基于深度信念网络的事件识别

发布时间：2018-04-16 22:41

  本文选题：事件识别 + 深度学习　； 参考：《电子学报》2017年06期

【摘要】：事件识别是信息抽取的重要基础.为了克服现有事件识别方法的缺陷,本文提出一种基于深度学习的事件识别模型.首先,我们通过分词系统获得候选词并将它们分为五种类型.然后选择六种识别特征并制定相应的特征表示规则用来将词转化为向量样例.最后我们通过深度信念网络抽取词的深层语义信息,并由Back-Propagation(BP)神经网络识别事件.实验显示模型最高F值达85.17%.同时,本文还提出了一种融合无监督和有监督两种学习方式的混合监督深度信念网络,该网络能够提高识别效果(F值达89.2%)并控制训练时间(增加27.50%).
[Abstract]:Event recognition is an important basis for information extraction.In order to overcome the shortcomings of existing event recognition methods, this paper presents an event recognition model based on deep learning.First, we obtain candidate words through word segmentation system and divide them into five types.Then six recognition features are selected and corresponding feature representation rules are made to transform words into vector samples.Finally, the deep semantic information of words is extracted by deep belief network, and the event is identified by Back-Propagation BP neural network.The experiment shows that the maximum F value of the model is 85.17.At the same time, this paper proposes a hybrid supervised depth belief network which combines unsupervised and supervised learning methods. The network can improve the recognition effect and control the training time (increase 27.50%).
【作者单位】： 上海大学计算机工程与科学学院;
【基金】：国家自然科学基金项目(No.61273328,No.61305053,No.71203135)
【分类号】：TP18;TP391.1
，

本文编号：1760922