电子学报 ›› 2018, Vol. 46 ›› Issue (7): 1691-1699.DOI: 10.3969/j.issn.0372-2112.2018.07.022

• 学术论文 • 上一篇    下一篇

基于Bi-LSTM的维吾尔语人称代词指代消解

田生伟1, 秦越2, 禹龙3, 吐尔根·依布拉音2, 冯冠军4   

  1. 1. 新疆大学软件学院, 新疆乌鲁木齐 830008;
    2. 新疆大学信息科学与工程学院, 新疆乌鲁木齐 830046;
    3. 新疆大学网络中心, 新疆乌鲁木齐 830046;
    4. 新疆大学人文学院, 新疆乌鲁木齐 830046
  • 收稿日期:2017-03-30 修回日期:2017-09-09 出版日期:2018-07-25 发布日期:2018-07-25
  • 通讯作者: 禹龙
  • 作者简介:田生伟,男,1973年出生,新疆乌鲁木齐人,博士,现为新疆大学教授,主要研究领域为计算机智能技术和自然语言处理.E-mail:tianshengwei@163.com;秦越,女,1992年出生,新疆昌吉人,现为新疆大学硕士研究生,主要研究领域为自然语言处理.E-mail:qinyue_xju@163.com
  • 基金资助:
    国家自然科学基金(No.61563051,No.61662074,No.61262064);国家自然科学基金重点项目(No.61331011);新疆维吾尔自治区科技人才培养项目(No.QN2016YX0051)

Anaphora Resolution of Uyghur Personal Pronouns Based on Bi-LSTM

TIAN Sheng-wei1, QIN Yue2, YU Long3, Turgun Ibrahim2, FENG Guan-jun4   

  1. 1. College of Software, Xinjiang University, Urumqi, Xinjiang 830008, China;
    2. College of Information Science and Technology, Xinjiang University, Urumqi, Xinjiang 830046;China;
    3. Network Center, Xinjiang University, Urumqi, Xinjiang 830046, China;
    4. College of Humanities, Xinjiang University, Urumqi, Xinjiang 830046, China
  • Received:2017-03-30 Revised:2017-09-09 Online:2018-07-25 Published:2018-07-25

摘要: 针对维吾尔语人称代词指代现象,提出利用双向长短时记忆网络(Bi-directional long short term memory,Bi-LSTM)的深度学习机制进行基于深层语义信息的维吾尔语人称代词指代消解.首先将富含语义和句法信息的word embedding向量作为Bi-LSTM的输入,挖掘维吾尔语隐含的上下文语义层面特征;其次对维吾尔语人称代词指代现象进行探索,提取针对人称代词指代研究的24个hand-crafted特征;然后利用多层感知器(multilayer perception,MLP)融合Bi-LSTM学习到的上下文语义层面特征与hand-crafted特征;最后使用融合的两类特征训练softmax分类器完成维吾尔语人称代词指代消解任务.实验结果表明,充分利用两类特征的优势,维吾尔语人称代词指代消解的F1值达到76.86%.实验验证了Bi-LSTM与单向LSTM、浅层机器学习算法的SVM和ANN相比更具备挖掘隐含上下文深层语义信息的能力,而hand-crafted层面特征的引入,则有效提高指代消解性能.

关键词: 指代消解, 双向长短时记忆网络, 词向量, 深度学习, 维吾尔语, 自然语言处理

Abstract: Specific to the anaphora phenomenon of Uyghur personal pronouns,a deep learning mechanism of Bi-LSTM (Bi-directional long short term memory) network is proposed,which is based on the deep semantic information to resolve anaphora resolution problem in Uyghur personal pronouns.Firstly,make the word embedding which contain semantic and syntactic information as the input of Bi-LSTM,to excavate the implicit semantic features of Uyghur.Secondly,explore the anaphora phenomenon in Uyghur and extract 24 hand-crafted features.Then,use multilayer perception(MLP)to concatenate hand-crafted features and context semantic features.Finally,two types of features are used to train the softmax classifier to complete the task.The experimental results show that,on the basis of full utilization of the advantages of two types of features,the F1 value of anaphora resolution is 76.86%.It is proved that Bi-LSTM is more capable of mining implicit context semantic information than LSTM、SVM as well as ANN,and the introduction of hand-crafted features can effectively improve the performance.

Key words: anaphora resolution, Bi-LSTM, word embedding, deep learning, Uyghur, natural language processing

中图分类号: