Prune or Retrain: Optimizing the Vocabulary of Multilingual Models for Estonian

TL;DR

This paper investigates how vocabulary modification in multilingual models, through retraining or pruning, impacts Estonian NER performance, revealing that pruning maintains performance while retraining can degrade it.

Contribution

It compares vocabulary pruning and retraining approaches for multilingual models, highlighting that pruning preserves performance and reduces computational costs.

Findings

Pruning does not harm NER performance.

Retraining the tokenizer can degrade NER results.

Vocabulary pruning reduces model size and input length.

Abstract

Adapting multilingual language models to specific languages can enhance both their efficiency and performance. In this study, we explore how modifying the vocabulary of a multilingual encoder model to better suit the Estonian language affects its downstream performance on the Named Entity Recognition (NER) task. The motivations for adjusting the vocabulary are twofold: practical benefits affecting the computational cost, such as reducing the input sequence length and the model size, and performance enhancements by tailoring the vocabulary to the particular language. We evaluate the effectiveness of two vocabulary adaptation approaches -- retraining the tokenizer and pruning unused tokens -- and assess their impact on the model's performance, particularly after continual training. While retraining the tokenizer degraded the performance of the NER task, suggesting that longer embedding tuning might be needed, we observed no negative effects on pruning.