国家重点研发计划 (2020YFB1313601); 国家自然科学基金 (62076174, 61836007)
常识问答是一项重要的自然语言理解任务, 旨在利用常识知识对自然语言问句进行自动求解, 以得到准确答案. 常识问答在虚拟助手或社交聊天机器人等领域有着广泛的应用前景, 且其蕴涵了知识挖掘与表示、语言理解与计算、答案推理和生成等关键科学问题, 因而受到工业界和学术界的广泛关注. 首先介绍常识问答领域的主要数据集; 其次, 归纳不同常识知识源在构建方式、常识来源和表现形式上的区别; 同时, 重点分析并对比前沿常识问答模型, 以及融合常识知识的特色方法. 特别地, 根据不同问答任务场景中常识知识的共性和特性, 建立包含属性、语义、因果、语境、抽象和意图6大类的知识分类体系. 以此为支撑, 针对常识知识数据集建设, 感知知识融合和预训练语言模型的协作机制, 以及在此基础上的常识知识预分类技术, 进行前瞻性的研究, 并具体报告上述模型在跨数据集迁移场景下的性能变化, 及其在常识答案推理中的潜在贡献. 总体上, 包含对现有数据和前沿技术的回顾, 也包含面向跨数据知识体系建设、技术迁移与通用化的预研内容, 借以在汇报领域技术积累的前提下, 为其理论和技术的进一步发展提供参考意见.
Commonsense question answering is an essential natural language understanding task that aims to solve natural language questions automatically by using commonsense knowledge to obtain accurate answers. It has a broad application prospect in areas such as virtual assistants or social chatbots and contains crucial scientific issues such as knowledge mining and representation, language understanding and computation, and answer reasoning and generation. Therefore, it has received wide attention from industry and academia. This study first introduces the main datasets in commonsense question answering. Secondly, it summarizes the distinctions between different sources of commonsense knowledge in terms of construction methods, knowledge sources, and presentation forms. Meanwhile, the study focuses on the analysis and comparison of the state-of-the-art commonsense question answering models, as well as the characteristic methods fusing commonsense knowledge. Particularly, based on the commonalities and characteristics of commonsense knowledge in different question answering task scenarios, this study establishes a commonsense knowledge classification system containing attribute, semantic, causal, context, abstract, and intention. On this basis, it conducts prospective research on the construction of commonsense knowledge datasets, the collaboration mechanism of perceptual knowledge fusion and pre-trained language models, and corresponding commonsense knowledge pre-classification techniques. Furthermore, the study reports specifically on the performance changes in the above models under cross-dataset migration scenarios and their potential contributions in commonsense answer reasoning. On the whole, this study gives a comprehensive review of existing data and state-of-the-art technologies, as well as a pre-research for the construction of cross-data knowledge systems, technology migration, and generalization, so as to provide references for the further development of theories and technologies while reporting on the existing technologies in the field.