Multi-Model Ensemble with Rubric Signals And Concept Coverage Features For Automatic Short Answer Grading

Abstract

Automatic Short Answer Grading (ASAG) plays an important part in scalable and consistent digital tests, even though conventional methods are probably unlikely to provide a tradeoff between semantic knowledge, conceptual richness and trustworthiness of the grader. Lexical overlap techniques have no means of detecting paraphrasing and the embedding techniques do not explicitly impose concepts. Standalone Large Language Model (LLM) grading enhances the reasoning assessment and can create inconsistency and calibration.
The article suggests CRANE-ASAG(Concept-aware Rubric-guided AI Neural Ensemble) a hybrid multi-model ensemble approach, that entails weighted concept coverage, rubric-based LLM scoring, and supervised meta-learning as a hybrid multi-model ensemble approach. The formulation of a grading problem is a bounded regression problem which combines symbolic, neural and rubric based signals to a common architecture. Experimental performance Benchmark testing demonstrates that it is more compatible with human scorers, has higher Quadratic Weighted Kappa, and has lower cuts of prediction error than the conventional baselines. This shall be put forward as a solution that is interpretable, robust and deployment of low resource ASAG system solution.