Data enhancement method based on adversarial training in text classification scene

Abstract

The invention discloses a data enhancement method based on adversarial training in a text classification scene, which comprises the following steps of: translating an original sample xi in a training set back to generate a sample sequence set D '= {x'i, y'i} 1... N; performing a word embedding process on the paraphrasing sample x'i to obtain p theta (x'i); carrying out adversarial training on the paraphrase sample x'i to obtain an adversarial sample original sample xi, carrying out a word embedding process to obtain p theta (xi), and carrying out a word embedding process of adding disturbance on the adversarial sample to obtain comparison loss of the original sample and the adversarial sample through calculation; and if the comparison loss obtained by calculation meets a preset threshold range, adding the generated adversarial sample into the original data set, and sending the adversarial sample and the original data set to a classification model for simulation to obtain a simulation classification accuracy result. According to the method, adversarial training is combined with data enhancement technologies such as back translation, random noise injection, cross enhancement and the like, so that enhanced samples with the highest quality and the highest diversity are obtained.

Description

technical field [0001] The invention belongs to the field of natural language processing, and more specifically relates to a data enhancement method based on adversarial training in a text classification scenario. Background technique [0002] There are generally two forms of data enhancement, the first is data expansion, and the second is feature enhancement. Either way, the essence is to generate new samples on the basis of original samples with the help of auxiliary information or auxiliary data. Increase sample diversity. Data augmentation is the addition of newly generated unlabeled or labeled samples to the original dataset. Feature enhancement is to amplify its data features on the basis of the original sample, so that the classification model can better identify it. [0003] Most of the existing research on data enhancement methods is oriented to the task of Computer Vision (CV) and has been able to play a very significant role in large-scale neural network models,...

AI Technical Summary

Problems solved by technology

[0003] Most of the existing research on data enhancement methods is oriented to the task of Computer Vision (CV) and has been able to play a very significant role in large-scale neural network models, but in Natural Language Processing (Natural Language Processing, referred to as NLP) task, but has been ignored by researchers

On the one hand, the key to data enhancement lies in the preserved transformation of labels, but natural language is discrete and cannot be applied to practical tasks through simple operations like image data

On the other hand, the scale of existing neural network models is generally large and has the characteristics of over-parameterization. For such models, how to fine-tune them has become the focus of research, especially when a specific In the task, when the data set contains too little label data, some slight changes may cause the model to collapse

Images