How do different tokenizers perform on downstream tasks in scriptio continua languages?: A case study in Japanese

TL;DR

This study systematically evaluates how different combinations of tokenizers affect the performance of pretrained language models on various downstream tasks in Japanese, a scriptio continua language, revealing task-specific optimal tokenizers and preferred subword methods.

Contribution

It provides a comprehensive analysis of tokenizer combinations in Japanese, identifying optimal pairs for different tasks and recommending subword methods for better downstream performance.

Findings

Different tasks have different optimal morphological analyzers.

Byte-Pair-Encoding and Unigram outperform WordPiece as subword tokenizers.

Using comprehensive tokenizer sets improves understanding of downstream task performance.

Abstract

This paper investigates the effect of tokenizers on the downstream performance of pretrained language models (PLMs) in scriptio continua languages where no explicit spaces exist between words, using Japanese as a case study. The tokenizer for such languages often consists of a morphological analyzer and a subword tokenizer, requiring us to conduct a comprehensive study of all possible pairs. However, previous studies lack this comprehensiveness. We therefore train extensive sets of tokenizers, build a PLM using each, and measure the downstream performance on a wide range of tasks. Our results demonstrate that each downstream task has a different optimal morphological analyzer, and that it is better to use Byte-Pair-Encoding or Unigram rather than WordPiece as a subword tokenizer, regardless of the type of task.