Manually reviewing subjective articles is a laborious process fraught with challenges related to data understanding and acceptance, which impedes the effective use of AI for evaluation. Many have explored using computer technology to assess student answers, often relying on traditional methods or specific terminology, yet validated datasets remain scarce. This paper proposes a novel approach to automatically evaluate descriptive responses by integrating machine learning, natural language processing, and toolkits such as Wordnet, Word2vec, WMD, cosine similarity, MNB, and TF-IDF. Solution statements and keywords are leveraged for evaluation, and a machine learning model is trained to predict grades. Results indicate that WMD outperforms cosine similarity in effectiveness. With sufficient training, the machine learning model can operate independently. Experimental findings demonstrate an 88% accuracy without MNB, which improves by 1.3% when MNB is included.

Subjective article evaluation, automatic assessment, machine learning, natural language processing, Wordnet, Word2vec, WMD, cosine similarity, MNB, TF-IDF.