基于深度信念网络的事件识别

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%).

Event Recognition Based on Deep Belief Network

Event recognition is critical to information extraction.To overcome limitations of the exiting event recognition approaches,we proposed an event recognition model based on deep learning (DL-ERM).Firstly,we acquired candidate words through a word segmentation system and classified them into five categories.Then,we selected six recognition feature layers and constructed corresponding feature representation rules to convert words into vector samples.Finally,we employed a deep belief network (DBN) to extract deep semantic features of words,and used a back propagation neural network to identify events.The results of experiments show that the maximum F-measure is 85.17%.Furthermore,we presented a hybrid-supervised DBN,which combines the unsupervised and supervised learning.The novel DBN improves the recognition performance (89.2% F-measure) and effectively controls the training time (increased by 27.50%).