基于自注意力机制的会话式机器阅读理解研究  

A Study of Conversational Machine Reading Comprehension Model Based on the Self-Attention Mechanism

作　　者：李堃 LI Kun(Inner Mongolia Electronic Information Vocational Technical College,Hohhot Inner Mongolia,China)

机构地区：[1]内蒙古电子信息职业技术学院,内蒙古呼和浩特010000

出　　处：《信息与电脑》2024年第22期171-173,共3页Information & Computer

基　　金：基于自注意力机制的会话式机器阅读理解研究(项目编号:KZY2024013)。

摘　　要：当前,会话式机器阅读理解(Conversational Machine Comprehension,CMC)研究主要使用单轮机器阅读理解模型框架构建模型。其模型的输入包括“文章段落”以及由会话历史与当前问题拼接组成的“问题”两部分构成,这使得会话历史与当前问题由同一个模块处理。缺乏单独的会话历史理解模块(History Conversation Understanding,HCU),导致模型无法高效理解和抽取会话历史中与当前问题最相关的语义信息。针对上述问题,文章提出了一种基于自注意力机制的会话历史理解模块,并探索了该模块在SDNet模型的不同位置下的性能表现。结果显示,在基准数据集Co QA下,使用相同训练资源,SDNet-Attention模型的总F1值相较于SDNet模型有所提高。Currently,conversational machine comprehension(CMC)research mainly uses a single-round machine reading comprehension modeling framework to build models.The input to the model consists of a“text passage”and a“question”consisting of a patchwork of conversational history and the current question,which is processed by the same module as the conversation history and the current question.The lack of a separate history conversation understanding(HCU)module makes it impossible for the model to efficiently understand and extract the most relevant semantic information from the conversation history to the current question.To address the above problems,the article proposes a session history understanding module based on a self-attention mechanism and explores the performance of the module under different positions of the SDNet model.The results show that under the benchmark dataset CoQA,using the same training resources,the total F1 value of the SDNet-Attention model is improved compared to the SDNet model.

关 键 词：会话式机器阅读理解 自注意力机制 会话历史理解 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]