情感信息抽取的资源建设及关键技术研究

One key task in sentiment analysis is sentiment-oriented information extraction which aims to extract various types of valuable information related to opinion from the text. This task benefits a great number of applications in many real situations such as E-commerce and information security. Meanwhile, this task is an important component of the research on deep understanding of sentences. Currently, studies on this task face some important problems such as the shortage of the sentiment-oriented information, difficulty of building the extraction model and domain dependence of the resource. To deal with these problems, we conduct the studies as follows. First, we propose a novel category framework, under which, we build a corpus annotated with various types of sentiment-oriented information from multiple domains.Second, we use shallow semantic parsing to model the task of sentiment-oriented information extraction. Especially, the extraction objects are considered as semantic roles related to opinion expression. The proposed model can well employ the sturcture knowledge from the parse tree and model various types of opinion-related information into a union framework. Third, we present a multi-domain fusion method to fully use the limited resources from avaiable domains to make them help each other. And finally, we propose a hierarchy active learning-based domain adaptation approach to overcome the shortage of labeled data in target domains.

情感信息抽取是情感分析的一个核心任务，旨在从文本中抽取出各种情感信息。该任务在电子商务、信息安全等领域具有广泛的应用需求,同时也是句子级深层次语义理解的一个重要组成部分。针对目前情感信息抽取研究中存在的信息类型匮乏、情感信息建模困难及语料资源领域依赖等问题，本项目将开展以下研究：1）定义包含丰富类型的情感信息分类体系，并在该分类体系下，构建覆盖多个领域一定规模的高质量标注语料；2）使用浅层语义分析模型对多种情感信息类型抽进行联合建模，将情感信息抽取的对象当作情感表达的语义角色，以充分利用各种结构化句法信息；3）提出基于元学习的多领域融合方法以有效集成多个领域的分类信息，充分使用来自所有领域的有限标注资源；4）提出基于层级主动学习的领域自适应情感信息抽取方法，有效解决目标领域中的标注语料缺乏问题。

本课题旨在进行情感信息抽取的资源建设及关键技术研究，重点研究细粒度与粗粒度的情感分类融合方法，并提出了一种基于概率图模型的细粒度情绪分类方法。此外，本课题还提出了一种联合学习方法可以有效提升跨领域情感回归任务的性能，为情感分类的进一步研究打下了坚实的基础。.四年来，课题总体进展顺利，所有研究计划已按照要求完成，达到预期目标。特别是：（1）在细粒度与粗粒度的情感分类融合方法方面，提出了基于整型线性规划方法，能够让情绪分类和情感分类的结果互相帮助，进一步提升各自的分类性能；（2）在细粒度情绪分类任务上，提出了基于概率图模型的分类方法，能充分利用句子的上下文信息及情绪标签之间的依赖信息，提高句子级别情绪分类的性能；（3）在跨领域情感回归任务方面，提出了一种联合学习方法，有效提升情感回归的领域适应性能。.四年来，发表国际顶级会议IJCAI/ACL/EMNLP/COLING/DASFAA论文14篇。