Neural Grapheme-to-Phoneme Conversion with Pre-trained Grapheme Models

TL;DR

This paper introduces GBERT, a pre-trained grapheme model inspired by BERT, which enhances neural G2P conversion, especially in low-resource language scenarios, by integrating self-supervised learning with Transformer models.

Contribution

The paper proposes GBERT, a novel pre-trained grapheme model, and demonstrates its effective integration into G2P systems for multiple languages with limited data.

Findings

GBERT improves G2P accuracy in low-resource settings.

Incorporating GBERT via fine-tuning or attention fusion enhances performance.

Experimental results confirm effectiveness across diverse languages.

Abstract

Neural network models have achieved state-of-the-art performance on grapheme-to-phoneme (G2P) conversion. However, their performance relies on large-scale pronunciation dictionaries, which may not be available for a lot of languages. Inspired by the success of the pre-trained language model BERT, this paper proposes a pre-trained grapheme model called grapheme BERT (GBERT), which is built by self-supervised training on a large, language-specific word list with only grapheme information. Furthermore, two approaches are developed to incorporate GBERT into the state-of-the-art Transformer-based G2P model, i.e., fine-tuning GBERT or fusing GBERT into the Transformer model by attention. Experimental results on the Dutch, Serbo-Croatian, Bulgarian and Korean datasets of the SIGMORPHON 2021 G2P task confirm the effectiveness of our GBERT-based G2P models under both medium-resource and low-resource data conditions.