The convenience of social media in communication and information dissemination has made it an ideal place for spreading rumor events, which raises a higher requirement for automatic debunking of rumor events. Meanwhile, the traditional rumor classification approaches relying on manual labeled features have to face a daunting number of manual efforts. In general, when facing a dubious claim, people can authenticate and verify the realness of an event with the contents of continuous posts, such as source credibility, public sentiments, propagation structures, and so on. In this paper, we pay more attention to the emotional expressions of posts host, especially the fine-grained sentiments, which are effective for rumor events detection. Thus, this paper presents a novel two-layer GRU model for rumor events detection based on a Sentiment Dictionary (SD) and a dynamic time series (DTS) algorithm, named as SD-DTS-GRU. The model learns continuous representations of microblog events in a better manner by making use of the SD to identify fine-grained human emotional expressions of each event and retaining the time distribution of social events by the DTS algorithm. The experimental results on Sina Weibo datasets show that our model achieves a high accuracy of 95.2% and demonstrate that our proposed SD-DTS-GRU model outperforms latest explorations on rumor events detection. INDEX TERMS Rumor events detection, sentiment dictionary, dynamic time series, GRU, online social networks.