The WMT 2023 Shared Task on Low-Resource Indic Language Translation featured to and from Assamese, Khasi, Manipuri, Mizo on one side and English on the other. We submitted systems supervised neural machine translation systems for each pair and direction and experimented with different configurations and settings for both preprocessing and training. Even if most of them did not reach competitive performance, our experiments uncovered some interesting points for further investigation, namely the relation between dataset and model size, and the impact of the training framework. Moreover, the results of some of our preliminary experiments on the use of word embeddings initialization, backtranslation, and model depth were in contrast with previous work. The final results also show some disagreement in the automated metrics employed in the evaluation.

Parallel corpora are still crucial to train effective Machine Translation systems. This is even more true for low-resource language pairs, for which Neural Machine Translation has been shown to be less robust to domain mismatch and noise. Due to time and resource constraints, parallel corpora are mostly created with sentence alignment methods which automatically infer alignments. Recent work focused on state-of-the-art pre-trained sentence embeddings-based methods which are available only for a tiny fraction of the world’s languages. In this paper, we evaluate the performance of four widely used algorithms on the low-resource English-Yorùbá language pair against a multidomain benchmark parallel corpus on two experiments involving 1-to-1 alignments with and without reordering. We find that, at least for this language pair, earlier and simpler methods are more suited to the task, all the while not requiring additional data or resources. We also report that the methods we evaluated perform differently across distinct domains, thus indicating that some approach may be better for a specific domain or textual structure.

Tokenization has been shown to impact the quality of downstream tasks, such as Neural Machine Translation (NMT), which is susceptible to out-of-vocabulary words and low frequency training data. Current state-of-the-art algorithms have been helpful in addressing the issues of out-of-vocabulary words, bigger vocabulary sizes and token frequency by implementing subword segmentation. We argue, however, that there is still room for improvement, in particular regarding low-frequency tokens in the training data. In this paper, we present “High Frequency Tokenizer”, or HFT, a new language-independent subword segmentation algorithm that addresses this issue. We also propose a new metric to measure the frequency coverage of a tokenizer’s vocabulary, based on a frequency rank weighted average of the frequency values of its items. We experiment with a diverse set of language corpora, vocabulary sizes, and writing systems and report improvements on both frequency statistics and on the average length of the output. We also observe a positive impact on downstream NMT.

We describe our neural machine translation systems for the WMT22 shared task on unsupervised MT and very low resource supervised MT. We submit supervised NMT systems for Lower Sorbian-German and Lower Sorbian-Upper Sorbian translation in both directions. By using a novel tokenization algorithm, data augmentation techniques, such as Data Diversification (DD), and parameter optimization we improve on our baselines by 10.5-10.77 BLEU for Lower Sorbian-German and by 1.52-1.88 BLEU for Lower Sorbian-Upper Sorbian.