面向武器装备领域的实体及关系抽取方法研究

Research on Entity and Relation Extraction for Weapons and Equipment Field

实体及关系抽取是实现海量数据知识化的关键,而现有实体及关系抽取方法应用于垂直领域时,表现出的效果很难达到实装应用水平。针对武器装备领域,文中在分析该领域文本数据特征的基础上,提出基于预训练模型与规则知识结合的武器装备实体及关系抽取方法,由实体抽取和关系抽取两个阶段组成。在实体抽取阶段,首先,利用BERT+BiLSTM+CRF模型完成武器装备实体的识别;然后,通过规则知识对领域性实体补充抽取。在关系抽取阶段,首先,利用BERT+BiGRU+CNN模型抽取武器装备实体间关系;然后,经过滤调模块对实体间关系抽取结果过滤和调整;最后,设计强领域性的关系抽取规则,用于实体间关系的补充抽取。在仿真数据集上对本文方法评测,结果表明在实体识别和关系抽取上的F1值分别为96.4%和95.1%,与基线相比均提升了约10%。同时,文中提出的实体及关系抽取方法可作为一种通用解决方案,推广至其他垂直领域。

Entity and relation extraction is the key to realize the knowledge of massive data.However,when the existing entity and relation extraction methods are applied to the vertical field,the results can hardly reach the level of practical application.Aiming at the field of weapon equipment,this paper proposes a method of extracting weapon equipment entities and relations based on the combination of pre-training model and rule-based knowledge on the analysis of the characteristics of text datasets in this field.The method consists of two stages:entity extraction and relation extraction.In the entity extraction stage,first,using BERT+BiLSTM+CRF model to complete the identification of weapon and equipment entities,secondly using rule-based knowledge to supplement the extraction of domain entities.In the relation extraction stage,first,the BERT+BiGRU+CNN model is used to extract the relation between the entities of weapons and equipmen,secondly,the filtering and adjustment module is used to filter and adjust the relation extracted in the previous step,finally,the strong domain-oriented rules are designed to extract the sematic relation between entities.The proposed extraction method was evaluated on the simulation data set.The F1 values of entity recognition and relationship extraction in the proposed method are 96.4%and 95.1%,respectively,both of which are improved by about 10%compared with the baseline.Meanwhile,the entity and relationship extraction method proposed in this paper can be used as a general solution and extended to other verticals.