Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (11): 140-147.DOI: 10.3778/j.issn.1002-8331.2007-0450

Previous Articles     Next Articles

Analysis of Fine-Grained Commodity Evaluation for Deep Learning Network

KANG Yue, XUE Huizhen, HUA Bin   

  1. School of Science and Technology, Tianjin University of Finance & Economics, Tianjin 300222, China
  • Online:2021-06-01 Published:2021-05-31

面向深度学习网络的细粒度商品评价分析

康月,薛惠珍,华斌   

  1. 天津财经大学 理工学院,天津 300222

Abstract:

Syntactic features and BERT word embedding model is integrated into the deep learning network to achieve the fine-grained commodity evaluation analysis by taking advantage of the BERT pre-training model. A two-stage fine-grained commodity evaluation sentiment analysis model based on deep learning is proposed. ?Firstly, the BILSTM-CRF attention mechanism model which combines syntactic features and BERT word embedding is used to extract commodity entities, attributes, and emotional words in user reviews. Then, the BILSTM model is applied to analyze the sentiment of the extracted results. The F1 value of feature extraction on SemEval-2016 Task 5 and COAE Task3 commodity evaluation dataset reaches 88.2%, which is 4.8 percentage points and 2.3 percentage points higher than that of the BILSTM model and BILSTM-CRF model, respectively. The accuracy of sentiment classification is up to 88.5%, which is 8 percentage points higher than ordinary RNN, and 15 percentage points higher than traditional machine learning methods, such as support vector machine and naive Bayes. To corroborate the deep learning model which integrates syntactic features and BERT word embedding is superior in the sentiment analysis of fine-grained commodity evaluation by analyzing the complexity of the model.

Key words: sentiment analysis, deep learning, BILSTM-CRF model, BERT, attention mechanism

摘要:

利用BERT预训练模型的优势,将句法特征与BERT词嵌入模型融入到深度学习网络中,实现细粒度的商品评价分析。提出一种基于深度学习的两阶段细粒度商品评价情感分析模型,利用融合句法特征与BERT词嵌入的BILSTM-CRF注意力机制模型提取用户评论中的商品实体、属性与情感词;运用BILSTM模型对提取的结果进行情感分析。在SemEval-2016 Task 5和COAE Task3商品评价数据集上的特征提取F1值达到88.2%,分别高出BILSTM模型、BILSTM-CRF模型4.8个百分点、2.3个百分点;情感分类精度达到88.5%,比普通的RNN高出8个百分点,比支持向量机、朴素贝叶斯等传统机器学习方法高出15个百分点。通过模型的复杂度分析,进一步证明融合句法特征与BERT词嵌入后的深度学习模型,在细粒度商品评价情感分析上的优势。

关键词: 情感分析, 深度学习, BILSTM-CRF模型, BERT, 注意力机制