Document-term matching, traditionally framed as a supervised text classification task, faces critical limitations in real-world applications: reliance on labeled data, high computational costs, and an inability to resolve the semantic ambiguity of domain-specific terminologies. Current methods struggle with the granularity mismatch between macro-level terms (e.g., “inflammation”) and micro-level document contexts (e.g., “atopic dermatitis”), as well as the definitional complexity (e.g., “gauge invariance”) and contextual nuances (e.g., “collateralized debt obligation”) inherent to specialized domains. To address these challenges, we propose DTM-MC, an unsupervised cross-domain document term matching frame-work based on meta-contrastive learning. DTM-MC operates through three pillars: It first addresses the problem of insufficient case documents and domain-specific term data in few-shot domains by using automated annotation tools to obtain relevant data. Secondly, it introduces contrastive learning to classify the importance of data, markedly boosting the accuracy of document-term matching. Finally, it employs meta-learning to construct a document-term matching model that is universally applicable across multiple domains enabling rapid model transfer across specialized fields. Experimental results demonstrate that this method achieves optimal detection performance in the transfer to few shot specialized domains.