BERT, mBERT, or BiBERT? A Study on Contextualized Embeddings for Neural Machine Translation (2109.04588v1)

Abstract: The success of bidirectional encoders using masked LLMs, such as BERT, on numerous natural language processing tasks has prompted researchers to attempt to incorporate these pre-trained models into neural machine translation (NMT) systems. However, proposed methods for incorporating pre-trained models are non-trivial and mainly focus on BERT, which lacks a comparison of the impact that other pre-trained models may have on translation performance. In this paper, we demonstrate that simply using the output (contextualized embeddings) of a tailored and suitable bilingual pre-trained LLM (dubbed BiBERT) as the input of the NMT encoder achieves state-of-the-art translation performance. Moreover, we also propose a stochastic layer selection approach and a concept of dual-directional translation model to ensure the sufficient utilization of contextualized embeddings. In the case of without using back translation, our best models achieve BLEU scores of 30.45 for En->De and 38.61 for De->En on the IWSLT'14 dataset, and 31.26 for En->De and 34.94 for De->En on the WMT'14 dataset, which exceeds all published numbers.