In today’s educational environment, evaluating students’ descriptive responses is an essential duty, but labor-intensive, time-consuming, sided scoring of the papers, scoring only focusing on the answer of a question manual grading systems can often make it more difficult. Automatic evaluation systems have been created in response to these issues; yet, most of these algorithms find it difficult to effectively assess descriptive replies because of the complexity of natural language. This research aims to enhance automatic evaluation systems by utilizing advanced neural network models, which show great potential in natural language processing applications. The primary objective is to improve the accuracy and reliability of the systems in understanding the context and coherence of descriptive answers. Even with these improvements, it is still challenging to fully comprehend the context and nuances of descriptive replies and to ensure that models perform well across a range of subjects and question types. The study concentrates on advanced neural networks, namely SBERT, to encode text inputs into embeddings and compute cosine similarity to produce text similarity scores. Their performance is validated using an extensive dataset and evaluation metrics including RMSE and Pearson and Spearman correlation. The results show that the model consistently performs well in identifying comparable text pairs, indicating its reliability and efficacy as a tool for educational evaluations.