主要针对中文网页语料研究人物关系抽取,提出一种融合最近距离和人名窗口信息的人物关系抽取方法.首先利用远程监督的方法构建人物关系库,提取关系候选语料,通过打分函数过滤掉语料中的噪音数据以提高语料质量;然后在卷积神经网络中引入最近距离,将词与人名之间的距离信息加入到网络中;在循环神经网络中以人名窗口内词向量代替整句词向量作为网络的输入.最后融合两部分网络信息并对网络模型进行训练.结果显示,该方法比传统基于SVM的中文人物关系抽取方法和一些其他的神经网络模型F1值提高3个以上百分点.%Based on the existing methods,this paper proposed a deep learning based mode to extract character relations from the Chinese web corpus.A character extraction strategy incorporated the nearest distance and name entity window information.Firstly,in the pre-processing stage,we utilized the remote-supervised method to construct the character relation set,then a carefully designed scoring function is introduced to improve the quality of candidate corpus.In the model learning stage,we integrated the nearest distance information between term and name entities into the convolutional neural network.Besides,instead of utilizing the embedding vector of the entire sentence as the input of recurrent neural network,we generated the input vector based on the terms located in a fixed window size around the name entity.The mentioned CNN and RNN models are learned under a unified framework.The results show that comparing to the traditional method based on SVM model and other methods based on neural network,the proposed framework improved more than three percent in F1 value.
展开▼
