信息工程与人工智能学院知识库

情感分析可以在文本数据中辨识出用户所表达的情感意向，是自然语言处理中的热门领域。但传统的情感分析大多是基于粗粒度层面的研究，以探讨产品整体情感极性为主，因此无法掌握产品特征及对应的细粒度情感。为弥补传统情感分析的不足，本文从产品特征和情感分类两个角度展开研究。其中，产品特征角度的研究是对产品具体特征进行分析，明确消费者重视的产品属性；情感分类角度的研究是将情感词典进一步细分，不再沿袭传统的情感极性二分类的方法，而是将情感分类划分的更加细致，并计算情感词的情感强度，明确消费者情感的细分倾向。将产品特征与情感细粒度结合，能够更加清楚的了解消费者对产品特征的具体要求，从而帮助商家更好地提供个性化服务。

本文的主要研究内容包括以下四部分：

（1）构建了手机产品关键短语集合。本文基于关键短语结构，以手机特征词典为基础，构建了手机关键短语集合，为手机特征和情感词的高精度提取创造了前提条件。

（2）构建了BERT-LDA模型。首先对BERT（Bidirectional Encoder Representations from Transformers）模型中相似度值域进行训练；然后应用BERT模型进行相似短语的提取；最后将提取结果输入到LDA模型中，获得评价对象和评价词。高质量的相似短语在降低LDA模型困惑度上发挥了极大作用，也有利于提高LDA模型主题提取的准确率。

（3）建立了基于手机特征词的细粒度情感词典。为实现对情感词的分类与情感强度的计算，采用由大连理工大学信息检索研究室整理的中文情感词汇本体库（DUTIR），以本文所建立的手机关键短语集合和手机特征词典为基础，构建了基于手机产品评价特征的细粒度情感词典，并增加了代表“疑惑”的“疑”的情感，构成8类情感词，同时对修饰词与否定词进行扩充。

（4）应用了细粒度情感分析模型。以某品牌的手机评论文本为例，应用BERT-LDA模型，找出消费者关注度较高的特征及情感词，并应用细粒度情感计算方法，计算特征的情感分类与情感强度，从而详细了解消费者对不同特征的情感倾向，有利于店铺提升口碑、增加销量，更好地提供给个性化服务。

Emotion analysis can identify the emotional intention expressed by users in text data, and is a popular area in natural language processing.However, the traditional emotion analysis is mostly based on the coarse-grained research, mainly to explore the overall emotional polarity of the product, so it is impossible to grasp the product characteristics and the corresponding fine-grained emotion.In order to make up for the deficiency of traditional emotion analysis, this paper studies from the perspectives of product characteristics and emotion classification.Among them, the study of product characteristics is to analyze the product attributes that consumers value; the study of emotion classification is to further subdivide the traditional method of emotion polar classification, but to divide the emotion classification more carefully, calculate the emotion intensity of emotion words, and clarify the segmentation tendency of consumer emotion.The combination of product characteristics and emotional grancan have a clearer understanding of consumers' specific requirements for product characteristics, so as to help businesses better provide personalized services.

The main research content of this article includes the following four parts:

(1)Build a collection of key phrases of mobile phone products.Based on the structure of key phrases and the dictionary of mobile phone features, this paper creates the preconditions for the high-precision extraction of mobile phone features and emotional words.

(2)The BERT-LDA model was constructed.First, the similarity value domain in BERT (Bidirectional Encoder Representations from Transformers) model is trained; then BERT model is applied to extract similar phrases; finally, the extraction results are input into LDA model to obtain the evaluation object and evaluation words.High-quality similarity phrases play a great role in reducing the confusion of the LDA model, and also help to improve the accuracy of the LDA model topic extraction.

(3)A fine-grained emotion dictionary based on mobile phone feature words is established.For the calculation of the classification of emotional words and emotional strength, organized by Dalian university of technology information retrieval Chinese emotional vocabulary ontology library (DUTIR), on the basis of the phone key phrases and mobile features dictionary, built based on the evaluation of mobile product emotion dictionary, and increase the "doubt" of "doubt" emotion, constitute eight types of emotional words, and expand the modifier and negative words.

(4)A fine-grained emotion analysis model was applied.With a brand of mobile phone review text, for example, the application of BERT-LDA model, find out the characteristics of consumer attention and emotional words, and apply fine-grained emotion calculation method, calculate the characteristics of emotional classification and emotional intensity, to understand the emotional tendency of different characteristics, to store word of mouth, increase sales, better provide personalized service.