ByteSpan: Information-Driven Subword Tokenisation

TL;DR

ByteSpan is a novel subword tokenization method that leverages an external language model to identify predictable byte sequences, creating efficient vocabularies that improve morphological alignment and multilingual compression.

Contribution

It introduces ByteSpan, an information-driven subword tokenizer that groups predictable bytes using an external language model, advancing subword segmentation techniques.

Findings

ByteSpan produces vocabularies with higher morphological alignment than BPE for English.

It achieves similar compression and efficiency across 25 languages.

ByteSpan outperforms traditional methods in morphological and multilingual tasks.

Abstract

Recent dynamic tokenisation methods operate directly on bytes and pool their latent representations into patches. This bears similarities to computational models of word segmentation that determine lexical boundaries using spikes in an autoregressive model's prediction error. Inspired by this connection, we explore whether grouping predictable bytes - rather than pooling their representations - can yield a useful fixed subword vocabulary. We propose a new information-driven subword tokeniser, ByteSpan, that uses an external byte-level LM during training to identify contiguous predictable byte sequences and group them into subwords. Experiments show that ByteSpan yields efficient vocabularies with higher morphological alignment scores than BPE for English. Multilingual experiments show similar compression and R\'enyi efficiency for 25 languages.