Method and device for expanding training samples

Abstract

The embodiments of this specification provide a method and device for expanding training samples. The method includes: obtaining the initial training sample group to be expanded, including the first number of training samples, the training samples include the original machine sentences and original user sentences in the historical dialogue group, and the category labels corresponding to the dialogue group, and the initial training sample group Each training sample of has a first category label; Obtain a second number of training samples from the initial training sample group; For each dialogue group in the second number of training samples, the first class associated with the original machine sentence in each dialogue group A machine statement is input into the dialog generation model of the pre-trained first category label, and the first user statement corresponding to the first machine statement in each dialogue group is generated; the first machine statement and the corresponding first machine statement in each dialogue group are The user sentence is added to the initial training sample group as an extended training sample to obtain an extended training sample group. Able to achieve data equalization of training samples.

Description

technical field [0001] One or more embodiments of this specification relate to the field of computers, and in particular to methods and devices for expanding training samples. Background technique [0002] When the robot answers the user's question, after the dialogue between the robot and the user, classify the dialogue groups including machine sentences and user sentences, and determine the user's request according to the classification results. The above classification may include determining a standard question sentence corresponding to the dialogue group, so that the robot provides an answer corresponding to the standard question sentence. Wherein, the standard questions are also referred to as standard questions, which are sorted out questions that some users may ask. Each question has a question ID. [0003] In the prior art, the classification model is usually trained by using historical conversation groups as training samples, and then the current conversation gro...

AI Technical Summary

Problems solved by technology

Data imbalance will cause the classification model to be biased to determine the dialogue group as the category of training samples with a large amount of data, so how to solve the problem of data imbalance has become the core issue of improving classification accuracy

Images