信息工程与人工智能学院知识库

情感分析作为自然语言处理的重要内容，在电子商务快速发展的今天有着广泛应用。然而，传统情感分析利用情感词典的方法，给予整条综合性评论单一情感极性，忽略了对多个不同角度情感的细化分析。因此，为进一步准确高效地分析综合性评论不同方面的情感，本文基于深度学习方法，构建了多通道餐饮评论细粒度情感分析模型，从优化细粒度情感分析模型和特征提取准确率两个角度展开研究。其中，情感分析模型的研究是粒度化餐饮评论情感极性，采用多通道方式提高各粒度情感分析的准确率，进而有针对性的挖掘综合性评论中的潜在价值；优化特征提取的研究是改进Bert预训练模型，构建餐饮评论词向量表，引入开源知识图谱，提高对隐式特征的抽取效果。本文主要工作：

（1）构建了Bert-BiSRU-Att单通道餐饮评论细粒度情感分析模型。针对多粒度多极性的多维度情感分析问题，本文提出了Bert-BiSRU-Att模型，将Bert预训练模型获得的动态词向量输入到单通道BiSRU-Att中，从而获得丰富的语义信息。为下阶段优化情感分析模型创造基础。

（2）构建了Bert-BiSRU-Att-TextCNN多通道餐饮评论细粒度情感分析模型。改变单通道模型输入方式，将词句向量以三通道形式分别输入到BiSRU-Att、TextCNN和句向量通道中，保留各通道提取到的特征信息、主题词与情感词之间的语义关系，最后输出13个粒度的情感极性。实验结果表明，本文所提出的多通道模型较其他模型在细粒度情感分析准确率方面有明显提升。

（3）构建了KW-Bert-BiSRU-Att-TextCNN多通道餐饮评论细粒度情感分析模型。为进一步提高情感分析的准确率，首先优化了预训练模型，引入带有开源知识图谱的预训练模型K-Bert，标注特殊词汇增加额外特征信息，再基于餐饮评论词向量表，借助kd-tree对词向量进行相似度检验，从而高效识别特殊领域的词汇，丰富下游情感分析模型的语义和语法信息；然后将改进后的预训练模型与多通道情感分析模型共同应用于餐饮评论数据集中，从而进一步提高情感分析的准确率；最后通过对比实验，验证本文提出的KW-Bert-BiSRU-Att-TextCNN模型的有效性。

As an important part of natural language processing, emotion analysis is widely used in today's rapid development of e-commerce. However, traditional sentiment analysis utilizes the method of sentiment dictionaries to provide a single emotional polarity for the entire comprehensive comment, neglecting the detailed analysis of multiple emotions from different perspectives. Therefore, in order to further accurately and efficiently analyze the emotions of different aspects of comprehensive reviews, this article constructs a multi-channel fine-grained sentiment analysis model for restaurant reviews based on deep learning methods, and conducts research from two perspectives: optimizing the fine-grained sentiment analysis model and feature extraction accuracy. Among them, the research on sentiment analysis models focuses on granular emotional polarity in food and beverage reviews, using a multi-channel approach to improve the accuracy of sentiment analysis at each granularity, and then targeted exploration of the potential value in comprehensive reviews; The research on optimizing feature extraction is to improve the Bert pre training model, construct a vector table of restaurant comment words, introduce open-source knowledge graph, and improve the extraction effect of implicit features. The main work of this article is:

(1) A Bert-BiSRU-Att single-channel fine-grained emotional analysis model for restaurant reviews was constructed. In order to solve the problem of multi-dimensional emotion analysis with multi granularity and multi polarity, this paper proposes a Bert-BiSRU-Att model, which inputs the dynamic word vector obtained by Bert pre training model into a single channel BiSRU-Att to obtain rich semantic information. Create a foundation for optimizing the emotional analysis model in the next stage.

(2) A Bert-BiSRU-Att-TextCNN multi-channel fine-grained emotional analysis model for restaurant reviews was constructed. Change the input method of the single channel model, input the phrase and sentence vectors into the BiSRU-Att, TextCNN, and sentence vector channels in three channels, preserve the feature information extracted from each channel, and the semantic relationship between the subject word and the emotional word. Finally, output 13 granularity emotional polarity. The experimental results show that the multi-channel model proposed in this paper significantly improves the accuracy of fine-grained emotional analysis compared to other models.

(3) A KW-Bert-BiSRU-At-TextCNN multi-channel fine-grained emotional analysis model for restaurant reviews was constructed. In order to further improve the accuracy of emotion analysis, the pre training model was first optimized, and the pre training model K-Bert with an open source knowledge map was introduced. Special words were labeled to add additional feature information. Then, based on the food and beverage review word vector table, word vectors were tested for similarity using kd-tree, thereby efficiently identifying words in special areas and enriching the semantic and grammatical information of downstream emotion analysis models; Then, the improved pre training model and multi-channel emotional analysis model are applied to the food and beverage review data set to further improve the accuracy of emotional analysis; Finally, through comparative experiments, the validity of the KW-Bert-BiSRU-At-TextCNN model proposed in this paper is verified.