融入软掩码机制增强数值表示的表格-文本混合问答  

Enhancing Numerical Representation by Fused Soft Mask Mechanism for Hybrid Tabular-textual Question Answering

作　　者：张腾勋 许鸿飞 约瑟夫·范·吉纳比斯 熊德意 昝红英[1] ZHANG Tengxun;XU Hongfei;Josef Van Genabith;XIONG Deyi;ZAN Hongying(School of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450000,China;School of Language Science and Technology,DFKI and Saarland University,Saarland 66123,Germany;School of Computer Science and Technology,Tianjin University,Tianjin 300000,China)

机构地区：[1]郑州大学计算机与人工智能学院,郑州450000 [2]萨尔大学语言科学与技术学院,萨尔66123 [3]天津大学计算机科学与技术学院,天津300000

出　　处：《小型微型计算机系统》2024年第6期1292-1300,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金青年项目(62306284)资助;国家自然科学基金区域联合重点项目(U23A20316)资助;中国博士后科学基金第74批(2023M743189)资助;德国联邦教育与研究部CORA4NLP(01IW20010)资助;云南省重点研发计划项目(202203AA080004)资助;河南省自然科学基金青年科学基金项目(232300421386)资助;河南省科技攻关项目(232102211041)资助。

摘　　要：表格-文本混合问答需要从异质数据中进行数值推理得到答案,当前的研究将问题、表格、文本拼接作为输入,但其中包含太多不相关的数值和文本会降低模型的性能.使用硬掩码机制将噪音去除以增强正确的数值表示会产生错误传播问题,本文提出一种新的软掩码机制对原始输入表示进行缩放,从而增强正确的数值表示,同时减轻错误传播问题,并且该机制可以灵活地运用到各种模型以及任务中.本文在FinQA、ConvFinQA和MultiHiertt数据集上进行实验,结果表明通过软掩码机制增强数值表示可显著提升模型的性能(+1.27/+1.65Exe/Prog Acc、+3.09/+2.96Exe/Prog Acc和+5.29/+3.45EM/F1).本文还通过消融实验分析软掩码相对于硬掩码的优势,以及不同软掩码向量的影响,并讨论了当前表格-文本混合问答的局限性与未来方向.Hybrid Tabular-Textual Question Answering requires numerical reasoning from heterogeneous data to infer answers.Current research concatenates questions,tables,and texts as input,but including too many irrelevant values and texts will reduce the performance of the model.Using a hard mask mechanism to remove noise to enhance the correct numerical representation will cause error propagation problems.This paper proposes a new soft mask mechanism to scale the original input representation,thereby enhancing the correct numerical representation while alleviating the error propagation problem.And it can be flexibly applied to various models and tasks.Experiments on the FinQA,ConvFinQA and MultiHiertt datasets show that enhancing the numerical representation through soft masking can significantly improve the performance of the model(+1.27/+1.65 Exe/Prog Acc,+3.09/+2.96 Exe/Prog Acc and+5.29/+3.45 EM/F1).We also conduct relevant ablation experiments to analyze the advantages of soft mask over hard mask and the effect of different soft masking vectors,and discuss the limitations and future directions of the current hybrid tabular-textual question answering.

关 键 词：自然语言处理 表格-文本问答 软掩码 数值推理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]