choice questions is trivially simple — a regex match against the answer key. Auto-marking for subjective responses is a different matter entirely, and the technology has matured significantly in the last three years.
Keyword and concept matching
The simplest form of subjective auto-marking checks whether the student's answer contains a set of required keywords or concepts defined by the examiner. This works well for factual recall questions but breaks down for higher-order thinking questions where the correct answer can be expressed in many different ways.
NLP-based semantic matching
Modern auto-marking systems use large language model embeddings to compare the semantic meaning of a student's answer against a model answer, not just the exact words. A student who writes "the mitochondria generates ATP through oxidative phosphorylation" and another who writes "mitochondria make energy using oxygen" will both score well on a well-calibrated semantic matcher.
Where human review is still essential
For marks worth more than 5 per question, for answers that require evaluating argument quality (as in law or philosophy exams), or for any case where a student disputes an AI mark, human review is non-negotiable. Auto-marking should be positioned as a first-pass efficiency tool, not a replacement for examiner judgement on high-stakes subjective responses.
Calibration
The quality of NLP-based auto-marking is only as good as the calibration. ExamRankers requires at least 30 human-marked sample answers per question before the auto-marking model is deployed. The system then shows its confidence score alongside every mark it assigns — low-confidence marks are automatically routed to a human examiner.