Model training method and device

The invention discloses a model training method and device. In a pre-training stage, a first training sample is determined for a search result and a pair of search results corresponding to the search result, and a first feature of each search result in the first sample and two second features corresponding to a query text are determined through a correlation model of a double-tower structure; and determining comparison loss based on the feature distance between the first features and the feature distance between the second features of each first sample so as to pre-train the correlation model. And respectively combining the query text with the corresponding search results to obtain second samples, taking the relevancy between the query text and the search results in the second samples as labels, and continuing to train the relevancy model. According to the method, the relevancy model can accurately represent the characteristics of the query text and the result text of the search result through comparative lea