Automatic Scoring Model for English Compositions Using Bidirectional Encoder Representations from Transformer with Pearson Correlation Coefficient

In recent days, automatic scoring models are playing a crucial role in evaluating essays which have been utilized in the large-scale writing compositions. The existing models for the automatic scoring model for English compositions including Automatic Evaluation System (AES) suffered with contextual understanding. To overcome the issues, Bidirectional Encoder Representations from Transformer with Pearson Correlation Coefficient (BERT -PCC) evaluation model is proposed in this research. The pre-trained BERT model improves the contextual understanding and enhances the semantic representation of the compositions. Initially, the input data is taken from the TOEFL11 dataset and further processed for preprocessing with various techniques including stop words, punctuations removal and lemmatization. After that, a pre-trained BERT model is employed to extract the features from preprocessed data with feature extraction layer from sequence, token and positional embedding's. At last, the PCC based evaluation model is introduced to evaluate the English compositions based on similarity of two variables with three manual scoring metrics. From the results, the proposed BERT-PCC model achieved better results in terms of score relevance of 0.60, average score error of 4.07 and average score accurate of 80.21% respectively when compared with existing Automatic Evaluation System (AES).