In the process of English translation, how to use monolingual corpus which is relatively easy to obtain and realize corpus expansion is the premise of applying intelligent term matching method. In order to improve the accuracy of automatic matching of technical terms in English translation, this paper proposes a semi-supervised learning translation method which incorporates the bottleneck of variational information. This method uses small-scale parallel corpus training to obtain the basic translation model and uses back translation to translate large-scale monolingual corpus into pseudo-parallel corpus, and then merges the two parallel corpus, so that the corpus scale can be well applied to variational method. The experimental results of semi-supervised learning on multiple translation data sets show that compared with the benchmark system, the proposed method can effectively improve the translation quality while maintaining the translation sentence length, and solve the problem of over-translation in traditional neural machine translation to a certain extent.