ExASAG: Explainable Framework for Automatic Short Answer Grading

Maximilian Tornqvist, Mosleh Mahamud, Erick Mendez Guzman, Alexandra FarazouliACL 2023

TL;DR

This paper introduces ExASAG, a novel framework that combines deep learning models with explainability techniques to provide not only automatic grading of short answers but also human-readable justifications for those grades. This is crucial for building trust and enabling the deployment of automated grading systems in educational settings.

What Problem Does This Paper Address?

Grading short answers automatically (Automatic Short Answer Grading or ASAG) is a challenging task. While deep learning models have significantly improved grading accuracy, they often act as "black boxes," making it difficult for educators and students to understand why a particular grade was assigned. This lack of transparency hinders trust and adoption. Furthermore, manual grading is time-consuming and can lead to inconsistencies. This paper aims to bridge this gap by developing an ASAG system that is both accurate and explainable, providing natural language justifications for its grading decisions.

Methods

The ExASAG framework consists of two main components: an encoder for generating explanations and a decoder for performing the binary classification (grading).

  • Encoder (Explainability Module): This component leverages post-hoc explainability methods, specifically SHAP (SHapley Additive exPlanations) and Integrated Gradients (IG). These methods analyze a fine-tuned Transformer model (SciBERT) to identify which parts of a student's answer are most influential in the model's prediction. The framework then processes these token-level attributions to generate Sentence Level Attributions (SLAs), which are essentially importance scores for each sentence in the answer.
  • Decoder (Grading Module): This component uses a fine-tuned SciBERT model, a Transformer variant pre-trained on scientific text, to perform the binary classification of student answers into "Satisfactory" or "Non-Satisfactory" grades.

The novelty lies in the integration of these explainability methods to generate human-understandable language explanations that justify the model's grading decisions, rather than just providing a score.

Datasets

The study utilizes a dataset of 1131 question-answer pairs collected from graded exams in a Data Mining course at Stockholm University. This dataset is characterized as "low-resource" due to its limited number of iterations and questions. The answers are often scenario-based and longer than those in typical ASAG datasets. The original grading scales were converted to binary labels: "Satisfactory" (1) for answers achieving 50% or more of the maximum grade, and "Non-Satisfactory" (0) for those below 50%. This resulted in 667 satisfactory and 464 non-satisfactory answers. Human rationales (justifications for the labels) were also collected from annotators.

Key Insights and Ideas

  • Explainability is Key for ASAG Deployment: The paper strongly advocates that for ASAG systems to be trusted and adopted in real-world educational settings, they must provide clear, human-understandable explanations for their predictions.
  • Transformer Models Adapt Well to Domain-Specific Data: Even with a relatively small, domain-specific dataset, fine-tuning a Transformer model like SciBERT can achieve competitive classification performance.
  • SHAP and Integrated Gradients Perform Similarly: Both SHAP and IG were found to be effective in generating explanations that showed reasonable agreement with human annotations, suggesting that either can be a viable choice for this task.
  • Plausibility as an Evaluation Metric: The paper highlights the importance of evaluating the quality of explanations, not just the grading accuracy. They propose using Intersection-Over-Union (IOU) to measure the agreement between machine-generated rationales and human-annotated ones.

Key Results

  • Classification Performance: The fine-tuned SciBERT model achieved an F1-score of 0.76 for both "Satisfactory" and "Non-Satisfactory" labels (micro-averaged).
  • Explanation Quality (Plausibility): The framework's explanations showed a degree of alignment with domain-expert judgment. When evaluating the agreement between model-generated rationales and human rationales using IOU with a threshold of 0.5:
  • Integrated Gradients (IG) achieved an F1-score of 0.62 and a recall of 0.45.
  • SHAP achieved an F1-score of 0.63 and a recall of 0.46. These results indicate that the generated explanations are somewhat representative of human reasoning.

Limitations and Future Work

  • Data Size and Complexity: The dataset is acknowledged as "low-resource" and complex, which might limit the generalizability of the findings. The presence of arithmetic in some answers also posed a challenge.
  • Stopword Handling: The current method of excluding stopwords from sentence-level attribution calculations might overlook their potential importance in certain contexts.
  • Preprocessing Impact: The paper questions the impact of preprocessing steps (like removing non-alphabetic characters) on the fidelity of explanations, suggesting a trade-off between explanation usability and precise representation of model attention.
  • Future Work: The authors suggest exploring the incorporation of human-generated rationales during model training to potentially boost performance and explanation quality. They also plan to investigate multi-task learning approaches and evaluate rationales beyond simple plausibility. Further research on a larger sample size is needed to assess the feasibility of real-world implementation.

Summary updated

BibTeX
@inproceedings{tornqvist2023exasag,
  author = {Maximilian Tornqvist and Mosleh Mahamud and Erick Mendez Guzman and Alexandra Farazouli},
  title = {{ExASAG: Explainable Framework for Automatic Short Answer Grading}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure