The identification and classification of event trigger plays a decisive role in event extraction. Usually, the trigger words are extracted based on supervised machine learning methods such as CRF. However, since these methods rely on expensive manual annotation and require predefined event types, they are not sufficient for open domain application. In this paper, we present an unsupervised method for event trigger identification and classification. First, we run a topic model to obtain the topic distribution over each candidate trigger word. Then, an improved twostate automaton is proposed to detect the real trigger word and capture the corresponding topics. The experiment on a large unlabeled corpus shows our unsupervised model is very inspiring.