This paper presents an unsupervised method of extracting entity relation from large-scale corpus which is based on the hypothesis that a named entity with the same relation has a similar context, analyzes the co-reference relation between the co-reference substance to be tested and the object to be resolved, completes the construction of the entity according to the adjacent principle of the type entity and the core word principle, and uses the relative position restriction rule to combine the context window method to extract the feature and construct a feature sequence. In the end, the completion of the entity relation extraction task is based on the improved K-means clustering algorithm. The experimental results show that the new method can effectively improve the effect of entity relation extraction with a certain practical value.