摘要
针对已有方面信息提取方法存在信息利用效率低、易受错误传递影响的问题,提出一种基于编码器-解码器架构的序列标注反馈模型,将文本分类特征分为语义特征、词性特征、依赖特征3类,通过双路编解码、门控机制,将语义、词性、标签等信息多次融合,获取并提高最终特征表示能力;为降低错误传递问题对模型性能的不良影响,提出数据增强与反馈方法,将模型判断错误的样本经变换后生成新样本,反馈到训练样本集合中并融入下一轮训练流程,提高模型对各种语言现象的识别接受能力。在两个数据集上进行实验并与现有方法进行对比,实验结果表明,该方法能有效强化信息利用水平,降低错误传递问题的影响,具有更好的性能。
To solve the problem of low information utilization efficiency and vulnerability caused by error transmission in the existing method of aspect extraction,a sequential label feedback model based on encoder-decoder architecture was proposed.Text classification features were divided into semantic features,part-of-speech features and dependent features,final feature representation was got and enhanced by fusing information such as semantics,part-of-speech,and labels multiple times through dual codec and gate controlling mechanism.To overcome the problem of error transmission,a sample generation mechanism was proposed,and the samples that the model judged wrongly were transformed to generate new samples,they were fed back to the training sample set and integrated into the next training process to improve the model’s ability to recognize and accept various language phenomena.Results of experiments on two data sets and comparison with existing methods show that the proposed method can effectively enhance the level of information utilization,reduce the impact of the problem of error transmission,and has better performance.
作者
范守祥
姚俊萍
李晓军
马可欣
FAN Shou-xiang;YAO Jun-ping;LI Xiao-jun;MA Ke-xin(No.301 Faculty,Rocket Force University of Engineering,Xi’an 710025,China;61486 PLA Troops,Shanghai 200072,China)
出处
《计算机工程与设计》
北大核心
2020年第9期2643-2649,共7页
Computer Engineering and Design
关键词
方面提取
深度学习
数据增强
双路编解码
信息融合
aspect extraction
deep learning
data enhancement
dual codec
information fusion
作者简介
范守祥(1986-),男,辽宁丹东人,硕士研究生,研究方向为智能信息处理技术;通讯作者:姚俊萍(1978-),女,陕西渭南人,博士,副教授,硕士生导师,研究方向为信息系统与数据工程;李晓军(1981-),男,河北秦皇岛人,博士,副教授,硕士生导师,研究方向为数据质量技术、信息系统人因工程;马可欣(1990-),女,辽宁沈阳人,硕士,助理工程师,研究方向为大数据与信息系统。E-mail:junpingy200225@163.com。
