基于评论数据的大规模商品属性级观点知识挖掘与管理

Users' opinion expressed in massive review data is crucial for Word of Mouth about Products on the Web, and has significant value for both merchants and customers. However, the opinion's expressions depend on the context and usually are contained in unstructured texts, which brings a challenge for mining users' opinion as well as the systemic management. In this proposal, we focus on the large-scale review data in Web environment, and study the techniques on mining and managing the fine-grained opinion knowledge. Our goal is to construct a high quality opinion knowledge graph for products which includes the objective product taxonomy and the subjective user opinion. Firstly, we devise an extraction method based on ensemble learning to co-extract product aspects and opinion words from reviews, in order to have a high precision and recall for harvesting product opinion knowledge. Then, we study the method on constructing automatically a context-aware sentiment lexicon and the aspect-level opinion indentification technique by introducing the important features such as users' social network relationship, so that to address the problem of opinion expression depends on product domains and aspects. Finally, we explore a scalable knowledge representation model to represent the large-scale diverse product opinion knowledge, to achieve a structured management for the knowledge. In addition, we develop a prototype system to analyse and demonstrate the effectiveness of the proposed solutions. Our research outputs will be help to improve the usability of product opinion knowledge significantly in term of the depth and breadth of applications, to provide high quality information services for users.

海量评论数据中蕴含的大量用户观点对商品的网络口碑有决定性影响，对商户和顾客有重要的参考价值。然而，用户的观点表达依赖于上下文且隐藏在无结构的文本中，从评论数据中准确获取用户观点并进行系统性管理极具挑战。本项目以Web上大规模评论数据为研究对象，以构建涵盖客观性商品分类和主观性用户观点的高质量商品观点知识图为目标，系统地研究细粒度的商品观点知识挖掘及其结构化管理。首先，针对商品观点知识挖掘的高准确率和高召回率挑战，研究基于集成学习的商品属性和观点词协同抽取技术；然后，探讨上下文敏感的情感词典自动构造技术，并结合用户社交关系等特征研究属性级的观点识别技术，解决观点表达的领域/属性依赖问题；最后，设计高可扩展的知识表示模型来组织海量多样性商品观点知识，实现这类知识的结构化管理。此外，实现原型系统分析验证提出方法的有效性。研究成果将有助提高商品观点知识的应用深度和广度，为用户提供高质量的信息服务。

海量评论数据中蕴含的大量用户观点对商品的网络口碑有决定性影响，对商户和顾客有重要的参考价值。然而，用户的观点表达依赖于上下文且隐藏在无结构的文本中，从评论数据中准确获取用户观点并进行系统性管理极具挑战。本项目以Web上大规模评论数据为研究对象，以构建涵盖客观性商品分类和主观性用户观点的高质量商品观点知识图为目标，系统地研究细粒度的商品观点知识挖掘及其结构化管理。首先，针对商品观点知识挖掘的高准确率和高召回率挑战，研究观点目标-观点词的协同抽取技术，并探讨了众包计算技术应用于观点数据集样本标注任务；然后，研究用户评论质量分析和用户观点挖掘技术，解决用户观点表达的领域依赖问题；最后，针对商品知识的海量性和异质性特点，研究基于RDF的大规模商品知识表示框架，并研究了基于数据平衡的知识组织策略，实现海量商品知识的结构化管理和高效的分布式查询处理。此外，项目组成员构建一个大规模的商品知识图谱，并在所提出技术的基础上研发了一个面向海量商品知识管理的原型系统，实验证明该系统在针对商品知识的查询处理相对于现有的知识管理系统在查询效率上有较大的提升。本项目获得的研究成果将有助提高商品观点知识的应用深度和广度，为用户提供高质量的信息服务。