CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation

TL;DR

CANINE is a character-based neural encoder that eliminates the need for explicit tokenization, using downsampling and deep transformers to efficiently process language data and outperform traditional models on multilingual benchmarks.

Contribution

Introduces CANINE, a tokenization-free encoder operating directly on characters with a novel pre-training strategy and architecture that improves multilingual language understanding.

Findings

Outperforms mBERT by 2.8 F1 on TyDi QA

Uses 28% fewer parameters than comparable models

Operates without explicit tokenization or fixed vocabularies

Abstract

Pipelined NLP systems have largely been superseded by end-to-end neural modeling, yet nearly all commonly-used models still require an explicit tokenization step. While recent tokenization approaches based on data-derived subword lexicons are less brittle than manually engineered tokenizers, these techniques are not equally suited to all languages, and the use of any fixed vocabulary may limit a model's ability to adapt. In this paper, we present CANINE, a neural encoder that operates directly on character sequences, without explicit tokenization or vocabulary, and a pre-training strategy that operates either directly on characters or optionally uses subwords as a soft inductive bias. To use its finer-grained input effectively and efficiently, CANINE combines downsampling, which reduces the input sequence length, with a deep transformer stack, which encodes context. CANINE outperforms a comparable mBERT model by 2.8 F1 on TyDi QA, a challenging multilingual benchmark, despite having 28% fewer model parameters.