【作者】 王晶；

【导师】 华强胜；

【作者基本信息】 华中科技大学 ， 计算机技术， 2020， 硕士

【摘要】 在学术领域中常常需要通过关键字对论文进行分析,关键字提供了高度概括性的论文信息,是对论文进行信息检索、自动分类、自动聚类和观点挖掘的基础。随着网络信息量的迅速增加,手工标识关键字费时费力且主观性强,因此论文关键字自动生成成为了研究热点。传统的关键字自动生成算法主要通过统计数据进行关键字生成,没有考虑到文本中的语义信息。随着深度学习和自然语言处理学科的快速发展,现今的关键字生成算法能够利用序列到序列框架和注意力机制来理解文本,从而进行关键字自动生成。由于近年来转换器(Transformer)在自然语言处理领域中取得了优异表现,与注意力相关的机制也得到了快速发展,这使得关键字自动生成算法在精确度和多样性方面有了可以进一步提高的空间。本研究将基于Transformer的序列到序列框架、拷贝机制和覆盖机制进行了结合,设计了关键字自动生成系统。本研究分为以下三步:首先,关键字集合被预处理为不同的关键字序列,并作为解码器的输入进行训练。其次,系统修改了原始Transformer的结构,应用拷贝机制和覆盖机制来减轻关键字不在预测词表的问题和关键字生成多样性问题。最后,系统通过束搜索对目标关键字进行搜索,平衡了关键字的生成时间和质量。本研究主要贡献如下:首先,系统能够对文本语义进行理解并预测出在原文中未出现的关键字。其次,系统使用了关键字自动生成任务中常用的五个数据集,并在关键字自动生成的能力方面超过了Copy RNN模型。最后,系统从关键字顺序和框架等不同方面对基于Transformer的关键字生成模型进行研究,并对实验结果进行了说明和分析。更多还原

【Abstract】 In the academic field,it is often necessary to analyze the paper by keywords.The keywords provide highly general paper information,which is the basis of information retrieval,automatic classification,automatic clustering and viewpoint mining.With the rapid increase of network information,manual identification of keywords is time-consuming and subjective,so the automatic generation of keywords become a research hotpot.The traditional automatic keyword generation algorithm mainly uses statistical data to generate keywords without considering the semantic information in the text.With the rapid development of deep learning and natural language processing,the current keyword generation algorithm can use Sequence-to-Sequence framework and attention mechanism to understand the text,so as to generate keywords automatically.In recent years,due to the excellent performance of Transformer in the field of natural language processing and the rapid development of attention-related mechanisms,the automatic keyword generation algorithm has room for further improvement in accuracy and diversity.This research combines the Sequence-to-Sequence framework,copy mechanism and coverage mechanism based on Transformer,and designs the automatic keyword generation system.This research is divided into three steps: Firstly,the set of keywords is preprocessed into different sequences of keywords and trained as the input of decoder.Secondly,the system modifies the structure of the original Transformer,and applies copy mechanism and coverage mechanism to alleviate the problem that keywords are not in the prediction vocabulary and the problem of keyword generation diversity.Finally,the system searches the target keywords by beam search,balancing the generation time and quality of keywords.The main contributions of this research are as follows: Firstly,the system can understand the text semantics and predict the keywords that do not appear in the original text.Secondly,the system uses five data sets commonly used in keyword automatic generation tasks,and exceeds the Copy RNN model in the ability of keyword automatic generation.Finally,the system studies the keyword generation model based on Transformer from different aspects such as order of keywords and framework,and explains and analyzes the experimental results.更多还原