【作者】 罗诚；

【导师】 吕建成；

【作者基本信息】 四川大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 随着深度学习的发展,出现了越来越多的基于深度学习的多轮对话系统,这些系统能根据上下文生成流畅的回复。但通常这类系统生成的回复是不可控且单调简单的,所以越来越多的研究人员已经不满足于仅依赖于上下文而生成的简单安全对话的系统,而是希望引入外部信息,即利用外部知识来生成更多包含真实信息的多样化句子,并同时让生成结果在一定范围内可控。对话系统利用知识进行生成是非常重要的,目前已经有一些方法来让对话系统利用外部知识进行相关对话生成,但其中也存在着一些挑战性。首先,对话系统需要找到合适的知识。其次,知识需要被有效地编码并生成流畅的对话。本文中,我们结合音乐相关数据提出了两个方法来解决上述挑战,并在此基础上搭建了基于音乐知识驱动的对话系统。主要工作如下:(1)为解决选择合适外部知识的挑战,我们提出了一个选择最合适外部知识的方法。该方法是将最佳知识的选择划分为了话题预测与知识选择两部分。针对话题预测,我们提出了一个话题预测器,其使用非深度学习(粗粒度)和深度学习(细粒度)相结合方式,形成一个粗召回和精排序的过程,并使用它们来预测对话话题。其核心是使用粗粒度算法来限制话题范围,再使用细粒度算法对该范围内的话题进行排序,从而获得最合适的话题。针对最佳知识选择,我们提出了一个知识选择器,旨在根据话题预测器过滤后的话题对相关知识进行重排序,从而找出最佳知识。实验结果表明比起传统方法,我们方法有一定的提升。(2)为解决利用外部知识生成流畅对话的挑战,我们提出了一个Bert2Transformer模型作为我们的对话生成器。它可以根据上下文和相关知识生成丰富流畅的对话。实验结果表明,与传统的基线方法相比,我们的模型取得了SOTA结果:在给定适当的知识和没有适当的知识的情况下,我们生成器的平均BLEU分数分别达到35.92和23.24,同时其Distinct-2分数分别达到16.32和15.93。(3)结合上述两个方法,我们部署了一个基于音乐知识驱动的对话系统,并进行了相关测评与分析,实验结果表明,对比其它对话系统,我们的对话系统性能具备极大优势。同时我们构造的相关训练数据将开源。更多还原

【Abstract】 With the development of deep learning,there are more and more deep learning-based multi-round conversation systems that can generate fluent responses based on the context.But usually,such generated responses are uncontrollable and monotonously simple.So more and more researchers are no longer satisfied with systems that generate simple secure conversations that rely only on context.Rather,they want to introduce external information,i.e.,use external knowledge to enable systems to generate more diverse sentences containing real information,and at the same time make the generation controllable within certain limits.It is very important for conversation systems to use knowledge for generation,and there are already some methods to allow conversation systems to use external knowledge for relevant conversation generation,but there are some challenges involved.First,the conversation system needs to find the right knowledge.Second,the knowledge needs to be efficiently coded and generate fluent conversations.In this paper,we propose two methods to address the above challenges by combining musicrelated data.And we build a music knowledge-driven conversation system.The main work is as follows:(1)To address the challenge of selecting appropriate external knowledge,we propose a method for selecting the most appropriate external knowledge.The method is to divide the best knowledge selection process into two parts: topic prediction and knowledge selection.For topic prediction,we propose a topic predictor that uses a combination of non-deep learning(coarse-grained)and deep learning(fine-grained)approaches to form a coarse recall and fineranking process and use them to predict conversation topics.It uses a combination of non-deep learning(coarse-grained)and deep learning(fine-grained)to form a rough recall and fine sorting process,and uses them to predict conversation topics.For the best knowledge selection,we propose a knowledge selector that aims to find the best knowledge by re-ranking the relevant knowledge based on the topics filtered by the topic predictor.The experimental results show an improvement of our method over the traditional method.(2)To address the challenge of generating fluent dialogues using external knowledge,we propose a Bert2 Transformer model as our dialogue generator.It can generate rich and fluent dialogues based on the context and relevant knowledge.The experimental results show that our model achieves SOTA results compared to the traditional baseline approach.The average BLEU scores of our generator reach 35.92 and 23.24 given appropriate knowledge and no appropriate knowledge,respectively,while its Distinct-2 scores reach 16.32 and 15.93,respectively.(3)Combining the above two methods,we build a music knowledge-driven dialogue system based on music knowledge and conducted related measurements and analysis.Experimental results show that our system has a significant performance advantage over other knowledgedriven dialogue systems.Also,the training corpus we constructed will be publicly available.更多还原