Automatic Chinese Composition Evaluation System Based on Distilled Bidirectional Encoder Representations from Transformers

In recent years, the Chinese compositions become critical and complex to evaluate since it consume more time and also required automated solutions to enhance the efficacy of the evaluation system. Previous researchers have suggested many ML models but still the system face challenges such as tonal variations, diverse vocabulary, and nuanced sentence structures which makes it difficult to build an effective system. Hence, this paper develops a robust evaluation system by incorporating numerous advanced Machine Learning (ML) models which are verified through the Sogou News cohort. Firstly, the Sogou News dataset is employed, and then the data is pre-processed to clean the data using various techniques that include tokenization, lowercasing, punctuation removal, and normalization. Further, from the processes data optimal features are extracted by imposing Term Frequency-Inverse Document Frequency (TFIDF) and Word2Vec to improve text classification. In the end Distilled Bidirectional Encoder Representations from Transformers (Distil-BERT) is employed to generate text embedding for effective evaluation of chinese composition. The suggested Distil-BERT model provides higher accuracy in the Chinese composition with higher accuracy value when compared existing models SGCSTC, RoBERT+HRL. As a result, the accuracy of $98.89 \%$, F 1 score is $98.88 \%$, precision is 98.64%, recall is 97.67%.