The embodiment of the invention provides a video description method, system and device, and the method comprises the steps: extracting the visual feature representation of a video frame at the currentmoment in a to-be-described video through a video encoder based on a convolutional neural network; writing the visual feature representation at the current moment into a visual memory at the currentmoment; reading attribute information from the attribute memory at the current moment according to the visual memory at the current moment and the text memory at the current moment; and generating predicted words by using a text decoder based on the long-short term memory network according to the word at the previous moment and the attribute information read at the current moment. Therefore, according to the embodiment, a multi-mode description method is adopted, and the flexibility of video description can be improved.