面向标注数据稀缺专利文献的科技实体抽取

Technology Entity Extraction of Patent Literature with Limited Annotated Data

专利中的科技实体是指专利文献中富含科技信息的词汇,抽取专利中的科技实体对科研工作者提高科研效率、企业布局专利体系都至关重要。提出一种基于半监督学习框架与命名实体识别模型相结合的科技实体抽取方法,半监督学习能够利用无标记数据的优势弥补标注数据稀缺的缺陷,利用大量的专利语料在通用领域的BERT模型基础上进行预训练,得到适用于专利领域的BERT模型BERT-Patent,有效提升模型对专利中科技实体的抽取性能。在专利数据集上的实验结果表明,提出的方法在准确率、召回率、F1值指标上分别提高了6.37%、2.99%、4.63%;在人民日报数据集上准确率、召回率、F1值分别提高了2.87%、1.24%、2.07%。

Technological information contained in patent documents was in the form of vocabulary.These vocabulary was called patent technology entity.Extracting the entity accurately from the patent was crucial for scientists to improve the efficiency of scientific research,and for enterprises to deploy the patent system.A method of extracting scientific and technological entity was proposed based on semi-supervised learning framework and named entity recognition model.It took advantage of semi-supervised learning to make up for the insufficiency of annotated data.At the same time,BERT-Patent model was pre-trained from the generic BERT model over a large patent corpus,in order to improve the feature extraction performance effectively in patent context.The proposed method had superior performance in terms of accuracy,recall rate,and F1 measure;specifically,it was scored 6.37%,2.99%,and 4.63%higher respectively on the patent dataset,and 2.87%,1.24%,and 2.07%higher respectively on People′s Daily dataset.