人物属性抽取旨在从人物介绍中抽取人物的各项属性，如性别、国籍等。已有抽取方法通常由序列标注模型对远程监督数据进行训练从而得到抽取模型，但是用该方式在数据上存在标注不准确和不同属性值重合的问题，在模型上缺少可扩展性和可泛化性能力。为解决此问题，该文提出将任务转化为阅读理解问题，通过阅读人物介绍来对人物属性表进行填写补全。为此，本文构造了一份基于阅读理解的文档级人物属性抽取数据集，并采用了基于Transformer算法的双向编码表征模型-机器阅读理解（BERT-MRC）和基于Transformer算法的双向编码表征模型-条件随机场-机器阅读理解（BERT-CRF-MRC）两种基线模型。研究结果表明BERT-CRF-MRC相比于BERT-MRC在F1值上高三个百分点，BERT-CRF-MRC的试验结果在短文本人物介绍中F1平均值约为92%，在长文本人物介绍中F1平均值约为75%。本文的新构建数据和代码已公开在Github上。

Person aspect term extraction aims to extract various attributes of individuals such as gender and nationality from their descriptions. Existing extraction methods typically train sequence labeling models on distantly-supervised data to obtain the extraction model. However, this approach has issues with inaccurate annotations and overlapping different attribute values in the data, and lacks scalability and generalizability in their models. To solve the problems, this article proposes to transform this task into a machine reading comprehension (MRC) problem, that is, to fill in the person attribute-value table by reading the person profile. This paper constructs a person attribute recognition data based on the reading comprehension framework from the person encyclopedia, and constructs two baseline models of bidirectional encoder representations from transformers-machine reading comprehension (BERT-MRC) and bidirectional encoder representations from transformers-conditional random field-machine reading comprehension (BERT-CRF-MRC). Among them, BERT-CRF-MRC is three percentage points higher than BERT-MRC on average in F1 score and the experimental results of BERT-CRF-MRC are about 92% F1 average in short text person profiles while about 75% in long text person profiles. The constructed data and code are exposed on Github.