融合ChineseBERT和双向注意力流的中文商品评论方面情感分析

准确分类电商平台中用户评论所包含的多个方面的情感极性，能够提升购买决策的有效性。为此，提出一种融合ChineseBERT和双向注意力流(Bidirectional Attention Flow,BiDAF)的中文商品评论方面情感分析模型。首先，通过融合拼音与字形的ChineseBERT预训练语言模型获得评论文本和方面文本的词嵌入，并采用从位置编码和内存压缩注意力两个方面改进的Transformer来表征评论文本和方面文本的语义信息。然后，使用双向注意力流学习评论文本与方面文本的关系，找出评论文本和方面文本中关键信息所对应的词语。最后，将Transformer和双向注意力流的输出同时输入到多层感知机(Multilayer Perceptron,MLP)中，进行信息级联和情感极性的分类输出。测试结果表明，提出的模型在两个数据集上的准确率分别为82.90%和71.08%，F1分数分别为82.81%和70.98%。

Aspect Sentiment Analysis for Chinese Commodity Reviews Based on ChineseBERT and Bidirectional Attention Flow

Accurately classifying the sentiment polarity of various aspects contained in user reviews in E-commerce platforms can improve the effectiveness of purchase decisions.Therefore,a sentiment analysis model of Chinese product reviews based on ChineseBERT and Bidirectional Attention Flow (BiDAF) is proposed.Firstly,the word embedding of the review text and the aspect text is obtained by the ChineseBERT pre-trained language model that integrates pinyin and glyph,and the semantic information of the review text and the aspect text is represented by the Transformer improved from two aspects of position coding and memory compression attention.Then,the bidirectional attention flow is used to learn the relationship between the review text and the aspect text,and find out the words corresponding to the key information in the review text and the aspect text.Finally,the outputs of Transformer and bidirectional attention flow are simultaneously input into Multilayer Perception (MLP) for information cascade and sentiment polarity classification output.The test results show that the accuracy of the proposed model on the two data sets is 82.90% and 71.08%,respectively,and the F1 scores are 82.81% and 70.98% respectively.