统计与数据科学学院知识库

随着互联网的普及和社交媒体的兴起，更多的游客选择撰写在线评论，分享自己的旅游体验和感受。这些互联网平台每天都会产生海量的评论数据，挖掘并分析这些评论数据的潜在情感倾向和偏好信息，可以帮助景区管理者了解游客的态度和需求，对于景区的服务改进和体验优化有着重要价值。

后冬奥时代，冰雪运动加速普及，冰雪产业加快发展，冰雪旅游持续升温，兴起了各项冰雪旅游的热潮。北疆“冰城”哈尔滨以其得天独厚的冰雪资源和丰富多彩的冰雪文化而火爆出圈，成为热度最高的话题之一，推动了当地旅游业的蓬勃发展。冰雪大世界作为这座“冰城”的标志性景点，人气火爆，网络评论数据丰富，选择冰雪大世界作为研究对象并收集在线旅游平台上的游客评论数据。

本研究旨在深入分析哈尔滨冰雪大世界的在线评论，首先采用LDA模型对评论中的正面与负面情感进行主题分析，总结不同主题词下的主题内容，从而揭示游客的旅游偏好，为景区管理者提供具有针对性的改进策略。然后分别使用情感词典、传统机器学习和深度学习的方法对评论数据进行情感分析，并且为提升情感分析的准确性和丰富性，提出了一种BERT-EW双通道情感分析模型。该模型巧妙地将评论的语义特征和情感词特征分开处理，通过语义通道和情感词通道分别提取特征，更有效地捕获评论中更丰富、更精确的情感信息。

根据分析结果，可以得出以下结论：（1）游客的正面评价主要集中在冰雪大世界的观赏性、订票和取票的便利性、门票价格的合理性和优惠政策以及景区交通的便捷性等方面；负面反馈主要针对导游和客服的服务质量、过长的排队等候时间以及门票价格和景区收费高昂等问题。（2）在情感分析中，使用传统机器学习方法要优于情感词典的方法，其中多项式朴素贝叶斯算法的效果较好；使用深度学习的方法进行情感分析效果显著提升，特别是BERT模型的效果远高于Word2Vec词向量模型。（3）本研究提出的基于BERT的双通道情感分析模型BERT-EW，成功地融合了评论的语义信息和情感词信息，在哈尔滨冰雪大世界评论数据上，该模型的表现效果要优于BERT-BiGRU模型。

With the popularity of the Internet and the rise of social media, more and more tourists choose to write online reviews to share their travel experiences and feelings. These Internet platforms generate massive comment data every day. Mining and analyzing the potential emotional tendencies and preferences of these comment data can help scenic spot managers understand the attitudes and needs of tourists, which is of great value for the service improvement and experience optimization of scenic spots.

In the post Winter Olympics era, the popularization of ice and snow sports has accelerated, the development of the ice and snow industry has accelerated, and ice and snow tourism has continued to heat up, leading to the rise of various ice and snow tourism trends. Harbin, known as the "Ice City" in northern Xinjiang, has gained popularity due to its unique ice and snow resources and rich and colorful ice and snow culture, becoming one of the hottest topics and promoting the vigorous development of local tourism. As a landmark attraction of this "ice city", the Ice and Snow World is popular and has rich online review data. We chose the Ice and Snow World as the research object and collected tourist review data on online travel platforms.

This study aims to conduct an in-depth analysis of online comments on Harbin Ice and Snow World. Firstly, the LDA model is used to analyze the positive and negative emotions in the comments, summarize the theme content under different theme words, and reveal tourists' travel preferences, providing targeted improvement strategies for scenic area managers. Then, sentiment analysis was performed on the comment data using sentiment dictionaries, traditional machine learning, and deep learning methods. To improve the accuracy and richness of sentiment analysis, a BERT-EW dual channel sentiment analysis model was proposed. This model cleverly separates the semantic features and emotional word features of comments, extracts features separately through semantic and emotional word channels, and more effectively captures richer and more accurate emotional information in comments.

Based on the analysis results, the following conclusions can be drawn: (1) Positive evaluations from tourists mainly focus on the viewing value of the ice and snow world, the convenience of booking and collecting tickets, the rationality of ticket prices and preferential policies, and the convenience of transportation in scenic areas; Negative feedback mainly targets issues such as the service quality of tour guides and customer service, long waiting times in queues, and high ticket prices and scenic area fees. (2) In sentiment analysis, traditional machine learning methods are superior to sentiment lexicon methods, with polynomial naive Bayes algorithm performing better; The use of deep learning methods for sentiment analysis has significantly improved the performance, especially the BERT model, which is much better than the Word2Vec word vector model. (3) The dual channel sentiment analysis model BERT-EW based on BERT proposed in this study successfully integrates semantic information and sentiment word information of comments. On the Harbin Ice and Snow World comment data, the performance of this model is superior to the BERT-BiGRU model.