The traditional bipartite recommendation model only considers the historical interaction behavior of users and items. In order to provide more accurate, diverse and interpretable recommendations, it is necessary to fully consider the label auxiliary information and the calculation method of weights on the basis of user-item interactive modeling. This paper proposed a recommendation algorithm (LWV) based on natural language processing for tag similarity auxiliary edge optimization. This method combined user historical behavior and tag assistance information to generate new edges for node interaction between nodes through word2vec and constructed the weight of the edges to update the recommendation list of the basic recommendation algorithm. A comparison between this algorithm and the benchmark algorithm in six public evaluation standards on the public data set shows that the updated recommendation algorithm of LWV achieves a better balance in terms of accuracy, diversity and novelty than the original algorithm.