【作者】 杨磊；

【导师】 吴偶；

【作者基本信息】 天津大学 ， 计算数学， 2020， 硕士

【摘要】 与传统的信息检索系统相比,问答系统提供了一种更为自然的人机交互方式,使得用户能够通过自然语言的方式与机器交流来获取信息。但在很多的实际应用中,机器需要主动地向用户提问,通过理解用户给出的答案来确定用户的真实意图,这个过程被称作反向问答。然而,当前几乎所有的解决方案都通过强制用户点击选项来避免自动的答案理解,这极大地破坏了用户的体验。针对答案理解任务,文章设计了相应的深度学习模型,旨在为用户提供一种自然的人机交互方式。深度学习是自然语言处理中最为主流的研究方向,而深度学习的成功与大规模的数据密不可分。在此背景下,本文探索了一种自动生成大规模语料的方法,即反向问答中的答案生成。为了更加深入地研究反向问答,我们构建了两个数据集TData和MData。本文重点研究了反向问答中的答案生成和答案理解。在答案生成中,通过给定问题和目标答案标签,来生成对应的答案。在编码部分,我们利用富特征编码器对输入信息进行编码,并利用带有注意力机制的解码器进行解码。对于答案理解,本文充分挖掘问题和答案之间的关联,设计了一个新的神经网络模型。模型包含三个关键模块:问题的骨架抽取,通过无监督的方式来挖掘问题中的关键信息;答案的相关性奖励表示,用来显式地建模问题与答案之间的相关性;多跳注意力的融合机制则为了充分融合已有的信息。对于这两个问题,我们在构建的数据集上验证了提出方法的合理性和有效性。更多还原

【Abstract】 Compared with traditional information retrieval systems,question answering systems provide a more natural way of human-machine interaction,allowing users to obtain information in a natural language.However,in many practical applications,the machine needs to proactively raise questions and understand the user’s answer to determine the user’s needs.This procedure is called reverse question answering(reverse QA).In almost all current solutions,the options are fixed,and human only allowed to select one by clicking buttons,which greatly harm the user’s experience.To this end,this study proposed a novel deep network for answer understanding in reverse QA,which aims to provide users with a natural human-machine interaction method.Deep learning has achieved successes in natural language processing(NLP),and it is based on large-scale data.Therefore,this study explored a method for automatically generating data,that is,answer generation in reverse QA.In order to study the reverse QA conveniently,we constructed two data sets,namely TData and MData.This study focuses on the answer understanding and answer generation in reverse QA.For answer generation,the corresponding answer is generated by giving the question and the label.Feature-rich encoder is used to encode input information,and then decode it with attention mechanism.For answer understanding,this study fully models the correlation between questions and answers.The model including three modules: skeleton extraction for questions,relevance-aware representation of answers,and multi-hop based fusion.The goal of skeleton extraction for questions is mining important information in questions through unsupervised methods.Relevance-aware representation of answers is used to model the relationship of between questions and answers explicitly.Multi-hop based fusion is proposed for fusing the input information more fully.For these two problems,we verify the rationality and effectiveness of the proposed method on the constructed data set.更多还原