Traditional Automated Essay Scoring (AES) and Short Answer Grading (SAG) systems typically operate as black boxes. They ingest a student's response and output a single numerical score. However, as noted by Kentaro Inui et al. in Frontiers in Explainable Automated Writing Evaluation, educational feedback is inherently a communicative act; a raw number fails to provide the qualitative reasoning necessary for formative improvement. Without transparency, students cannot identify their specific mistakes, and educators cannot verify whether a system is prioritizing superficial patterns—such as text length or keyword density—over actual subject matter mastery (Jiazheng Li et al., Distilling ChatGPT for Explainable Automated Student Answer Assessment).
To bridge this gap, modern research focuses on explainable automated assessment. By forcing models to ground their evaluations in human-readable rubrics, intermediate traits, or natural language rationales, these systems transform opaque grading algorithms into transparent, actionable tools for learning.
Decomposing the Grade via Rubric Alignment
To build trust and utility, recent frameworks shift the paradigm from direct score prediction to explicit rubric alignment. Instead of training a model to map an entire student response directly to a holistic score, systems are designed to evaluate how well a submission satisfies discrete grading criteria.
For instance, Naoya Inoue et al. introduced a rubric-aware mechanism in Inject Rubrics into Short Answer Grading System that uses word-level attention to explicitly link student answers to required "key elements" defined by teachers. Taking this a step further, Shashank Sonkar et al. formulated grading as a logic verification problem in Automated Long Answer Grading with RiceChem Dataset. Their "rubric entailment" framework verifies whether a student's text logically entails each individual rubric item. This creates a clear, verifiable audit trail: a student can see exactly which components of a multi-faceted answer satisfied the educational requirements and which fell short.
Generating Rationales via LLM Distillation
With the advent of Large Language Models (LLMs), researchers can now generate natural language justifications alongside scores. However, deploying massive LLMs for every student submission is computationally expensive and slow. To solve this, researchers leverage knowledge distillation to transfer an LLM's explanatory capabilities into smaller, production-ready models.
In the Rationale-based Multi-Trait Scoring (RMTS) framework proposed by SeongYeub Chu et al. in Rationale Behind Essay Scores: Enhancing S-LLM's Multi-Trait Essay Scoring with Rationale Generated by LLMs, a powerful LLM generates rubric-aligned qualitative rationales across multiple traits (such as organization, voice, and content). These rationales are then used to train smaller, highly efficient student models.
Similarly, Jiazheng Li et al. demonstrated that distilling ChatGPT's grading rationales into a smaller Long T5 model not only provides natural language explanations but actually improves grading accuracy over the teacher model. To ensure these rationales are pedagogically sound, Clayton Cohn et al. in A Chain-of-Thought Prompting Approach with LLMs for Evaluating Students' Formative Assessment Responses in Science combined Chain-of-Thought (CoT) prompting with active learning, forcing the model to output its step-by-step deductive process before arriving at a final grade.
Why Explainable Assessment Works
The success of explainable automated assessment relies on a fundamental machine learning principle: constraining the hypothesis space through intermediate representations.
Traditional Black-Box Scoring:
Student Text ───────────────────────────────────────────► Numerical Score (Opaque)
Explainable Rubric/Trait Scoring:
Student Text ──► Intermediate Traits / Rubric Entailment ──► Numerical Score + Rationale (Verifiable)
When a model maps a complex student essay directly to a single scalar score, it faces an unconstrained optimization problem. It can easily overfit to spurious correlations, like vocabulary diversity or sentence length, without understanding the underlying logic.
By forcing the model to first predict intermediate traits (Rahul Kumar et al., Many Hands Make Light Work: Using Essay Traits to Automatically Score Essays) or verify rubric entailment, the system is forced to ground its final score in structured, human-readable evidence. This multi-task or pipeline structure acts as a regularizer, ensuring that the final score is a logical consequence of identified textual features rather than an arbitrary black-box output.
Extending the Paradigm Beyond Text
This drive for explainable, component-based feedback extends beyond written text into multimodal educational domains, such as sign language learning. Traditional systems provide only binary feedback (correct/incorrect), leaving learners blind to their specific physical mistakes.
To address this, Prajwal Paudyal et al. developed Learn2Sign: Explainable AI for Sign Language Learning, which breaks down signs into their phonetic components—such as hand-shape, movement, and location—to tell students exactly why a sign was incorrect. Similarly, Dylan Pallickara et al. in A Wireframe-Based Approach for Classifying and Acquiring Proficiency in the American Sign Language (Student Abstract) distilled raw video data into minimalist joint-angle wireframes. By evaluating these discrete geometric features rather than raw pixels, the system provides granular, real-time feedback on postural mistakes while eliminating demographic and background biases.
Key Foundational Work
A Reliable Approach to Automatic Assessment of Short Answer Free Responses (2002) — contributed 2 concept elements to this direction.
An Expert System for Understanding Expressions from Electric Circuit Analysis (1985) — contributed 1 concept elements to this direction.