查询结果:   白静,李霏,姬东鸿.基于注意力的BiLSTM-CNN中文微博立场检测模型[J].计算机应用与软件,2018,35(3):266 - 274.
中文标题
基于注意力的BiLSTM-CNN中文微博立场检测模型
发表栏目
算法
摘要点击数
809
英文标题
ATTENTION BASED BILSTM-CNN CHINESE MICROBLOGGING POSITION DETECTION MODEL
作 者
白静 李霏 姬东鸿 Bai Jing Li Fei Ji Donghong
作者单位
武汉大学计算机学院 湖北 武汉 430072     
英文单位
School of Computer, Wuhan University, Wuhan 430072, Hubei, China     
关键词
立场检测 微博 神经网络 注意力机制
Keywords
Stance detection Micro-blog Neural networks Attention mechanism
基金项目
作者资料
白静,硕士生,主研领域:自然语言处理,情感分析。李霏,博士生。姬东鸿,教授。 。
文章摘要
针对海量社交网络数据,挖掘其中蕴含的立场信息逐渐成为一个重要的研究方向。第五届自然语言处理与中文计算会议(Nlpcc2016)提出了针对中文微博的立场检测任务。已有的立场检测任务工作中,研究者主要通过手工构建特征工程,添加情感词典和专家知识等方式挖掘语义特征,但这种方式需要花费大量人力在特征设计上。另一些研究者将深度学习应用于立场检测领域,但是没有考虑到句子中不同词对立场倾向有不同影响力。注意力机制由于能够凸显出有价值的特征常常被用于优化神经网络模型。提出一种基于注意力的BiLSTM-CNN中文微博立场检测方法,首先使用双向(Bi-directional)长短期记忆神经网络(LSTM)和卷积神经网络(CNN)分别获取文本表示向量和局部卷积特征,然后通过注意力机制(Attention Mechanisms)在局部卷积特征中加入影响力权重信息,最终将两种特征融合进行分类。针对Nlpcc语料的实验表明,该方法取得了较好的立场检测效果,注意力机制的添加可以有效地提升立场检测的准确性。
Abstract
For a large number of social network data, mining the information contained in the position has gradually become an important research direction. The Fifth Session of Natural Language Processing and Chinese Computing (Nlpcc2016) proposed the stance detection task for Chinese microblogging. In existing studies, researchers always use feature engineering such as adding emotional dictionaries and expert knowledge. But this approach requires a lot of manpower. Other researchers use the deep learning to detect the stance information. But they do not consider the fact that different words in sentences have different influences on position tendencies. Attention mechanisms are often used to optimize neural network models because they highlight the valuable features. This paper presented an attention-based Chinese microblogging position detection method for BiLSTM-CNN. Firstly, the text representation vector and the local convolution feature were obtained by Bi-directional long-short memory neural network (LSTM) and convolutional neural network (CNN) respectively. Then we added influence weight information to the local convolution features through Attention Mechanisms, and finally combined the two features to classify them. Experiments on Nlpcc corpus showed that the method of this article had achieved a good effect of position detection. The addition of attention mechanism could effectively improve the accuracy of position detection.
下载PDF全文