联合词汇增强的中文细粒度地理命名实体识别模型研究  

Research on Chinese Fine-grained Geographic Entity Recognition Model based on Joint Lexicon Enhancement

作　　者：李发东 王海起[1] 孔浩然 刘峰[1] 王志海 王琼[1] 徐建波 单宇飞 周啸宇 闫峰 LI Fadong;WANG Haiqi;KONG Haoran;LIU Feng;WANG Zhihai;WANG Qiong;XU Jianbo;SHAN Yufei;ZHOU Xiaoyu;YAN Feng(College of Oceanography and Space Informatics,China University of Petroleum(East China),Qingdao 266580,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,青岛266580

出　　处：《地球信息科学学报》2023年第6期1106-1120,共15页Journal of Geo-information Science

基　　金：山东省自然科学基金项目(ZR2021MD068)。

摘　　要：命名实体识别(NER)是自然语言处理众多研究基础,其可以被定义为分类任务,旨在从非结构化文本中定位出命名实体,同时将命名实体分类成预定义类别。与英文相比,中文构词灵活、不具有边界性,且缺乏高质量中文NER数据集,导致中文命名实体识别难度较大。细粒度实体是粗粒度实体的细分类型,中文细粒度命名实体尤其是地理命名实体识别难度更大。中文地理命名实体识别无法同时兼顾精度和召回率,改善中文细粒度地理命名实体识别性能至关重要。因此,本文提出2种联合词汇增强模型的中文细粒度地理命名实体识别模型。首先,将词汇作为“知识”注入模型,基于词汇增强方式探究适合细粒度命名实体识别方法,并找出适合细粒度命名实体识别方法BERT-FLAT以及LEBERT;其次,为进一步提升细粒度地理命名实体识别性能,针对上述2种方法在预训练模型、对抗训练以及随机权重平均3个方面进行改进,形成联合词汇增强模型RoBERTa-wwm-FLAT以及LE-RoBERTa-wwm;最后,对联合词汇增强模型进行消融实验,探究不同改进策略对于地理命名实体识别性能影响。基于CLUENER数据集和1个微博数据集的实验表明:(1)与无词汇增强功能模型相比,具有词汇增强功能模型在细粒度命名实体识别任务中F1-score提升了10%左右;(2)针对词汇增强方法进行的3处改进使模型在细粒度地理命名实体识别任务中F1-score提升了0.36%~2.35%;(3)与对抗训练改进、随机权重平均改进相比,预训练模型改进对地理命名实体识别精度的影响最大。Named Entity Recognition(NER)is the basis of many researches in natural language processing.NER can be defined as a classification task.The aim of NER is to locate named entities from unstructured texts and classify them into different predefined categories.Compared with English,Chinese have the features of flexible formation and no exact boundaries.Because of the features of Chinese and the lack of high-quality Chinese named entity datasets,the recognition of Chinese named entities is more difficult than English named entities.Fine-grained entities are subdivisions of coarse-grained entities.The recognition of Chinese finegrained named entities especially Chinese fine-grained geographic entities is even more difficult than that of Chinese named entities.It is a great hardship for Chinese geographic entity recognition to take both accuracy and recall rate into account.Therefore,improving the performance of Chinese fine-grained geographic entities recognition is quite necessary for us.In this paper we proposed two Chinese fine-grained geographic entity recognition models.These two models are based on joint lexical enhancement.Firstly,we injected the vocabulary into the experimental models.The vocabulary was considered as the'knowledge'in the models.Then we explored the appropriate fine-grained named entity recognition method based on vocabulary enhancement.And we found two models,BERT-FLAT and LEBERT,that were suitable for fine-grained named entity recognition.Secondly,to further improve the performance of these two models in fine-grained geographical named entities recognition,we improved the above two models with lexical enhancement function in three aspects:pre-training model,adversarial training,and stochastic weight averaging.with these improvements,we developed two joint lexical enhancement models:RoBERTa-wwm-FLAT and LE-RoBERTtawwm.Finally,we conducted an ablation experiment using these two joint lexical enhancement models.We explored the impacts of different improvement strategies on geographic entity reco

关 键 词：命名实体识别 自然语言处理 中文细粒度实体 地理命名实体识别 词汇增强 预训练模型 对抗训练 随机权重平均 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]