Saved in:
Bibliographic Details
Main Authors: Seo, Jean, Kim, Jaeyoon, Byun, SungJoo, Shin, Hyopil
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2502.12560
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913701909495808
author Seo, Jean
Kim, Jaeyoon
Byun, SungJoo
Shin, Hyopil
author_facet Seo, Jean
Kim, Jaeyoon
Byun, SungJoo
Shin, Hyopil
contents The necessity of language-specific tokenizers intuitively appears crucial for effective natural language processing, yet empirical analyses on their significance and underlying reasons are lacking. This study explores how language-specific tokenizers influence the behavior of Large Language Models predominantly trained with English text data, through the case study of Korean. The research unfolds in two main stages: (1) the development of a Korean-specific extended tokenizer and (2) experiments to compare models with the basic tokenizer and the extended tokenizer through various Next Token Prediction tasks. Our in-depth analysis reveals that the extended tokenizer decreases confidence in incorrect predictions during generation and reduces cross-entropy in complex tasks, indicating a tendency to produce less nonsensical outputs. Consequently, the extended tokenizer provides stability during generation, potentially leading to higher performance in downstream tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2502_12560
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle How does a Language-Specific Tokenizer affect LLMs?
Seo, Jean
Kim, Jaeyoon
Byun, SungJoo
Shin, Hyopil
Computation and Language
The necessity of language-specific tokenizers intuitively appears crucial for effective natural language processing, yet empirical analyses on their significance and underlying reasons are lacking. This study explores how language-specific tokenizers influence the behavior of Large Language Models predominantly trained with English text data, through the case study of Korean. The research unfolds in two main stages: (1) the development of a Korean-specific extended tokenizer and (2) experiments to compare models with the basic tokenizer and the extended tokenizer through various Next Token Prediction tasks. Our in-depth analysis reveals that the extended tokenizer decreases confidence in incorrect predictions during generation and reduces cross-entropy in complex tasks, indicating a tendency to produce less nonsensical outputs. Consequently, the extended tokenizer provides stability during generation, potentially leading to higher performance in downstream tasks.
title How does a Language-Specific Tokenizer affect LLMs?
topic Computation and Language
url https://arxiv.org/abs/2502.12560