Domain Adaptation for Medical Semantic Textual Similarity

Semantic textual similarity is a common task to determine whether two sentences in a pair own the same meaning. In the medical domain, the annotated data is limited and sparse, which brings great difficulty to obtain accurate semantic information from it. In this paper, we propose a two-stream model to adapt knowledge learned from other domains to the medical domain. To optimize and reduce the computation, we further compress the proposed model by knowledge distillation. Experimental results show that our proposed method achieves better performance than the baseline methods.