MT2023@UdS

Wayúunaiki to Spanish translation

Wayúunaiki is the native language spoken in the Wayúu community, located in the Caribbean region connecting Colombia and Venezuela, where the language coexists with Spanish. In this challenge we want to create translation resources for the Wayúu community.

The shared task

Phase I, development: We provide training data extracted from the Tatoeba challenge [1,2]. It mostly belongs to the relogious domain. We substracted a set from the original Tatoeba corpus: the in-domain test set that is going to be used for evaluating your MT engine in this phase. Together with the parallel corpus we provide, you can use any other data you can find to train your MT engine, but please, don't use the original Tatoeba corpus which would also contain the test set. Upload the translation into Spanish of the test set to Codalab. The translation will be evaluated using BLEU, TER, chrF and COMET. chrF is the official evaluation metric of the challenge.

Phase II, evaluation: We provide an out-of-domain test set (general domain) in Wayúunaiki and you will have a week to upload its translation into Spanish to Codalab. The translation will be evaluated using BLEU, TER, chrF and COMET but only chrF is the official evaluation metric of the challenge. Notice that the best system in an in-domain test set is not necessarily the best system in a general domain.

What can you do?

Baseline: You can reproduce the SMT and/or the NMT system developed in class. Finetuning the NMT hyperparameters is relevant

Data: The training data is not clean at all and it is insuficent. Can you improve on that?

Languages: Spanish in this dataset comes from Latin America. Wayuu is an aglutinative language while Spanish is a fusional language. Can you ease the training in such a difficult context?

Advanced: Multilingualism. Any other language pair with data available that can help with joint training? Any close language even monolingually?

What is other people doing?

• Survey of Low-Resource Machine Translation [3]
• Shared Task: General Machine Translation at WMT (not low-resourced) [4]
• AmericasNLP 2023 Shared Task on Machine Translation into Indigenous Languages (and previous years) [5]

Timeline

10/07/2023: Competition set up

18/07/2023: Competition starts

11/09/2023: Test phase starts

17/09/2023: Test phase ends

20/09/2023: Deadline for filling the form

References

[1] Jörg Tiedemann. 2020. The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT. In Proceedings of the Fifth Conference on Machine Translation, pages 1174–1182, Online. Association for Computational Linguistics.

[2] https://github.com/Helsinki-NLP/Tatoeba-Challenge

[3] Barry Haddow, Rachel Bawden, Antonio Valerio Miceli Barone, Jindřich Helcl, and Alexandra Birch. 2022. Survey of Low-Resource Machine Translation. Computational Linguistics, 48(3):673–732.

[4] https://www.statmt.org/wmt22/translation-task.html

[5] https://turing.iimas.unam.mx/americasnlp/2023_st.html