Investigating the BERT Capabilities with GRU Model in Semantic Extraction for Short Answer Grading Tasks: A Regression Problem

Assessment and quantification of knowledge gain are fundamental aspects of evaluating the quality of the learning process. The extensive use of remote learning led to the necessity for reliable and effective online exam assessment techniques. Automatic answer grading is a challenging task that requires a deep understanding of the answer text. Pretrained models like Bidirectional Encoder Representations from Transformers (BERT) capture the nuances of meaning that arise from different sentence structures and word usage, which is crucial for semantic extraction. This paper investigates the effectiveness of combining BERT with Gated Recurrent Units (GRU) for regression tasks in the context of automated student answer evaluation. Specifically, the performance of BERT embeddings and GRU's sequence modeling abilities is evaluated on a small dataset, where overfitting is a major concern. The various architectural combinations, including BERT with fine-tuning, GRU models for sequential dependency modeling, and hybrid approaches where BERT encodings serve as inputs to GRU layer are evaluated. The models are assessed based on their ability to predict numerical scores (ranging from 0 to 5) for student answers. Results suggest that hybrid models effectively leverage the strengths of both transformer-based and recurrent architectures, offering a promising approach for automated answer evaluation in small-scale educational datasets.