The two most important tasks in entity information summarization from the Web are named entity recognition and relation extraction. Little work has been done toward an integrated statistical model for understanding both named entities and their relationships. Most of the previous works on relation extraction assume the named entities are pre-given. The drawbacks of these sequential models are that the results of relation extraction cannot be used to guide the named entity recognition, which have been proven useful. This paper proposed a novel integrate framework called EntSum, which enables bidirectional integration of named entity recognition and relation extraction using iterative optimization. Experiments on a one million large real Web data set show that EntSum achieves much better performance on both tasks than sequential methods.