There has been recent success in pre-training on monolingual data and fine-tuning on Machine Translation (MT), but it remains unclear how to best leverage a pre-trained model for a given MT task. This paper investigates the benefits and drawbacks of freezing parameters, and adding new ones, when fine-tuning a pre-trained model on MT. We focus on 1) Fine-tuning a model trained only on English monolingual data, BART. 2) Fine-tuning a model trained on monolingual data from 25 languages, mBART. For BART we get the best performance by freezing most of the model parameters, and adding extra positional embeddings. For mBART we match or outperform the performance of naive fine-tuning for most language pairs with the encoder, and most of the decoder, frozen. The encoder-decoder attention parameters are most important to fine-tune. When constraining ourselves to an out-of-domain training set for Vietnamese to English we see the largest improvements over the fine-tuning baseline.

Many recent works use ‘consistency regularisation’ to improve the generalisation of fine-tuned pre-trained models, both multilingual and English-only. These works encourage model outputs to be similar between a perturbed and normal version of the input, usually via penalising the Kullback–Leibler (KL) divergence between the probability distribution of the perturbed and normal model. We believe that consistency losses may be implicitly regularizing the loss landscape. In particular, we build on work hypothesising that implicitly or explicitly regularizing trace of the Fisher Information Matrix (FIM), amplifies the implicit bias of SGD to avoid memorization. Our initial results show both empirically and theoretically that consistency losses are related to the FIM, and show that the flat minima implied by a small trace of the FIM improves performance when fine-tuning a multilingual model on additional languages. We aim to confirm these initial results on more datasets, and use our insights to develop better multilingual fine-tuning techniques.

Adapter layers are lightweight, learnable units inserted between transformer layers. Recent work explores using such layers for neural machine translation (NMT), to adapt pre-trained models to new domains or language pairs, training only a small set of parameters for each new setting (language pair or domain). In this work we study the compositionality of language and domain adapters in the context of Machine Translation. We aim to study, 1) parameter-efficient adaptation to multiple domains and languages simultaneously (full-resource scenario) and 2) cross-lingual transfer in domains where parallel data is unavailable for certain language pairs (partial-resource scenario). We find that in the partial resource scenario a naive combination of domain-specific and language-specific adapters often results in ‘catastrophic forgetting’ of the missing languages. We study other ways to combine the adapters to alleviate this issue and maximize cross-lingual transfer. With our best adapter combinations, we obtain improvements of 3-4 BLEU on average for source languages that do not have in-domain data. For target languages without in-domain data, we achieve a similar improvement by combining adapters with back-translation. Supplementary material is available at https://tinyurl.com/r66stbxj.