Adaptation of Deep Bidirectional Multilingual Transformers for Russian Language

TL;DR

The paper tackles improving Russian NLP by leveraging multilingual pretraining to bootstrap a monolingual Russian model. It demonstrates that initializing a RuBERT from a multilingual BERT, coupled with a Russian-specific subword vocabulary, yields superior performance and faster convergence across paraphrase, sentiment, and QA tasks. Key contributions include the vocabulary reconstruction method, empirical gains over multilingual baselines, and open-sourcing the RuBERT model. This approach offers a practical path to robust Russian language models with reduced training cost and broader accessibility for downstream tasks.

Abstract

The paper introduces methods of adaptation of multilingual masked language models for a specific language. Pre-trained bidirectional language models show state-of-the-art performance on a wide range of tasks including reading comprehension, natural language inference, and sentiment analysis. At the moment there are two alternative approaches to train such models: monolingual and multilingual. While language specific models show superior performance, multilingual models allow to perform a transfer from one language to another and solve tasks for different languages simultaneously. This work shows that transfer learning from a multilingual model to monolingual model results in significant growth of performance on such tasks as reading comprehension, paraphrase detection, and sentiment analysis. Furthermore, multilingual initialization of monolingual model substantially reduces training time. Pre-trained models for the Russian language are open sourced.

Figures (2)

• Figure 1: Distribution of lengths in subtokens of contexts with their questions (SDSJ Task B dataset). Red vertical lines represent mean values.
• Figure 2: Models training dynamics to get to the same value of loss.