APP下载

实体—属性抽取的GRU+CRF方法

2018-12-08王仁武孟现茹孔琦

现代情报 2018年10期

王仁武 孟现茹 孔琦

〔摘 要〕[目的/意义]研究利用深度学习的循环神经网络GRU结合条件随机场CRF对标注的中文文本序列进行预测,来抽取在线评论文本中的实体—属性。[方法/过程]首先根据设计好的文本序列标注规范,对评论语料分词后进行实体及其属性的命名实体标注,得到单词序列、词性序列和标注序列;然后将单词序列、词性序列转为分布式词向量表示并用于GRU循环神经网络的输入;最后输出层采用条件随机场CRF,输出标签即是实体或属性。[结果/结论]实验结果表明,本文的方法将实体—属性抽取简化为命名实体标注,并利用深度学习的GRU捕获输入数据的上下文语义以及条件随机场CRF获取输出标签的前后关系,比传统的基于规则或一般的机器学习方法具有较大的应用优势。

〔关键词〕实体属性抽取;GRU;循环神经网络;条件随机场;命名实体识别

DOI:10.3969/j.issn.1008-0821.2018.10.009

〔中图分类号〕TP391.1 〔文献标识码〕A 〔文章编号〕1008-0821(2018)10-0057-08

〔Abstract〕[Purpose/Significance]The study used the recurrent neural network GRU combined conditional random field CRF to predict the annotated Chinese sequence text to extract the entity-attribute in the online review text.[Method/Process]Firstly,according to the designed annotation specification to a text sequence,the paper made name entity annotations for entities and their attributes after the segmentation of corpus,and got word sequence,part of speech sequence and annotation sequence;Then the word sequence and part-of-speech sequence were converted into distributed word vector representation and used for input of GRU recurrent neural network;finally,the output layer used the conditional random field CRF and the output label was the entity or attribute.[Result/Conclusion]The method in this paper simplified entity-attribute extraction to named entity annotation,and used GRU to capture the contextual semantics of input data and conditional random field CRF to obtain the output label context,which had a larger application advantage than the traditional rule based or general machine learning method.

〔Key words〕entity attribute extraction;GRU;RNN;CRF;NER

实体—属性抽取是从非结构的文本数据中抽取有价值的语义单元的重要手段,是信息抽取、觀点挖掘、智能检索、自动问答、知识图谱等构建任务的基础。实体—属性抽取属于自动内容抽取(Automatic Content Extraction,ACE)的研究范畴。美国国家标准与技术研究院(National Institute of Standards and Technology,NIST)组织开展了系列ACE测评会议。ACE会议旨在研究自动抽取出新闻语料中的实体、关系以及事件等内容[1]。ACE的目标是发展包括自动识别和标识文本在内的自动内容抽取技术,以支持对语料库的自动处理[2]。

本文研究的实体是文本中描述的对象,或者说是文本中包含的特定事实信息,例如产品评论中的产品、服务、商家等。……

登录APP查看全文