数据到文本生成研究综述  

Review of Data-to-text Generation

作　　者：曹娟 龚隽鹏[2] 张鹏洲[2] CAO Juan;GONG Jun-peng;ZHANG Peng-zhou(New Media Institute,Communication University of China,Beijing 100024,China;Faulty of Science and Technology,Communication University of China,Beijing 100024,China)

机构地区：[1]中国传媒大学新媒体研究院,北京100024 [2]中国传媒大学理工学部,北京100024

出　　处：《计算机技术与发展》2019年第1期80-84,89,共6页Computer Technology and Development

基　　金：国家科技支撑计划项目(2014BAK10B01)

摘　　要：机器新闻写作,作为人工智能在传媒业的一种应用,越来越受到学界和业界的关注,目前主要用于体育、财经、气象地质和健康等领域。机器新闻写作的核心是自然语言生成技术,而数据到文本生成是自然语言生成领域的典型技术,是实现机器新闻写作的关键技术之一。为了更好地研究数据到文本生成技术并将其应用于机器新闻写作领域,以内容选择和表层实现为重点,梳理了近年来数据到文本生成的发展脉络,并比较了基于规则和数据驱动两种研究方法,归纳了不同领域的可用数据集,总结了内在和外在两类评价方法,分析了数据到文本生成技术当前存在的问题,以及探讨了其未来可能的研究方向。As an application of artificial intelligence in the media industry,robot journalism has attracted more and more attention from academia and industry.It is mainly used in sports,finance,meteorology,geology,health and other fields.The core of robot journalism is natural language generation,and data-to-text generation is a typical technology in the field of natural language generation,and also one of the key technologies to realize robot journalism.In order to study data-to-text generation technology better and apply it to the field of robot journalism,focusing on content selection and surface realization,we comb the development of data-to-text generation in recent years,compare two research methods based on rule and data-driven,sum up the available datasets in different areas,summarize the intrinsic and extrinsic evaluation methods,analyze the existing problems in data-to-text generation technology,and discuss the possible future research direction.

关 键 词：数据到文本生成 机器新闻写作 自然语言生成 内容选择 表层实现 神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]