| 摘 要: 目的 针对基于网络评价的产品设计信息挖掘中存在的话题指向不明、语义模糊及人工依赖性强等问题，提出一种融合社交关系与多模型链式处理的话题定位新方法。方法 首先，利用社交关系构建社交网络图。其次，采用Louvain社区发现算法对社交网络进行高模块度聚类，实现以用户为节点的话题聚类。最后，构建TextRank-BERT-SVM多模型链式处理流程，从社区文本中自动提取并筛选出符合研究目标的话题。结果 以汽车外观设计为目标，采集微博平台500个车型的859万条评论数据进行实验验证。社区发现的模块度普遍高于0.73，目标话题提取的准确率达87.9%，显著优于传统基线模型(提升8.76%)。结论 应用该方法无需人工干预，可有效定位大规模社交媒体中的产品设计相关话题，为后续情感意象挖掘提供精准的数据基础。 |
| 关键词: 社交关系网络 多模型融合 深度学习 话题定位 在线评论 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| A Social Relation Network and Multi-model Fusion Based Accurate Topic Localization Approach for Product Design |
|
TaoYuehao, Lin Li
|
Guizhou university
|
| Abstract: Objective Aiming at the problems of unclear topic orientation, semantic ambiguity and high manual dependency in product design information mining based on online reviews, this paper proposes a novel topic localization method integrating social relations and multi-model chain processing. Methods First, a social network graph is constructed using comments, @mentions and co-occurrence relationships between users, breaking through the traditional idea of only analyzing isolated texts. Second, the Louvain community detection algorithm is adopted to perform high-modularity clustering on the social network, automatically aggregating users with similar concerns into topic communities. Finally, a TextRank-BERT-SVM multi-model chain processing pipeline is constructed to automatically extract and filter topics that meet the research objectives from community texts. Taking automobile exterior design as the research target, 8.59 million comment data from 500 vehicle models on the Weibo platform were collected for experimental verification. Results The modularity of community detection was generally higher than 0.73, and the accuracy of target topic extraction reached 87.9%, which was significantly better than the traditional random forest baseline model (an improvement of 8.76%). Conclusion Using this method requires no manual intervention, can effectively locate product design-related topics in large-scale social media, and provides an accurate data foundation for subsequent emotional image mining. |
| Keywords: Social Relation Network Multi-model Fusion Deep Learning Topic Localization Online Reviews |
