The construction of the evaluation system of English translation teaching in colleges and universities under the cross-cultural perspective to improve the quality of translation education. The focus is on integrating the LMBP neural network algorithm and quantitative judgment indexes to provide an adequate evaluation model for English translation teaching. The study adopts the neural network algorithm based on LMBP, combined with the quantitative processing methods of “AHP+DEA” and “Entropy Assignment Method+Euclidean Distance Method”, to systematically construct and empirically analyze the evaluation indexes of teaching. It is found that teaching construction, application and effect are the core indicators of evaluation. Regarding teaching construction, classroom content integrity (correlation coefficient 0.428) and learning resources integrity (correlation coefficient 0.439) are the key factors. Teaching application analysis showed that hybrid teaching positively correlated with student achievement (r=0.569). Conducting effectiveness analysis showed that student discussion and in-class performance were significantly and positively correlated with teaching evaluation (p<0.005). These results prove the effectiveness and usefulness of the constructed evaluation system.