How Do Language Models Acquire Character-Level Information?

TL;DR

This paper investigates how language models implicitly learn character-level information, analyzing the roles of tokenization, orthographic constraints, and semantic associations in this process.

Contribution

The study provides a detailed analysis of mechanisms behind character-level knowledge acquisition in language models, distinguishing tokenization-dependent and independent factors.

Findings

Merge rules and orthographic constraints are primary tokenization-dependent factors.

Semantic associations and syntactic information are key tokenization-independent factors.

Analysis reveals multiple mechanisms contribute to character-level knowledge in LMs.

Abstract

Language models (LMs) have been reported to implicitly encode character-level information, despite not being explicitly provided during training. However, the mechanisms underlying this phenomenon remain largely unexplored. To reveal the mechanisms, we analyze how models acquire character-level knowledge by comparing LMs trained under controlled settings, such as specifying the pre-training dataset or tokenizer, with those trained under standard settings. We categorize the contributing factors into those independent of tokenization. Our analysis reveals that merge rules and orthographic constraints constitute primary factors arising from tokenization, whereas semantic associations of substrings and syntactic information function as key factors independent of tokenization.