Darknet markets use anonymity technologies to conduct illicit transactions, making them hard to monitor. These markets feature diverse transaction categories, low-resource data, and semantic ambiguity, which traditional methods struggle to handle. To address these issues, we propose the FCA-XLNet-BiGRU-MultiTask (Formal Concept Analysis integrated XLNet and Bidirectional GRU with Multi-Task Learning) framework, a multi-task learning model that addresses closed-set classification, open-set detection, and low-resource adaptation. It integrates XLNet, BiGRU, AW-Attention, and FCA-based domain knowledge to improve feature representation. We preprocess 14,744 transactions to create a high-quality dataset. Experiments show our model outperforms BERT, RoBERTa, DeepSeek-R1, and Falcon in classification and detection tasks, especially in low-resource settings. This study provides a robust solution for monitoring darknet activities and has significant cybersecurity applications.