Large Language Models Lack Understanding of Character Composition of Words

2024/05/18 by Andrew Shin, Kunitake Kaneko, Shin, Andrew +1 · 1 voice · 6 citations
Computer Science · Mathematics · #Character (mathematics) #Composition (language) #Computer science #Linguistics #Mathematics #Natural language processing #Philosophy #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2405.11357

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2024/05/18 · openalex created_date 2024/05/22 · openalex updated_date 2026/07/28

Abstract

Large language models (LLMs) have demonstrated remarkable performances on a wide range of natural language tasks. Yet, LLMs' successes have been largely restricted to tasks concerning words, sentences, or documents, and it remains questionable how much they understand the minimal units of text, namely characters. In this paper, we examine contemporary LLMs regarding their ability to understand character composition of words, and show that most of them fail to reliably carry out even the simple tasks that can be handled by humans with perfection. We analyze their behaviors with comparison to token level performances, and discuss the potential directions for future research.

Cited by

• Spelling Bee Embeddings for Language Modeling
• Prompt-Based One-Shot Exact Length-Controlled Generation with LLMs
• CharBench: Evaluating the Role of Tokenization in Character-Level Tasks
• Spelling-out is not Straightforward: LLMs' Capability of Tokenization from Token to Characters
• StochasTok: Improving Fine-Grained Subword Understanding in LLMs
• The Strawberry Problem: Emergence of Character-level Understanding in Tokenized Language Models

Discussions

• There's now lots of posts about ChatGPT failures to count R's in strawberry but it's a well known limitation, largely caused by their internal units being multi-letter tokens instead of individual let [bsky, 17 points, 2 comments]

Related

• CharBench: Evaluating the Role of Tokenization in Character-Level Tasks
• Spelling-out is not Straightforward: LLMs' Capability of Tokenization from Token to Characters
• What do tokens know about their characters and how do they know it?
• CUTE: Measuring LLMs' Understanding of Their Tokens
• LLMs' Understanding of Natural Language Revealed