Language models can greatly benefit from efficient tokenization. However, they still mostly utilize the classical Byte-Pair Encoding (BPE) algorithm, a simple and reliable method. BPE has been shown to cause such issues as under-trained tokens and sub-optimal compression that may affect the downstream performance. We introduce PickyBPE, a modified BPE algorithm that carries out vocabulary refinement during tokenizer training by removing merges that leave intermediate “junk” tokens. Our method improves vocabulary efficiency, eliminates under-trained tokens, and does not compromise text compression. Our experiments show that this method either improves downstream performance or does not harm it.

Machine translation (MT) research is most typically English-centric. In recent years, massively multilingual translation systems have also been increasingly popular. However, efforts purposefully focused on less-resourced languages are less widespread. In this paper, we focus on MT from and into the Estonian language. First, emphasizing the importance of data availability, we generate and publicly release a back-translation corpus of over 2 billion sentence pairs. Second, using these novel data, we create MT models covering 18 translation directions, all either from or into Estonian. We re-use the encoder of the NLLB multilingual model and train modular decoders separately for each language, surpassing the original NLLB quality. Our resulting MT models largely outperform other open-source MT systems, including previous Estonian-focused efforts, and are released as part of this submission.

Grammatical Error Correction (GEC) enhances language proficiency and promotes effective communication, but research has primarily centered around English. We propose a simple approach to multilingual and low-resource GEC by exploring the potential of multilingual machine translation (MT) models for error correction. We show that MT models are not only capable of error correction out-of-the-box, but that they can also be fine-tuned to even better correction quality. Results show the effectiveness of this approach, with our multilingual model outperforming similar-sized mT5-based models and even competing favourably with larger models.

Machine translation quality is highly reliant on large amounts of training data, and, when a limited amount of parallel data is available, synthetic back-translated or multilingual data can be used in addition. In this work, we introduce SynEst, a synthetic corpus of translations from 11 languages into Estonian which totals over 1 billion sentence pairs. Using this corpus, we investigate whether adding synthetic or English-centric additional data yields better translation quality for translation directions that do not include English. Our results show that while both strategies are effective, synthetic data gives better results. Our final models improve the performance of the baseline No Language Left Behind model while retaining its source-side multilinguality.

This paper explores knowledge distillation for multi-domain neural machine translation (NMT). We focus on the Estonian-English translation direction and experiment with distilling the knowledge of multiple domain-specific teacher models into a single student model that is tiny and efficient. Our experiments use a large parallel dataset of 18 million sentence pairs, consisting of 10 corpora, divided into 6 domain groups based on source similarity, and incorporate forward-translated monolingual data. Results show that tiny student models can cope with multiple domains even in case of large corpora, with different approaches benefiting frequent and low-resource domains.

Many works proposed methods to improve the performance of Neural Machine Translation (NMT) models in a domain/multi-domain adaptation scenario. However, an understanding of how NMT baselines represent text domain information internally is still lacking. Here we analyze the sentence representations learned by NMT Transformers and show that these explicitly include the information on text domains, even after only seeing the input sentences without domains labels. Furthermore, we show that this internal information is enough to cluster sentences by their underlying domains without supervision. We show that NMT models produce clusters better aligned to the actual domains compared to pre-trained language models (LMs). Notably, when computed on document-level, NMT cluster-to-domain correspondence nears 100%. We use these findings together with an approach to NMT domain adaptation using automatically extracted domains. Whereas previous work relied on external LMs for text clustering, we propose re-using the NMT model as a source of unsupervised clusters. We perform an extensive experimental study comparing two approaches across two data scenarios, three language pairs, and both sentence-level and document-level clustering, showing equal or significantly superior performance compared to LMs.

This paper describes the University of Tartu’s submission to the news translation shared task of WMT19, where the core idea was to train a single multilingual system to cover several language pairs of the shared task and submit its results. We only used the constrained data from the shared task. We describe our approach and its results and discuss the technical issues we faced.