Saved in:
Bibliographic Details
Main Authors: Goriely, Zébulon, Salhan, Suchir, Lesci, Pietro, Cheng, Julius, Buttery, Paula
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2506.18639
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913907736576000
author Goriely, Zébulon
Salhan, Suchir
Lesci, Pietro
Cheng, Julius
Buttery, Paula
author_facet Goriely, Zébulon
Salhan, Suchir
Lesci, Pietro
Cheng, Julius
Buttery, Paula
contents 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 Rényi efficiency for 25 languages.
format Preprint
id arxiv_https___arxiv_org_abs_2506_18639
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ByteSpan: Information-Driven Subword Tokenisation
Goriely, Zébulon
Salhan, Suchir
Lesci, Pietro
Cheng, Julius
Buttery, Paula
Computation and Language
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 Rényi efficiency for 25 languages.
title ByteSpan: Information-Driven Subword Tokenisation
topic Computation and Language
url https://arxiv.org/abs/2506.18639