Abstract: The sudden outbreak of COVID-19 has stimulated consumers’ online purchasing behavior, resulting in an explosive growth in the scale of pharmaceutical e-commerce transactions and an increasingly rich content of online comments on pharmaceutical e-commerce. The online reviews of pharmaceutical e-commerce contain a variety of information, including not only the overall star rating of consumers’ purchasing experience, but also detailed text comments, which hide consumers’ subjective feelings and consumption needs for product purchases. In order to promote the healthy development of pharmaceutical e-commerce and better meet the medication demand of consumers during the epidemic, it is urgent to carry out research on the demand of pharmaceutical online consumers under the COVID-19. From a theoretical perspective, this study focuses on online reviews of pharmaceutical e-commerce, and expands the application fields of text mining methods. From a practical perspective, this article studies the information contained in online reviews of pharmaceutical e-commerce, which can help pharmaceutical e-commerce better catch the sour spot consumer demand, timely identify problems in operating pharmaceutical e-commerce platforms, provide practical suggestions for platform operation and development, and improve consumer purchasing experience and service quality.
   This study uses the Python crawler tool to collect online comment data from a certain pharmaceutical e-commerce platform in 2019, 2020, and 2021, and captures a total of 176602 data from 17 categories of products. By processing data cleaning, word segmentation, and word frequency statistics, high-frequency words in online reviews of pharmaceutical e-commerce are extracted and displayed through word cloud maps. Then, the LDA theme model is used to further analyze the semantic relationships behind high-frequency words, in order to better understand the connections between high-frequency words. By summarizing each theme, the concerns and needs of consumers are clarified. Next, we construct a sentiment analysis model to classify emotions in online comments. The first step is to calculate the sentiment value of the text based on the Boson NLP sentiment dictionary. The second step is to train text at the word and word levels based on the BERT model beforehand. The third step is to connect the sentence vectors obtained in the first two steps and input the new sentence vectors into the SVM classifier for classification. The fourth step is to test the emotional classification performance of this model. The fifth step is to perform sentiment classification on all online comment data, including sentiment classification for individual text comments and individual topics. This study focuses on analyzing online reviews of negative emotions, as negative reviews often contain more suggestions related to products or services, which can help pharmaceutical e-commerce understand consumer sour spot and improve service levels.
   The main conclusions of this study are as follows: Firstly, consumers always pay attention to the effectiveness of medication use, logistics services, product prices, platform reliability and safety when purchasing pharmaceutical products online. Secondly, by comparing and analyzing the high-frequency words in online comments throughout this three-year epidemic, it can be found that before the COVID-19 broke out in 2019, consumers paid great attention to service attitude, commodity brand and purchase convenience. After the COVID-19 just broke out in 2020, consumers paid more attention to cold and vitamin medications, which may be because these medications help to prevent, control and cure COVID-19. In addition, the outbreak of the epidemic will affect consumers’ medication purchase decisions, and gradually cultivate consumers’ habit of purchasing medications online. In the late stage of the epidemic in 2021, consumers were more concerned about the cost-effectiveness of goods, the speed of purchase, and the quality of medication. Thirdly, consumers generally show a positive emotional attitude towards purchasing medications on pharmaceutical e-commerce platforms, with over 85% of positive comments. Fourthly,the negative online comments are mostly about medication prices,efficacy,quality, purchase of prescription, platform reliability, logistics packaging, and purchasing experience during the epidemic.
   Although this study has achieved certain research results, there are still certain limitations. For example, taking online reviews of pharmaceutical e-commerce as the research object, the amount of online review data obtained from this pharmaceutical e-commerce is sufficient but the subject is single. In addition, the emotional analysis model lacks comparative experiments to further verify the superiority of the model, and these issues can be continuously explored and improved in subsequent research.

Key words: online review, text mining, sentiment analysis, LDA topic model, COVID-19

摘要： 基于新冠肺炎疫情下医药电商交易规模的爆炸式增长,对医药电商在线评论进行文本分析,以某B2C医药电商平台2019—2021年在线评论数据为样本,利用LDA主题模型提取在线评论蕴含的主题,并构建情感词典融合深度学习的情感分析模型,对评论和主题词进行情感分析。研究结果显示:1)消费者网购医药商品始终关注平台的可靠性、物流服务、商品价格、药品的使用效果;2)新冠肺炎疫情爆发之前,消费者对服务态度、商品品牌、购买便捷性有很大关注度;疫情爆发后对感冒类和维生素类药品关注度更高,疫情的爆发会影响消费者的购药决策;后疫情时代,消费者更关注商品性价比、购买快捷性以及药品的品质;3)消费者对于在医药电商平台进行购药整体上表现出积极正面的情感态度;4)负面在线评论主要集中在价格、药效、处方药购买、虚假宣传、物流包装、限购等方面。本研究挖掘出疫情下消费者对于网购医药商品的需求重点和痛点,对医药电商平台改善服务质量提供建设性意见。

关键词: 在线评论, 文本挖掘, 情感分析, LDA主题模型, COVID-19