Detecting malicious URL is important for defending against cyber attacks. In view of the problem that supervised learning requires a large number of labeled samples, this study uses a semi-supervised learning method to train malicious URL detection models, which reduces the cost overhead of labeling data. We propose an improved algorithm based on the traditional co-training. Two kinds of classifiers are trained by using expert knowledge and Doc2Vec pre-processed data, and the data with the same prediction result and the high confidence of the two classifiers are screened and used for classifiers learning after being pseudo-labeled. The experimental results show that the proposed method can train two different types of classifiers with detection precision of 99.42% and 95.23% with only 0.67% of labeled data, which is similar to supervised learning performance and performs better than self-training and co-training.