一种基于Seq2Seq框架和文本数据处理的问答对生成方法  

A Method of Generating"Q&A"Based on Seq2Seq Framework and Text Data Processing

作　　者：周艳平[1] 朱小虎 ZHOU Yanping;ZHU Xiaohu(College of Information Science and Technology,Qingdao University of Science&Technology,Qingdao 266061)

机构地区：[1]青岛科技大学信息科学技术学院,青岛266061

出　　处：《计算机与数字工程》2022年第11期2515-2520,共6页Computer & Digital Engineering

摘　　要：为解决从未标记文本中自动化的生成问答对(Q&A),提出了一种基于Seq2Seq框架和自动化构建数据的问答对生成方法(Seq2Seq-TDP-QAG)。首先,已知段落、问题、答案的阅读理解数据集中,通过构造函数从中获得线索、问题类型信息构造训练数据集,并用数据训练基于Seq2Seq框架的问题生成模型;其次,采用三步取样法从未标记文本中获取答案、线索、问题类型信息,构造问题生成模型的输入数据集,并输入模型中初步生成问题信息;最后,通过数据过滤器来控制问题的质量,生成高质量的问题信息,并和答案组成问答对。将提出的方法在SQuAD数据集上进行实验,与PCFG-Trans模型相比ROUGE-L和METEOR分别提升18.68和6.13,结果证明,Seq2Seq-TDP-QAG有能够生成质量更高的问答对。In order to solve the problem of automatically generating question-and-answer pairs(Q&A)from unmarked text,a question-and-answer pair generation method based on the Seq2Seq framework and automated sample construction(Seq2Seq-TDP-QAG)is proposed.Firstly,the reading comprehension data set of known paragraphs,questions,and answers is used to obtain clues and question type information from the constructor to construct the training data set,and use the data to train the question generation model based on the Seq2Seq framework.Secondly,the three-step sampling method is never used.Answers,clues,question type information are obtained from the marked text,the input data set of the question generation model is constructed,and the preliminary generated question information is input into the model.Finally,the data filter is used to control the quality of the question,generate high-quality question information,and a question-and-answer pair is composed with the answer.Experiments with the proposed method on the SQuAD dataset,compared with the PCFG-Trans model,ROUGE-L and METEOR increase 18.68 and 6.13,respectively.The results prove that Seq2Seq-TDP-QAG can generate higher quality question and answer pairs.

关 键 词：问答对生成 Seq2Seq框架 数据处理 问答对质量 

分 类 号：P413[天文地球—大气科学及气象学]