In order to solve the problem that the traditional keyword matching based retrieval method cannot meet people’s demand for information retrieval, the keyword extraction algorithm based on deep learning is proposed in legal information retrieval system. In order to solve the problem that the traditional keyword matching based retrieval method cannot meet people’s demand for information retrieval, a deep learning-based keyword extraction algorithm is proposed in the legal information retrieval system. In this paper, for the shortcomings of the statistical based method that can only extract two-word words, we consider to add the rule-based method to extract the compound words, and we propose the improved method for the concepts of the existence of the relationship of synonyms and the relationship between the whole and the part. The experimental results show that: the size of the corpus has a great influence on the experimental results, this paper selects 100 documents of each domain, the scale is not too large, with the increase of the corpus size, the accuracy and recall rate can be improved. With the addition of correlation rules, some compound words can be extracted, and the domain relevance and domain consistency formulas are improved to extract concepts with synonym relationship and low-frequency concepts with part-of-the-whole relationship.