面向Web主观性文本意见挖掘研究

Subjective Web text oriented opinion mining aims at automatically mining and analyzing the subjective types of text including blogs, micro blogs and online reviews on the Web. It has become a hot research topic in the fields of intelligent information processing, data mining and computational linguistics etc., and has being confronted with the problems of huge opinion feature space, sparsity of effective features and difficulty of domain-specific sentiment lexicon construction etc. This proposal will focus on the critical techniques of subjective text opinion mining including fine-grained opinion mining model, opinionated targets clustering and automatic domain-specific sentiment lexicon construction. In accordance with the shortage of effective features and unified mining framework, a sequence labeling based unified model for fine-grained opinion mining is proposed. In accordance with the shortage of effective semantic associations in opinionated targets clustering, an adaptive opinionated targets clustering approach based on constrained spectral clustering is also proposed. With regard to the domain-dependent characteristic and shortage of sentiment associations between candidate sentiment terms in domain sentiment lexicon construction, how to automatically acquire domain-specific prior knowledge and enhance the sentiment associations between candidate sentiment terms will be studied. An automatic domain-specific sentiment lexicon construction strategy based on constrained label propagation is also proposed, this method avoids the need for manual annotated domain corpus in existing sentiment lexicon construction methods, so it shows better domain-adaption ability.

面向Web主观性文本意见挖掘通过自动的方法对博客、微博、在线评论等Web主观性文本信息进行挖掘和分析。它是智能信息处理、数据挖掘、计算语言学等领域的前沿性课题，通常面临着情感特征空间巨大，有效特征稀疏、情感词典构建困难等问题。本项目主要研究Web主观性文本细粒度意见挖掘、自适应评价特征聚类和领域情感词典自动构建等关键技术。针对细粒度意见挖掘中存在有效特征稀疏和缺乏统一框架的问题，研究基于序列标注学习的融合多级特征的细粒度挖掘模型。针对评价特征聚类中存在的语义关联信息不足的问题，提出基于约束谱聚类的自适应评价特征聚类算法。针对情感词典自动构建中存在的领域依赖性和情感关联信息不足问题，研究如何自动获取领域先验知识和增强候选情感词之间的情感关联信息，提出基于约束标签传递的领域情感自动构建算法，该方法可解决传统领域情感词典构建中需要人工标注领域数据的问题，具有良好的领域适应性。

面向Web主观性文本意见挖掘通过自动的方法对博客、微博、在线评论等Web主观性文本信息进行挖掘和分析。它是智能信息处理、数据挖掘、计算语言学等领域的前沿性课题,通常面临着情感特征空间巨大,有效特征稀疏、情感词典构建困难等问题。本项目主要研究Web主观性文本细粒度意见挖掘、自适应评价特征聚类和领域情感词典自动构建等关键技术。项目针对细粒度意见挖掘中存在有效特征稀疏和缺乏统一框架的问题，研究基于序列标注学习的融合多级特征的细粒度挖掘模型；针对评价特征聚类中存在的语义关联信息不足的问题，提出基于约束谱聚类的自适应评价特征聚类算法；针对情感词典自动构建中存在的领域依赖性和情感关联信息不足问题，研究如何自动获取领域先验知识和增强候选情感词之间的情感关联信息，提出基于约束标签传递的领域情感自动构建算法，该方法可解决传统领域情感词典构建中需要人工标注领域数据的问题，具有良好的领域适应性。此外，针对传统两阶段方法中主题得分和意见得分的融合问题，完善所提出的将生成模型和混合模型进行结合的博客意见检索方法；结合学术文献检索，提出一种基于本体的文献语义索引，并开展文献情感分析工作；针对由于事件发现算法没有考虑用户偏好,个性化推荐算法没有考虑新闻的事件相关性问题,提出一种基于分布式的新闻事件混合推荐算法；针对数据实体抽取、不平衡性数据分类、Top N推荐、新闻事件意见检索等课题相关问题开展研究。围绕课题目标和任务，课题组已在Journal of Intelligent Information Systems 、Knowledge-based System、Frontiers in Computational Neuroscience 、Future Generation Comp. Syst.、Eng. Appl. of AI、Neurocomputing、Artificial Intelligent Review 等主流SCI期刊录用发表论文10多篇、在WISE、WAIM 、KSEM 等主流国际会议录用发表学术论文10多篇；申请专利4项，培养博士后1名，博士生5名、硕士生10多名，完成了项目预期目标。