基于自监督学习的维基百科家庭关系抽取

Family relation extraction from Wikipedia by self-supervised learning

传统有监督的关系抽取方法需要大量人工标注的训练语料,而半监督方法则召回率较低,对此提出了一种基于自监督学习来抽取人物家庭关系的方法。该方法首先将中文维基百科的半结构化信息——家庭关系三元组映射到自由文本中,从而自动生成已标注的训练语料;然后,使用基于特征的关系抽取方法从中文维基百科的文本中获取人物间的家庭关系。在一个人工标注的家庭关系网络测试集上的实验结果表明,该方法优于自举方法,其F1指数达到77%,说明自监督学习可以较为有效地抽取人物家庭关系。

Traditional supervised relation extraction demands a large scale of manually annotated training data while semisupervised learning suffers from low recall. A self-supervised learning based approach was proposed to extract personal family relationships. First, semi-structured information（ family relation triples） was mapped to the free text in Chinese Wikipedia to automatically generate annotated training data. Then family relations between person entities were extracted from Wikipedia text with feature-based relation extraction method. The experimental results on a manually annotated test family network show that this method outperforms Bootstrapping with F1-measure of 77%, implying that self-supervised learning can effectively extract personal family relationships.