基于多知识源融合的级联式常识问答方法  

Cascading Commonsense Question Answering Method Base on Multi-Source Knowledge Fusion

作　　者：李志峰 邹博伟 李烨秋 金志凌 洪宇[1] LI Zhifeng;ZOU Bowei;LI Yeqiu;JIN Zhiling;HONG Yu(School of Computer Science and Technology,Soochow University,Suzhou 215000,China;Institute for Infocomm Research,Singapore 999002,Singapore)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215000 [2]新加坡资讯通信研究院,新加坡999002

出　　处：《山西大学学报（自然科学版）》2022年第2期264-273,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：科技部重大专项课题(2020YFB1313601);国家自然科学基金(62076174)。

摘　　要：常识问答是一项重要的自然语言处理任务,旨在让模型结合常识对给定的问题进行解答。然而,现有模型容易受到知识源中噪声知识的干扰,且对多知识源的融合与利用不够充分,所以问答效率低。为此,文章从筛选知识和利用多知识源的角度出发,提出了由多知识源信息获取和多通道融合推理级联而成的常识问答方法。该方法首先从多个外部知识源检索或匹配知识,再根据贪心筛选策略对噪声知识进行过滤,最后利用多个编码通道对知识和问题进行编码,融合各知识源不同类型的信息并推断出答案。算法在数据集Commonsense QA上的结果表明,本文方法在测试集上取得了78.3%的准确率(Accuracy),与不引入外部知识和引入多种外部知识的两个基线方法相比,分别提升了4.8%和3.0%。表明文章提出的知识筛选策略能有效地过滤噪声,且多通道信息融合推理框架能充分融合多知识源的信息进行常识推理。Commonsense Question Answering(CSQA) is a challenging task in natural language processing,which requires the model to combine knowledge and generate or select the correct answer on a given question.The existing CSQA models often suffer from the interference of noise and do not fully utilize multiple knowledge sources.From the perspective of knowledge filtering and multiple knowledge sources,this paper proposes a cascading Commonsense Question Answering(CSQA) method.This method leverages a greedy screening strategy to filter noise from the two external knowledge sources,and applies a multi-channel knowledge integration framework to combine different types of information and conduct inference.The results on the Commonsense QA dataset show that this method achieves an accuracy of 78.3% on the test set.Compared with the two baseline methods that introduce no external knowledge and multiple knowledge sources,this method improves by 4.8% and 3.0%,respectively.It suggests that the screening strategy can effectively filter noise,and the multi-channel fusion framework can fully integrate external information for common sense reasoning.

关 键 词：常识问答 外部知识 常识推理 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]