Purely character-based language models (LMs) have been lagging in quality on large scale datasets, and current state-of-the-art LMs rely on word tokenization. It has been assumed that injecting the prior knowledge of a tokenizer into the model is essential to achieving competitive results. In this paper, we show that contrary to this conventional wisdom, tokenizer-free LMs with sufficient capacity can achieve competitive performance on a large scale dataset. We train a vanilla transformer network with 40 self-attention layers on the One Billion Word (lm1b) benchmark and achieve a new state of the art for tokenizer-free LMs, pushing these models to be on par with their word-based counterparts.
@article{arxiv.1908.10322,
title = {Bridging the Gap for Tokenizer-Free Language Models},
author = {Dokook Choe and Rami Al-Rfou and Mandy Guo and Heeyoung Lee and Noah Constant},
journal= {arXiv preprint arXiv:1908.10322},
year = {2019}
}