网购产品评论可信计算模型、方法及评价研究

Along with the increasingly prominent role of product reviews data in recommendation, decision making and other economic activities, the phenomenon of false comments will also become more and more salient, the effective recognition of deceptive reviews has gradually attracted much attention. The project will focus on the theoretical model, algorithm and evaluation method of the credibility calculation of product reviews, find abnormal reviews from the angle of both the behavior analysis and content expression analysis. Firstly, based on the domestic and foreign research results, the multi-dimension evaluation system and opinion review deviation model are established and thuas we can make breakthrough from the perspective of reviewer analysis. Secondly, from the perspective of text content, the research is also based on the anomaly detection, the credibility of the subject and the reliability of the method based on the iteration. Then, a deep trust network for multi modality data is constructed to form the theoretical framework of the fusion of product reviews credibility, and thus to evaluate the practical effect of different machine learning methods and characteristics. Finally, the evaluation method of the credibility evaluation method of product reviews from different perspectives is constructed, and the corresponding platform will be built. The research results can further enrich the existing theory and method of false comments, and help to provide real and effective data for management and decision making driven by big review data, and the forecast and early warning of review attack.

随着产品评论大数据在推荐、决策等活动中起到的作用日益突出，虚假评论等现象也相应增多，其有效识别问题逐渐引起重视。项目围绕产品评论可信计算涉及的理论模型、方法及评价等展开，并从行为分析及内容表达两个角度进行评论异常发现。首先在借鉴国内外研究成果的基础上，建立网购产品评论可信计算的多维度指标评估体系及评价偏离模型并重点从用户可信度计算角度进行综合、突破。其次从文本内容角度研究基于细颗粒度情感计算的异常评论偏离发现、可信度主题判别及基于迭代的评论可信度等相关计算模型及方法。然后构建面向多模态数据的深度信任网络以形成产品评论可信度融合计算框架，评价不同机器学习方法和特征的实际作用。最后从不同角度构建产品评论可信度计算的评价方法及进行相应平台建设，为研究展开和效果提升打下基础。研究成果有助于丰富现有虚假评论识别理论与方法，提高评论数据驱动决策、评论攻击预测预警等相关研究的数据源可信甄别能力。

网购产品评论可信计算目前面临的问题主要包括：1) 虚假评论的辨识模型与方法体系需进一步完善；2) 缺乏实验语料和研究平台。为此，本项目主要围绕网购产品评论可信计算的模型、方法以及评价技术展开研究。四年来的主要研究内容主要侧重在：1）基于文本N-gram特征的评论可信计算模型及方法研究，这部分工作主要是在Ott数据集上进行的理论研究和实验工作 ； 2）基于评价偏离的用户可信计算模型；3）基于标记扩展和神经网络相结合的评论可信计算方法；4）基于注意力机制和图模型相结合的用户可信计算模型等。发表论文13篇，其中SCI 6篇，CCF A类会议1篇（ACL 2019），CCF B类会议1篇(COLING 2020)。另有若干论文在审稿中。获得了软件著作权2项。资助项目执行期间培养博士研究生2人、硕士研究生11人。参照了同类研究的做法（Mukherjee et al.，2013；Rayana et al.，2015），基于国内大型点评网站Dianping的过滤结果，整理形成了两个关于餐馆和酒店的大型中文垃圾评论语料，并围绕该语料进行了研究。本项目的研究有助于为相关应用如点评网站虚假评论过滤、基于评论的决策支持、评论异构信息处理与集成等打下坚实的基础。