基于序列模型的作战文书知识抽取技术研究

Research on Knowledge Extraction Technology of Combat Document Based on Sequence Model

作战文书具有实体名称复杂多样但结构规范的特点,并且在句子中有大量的重叠实体关系。对于作战文书的知识抽取,现有的方法中采用的流水线模型有误差传播以及关系冗余的问题造成关系抽取能力较差,并且现有的流水线模型无法抽取作战文书中复杂的重叠实体关系。针对这些问题,文中提出了一种基于序列生成模型并结合位置注意力机制的实体与关系联合抽取模型。通过使用作战文书作为数据集并与其他知识抽取模型做对比实验,论文模型既提高了识别非重叠实体关系的准确率,又实现了对重叠实体关系的抽取,从而提高了作战文书知识抽取的整体效果。

Combat documents have the characteristics of complex and diverse entity names but standardized structure,and there are a large number of overlapping entity relationships in the sentence.For the knowledge extraction of combat documents,the pipeline model adopted in the existing method has the problem of error propagation and relationship redundancy,which results in poor relation extraction capabilities,and the existing pipeline model cannot extract complex overlapping entity relationships in com⁃bat documents.In response to these problems,the paper proposes a joint generation model of entities and relationships based on a sequence generation model combined with a location attention mechanism to achieve overlapping entities relation extraction.By us⁃ing combat documents as a data set and other knowledge extraction models for comparison experiments,this paper improves the ac⁃curacy of identifying overlapping entity relationships.