面向问答社区的中文描述性答案融合框架及融合方法研究

The emergence and rise of community-driven question answering (CQA) services has indicated the human beings' demand for knowledge acquisition and sharing. With the development of the CQA portals, the amount of the answers has been growing rapidly. Thus users are not satisfied with the single answer to the given question any longer, but wish to obtain the comprehensive knowledge from the high-quality answers. In this case, it is necessary to fuse the several correct answers to the given question reasonably. In this project, an integrative framework for answer fusion of user generated Chinese descriptive (complex) answers is proposed. Noticing the short-text characteristics of the answers, we propose a uniform strategy for the fusions of the different types of answers by presenting the concept of the semantic information element (SIE) and distributed representation of the answer. Meanwhile, the deep learning based approaches are given to address the content analyzing and semantic relevance quantifying of the answers, together with a method for generating the answer fusion results based on the analysis of the SIEs' relationships. This research is a deep exploration on the semantic analysis and fusion of the short texts, with the studies on the content understanding, abstracting and reconstructing of the short texts, the essential techniques of which can be directly utilized by the information mining and fusing applications on many new kinds of web services.

网络问答社区的出现和兴起体现了人们对知识获取和知识分享的客观需求。随着问答社区的不断发展，答案信息的规模飞速增长，用户已经不再满足于对问题的单一答案的获取，而希望从更高质量的答案中得到全面的知识，因此有必要对同一问题的多个正确答案进行合理的融合。本项目提出了一种面向问答社区的描述性（复杂）答案的融合框架，针对答案信息的短文本语言特点，通过引入语义信息要素和分布式表示，为不同类型答案的融合任务给出了统一的解决策略。同时提出了基于深度学习的答案内容分析和语义相关性量化方法，以及基于语义信息要素关系的融合结果生成方法。本项研究是对短文本语义分析和融合的一次深入探索，研究内容涵盖短文本内容理解、抽象、重构等多个关键问题，其研究成果可以直接服务于多种新型网络信息平台的信息挖掘和融合应用。

本项目提出了一种面向问答社区的描述性（复杂）答案的融合框架。为不同类型答案的融合任务给出了统一的解决策略。首先对答案融合的相关研究尤其是深度学习方法在短文本语义表示方面最新的研究成果进行了深入的分析，给出了答案融合的定义，同时收集大规模CQA答案语料、制定相关的标注标准、获取答案语义信息要素和建立答案分布式表示模型；然后对答案语句之间的关系进行量化、使用注意力机制和问答社区的非文本特征来完成答案自动抽取任务；最终研究基于语义信息要素的语句生成算法、初步融合结果生成算法以及答案融合结果的平滑方法，并给出答案融合的自动评测方法。本项研究是对短文本语义分析和融合的一次深入探索，其研究成果可以直接服务于多种新型网络信息平台的信息挖掘和融合应用。.本项目执行期间共发表期刊和会议论文28篇；培养博士研究生2名，硕士研究生14名；获得黑龙江省自然科学二等奖1项。