The present disclosure provides a method, an apparatus, a device and a system for sign language translation, where a server receives video information sent by a terminal device, and preprocesses the video information to obtain at least one sign language action; the at least one sign language action is input into a sign language model for classification and prediction to obtain a word corresponding to the at least one sign language action; each word is input into a language model to determine whether an intention expression is complete; and each word is sent to the terminal device when the intention expression is complete, so that the terminal device displays each word, thereby realizing the translation of the sign language action into text, enabling the ordinary persons to better understand intentions of the hearing impaired, thus improving efficiency of communications.