How Do Language Models Acquire Character-Level Information?

TL;DR

The paper investigates how language models acquire character-level information despite training on subword units. It uses probing tasks, controlled tokenizer experiments, and input-string transformations to separate tokenization-driven cues (merge rules, orthographic constraints) from tokenization-independent cues (semantic substring associations, syntactic information). Key findings show that both families of cues contribute to character-level knowledge, with orthographic constraints and merge-rule structure playing major roles, and semantic/syntactic cues providing additional, independent support. These insights illuminate the mechanisms by which LMs exhibit spelling-like capabilities and character-level sensitivity, informing tokenizer design and pretraining strategies for more robust language understanding.

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.

Figures (11)

• Figure 1: Factors contributing to character-level information acquisition in LMs.
• Figure 2: Overview of the probing task, which evaluates whether token embeddings encode character-level information in LMs. Ġ is a symbol used in BPE tokenization to indicate the space preceding a word.
• Figure 3: Case where merge (b,c) has higher priority than (c,d).
• Figure 4: Case where merge (c,d) has higher priority than (b,c).
• Figure 5: The case where G$\alpha\beta$, XY$\gamma$ is segmented as G$\alpha\beta$ and XY$\gamma$.