New dialogue intention recognition method based on pseudo label self-training and source domain retraining

Abstract

The invention discloses a new dialogue intention recognition method based on pseudo label self-training and source domain retraining and belongs to the technical field of language processing. According to the new dialogue intention recognition method based on pseudo label self-training and source domain retraining, pseudo labels are generated for unlabeled data containing new dialogue intentions, and model parameters are iteratively updated by utilizing a self-training method, so recognition accuracy is continuously improved; meanwhile, a retraining strategy is provided, so knowledge can be better migrated between a source domain and a target domain, and the expression ability of models is improved; and finally, the output of the three models is fused for integrated learning, so robustness of the models is improved.

Description

technical field [0001] The invention belongs to the technical field of language processing, and in particular relates to a new dialogue intent recognition method based on pseudo-label self-training and source domain retraining. Background technique [0002] The core module of the intelligent dialogue system is user intent recognition. New dialog intent recognition aims to discover newly generated dialog intents based on existing dialog intents, using only a small amount of labeled known intent data to discover and classify new intents from a large amount of unlabeled data. Since the data containing new intents are all unlabeled data, the existing dialog intent classification models cannot handle them, which leads to errors in the recognition of user intent and affects the subsequent response of the intelligent dialog system. [0003] In order to solve the above problems, the academic community currently mainly adopts two types of methods: 1. The method based on comparative ...

AI Technical Summary

Problems solved by technology

The above-mentioned technical solution has the following disadvantages: First, the pseudo-labels generated by the existing model for unlabeled data are relatively coarse-grained, and the model cannot be well trained to discover new dialogue intentions.

Second, existing models only use labeled data for model initialization, but fail to make full use of labeled data during training, making knowledge transfer difficult

Third, the existing model only uses the clustering model to generate the final prediction, and does not consider integrating other models for ensemble learning, which reduces the robustness of the model

Images