文档级关系抽取方法的研究进展

Survey on Document-Level Relation Extraction

关系抽取是自然语言处理领域的一项基础研究,抽取的结果可以用于知识图谱构建、人机问答、语义搜索等下游任务,具有广泛的应用场景和重要的研究价值。近年来,关系抽取研究取得了丰富的成果,但绝大多数研究局限于句子级关系抽取。研究表明,大量的关系无法通过单个句子提取,随着深度学习和自然语言处理技术的不断发展,文档级关系抽取研究工作迎来了新一轮的机遇和挑战。文中着重对近几年文档级关系抽取的研究进展进行分类和梳理,提炼出文档级关系抽取的一般技术路线图,分析文档级关系抽取研究的特征编码及特征聚合方法,并根据提取特征的不同,将文档级关系抽取方法概括为基于词汇特征、基于句法特征以及基于关系特征的3类方法;同时介绍常用文档级关系抽取数据集和评测指标,并对未来的研究趋势进行展望。

Relation extraction(RE)is one of the basic research in natural language processing.The result of RE can be applied to downstream missions such as construction of knowledge graphs,knowledge base question answering,semantic search et al,so RE has wide-ranging application scenarios and important research value.In recent years,RE has achieved fruitful results,but most studies are limited in sentence-level RE.Research shows that many relations can not be extracted from a single sentence.Under the development of deep learning and NLP,do-cument-level RE is facing new opportunities and challenges.This study reviewed the recent advances in document-level RE research,summarized a general technology roadmap of document-level RE,and analyzed the methods of feature encoding and feature aggregation used in researches.According to the different features of extraction,the document-level RE methods were divided into three categories,namely,extraction method based on lexical features,extraction method based on syntactic features,and extraction method based on relational features.The paper also introduced the common datasets and evaluation metrics of document-level RE and forecasted the future development trend of this task.