Evaluating language understanding in Danish LLMs based on semantic dictionaries

Bolette Sandford Pedersen, Nathalie C. Hau Sørensen, S. Olsen, Sanni Nimb
{"title":"Evaluating language understanding in Danish LLMs based on semantic dictionaries","authors":"Bolette Sandford Pedersen, Nathalie C. Hau Sørensen, S. Olsen, Sanni Nimb","doi":"10.7146/nys.v1i65.145394","DOIUrl":null,"url":null,"abstract":"In this paper, we describe how we have generated a number of evaluation datasets – a so-called benchmark – in order to evaluate certain reasoning and understanding capacities of Danish language models. We hypothesize that the semantic knowledge already given in existing Danish dictionaries can be conceived as a ‘ground truth’ for the semantics of the Danish vocabulary. Our method therefore regards turning the semantic dictionaries into a number of evaluation datasets that can be used for testing how well the models understand Danish. More specifically, we examine how well the models i) understand synonymy and semantic association between concepts, ii) make inferences in relation to conceptual knowledge and inheritance structures from super-con-cept to sub-concepts, iii) make correct inferences in relation to acts and events, iv) disambiguate correctly when words have multiple meanings, and v) treat ‘sentiment’, i.e. positive and negative connotation, in running text. We test our datasets on ChatGPT 3.5 turbo and ChatGPT 4.0 and find that the models are challenged in an adequate manner even if ChatGPT 4.0 does perform very well on several of the datasets.","PeriodicalId":509280,"journal":{"name":"NyS, Nydanske Sprogstudier","volume":"77 11","pages":""},"PeriodicalIF":0.0000,"publicationDate":"2024-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"NyS, Nydanske Sprogstudier","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.7146/nys.v1i65.145394","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0

Abstract

In this paper, we describe how we have generated a number of evaluation datasets – a so-called benchmark – in order to evaluate certain reasoning and understanding capacities of Danish language models. We hypothesize that the semantic knowledge already given in existing Danish dictionaries can be conceived as a ‘ground truth’ for the semantics of the Danish vocabulary. Our method therefore regards turning the semantic dictionaries into a number of evaluation datasets that can be used for testing how well the models understand Danish. More specifically, we examine how well the models i) understand synonymy and semantic association between concepts, ii) make inferences in relation to conceptual knowledge and inheritance structures from super-con-cept to sub-concepts, iii) make correct inferences in relation to acts and events, iv) disambiguate correctly when words have multiple meanings, and v) treat ‘sentiment’, i.e. positive and negative connotation, in running text. We test our datasets on ChatGPT 3.5 turbo and ChatGPT 4.0 and find that the models are challenged in an adequate manner even if ChatGPT 4.0 does perform very well on several of the datasets.
基于语义词典评估丹麦语学习者的语言理解能力
在本文中,我们将介绍如何生成一些评估数据集(即所谓的基准),以评估丹麦语模型的某些推理和理解能力。我们假设,现有丹麦语词典中已经给出的语义知识可以被视为丹麦语词汇语义的 "基本事实"。因此,我们的方法是将语义词典转化为一系列评估数据集,用于测试模型对丹麦语的理解程度。更具体地说,我们检验了这些模型在以下方面的能力:i) 理解同义词和概念之间的语义关联;ii) 就概念知识和从上概念到下概念的继承结构进行推断;iii) 就行为和事件进行正确推断;iv) 在单词具有多重含义时正确地进行歧义区分;v) 处理运行文本中的 "情感",即积极和消极内涵。我们在 ChatGPT 3.5 turbo 和 ChatGPT 4.0 上测试了我们的数据集,发现即使 ChatGPT 4.0 在几个数据集上表现非常出色,模型也受到了适当的挑战。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
求助全文
约1分钟内获得全文 求助全文
来源期刊
自引率
0.00%
发文量
0
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:481959085
Book学术官方微信