计算机应用 ›› 2021, Vol. 41 ›› Issue (7): 1891-1896.DOI: 10.11772/j.issn.1001-9081.2020091492

所属专题: 人工智能

• 人工智能 • 上一篇    下一篇

基于命名实体识别任务反馈增强的中文突发事件抽取方法

武国亮, 徐继宁   

  1. 北方工业大学 电气与控制工程学院, 北京 100144
  • 收稿日期:2020-09-25 修回日期:2020-12-04 出版日期:2021-07-10 发布日期:2020-12-14
  • 通讯作者: 徐继宁
  • 作者简介:武国亮(1996-),男,北京人,硕士研究生,主要研究方向:自然语言处理、事件抽取;徐继宁(1970-),女,陕西西安人,副教授,博士,主要研究方向:自动控制系统、现场总线。
  • 基金资助:
    国家重点研发计划项目(2018YFC0807000)。

Chinese emergency event extraction method based on named entity recognition task feedback enhancement

WU Guoliang, XU Jining   

  1. School of Electrical and Control Engineering, North China University of Technology, Beijing 100144, China
  • Received:2020-09-25 Revised:2020-12-04 Online:2021-07-10 Published:2020-12-14
  • Supported by:
    This work is partially supported by the National Key Research and Development Program of China (2018YFC0807000).

摘要: 针对基于双向长短期记忆网络-条件随机场(BiLSTM-CRF)的事件抽取模型仅能获取字粒度语义信息,可学习特征维度较低致使模型上限低的问题,以开放领域的中文公共突发事件数据为研究对象,提出了一种基于命名实体识别任务反馈增强的中文突发事件抽取方法FB-Latiice-BiLSTM-CRF。首先,将Lattice(点阵)机制融合双向长短期记忆(BiLSTM)网络作为模型的共享层,获取句子中的词语语义特征;其次,增加命名实体识别辅助任务,以联合学习和挖掘实体语义信息,同时将命名实体识别任务的输出反馈到输入端,提取其中实体对应的分词结果作为Lattice机制的外输入,以减少该机制自组词数量大带来的运算负荷并进一步强化对实体语义特征的提取;最后,通过最大化同方差不确定性的最大高斯似然估计方法计算模型总损失,从而解决多任务联合学习产生的损失不平衡问题。实验结果表明,FB-Latiice-BiLSTM-CRF在测试集上的准确率达到81.25%,召回率达到76.50%,F1值达到78.80%,较基准模型分别提升7.63、4.41和5.95个百分点,验证了该方法对基准模型改进的有效性。

关键词: 中文突发事件, 事件抽取, 命名实体识别, 多任务学习, 点阵双向长短期记忆网络, 损失平衡

Abstract: Aiming at the problem that the Bidirectional Long Short-Term Memory network-Conditional Random Field (BiLSTM-CRF) based event extraction model can only obtain the semantic information of character granularity, and the upper limit of the model is low due to the low dimensionality of learnable features, a Chinese emergency event extraction method based on named entity recognition task feedback enhancement was proposed by taking the Chinese public emergency event data in open field as the research object, namely FeedBack-Lattice-Bidirectional Long Short-Term Memory network-Conditional Random Field (FB-Latiice-BiLSTM-CRF). Firstly, the Lattice mechanism was integrated with Bidirectional Long Short-Term Memory network (BiLSTM) as the sharing layer of the model to obtain the semantic features of words in sentences. Secondly, the named entity recognition auxiliary task was added to jointly learn and mine entity semantic information. At the same time, the output of the named entity recognition task was fed back to the input end, and the word segmentation results corresponding to the entities were extracted as the external input of the Lattice mechanism, so as to reduce the computing overhead brought by the large number of self-formed words of the mechanism and further enhance the extraction of entity semantic features. Finally, the total loss of the model was calculated by the maximum Gaussian likelihood estimation method to maximize the homoscedasticity uncertainty, so as to solve the problem of loss imbalance caused by multi-task joint learning. Experimental results show that FB-Latiice-BiLSTM-CRF has the accuracy of 81.25%, the recall of 76.50%, and the F1 value of 78.80% on the test set, which are 7.63, 4.41 and 5.95 percentage points higher than those of the benchmark model, respectively, verifying the effectiveness of the improvement performing to the benchmark model.

Key words: Chinese emergency event, event extraction, named entity recognition, multi-task learning, Lattice-Bidirectional Long Short-Term Memory network (Lattice-BiLSTM), loss balance

中图分类号: