数据到文本生成研究综述-《计算机技术与发展》

文章信息/Info

Author(s):: CAO Juan1; GONG Jun-peng2; ZHANG Peng-zhou2; 1. New Media Institute,Communication University of China,Beijing 100024,China;2. Faulty of Science and Technology,Communication University of China,Beijing 100024,China

Keywords:: data-to-text generation; robot journalism; natural language generation; content selection; surface realization; neural network

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.

Abstract:: 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 ofrobot journalism,focusing on content selection and surface realization,we comb the development of data-to-text generation in recentyears,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 futureresearch direction.