How Do Language Models Acquire Character-Level Information?
This addresses a fundamental question in NLP for researchers, but it is incremental as it builds on prior observations without introducing new methods or broad SOTA results.
The paper tackled the problem of understanding how language models implicitly acquire character-level information without explicit training, and found that factors like merge rules and orthographic constraints from tokenization, along with semantic associations and syntactic information independent of tokenization, are key contributors.
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.