大型语言模型解析被戏谑化语言的惊人能力
计算与语言
2026-05-12 v2
摘要
我们展示了大型语言模型(LLM)在从严重退化的英文文本中恢复意义方面拥有惊人的能力。其中内容词被随机替换为无意义字符串的文本,例如“At the ghybe of the swuint, we are haiveed to Wourge Phrear-gwurr, who sproles into an ghitch flount with his crurp”,可以被翻译成符合语法的英文,且在许多情况下接近原文,例如“At the start of the story, we meet a man, Chow, who moves into an apartment building with his wife.”这些结果表明,结构线索(例如形态语法、封闭类词)对词汇意义的约束作用远超想象。尽管LLM理解“Jabberwockified”英文的能力显然超越人类,但这在理解语言结构方面具有高度意义,并暗示生物或人工语言处理系统可能受益于语法、词汇语义和一般世界知识高度集成的效率。
引用
@article{arxiv.2602.23928,
title = {The Astonishing Ability of Large Language Models to Parse Jabberwockified Language},
author = {Gary Lupyan and Senyi Yang},
journal= {arXiv preprint arXiv:2602.23928},
year = {2026}
}
备注
Submitted to the 2026 Annual Meeting of the Cognitive Science Society