基于改进生成对抗网络的文本生成模型

Text Generation Model Based on Improved Generative Adversarial Network

在大数据信息智能化时代下,人工智能技术贯穿于人们的日常语言交流之中,因此,自然语言生成的文本是许多学者所研究的内容之一,而后生成对抗网络广泛应用于文本生成领域且具有优异的性能。针对文本为离散型数据在生成对抗网络中判别器输出所带来的精度影响问题,提出了一种改进的生成对抗网络文本生成模型,在MaliGAN模型的基础上设计了一种Loss函数,既能让模型保留寻找全局最优解的优势,又能适当降低离散型变量带来的精度影响,以此提高文本生成效果。为了体现模型的可行性与有效性,在COCO图像字幕数据集和EMNLP2017新闻数据集上进行实验,通过提升BLEU指标说明该模型能够避免离散型数值所带来的精度下降,从而在文本生成任务中取得较好的效果。

In the era of big data information intelligence,artificial intelligence technology runs through people’s daily language communication.Therefore,the text generated by natural language is one of the contents studied by many scholars,and the post-generation adversarial network is widely used in the field of text generation and has excellent performance.Aiming at the problem of the accuracy impact of the discriminator output in the genera⁃tion of discrete data in adversarial network,an improved generation adversarial network text generation model was proposed.Based on the MaliGAN model,a loss function was designed,which not only keeps the advantages of the model in finding the global optimal solution,but also properly reduces the accuracy impact caused discrete variables,to improve the text generation effect.In order to reflect the feasibility and effectiveness of the model,experiments were carried out on COCO image subtitle data set and EMNLP2017 news data set.By improving the BLEU index,it is shown that the model can avoid the accuracy degradation caused by discrete values,and thus achieve good results in the text generation task.