基于LSTM网络的评价对象和评价词抽取  

Aspect and polarity words extraction based on LSTM Network

作　　者：李盛秋 赵妍妍[1] 秦兵[1] 刘挺[1] 

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001

出　　处：《智能计算机与应用》2017年第5期95-97,共3页Intelligent Computer and Applications

摘　　要：随着网络电商的兴起,越来越多的用户在网上对商品进行评价。所谓评价对象与评价词,指的是在产品评论中用户所表达的态度的对象以及所使用的极性词。本文将评价对象与评价词的抽取看做为序列标注任务,并介绍了如何使用循环神经网络解决序列标注任务。同时,还说明了循环神经网络的局限性,并使用LSTM对评价对象与评价词进行抽取。最后,本文还与基于规则的双向传播算法进行了对比。从实验结果可以看出,LSTM在评价对象与评价抽取任务上的性能较双向传播算法来说有非常大的提高。As more and more users do shopping online,product reviews become a big resource for NLP research. Aspects are the opinion targets in the reviews,and the polarities are the attitude of users towards aspects. The paper treats the aspect and polarity words extraction as a sequence labeling task,and introduces how to use Recurrent Neural Network to solve the sequence labeling task. Also,the paper points out the shortcoming of RNN and uses LSTM instead to extract the aspects and polarities. Finally,the results of experiments show that LSTM model can significantly improve the performance in contrast to the Double Propagation algorithm,a rule-based algorithm for aspects and polarities extraction.

关 键 词：循环神经网络 LSTM模型 评价对象 双向传播算法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]