融合位置注意力机制和改进BLSTM的食品评论情感分析-《郑州大学学报(工学版)》

文章信息/Info

Title:: Improved BLSTM Food Review Sentiment Analysis with Positional Attention Mechanisms

Author(s):: Li Yong; Jin Qingyu; Zhang Qingchuan; National Engineering Laboratory of Agricultural Product Quality and Safety Traceability Technology and Application, Beijing Technology and Business University

摘要:: 为了对食品评价的情感倾向进行更加精确的分类，在进行情感语义分析时，卷积神经网络（con-volutional neural networks,CNN)方法在提取目标的结构特征方面具有一定的优势，可以提取到多种局部特征，循环神经网络(recurrent neutal network,RNN)具有记忆功能，在序列特征提取方面具有一定的优势，双向长短时记忆网络(bidirectional long short-term memory,BLSTM)在提取远距离依赖序列语义特征方面可以得到很好的效果。在 BLSTM 的基础上，又引入基于食品领域的语义角色标注与位置相结合的位置注意力机制，来实现距离相关的序列语义特征提取，利用 CNN实现序列语义特征的情感语义分类, 从而构造出了一种基于BLSTM和位置注意力机制的食品评论情感分析模型。实验结果表明，设计的模型在情感分类方面取得了很好的分类效果，与之前的情感分类模型进行比较，在准确率结果上有所提高。

Abstract:: Sentiment analysis is a research hotspot in the field of natural language processing in recent years. However, the current deep learning model lacks the importance of studying the position of emotional words for the whole sentiment analysis in the emotional analysis of text sentences. In the sentiment semantic analysis of e-commerce commodity review data, CNN has certain advantages in extracting the structural features of the target, and can extract a variety of local features. RNN has memory function and has certain advantages in sequence feature extraction. Bidirectional Long Short-Term Memory (BLSTM) can achieve good results in extracting remote-dependent sequence semantic features. Based on BLSTM, this paper introduces a positional attention mechanism based on the combination of semantic role labeling and location in the food field. The distance-related sequence semantic feature extraction is realized, and the sentiment semantic classification of sequence semantic features is realized by CNN, and a food comment sentiment analysis model based on BLSTM and positional attention mechanism is constructed. Experimental results show that the model designed in this paper It has achieved good results in emotional classification, and has improved the accuracy rate results compared with the previous sentiment classification model