Regulatory authorities face challenges in efficiently extracting essential information from the extensive corporate business data distributed across various locations. This study proposes an enterprise hierarchical portrait scheme utilizing a deep learning model combining BERT and a bi-directional long and short-term memory network (BiLSTM) to address this issue. The process involves data preprocessing, utilizing BERT and BiLSTM for feature extraction to model text sequences and extract relevant semantic information, and generating hierarchical labels through a full connectivity layer to create a comprehensive company portrait. This approach aims to provide authorities with a framework for categorizing businesses effectively.