Implementation of an Automated Answer Evaluation System

Grading descriptive responses has always been seen as a time consuming job typically given to evaluators. That is because, unlike multiple-choice questions, descriptive answers require a comprehension system that can interpret the meaning of the sentences and judge whether it is relevant to the question asked. With an increasing number of students, this manual evaluation process becomes increasingly burdensome and prone to errors, highlighting the need for a more efficient and accurate approach. This paper proposes an implementation of an automated assessment system for descriptive answers. The system takes into account the various approaches used for assessing the answers and finally uses the best model that complies with the requirements of the system. The paper discusses the benefits and drawbacks of each approach and provides a tailored path for the choice of model for assessment based on different needs of size and performance. The paper specifically implements the SBERT architecture for generating vector embeddings of the sentences. This paper can serve as a guide for software developers to design an automated assessment system. The proposed system’s results have been compared against actual human moderator results across various domains of examination subjects. The measure of the system’s 95.1% accuracy while assessing theoretical answers shows that Sentence-BERT (Bi-directional Encoder Representation of Transformers) has the potential to revolutionize the assessment landscape. The utilization of sentence transformers and sentence embedding emerges as a cornerstone in achieving efficient and robust semantic analysis of textual responses.