基于多元信息嵌入与协同神经网络的力学实体识别算法  

Mechanics Entities Recognition Algorithm Based on Multi-Meta Information Embedding and Collaborative Neural Network

作　　者：张家熔 苑津莎[1] 许珈宁 罗志宏 ZHANG Jiarong;YUAN Jinsha;XU Jianing;LUO Zhihong(Department of Electronic and Communication Engineering,North China Electric Power University,Baoding 071000,Hebei,China;Power Dispatching Control Center,State Grid Fuxin Power Supply Company,Fuxin 123000,Liaoning,China;Department of Automation,North China Electric Power University,Baoding 071000,Hebei,China)

机构地区：[1]华北电力大学电子与通信工程系,河北保定071000 [2]国网阜新供电公司电力调度控制中心,辽宁阜新123000 [3]华北电力大学自动化系,河北保定071000

出　　处：《计算机工程》2023年第7期125-134,共10页Computer Engineering

基　　金：中央高校基本科研业务费专项资金(2020JG006)。

摘　　要：自动抽取力学问题中的关键实体是力学问题自动解答的重要手段。然而,与开放域相比,力学问题具有大量的专业词汇和较长的实体,其识别难度高,准确率低。针对该问题,基于图注意力网络(GAT)和Transformer编码器提出一种用于力学问题关键实体抽取的实体识别算法。针对汉语的特点设计一个包括字符信息、词汇信息和部首信息的多元信息嵌入用于增强中文句子表示。提出结构图和语境图两个图模型对中文句子进行建模,并设计一种协同架构,该架构使用两个独立的GAT整合多元信息并学习句子的上下文信息。为平衡词汇信息与部首信息对中文字符的影响,提出一种协同Transformer架构,该架构由字符-词汇Transformer与字符-部首Transformer构成,并增加词汇-部首注意力偏置,从而增强模型的识别能力。在自建的数据集上进行多组对比实验,结果表明,在力学问题实体识别任务中,相对于WCGCN算法,该算法在精度、召回率和F1值上分别提高1.92、0.99和1.44个百分点,能够有效提取力学问题中的关键信息。Automatic extraction of key entities is important for solving mechanical problems automatically.However,compared to open domains,mechanical problems entail a larger specialized vocabulary and longer entities,making their recognition difficult and thus lowering the accuracy.To address this issue,an entity recognition algorithm based on a Graph Attention Network(GAT)and Transformer encoder is proposed for key entity extraction in mechanical problems.A Multi-Meta Information Embedding(MMIE)that includes character information,lexical information,and radical information is designed to enhance Chinese sentence representation based on the characteristics of Chinese.Additionally,two graph models,Structural Graph(S-Graph)and Contextual Graph(C-Graph),are proposed to model Chinese sentences.To learn the contextual information of sentences,a collaborative architecture is designed that integrates multiple types of information using two independent GATs.To balance the impact of lexical information and radical information on Chinese characters,a collaborative Transformer architecture is proposed,consisting of a character-lexical Transformer and a character-radical Transformer.A lexical-radical attention bias is added to enhance the recognition ability of the model.Several comparative experiments were conducted on a self-built dataset,and the results show that the proposed algorithm improves precision,recall,and F1 values by 1.92,0.99,and 1.44 percentage points compared to Word-Character Graph Convolution Network(WCGCN)algorithm,respectively,for the entity recognition task in mechanical problems,and the algorithm effectively extracts key information in mechanical problems.

关 键 词：命名实体识别 多元信息嵌入 图注意力网络 Transformer架构 注意力机制 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]