Iterative incremental dialogue intention category identification method based on small sample

Abstract

The invention relates to an iterative incremental dialogue intention category identification method based on small samples. The method is based on a small sample data set. Training starts from a preliminary classification model, and along with the use of a model, the number of the preliminary models is continuously increased. Model accuracy is also gradually improved, and a training mode that a large number of samples are needed by a previous deep learning model is abandoned. According to the method, in the iterative training process, only a small number of samples are needed to train a new preliminary classification model each time. The weights of other existing historical preliminary classification models are not changed. Results of all the preliminary classification models are input into the secondary classification model for training, the calculation speed of the model cannot decrease along with increase of the number of samples, the similarity screening model can screen and removethe existing preliminary classification model, the performance is maintained under the condition that the accuracy is guaranteed. Compared with the prior art, the method is advantaged in that the number of training samples is small, the calculation performance is stable, and the model is easy to update and expand.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to an iterative incremental dialogue intent category recognition method based on small samples. Background technique [0002] In order to improve the quality of products and services, many companies have launched their own customer service systems to help users answer questions through manual customer service, so that users can enjoy better services and improve service quality and efficiency. However, with the increase in the number of product users, traditional The manual customer service cannot meet the needs of many users, and the manual customer service needs special training and learning for the business, which also brings a certain cost. The customer service hotline is busy for a long time, which will affect the user experience. Therefore, major companies have launched their own artificial customer service products, which can help users solve related businesses faster by lea...

AI Technical Summary

Problems solved by technology

The rule-based algorithm often roughly guesses the user's intention by counting the keyword information in the text. This classification method is slow in querying when the amount of data is large, and requires manual labeling, which is time-consuming and labor-intensive; probability-based The document classification algorithm of the statistical model has high requirements on the quality and distribution of the text, and the classification is inaccurate in the case of small samples; the text classification algorithm based on the machine learning classification model is better for short text classification, and is very difficult for long text. It is difficult to capture the dialogue context information, it is easy to answer the wrong question, and the model needs to be retrained for new predictions. As the number of samples increases, the training complexity is greater

Images