Abstract
Addressing the bottleneck issue of low accuracy and poor generalization in cargo risk intent detection caused by annotation scarcity in manual inspection scenarios within the import-export trade supervision domain, this study proposes an intent detection model named MBCS (Multi-task Learning with BERT for Classification and Semantic Similarity Comparison), designed for few-shot scenarios. To tackle the challenges of scarce domain-specific data and inadequate text representation capabilities, the research introduces a multi-task learning framework that integrates text classification with semantic similarity comparison. By incorporating semantic contrastive learning as an auxiliary task, the model's semantic representation capability is enhanced. Concurrently, an attention-weight-based synonym substitution strategy was introduced, replacing the highest-attention words in the sequence by integrating contextual information. Experiments conducted on real-world customs business datasets demonstrate that MBCS achieves significant accuracy improvements of over 4.19% (4.91%) in 5-shot (10-shot) scenarios, substantially outperforming baseline models. This method provides an optimized solution for intent detection tasks plagued by the annotation scarcity dilemma.