多轮对话是人工智能领域的一个重要分支.如何从多轮对话上下文中正确提取与问题相关的核心内容是多轮对话任务的关键问题.现有模型存在辅助任务低效,对全局与局部信息的筛选不够充分,对较短的多轮对话数据学习能力不足等问题.针对上述问...多轮对话是人工智能领域的一个重要分支.如何从多轮对话上下文中正确提取与问题相关的核心内容是多轮对话任务的关键问题.现有模型存在辅助任务低效,对全局与局部信息的筛选不够充分,对较短的多轮对话数据学习能力不足等问题.针对上述问题,本文提出了一种局部信息增强且能够感知对话结构的多轮对话模型(Structure-aware Dialogue Model with Fine-grained Local Information,SAFL).针对子任务训练代价大的问题,提出了随机滑动窗口回复预测任务,在多轮对话上下文中的不同位置与大小的窗口内进行回复预测,充分学习细粒度的局部对话语义.针对信息筛选不够充分的问题,提出了重点局部信息蒸馏机制,借助多门控融合方法从全局和局部信息之中蒸馏出重点信息,提升模型融合效果.针对模型对较短的多轮对话上下文学习能力不足的问题,提出阶段信息学习机制,在微调前加强预训练语言模型对短多轮对话数据的领域学习,降低微调阶段中对短多轮对话的学习难度.此外,SAFL设计了对话结构感知任务在对话结构方面进一步加强模型对对话上下文的理解能力.Ubuntu和E-commerce数据集上的实验结果表明,SAFL模型的总体性能优于对比模型.展开更多
In view of the serious threat to people's health caused by the Coronavirus Disease 2019(COVID-19),according to the relevant spirit of the Notice on Establishing the Cooperation Mechanism of traditional Chinese and...In view of the serious threat to people's health caused by the Coronavirus Disease 2019(COVID-19),according to the relevant spirit of the Notice on Establishing the Cooperation Mechanism of traditional Chinese and Western Medicine in the Prevention and Treatment of COVID-19 and other Infectious Diseases issued by the National Health Commission of the People's Republic of China,National Administration of Traditional Chinese Medicine of the People's Republic of China,on the basis of our early clinical research of using press needle embedding(intradermal acupuncture)to prevent and treat related diseases,the symptoms,etiology and pathogenesis of the disease were analyzed.We try to use the simple and easy intradermal acupuncture therapy to prevent and assist the treatment of the disease and its complications,hoping to make a modest contribution to help patients relieve part of suffering,and also provide advice for the doctors in the treatment of the disease.展开更多
文摘多轮对话是人工智能领域的一个重要分支.如何从多轮对话上下文中正确提取与问题相关的核心内容是多轮对话任务的关键问题.现有模型存在辅助任务低效,对全局与局部信息的筛选不够充分,对较短的多轮对话数据学习能力不足等问题.针对上述问题,本文提出了一种局部信息增强且能够感知对话结构的多轮对话模型(Structure-aware Dialogue Model with Fine-grained Local Information,SAFL).针对子任务训练代价大的问题,提出了随机滑动窗口回复预测任务,在多轮对话上下文中的不同位置与大小的窗口内进行回复预测,充分学习细粒度的局部对话语义.针对信息筛选不够充分的问题,提出了重点局部信息蒸馏机制,借助多门控融合方法从全局和局部信息之中蒸馏出重点信息,提升模型融合效果.针对模型对较短的多轮对话上下文学习能力不足的问题,提出阶段信息学习机制,在微调前加强预训练语言模型对短多轮对话数据的领域学习,降低微调阶段中对短多轮对话的学习难度.此外,SAFL设计了对话结构感知任务在对话结构方面进一步加强模型对对话上下文的理解能力.Ubuntu和E-commerce数据集上的实验结果表明,SAFL模型的总体性能优于对比模型.
基金Supported by 2011 Special Projects in Scientific Research of TCM Profession(201107006)2018 Preventive Treatment of Disease in Tianjin University of Traditional Chinese Medicine(XJ201801)。
文摘In view of the serious threat to people's health caused by the Coronavirus Disease 2019(COVID-19),according to the relevant spirit of the Notice on Establishing the Cooperation Mechanism of traditional Chinese and Western Medicine in the Prevention and Treatment of COVID-19 and other Infectious Diseases issued by the National Health Commission of the People's Republic of China,National Administration of Traditional Chinese Medicine of the People's Republic of China,on the basis of our early clinical research of using press needle embedding(intradermal acupuncture)to prevent and treat related diseases,the symptoms,etiology and pathogenesis of the disease were analyzed.We try to use the simple and easy intradermal acupuncture therapy to prevent and assist the treatment of the disease and its complications,hoping to make a modest contribution to help patients relieve part of suffering,and also provide advice for the doctors in the treatment of the disease.