The invention provides a method, a device, a storage medium and a terminal device for mining nearest neighbor semantics. The method comprises the following steps: acquiring natural sentences of nearest neighbor semantics to be mined; according to an encoder of a variational self-encoding model, encoding the natural sentence to obtain an implicit vector of the natural sentence, wherein the hidden vector is located at a mean point of the encoder output distribution; performing nearest neighbor semantic search on the implicit vector to obtain a nearest neighbor implicit vector having nearest neighbor semantics with the implicit vector; and decoding a nearest neighbor hidden vector according to a decoder of the variational self-coding model to obtain a statement having nearest neighbor semantics with the natural statement. Through the invention, a large number of sentences with semantic syntactic approximation can be excavated.