This paper proposes an algorithm based on deep learning, and designs and verifies its effectiveness for the problem of multi-target audience classification in cultural communication. The Transformer model framework is adopted, and feature extraction is performed through the pre-trained BERT model. Combined with multi-dimensional features such as social media text, user behavior and cultural preferences, the audience is accurately classified. The experimental results show that the proposed model performs well in the tasks of cultural preference, behavior pattern and interest tag classification, with the highest F1 score reaching 0.92 and the accuracy reaching 0.93. The design of the weighted multi-task loss function further balances the performance of different tasks, and the weight optimization significantly improves the overall performance of the model. Through visual analysis, the model shows good generalization ability and stability, which provides technical support for audience segmentation and precise content dissemination in cultural communication.