This research focuses on an architecture that vectorizes the names of various products found in daily life using BERT, followed by predicting product categories based on these embeddings. The architecture's performance is determined by the BERT model,...
This research focuses on an architecture that vectorizes the names of various products found in daily life using BERT, followed by predicting product categories based on these embeddings. The architecture's performance is determined by the BERT model, which extracts embeddings from product names, and the classifier that predicts categories from these embeddings. Consequently, this research initially aimed to identify a BERT model suitable for classifying product names and then find the most efficient combination of BERT model and classifier by applying various classifiers to the chosen BERT model. A simple CNN classifier was employed for the initial selection of a suitable BERT model, serving as a baseline for performance comparison with other classifiers. The architecture's effectiveness was quantified using precision, recall, f1 score, and accuracy for category predictions. Experimental results showed that the Sentence BERT model was more suitable for this task than a conventional BERT model. Additionally, classifiers enhanced with Residual Blocks demonstrated superior performance compared to the baseline combination of Sentence BERT and CNN. The Sentence BERT model used in this study, not trained on Korean data, suggests that further improvements could be achieved through Domain Adaptation by training with diverse Korean datasets.