ABSTRACT Amazon, eBay, IMDb as well as several websites provide a convenient platform where users share their opinions on any entities without hindrance. Though those opinions are too many to be examined one by one, this is why a general reputation value will help people make a decision toward a target entity (purchase, download, rent …). This fact makes reputation generation task very challenging because an inaccurate reputation system will directly damage the credibility and popularity of the target entity. This paper aims to improve a recent work that handles the task of generating reputation based on fuzing and mining opinions expressed in natural languages and user feedback ratings. Therefore, we have proposed a hybrid approach that, (i) separates reviews into positive and negative based on their sentiment polarity by applying the two classifiers Naïve Bayes and Linear Support Vector Machine (LSVM), (ii) groups positive and negative reviews into principal opinion sets based on their semantic relations, (iii) calculates a custom reputation value separately for positive and negative groups by considering some statistics of principal opinion sets and finally (iv) computes the final reputation value using Weighted Arithmetic Mean. Experimental results show a significant improvement with respect to recent work.

中文翻译：

---

一种基于观点融合和情感分析的生成声誉的混合方法

摘要 Amazon、eBay、IMDb 以及一些网站提供了一个方便的平台，用户可以在其中不受阻碍地分享他们对任何实体的意见。尽管这些意见太多而无法一一审查，但这就是为什么一般声誉值将帮助人们对目标实体做出决定（购买、下载、出租……）的原因。这一事实使得声誉生成任务非常具有挑战性，因为不准确的声誉系统将直接损害目标实体的可信度和受欢迎程度。本文旨在改进最近的一项工作，该工作处理基于自然语言表达的模糊和挖掘意见和用户反馈评级来生成声誉的任务。因此，我们提出了一种混合方法，(i) 通过应用两个分类器朴素贝叶斯和线性支持向量机 (LSVM)，根据情感极性将评论分为正面和负面，(ii) 根据语义关系将正面和负面评论分为主要意见集，(iii) ) 通过考虑主要意见集的一些统计数据分别为正面和负面群体计算自定义声誉值，最后 (iv) 使用加权算术平均值计算最终声誉值。实验结果表明相对于最近的工作有显着的改进。(iii) 通过考虑主要意见集的一些统计数据，分别为正面和负面群体计算自定义声誉值，最后 (iv) 使用加权算术平均值计算最终声誉值。实验结果表明相对于最近的工作有显着的改进。(iii) 通过考虑主要意见集的一些统计数据，分别为正面和负面群体计算自定义声誉值，最后 (iv) 使用加权算术平均值计算最终声誉值。实验结果表明相对于最近的工作有显着的改进。