数据到文本生成的神经网络模型研究-《计算机技术与发展》

文章信息/Info

Author(s):: CAO Juan; GONG Jun-peng; ZHANG Peng-zhou; Communication University of China,Beijing 100024,China

Keywords:: data-to-text generation; robot journalism; natural language generation; neural network

摘要:: 随着人工智能的发展,国内各行各业纷纷涌入人工智能领域,尤其是新闻行业,各大科技公司和科研机构研发自动化新闻生产,掀起了一股新闻机器人的热潮。机器新闻写作,实质上就是一个利用自然语言生成技术进行新闻写作的过程。数据到文本生成属于自然语言生成技术,是指采用结构化的数据例如一张表格作为输入,生成恰当而流畅的文本作为输出来描述数据,是实现机器新闻写作的关键技术之一。近年来随着数据到文本生成研究的不断深入,研究人员将神经网络方法引入该领域,主要基于循环神经网络构建模型,并已取得了不错的研究进展。文中首先介绍了用于数据到文本生成的神经网络模型,并梳理了近年来该领域的研究成果,然后介绍了相关数据集并对比了各模型的实验结果,最后分析了该领域研究存在的问题并提出了未来发展的建议。

Abstract:: With the development of artificial intelligence,various industries in China have flocked to the field of artificial intelligence,especially the news industry. The domestic technology companies and major scientific research institutions have independently developed automated news production,setting off a wave of news robots. Robot journalism is essentially a process of news writing using natural language generation technology. Data-to-text generation belongs to natural language generation technology which refers to using structured data such as a table as input and generating appropriate and fluent text as output to describe data and is one of the key technologies to realize robot journalism. In recent years,with the deepening of data-to-text generation research,researchers have introduced neural network into this field,mainly building models based on recurrent neural network,and have made excellent research progress. In this paper,we first introduce the neural network models used for data-to-text generation and review the research results in this field in recent years,then introduce the relevant datasets and compare the experimental results of each model. Finally we analyze the problems in this field and put forward suggestions for future development.