Abstract: A method and device for behavior control of a virtual image based on a text, and a medium are disclosed. The method includes inserting a symbol in a text, and generating a plurality of input vectors corresponding to the symbol and elements in the text; inputting the plurality of input vectors to a first encoder network, and determining a behavior trigger position in the text based on an attention vector of a network node corresponding to the symbol; determining behavior content based on a first encoded vector that is outputted from the first encoder network and corresponds to the symbol; and playing an audio corresponding to the text, and controlling the virtual image to present the behavior content when the audio is played to the behavior trigger position.

Abstract: A method and device for behavior control of a virtual image based on a text, and a medium are disclosed. The method includes inserting a symbol in a text, and generating a plurality of input vectors corresponding to the symbol and elements in the text; inputting the plurality of input vectors to a first encoder network, and determining a behavior trigger position in the text based on an attention vector of a network node corresponding to the symbol; determining behavior content based on a first encoded vector that is outputted from the first encoder network and corresponds to the symbol; and playing an audio corresponding to the text, and controlling the virtual image to present the behavior content when the audio is played to the behavior trigger position.

Abstract: A multi-document summary generation method includes obtaining a candidate sentence set, training each candidate sentence in the candidate sentence set using a cascaded attention mechanism and an unsupervised learning model in a preset network model, to obtain importance of each candidate sentence, selecting, based on the importance of each candidate sentence, a phrase that meets a preset condition from the candidate sentence set as a summary phrase set, and obtaining a summary of a plurality of candidate documents based on the summary phrase set.

Abstract: A multi-document summary generation method includes obtaining a candidate sentence set, training each candidate sentence in the candidate sentence set using a cascaded attention mechanism and an unsupervised learning model in a preset network model, to obtain importance of each candidate sentence, selecting, based on the importance of each candidate sentence, a phrase that meets a preset condition from the candidate sentence set as a summary phrase set, and obtaining a summary of a plurality of candidate documents based on the summary phrase set.