Provided are an electronic device, a method and a training method for natural language processing. The electronic device for natural language processing includes a processor configured to: for each of words obtained by segmenting a sentence in a training data set, obtain an attention parameter representing correlation between the word and each of one or more words of other words in the sentence, where each of the words is represented by a real vector; and train, based on each of the words in the sentence, information on the attention parameter acquired for the word and label information of the sentence in the training data set, a neural network for sentence classification.