基于异形同义现象的机器空间语义理解能力评测研究  

Evaluating Machine Capability in Spatial Cognition Through Synonymous Variants

作　　者：詹卫东[1] 秦宇航 肖力铭 Zhan Weidong;Qin Yuhang;Xiao Liming

机构地区：[1]北京大学中文系,北京100871

出　　处：《辞书研究》2024年第5期1-19,I0001,共20页Lexicographical Studies

基　　金：教育部人文社会科学重点研究基地重大项目“面向机器语言能力评测的综合型语言知识库研究”(项目编号22JJD740004)的支持。

摘　　要：不同的方位义词语可以用于表达相同的空间场景。文章考察了汉语空间表达“异形同义”现象的不同类型,并据此制作测试题,作为“异形同义判别”子任务,成为中文空间语义理解能力SpaCE评测基准的一个组成部分。针对大语言模型的评测结果显示,大语言模型在“异形同义判别”任务上与人类水平尚有较大差距,且机器在不同试题上的表现特点也与人类表现有所不同。从空间认知图式的角度讲,大语言模型基于语符分布学习到的人类语言知识,还没有转化为类人的空间认知图式理解能力。In Chinese,different spatial locative expressions can describe the same spatial scene.This phenomenon of“synonymous variants”,which are caused by different factors,can be categorized into various types.As part of the SpaCE benchmark series designed to evaluate machine spatial semantic understanding,a subtask known as“Synonymous Variants Discrimination”is created.Evaluation results on large language models show that there is still a significant gap between their performance on this task and human-level performance.Moreover,the characteristics of machine performance on different types of test items are noticeably different from those of humans.From the perspective of spatial cognitive schemas,the human language knowledge acquired by large language models based on symbol distribution has not yet transformed into human-like understanding capability for spatial information cognition.

关 键 词：空间表达 空间认知 异形同义 机器语言能力评测 大语言模型 

分 类 号：H146[语言文字—汉语]