问答系统构建及推理研究综述
A Review of Question Answering System Construction and Inference Research
姚奕 1尹瑞江 1陈朝阳1
作者信息
- 1. 中国人民解放军陆军工程大学 指挥控制工程学院,江苏 南京 210000
- 折叠
摘要
近年来对问答系统的研究提高了信息提取的质量,并取得了许多领域内的不错成果.传统方法构建的问答系统满足不了如今的需求,因此结合深度学习模型构建问答系统以提高检索能力成为当前研究的主流.而且面对越来越多限制条件的多跳问题,问答系统需要具备一定推理能力推导出更多的信息以准确的找到答案.该文讨论了基于语义解析和基于信息检索的两种实现问答系统的方法,这两种方法都可以有效地处理单一约束的简单问题,而且结合深度学习模型,可以更好地解决多约束的复杂问题.此外,针对在知识库中多次跳跃的问答,还讨论了基于图神经网络和强化学习等方法的问答推理技术,这些技术可以在知识库中进行多跳推理,补充问答中的缺失信息完成问答任务.最后,对两种构建方法的优缺点进行了总结,并展望了未来问答系统的发展前景.
Abstract
In recent years,the research on question answering system has improved the quality of information extraction,and achieved good results in many fields.The question-answering system constructed by traditional methods cannot meet today's needs,so it has become the mainstream of current research to build a question-answering system to improve retrieval ability by combining deep learning model.In addition,there are more and more multi-hop questions with more and more constraints.The question answering system needs to have certain reasoning ability to deduce more information to find the answer accurately.We discuss two methods of question answering system based on semantic parsing and information retrieval.Both of these methods can deal with simple problems with single constraints effectively,and combined with deep learning model,can solve complex problems with multiple constraints better.In addition,for the question answering with multiple hops in the knowledge base,the question answering reasoning techniques based on graph neural network and reinforcement learning are discussed.These techniques can perform multi-hop reasoning in the knowledge base and complete the question answering task by supplementing the missing information in the question answering.Finally,we summarize the advantages and disadvantages of the two methods,and look forward to the future development of question answering system.
关键词
问答系统/语义解析/信息检索/问答推理/深度学习Key words
question-answering system/semantic parsing/information retrieval/question-answering reasoning/deep learning引用本文复制引用
基金项目
国家社科基金军事学项目(公开)(2021-SKJJ-B-06)
出版年
2023
国家科技期刊平台
NSTL
维普
万方数据