Unified Machine Learning Approaches for Scoring Paper and Online K-12 Assessments: Bridging Traditional and Digital Testing with Intelligent Scoring Systems

Abstract

The modern education systems in the K-12 level are becoming more dependent on artificial intelligence-based assessment systems to resolve the long-standing issues of scoring efficiency, consistency, and equity between the traditional paper-based and modern digital tests. The merging of Optical Mark Recognition, Optical Character Recognition, computer vision, Natural Language Processing, and multimodal analytical systems facilitates the development of single pipelines that support the collection of various types of responses, such as multiple choices, handwritten text, constructed diagrams, essays, speech recordings, and video submissions. These parallel processing streams converge to a centralized analytics infrastructure that harmonizes heterogeneous data formats, allows tracking performance longitudinally, and enforces a continuous bias detection across demographic subgroups. Convolutional neural networks are capable of automated learning of features to classify bubbles and evaluate diagrams, whereas Transformer-based models can comprehend written responses in a context to score on par with humans. Graph Neural Networks process spatial relationships in visual constructs, and multimodal fusion models combine acoustic and visual signals to holistically assess oral delivery. The unified analytics layer integrates the scoring results across all modalities into stakeholder-specific dashboards that provide finer-grained information to teachers, school leaders, and policymakers, and is transparent with the ability to explain scoring decisions (through explainability facilities) and rubric alignment. To accomplish successful deployment, sociotechnical issues will have to be addressed, such as infrastructure differences, data privacy concerns, reduction of algorithmic bias with a variety of training data and fairness auditing, and creation of trust mechanisms with transparent documentation of model structures and validation evidence. Integration of machine learning in educational assessment is a paradigm shift to more efficient, fairer, and analytically advanced systems of evaluating students through a means that facilitates data-based instructional decision-making without compromising on adequate human judgments that can be made on consequential educational decisions about student movement and placement.