基于生成对抗网络的数据增强方法

Data Augmentation Method Based on Generative Adversarial Network

深度学习在分类任务上取得了革命性的突破,但是需要大量的有标签数据作为支撑.当数据匮乏的时候,神经网络极易出现过拟合的问题,这种现象在小规模数据集上尤为明显.针对这一难题,本文提出了一种基于生成对抗网络的数据增强方法,并将其应用于解决由于数据匮乏,神经网络难以训练的问题.实验结果表明,合成的数据和真实的数据相比既具有语义上的相似性,同时又能呈现出文本上的多样性;加入合成的数据后,神经网络能够更加稳定地训练,而且分类的准确度也有了进一步的提高.将提出的算法和其他一些数据增强的技术对比,我们的方法结果最好,从而证明了这种技术的可行性和有效性.

Deep learning has revolutionized the performance of classification, but meanwhile demands sufficient labeled data for training. Given insufficient data, neural network is apt to overfitting, which is quite general in low data regime. We propose a data augmentation technique based on generative adversarial network to address the network training and data shortage problem. The experimental results show that the synthesized data has semantic similarity compared with the real data, and at the same time it can present the diversity of the context. After adding the synthesized data, the neural network can be trained more stably, and the accuracy of the classification is further improved. Comparing the proposed algorithm with some other data augmentation techniques, the proposed method has the best performance, which proves the feasibility and effectiveness of this technique.