E-commerce platforms rely on structured metadata, such as attribute-value pairs, to enhance product search, recommendations, and customer experiences. However, sellers often provide incomplete product information, creating challenges in maintaining accurate metadata at scale. This paper addresses the problem of predicting detailed attributes such as product categories (L0 to L4) and brand name, from unstructured product descriptions. After analyzing the dataset, we identified hierarchical relationships between categories and developed two approaches: a standard BERT-based classification model and an improved hierarchical BERT approach. By predicting categories sequentially, the hierarchical method effectively utilized the dataset’s structure, improving accuracy from 80% to 87.28%. These findings demonstrate the importance of leveraging hierarchy for more accurate predictions, paving the way for better automation in e-commerce metadata extraction.