提升神经网络标题生成一致性方法研究

随着互联网络的快速发展，人们面临着越来越严重的信息过载问题，而标题是人们快速有效地获取信息的重要途径。随着深度学习技术的出现以及普及，神经网络标题生成方法因为其完全数据驱动以及无需额外人工定义特征的特点，获得了广泛的关注，在社会、经济以及科学研究层面，都具有十分广阔的应用前景和研究价值。由于神经网络标题生成涉及到自然语言的理解和生成，相对比较复杂，整个系统会出现很多不一致问题：不同文档之间主题不一致；不同任务之间建模方式不一致；不同语言文档之间数据分布不一致。本项目拟针对这三个问题分别提出解决方法。具体地，包括：1）针对不同文档之间主题不一致的问题，提出融合知识的神经网络标题生成模型；2）针对不同任务之间建模方式不一致的问题，提出基于重要性理论的神经网络标题生成模型；3）针对不同语言文档之间数据分布不一致的问题，提出基于领域自适应的跨语言神经网络标题生成模型。

With the rapid development of the Internet, people are facing more and more serious problems of information overload. Headlines provide people with an efficient way to obtain information. Therefore, it is of great significance to the study headline generation. Recently, the neural headline generation has drawn much attention since it is completely data-driven, does not need extra hand-crafted features, and automatically learns how to map a news article to a headline. However, the neural headline generation model is relatively complex, there are still some discrepancies in the whole system: the discrepancy between documents with different topics; the discrepancy between the modeling methods of different tasks; and the discrepancy between the data distribution of different languages. To alleviate these problems, we propose three solutions: 1) to minimize the discrepancy between documents with different topics, we propose knowledge enhanced neural headline generation; 2) to minimize the discrepancy between the modeling methods of different tasks, we propose neural headline generation with importance theory; 3) to minimize the discrepancy between the data distribution of different languages, we propose cross-lingual neural headline generation with domain adaptation.