With the rapid development of Internet technologies and applications, a lot of harmful video have spread out on Internet, it is enormously harmful to stability of social and people's physical and mental health. The means to extract video caption text is studied in this paper, and improvement method of text security detection. The proposed method first classifies caption text, then compares the result of classification with the library of user’s demands to determine whether to trigger alarms, through which the aim of monitoring harmful videos could be achieved. In this method, the text detection manner calculate the polarity of sentiment words by analyzing the context of those, meanwhile considers the effect of noun, then gets the orientation of the whole texts. Experiment has shown the method can monitor harmful video effectively.