基于弱监督与表示学习的关系抽取算法研究

发布时间：2018-09-04 18:56

【摘要】：我们生活在信息时代,但是从中获取有用和结构化的知识并非易事。近期的研究估计,人类已经存储了超过295艾字节(295*1018字节)的数据。大部分的文本数据以非结构化形式存在,如新闻文章、论坛与留言板、电子邮件、社交网络状态更新,和近十亿计的网页。如何有效的利用人类所产生的数据,进行深度数据挖掘与文本理解有重要意义,这些数据可以用于强化搜索结果、知识图谱补全,自动问答系统等。为了在信息中提炼出知识,本文研究了从海量非结构化文本进行关系抽取的方法,并提出了基于弱监督与表示学习的实体关系抽取方法,主要完成了以下几方面的工作:1.针对弱监督学习会引入噪声的问题,本文提出了基于模糊分类的多实例多标签学习算法。针对某个特定实体对,将多种关系和多个样本整合起来一同学习,并通过模糊分类的方法聚合句子级特征。相比其他弱监督学习方法,本文提出的方法能够提升2%以上的fl值。2.提出了基于词向量和循环神经网络进行自动特征抽取的方法。传统的特征抽取方法依赖人工经验与知识,并且是针对特定问题的设计,扩展性不强。神经网络能够自动进行特征抽取,分层对原始文本进行处理,并且不需要进行预处理。实验证明利用不同神经网络产生的特征,都对系统性能有一定提升。3.设计并实现了基于弱监督和神经网络特征提起的实体关系抽取平台。该实验平台整合了多项工具,包括文本检索、文本分析等,能够利用知识库对文本进行自动标注,学习关系抽取模型,最后输出关系抽取结果。在LDC与New York Times数据集上,本文提出的方法较传统方法,均有一定的效果提升。综上所述,本文提出的方法能够有效提升实体关系抽取系统性能,利用弱监督扩展了学习语料,通过表示学习增强了系统在不同领域的扩展性,同时能够减少预处理工作。这些内容对于关系抽取研究有一定的贡献与参考价值。
[Abstract]:We live in the information age, but it is not easy to gain useful and structured knowledge from it. Recent research has estimated that humans have stored more than 295 exabytes (295,1018 bytes) of data. Most text data exists in unstructured formats, such as news articles, forums and message boards, email, social network status updates, and nearly a billion pages. How to effectively utilize the data produced by human beings, and how to mine the deep data and understand the text is of great significance. These data can be used to strengthen the search results, complement the knowledge map, automatic question and answer system and so on. In order to extract knowledge from information, this paper studies the method of relational extraction from massive unstructured text, and proposes an entity relation extraction method based on weak supervision and representation learning. In this paper, a multi-instance and multi-label learning algorithm based on fuzzy classification is proposed to solve the problem of introducing noise into weakly supervised learning. For a particular entity pair, a variety of relationships and multiple samples are integrated together, and sentence level features are aggregated by fuzzy classification. Compared with other weakly supervised learning methods, the proposed method can increase the fl value by more than 2%. An automatic feature extraction method based on word vector and cyclic neural network is proposed. The traditional feature extraction method relies on artificial experience and knowledge, and is not extensible for specific problems. The neural network can automatically extract features and process the original text in layers without the need of preprocessing. Experiments show that using the characteristics of different neural networks, the system performance has a certain improvement. 3. 3. The entity relation extraction platform based on weak supervision and neural network feature is designed and implemented. The experimental platform integrates many tools, including text retrieval, text analysis and so on. It can use knowledge base to automatically annotate text, learn relational extraction model, and finally output the result of relational extraction. On the data sets of LDC and New York Times, the methods proposed in this paper are more effective than the traditional methods. To sum up, the proposed method can effectively improve the performance of entity relation extraction system, extend the learning corpus by using weak supervision, enhance the extensibility of the system in different fields by representation learning, and reduce the preprocessing work at the same time. These contents have certain contribution and reference value to the research of relation extraction.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1

【相似文献】