作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (8): 64-71. doi: 10.19678/j.issn.1000-3428.0057138

• 人工智能与模式识别 • 上一篇    下一篇

基于语言学特征与层次注意力机制的幽默识别

杨勇1a, 杨亮2, 邹艳波1b, 任鸽1a, 樊小超1a,2   

  1. 1. 新疆师范大学 a. 计算机科学技术学院;b. 物理与电子工程学院, 乌鲁木齐 830054;
    2. 大连理工大学 计算机科学与技术学院, 辽宁 大连 116024
  • 收稿日期:2020-01-06 修回日期:2020-03-20 发布日期:2020-03-25
  • 作者简介:杨勇(1979-),男,副教授、博士,主研方向为自然语言处理;杨亮,讲师、博士;邹艳波,副教授、博士;任鸽、樊小超(通信作者),讲师、硕士。
  • 基金资助:
    国家语委"十三五"科研规划项目(YB135-8);新疆维吾尔自治区高等学校科研计划(XJEDU2016S066);新疆师范大学博士科研启动基金(XJNUBS1609)。

Humor Recognition Based on Linguistic Features and Hierarchical Attention Mechanism

YANG Yong1a, YANG Liang2, ZOU Yanbo1b, REN Ge1a, FAN Xiaochao1a,2   

  1. 1a. School of Computer Science and Technology;1b. School of Physics and Electronic Engineering, Xinjiang Normal University, Urumqi 830054, China;
    2. School of Computer Science and Technology, Dalian University of Technology, Dalian, Liaoning 116024, China
  • Received:2020-01-06 Revised:2020-03-20 Published:2020-03-25

摘要: 结合英文幽默语言学特征,提出基于语音、字形和语义的层次注意力神经网络模型(PFSHAN)进行幽默识别。在特征提取阶段,将幽默文本表示为音素、字符以及携带歧义性等级信息的语义形式,分别采用卷积神经网络、双向门控循环单元和注意力机制提取PFSHAN模型的语音、字形和语义特征。在特征融合阶段,针对不同单词对幽默语言学特征的贡献程度不同,且不同幽默语言学特征和语句之间关联程度不同的问题,采用层次注意力机制调整不同幽默语言学特征对于PFSHAN模型性能的影响。在Puns和Onliner数据集上的实验结果表明,PFSHAN模型的F1值分别为91.03%和91.11%,能有效提高幽默识别性能。

关键词: 幽默识别, 注意力机制, 卷积神经网络, 双向门控循环单元, 语言学特征

Abstract: This paper proposes a hierarchical attention mechanism neural network model based on the features of pronunciation,font and semantics(PFSHAN) for humor recognition,which extracts the features of English humor linguistics.During the feature extraction stage,the humor texts are presented phoneme,character and semantic information that carries ambiguity level information,and then the features of pronunciation,font and semantics of the PFSHAN model are extracted by using the Convolutional Neural Network(CNN),Bi-directional Gated Recurrent Unit(Bi-GRU),and the attention mechanism.During the feature fusion stage,as words contribute differently to the linguistic features of humors,and the linguistic features of humors are also correlated differently to sentences,the hierarchical attention mechanism is used to adjust the influence of different linguistic features on performance of the PFSHAN model.Experimental results on datasets of Puns and Onliner show that the F1 scores of the PFSHAN model are 91.03% and 91.11% respectively,significantly improving the humor recognition performance.

Key words: humor recognition, attention mechanism, Convolutional Neural Network(CNN), Bi-directional Gated Recurrent Unit(Bi-GRU), linguistic feature

中图分类号: