Subjective Answer Grader using Semantic Similarity and Keyword Matching

Subjective exams and tests have been a staple metric of student evaluation. These tests require students to give brief theoretical responses to the questions. Evaluation of subjective responses is based on the presence of certain facts, keywords, phrases that need to be mentioned. Traditional assessment of these tests requires an examiner to affirm the understanding of concepts by students through their answers. The human-dependant evaluation, however, is tedious, time consuming and is rarely impartial. The time constraints often mean that the examiner is unable to provide a detailed feedback to each and every student. This calls for the need to develop an objective automated grader system, which can investigate and analyze the student performances by considering different facets of learning based on semantic similarity. The enhanced BERT model, which has been fine-tuned, is utilized to capture the semantic characteristics found in the student’s response. In order to achieve this, the model compares the student’s response with the model answer and generates a model score based on the comparison. This model score is boosted through a keyword matching algorithm, which emphasizes the factual information contained within the student’s response. These final similarity scores are employed to offer comprehensive feedback that illuminates the specific areas in which the student’s comprehension may be lacking. Finally, the proposed approach is compared to other renowned algorithms like Cosine Similarity, Word Mover’s Distance, Paraphrase-mpnet, and Multi-qa-mpnet. The proposed system yields an experimental Mean Absolute Error (MAE) of 2.11, which is significantly better as compared to other algorithms.