基于局部注意力机制的中文短文本实体链接

Entity Linking Based on Local Attention Mechanism for Chinese Short Text

实体链接是加强语义理解和连接知识信息与文本的有效方法,但目前多数模型对上下文语境的精准理解受限于文本长度,面向短文本的实体链接任务存在实体边界识别错误和实体语义理解错误的问题。针对中文短文本的实体链接任务,构建基于局部注意力机制的实体链接模型。在实体消歧的过程中,通过对待消歧文本与实体的知识描述文本进行拼接,将短文本转换为长文本,同时引入局部注意力机制,缓解长距离依赖问题并强化局部的上下文信息。实验结果表明,相比于传统加入BIO标注方法的模型,该模型在CCKS2019和CCKS2020数据集上的F1值分别提升了4.41%和1.52%。

Entity linking is an effective method for strengthening semantic understanding and linking knowledge information to the text.However,the accuracy of most existing models in understanding the context is limited by the text length,so entity linking for short texts suffer from errors in entity boundary recognition and entity semantic understanding.To deal with the entity linking tasks for Chinese short texts,this paper proposes an entity linking model based on local attention mechanism.In the process of entity disambiguation,the short text is transformed into long text by concatenating the to-be-disambiguated text and the knowledge description text of the entity.At the same time,the local attention mechanism is introduced to alleviate the long-distance dependence and strengthen the local context information.Experimental results show that,compared with the traditional model with BIO labeling model,the proposed model increases the F1 value by 4.41%on the CCKS2019 data set and 1.52%on the CCKS2020 data set.