Abstract: Aspects of the present disclosure describe improving neural network robustness through neighborhood preserving layers and learning weighted-average neighbor embeddings.

Abstract: A text classification apparatus determines a word vector corresponding to a keyword according to a word vector model, and determines a potential extended word of the keyword based on the word vector. Then, when an extension rule input is received, and an adding instruction is detected, the apparatus adds the potential extended word to the keyword library, and adds the extension rule to a matching rule library. The apparatus determines a first probability that a text belongs to each of multiple preset classes, wherein the determination is based on the keyword library and the matching rule library. The apparatus determines a class to which the text belongs from the multiple preset classes.