基于Self-Attention的多语言语义角色标注联合学习方法  

MULTI-LANGUAGE SEMANTIC ROLE TAGGING JOINT LEARNING METHOD BASED ON SELF-ATTENTION

作　　者：蒲相忠 梁春燕[1] 李鑫鑫 赵磊[1] 王栋 Pu Xiangzhong;Liang Chunyan;Li Xinxin;Zhao Lei;Wang Dong(College of Computer Science and Technology,Shandong University of Technology,Zibo 255049,Shandong,China)

机构地区：[1]山东理工大学计算机科学与技术学院,山东淄博255049

出　　处：《计算机应用与软件》2021年第12期174-178,共5页Computer Applications and Software

基　　金：国家自然科学基金项目(11704229)。

摘　　要：为解决文本语言输出标签序列过于模糊的问题,建立一种相对平稳的级联重排序模式,提出基于Self-Attention的多语言语义角色标注联合学习方法。按照卷积神经网络的框架连接需求,搭建卷积神经网络、处理文本词向量及提取分类特征实施多语言文本词的向量化处理,并根据分类特征的提取行为,完成基于Self-Attention理论的多语言文本分类调节。实验结果表明,该方法的文本语言输出标签序列的模糊性水平明显降低,而级联重显示指标却大幅提升,整个物理排序模式开始逐渐趋于稳定。In order to solve the problem of too vague output tag sequences in text language,a relatively stable cascade reordering mode is established,and a multi-language semantic role tag joint learning method based on self-attention is proposed.According to the framework connection requirements of the convolutional neural network,a convolutional neural network was built,text word vectors was processed,and classification features were extracted to implement vectorization of multilingual text words.Based on the extraction behavior of classification features,the self-attention theory-based multilingual text classification adjustment was completed.Experiments show that the ambiguity level of the text language output tag sequence is significantly reduced,the cascade re-display index is greatly improved,and the entire physical sorting mode has gradually stabilized.

关 键 词：多语言语义 角色标注 联合学习方法 卷积神经网络 

分 类 号：TP391.2[自动化与计算机技术—计算机应用技术]