A semantic annotation method is presented for massive videos based on Spark computing model. These massive videos are stored on Hadoop distributed file system (HDFS) and distributed to several nodes. These shot segmentationsin videos are realized by the fractal dimension method, and then the key frames of video shots are extracted based on color features,texture features and fractal features.The features in shots are trained using meta-learning strategy, and changed to video words and collected into the visual video dictionary.So video content is predicted and expressed by several video words according to the video dictionary. Then the video words are arranged according to importance sequence by Markov chains and the important words are described as video content prediction. Compared with the traditional distributed computing model, the Spark computing method illustrates the superiority from the correct rate, the average running time and the expansion efficiency.