面向社交群问答对获取的深度学习模型  

DEEP LEARNING MODEL FOR ACQUISITION OF Q&A PAIRS IN SOCIAL GROUPS

作　　者：张民航 蔡东风[1] 李绍鸣 Zhang Minhang;Cai Dongfeng;Li Shaoming(Human-Computer Intelligence Research Center,Shenyang Aerospace University,Shenyang 110135,Liaoning,China)

机构地区：[1]沈阳航空航天大学人机智能研究中心,辽宁沈阳110135

出　　处：《计算机应用与软件》2022年第3期172-179,共8页Computer Applications and Software

基　　金：教育部人文社会科学研究项目(17YJCZH003);辽宁省重点研发计划项目(2019JH2/10100020)。

摘　　要：关注社交群中的问答资源,提出面向社交群的问答对获取方法,主要包括问句识别和答案获取。分析了基于规则和深度学习及结合方法三种问句识别方法的特性;答案获取以深度学习模型为基础,将区分正反例回答同问题的相关度作为学习目标,对各个候选答案与问题的相关度打分排序。引入回答顺序和共现词特征对基础打分作调整进行二次打分排序。实验结果表明,问句识别方法在WebQA、Dbqa和真实小区群聊语料CMY上的F1值分别达到0.930、0.932和0.892;CMY上的问答对获取F1值达到了0.690。Focusing on QA resources in social groups,this paper proposes a QA pair acquisition method oriented to social groups,which includes question recognition and answer acquisition.The characteristics of three different methods including rule-based method,deep learning based method and combined method were analyzed.Answer acquisition was based on a deep learning model,and the correlation between positive and negative example answers and the question was taken as the learning objective,and the correlation between each candidate answer and the question was scored and sorted.The answer order and word co-occurrence features were introduced to adjust the basic scores for secondary scoring and ranking.Experimental results show that the F1 scores of the question recognition method are 0.930,0.932 and 0.892 on WebQA,Dbqa and real community group chatting corpus CMY respectively;the QA pair on CMY achieves an F1 score of 0.690.

关 键 词：问答对获取 问句识别 问答匹配 问答系统 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]