基于图神经网络的问答系统
Question answering systems based on graph neural networks
冯雨溪 1张燮弛 1黄正结 1王琳琳1
作者信息
- 1. 华东师范大学计算机科学与技术学院,上海 200062
- 折叠
摘要
针对目前问答模型对文本语义理解能力不足、缺少常识性知识,且无法有效利用现存海量文本信息的现状,研究一种基于图神经网络的问答系统.通过预训练技术和图注意力算法的有效结合,融合来自常识知识库中的结构化知识,通过准确的知识推理获得对文本深层次语义的理解,最终作出正确的回答.在对话文本和抽象语义文本等数据集上进行的一系列对比实验验证了该问答系统的有效性,表明其具备在对话文本及抽象语义概念上的理解能力.
Abstract
Aiming at the current situation that the question and answer models have insufficient semantic understanding of text,lack commonsense knowledge,and cannot effectively use the existing massive text information,a question and answer system based on graph neural network was studied.Through the effective combination of pre-training technology and graph attention algorithm,the structured knowledge from commonsense knowledge base was integrated,a deeper semantic understanding was obtained after accurate knowledge reasoning,and the correct answers were made.The effectiveness of the proposed question answering system was verified by a series of comparative experiments on datasets such as dialogue text and abstract semantic text,and the ability to understand dialogue text and abstract semantic concepts of this system was indicated.
关键词
图神经网络/预训练范式/知识增强模块/知识推理/问答系统/机器阅读理解/自然语言处理Key words
graph neural networks/pre-training paradigm/knowledge enhancement module/knowledge reasoning/question answering system/machine reading comprehension/natural language process引用本文复制引用
基金项目
上海市扬帆计划(启明星培育扬帆专项)(20YF1411800)
国家自然科学基金(62006077)
出版年
2024
NETL
NSTL
维普
万方数据