Adam Mickiewicz University at WMT 2022: NER-Assisted and Quality-Aware Neural Machine Translation

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 326 - 334 December 7-8, 2022 ©2022 Association for Computational Linguistics Adam Mickiewicz University at WMT 2022: NER-Assisted and Quality-Aware Neural Machine Translation Artur Nowakowski * 1,2 and Gabriela Pałka * 1,3 and Kamil Guttmann † 1,2 and Mikołaj Pokrywka † 1,2 1 Adam Mickiewicz University, Pozna´n, Poland 2 Poleng, Pozna´n, Poland 3 Applica.ai, Warsaw, Poland {artur.nowakowski,gabriela.palka}@amu.edu.pl, {kamgut,mikpok1}@st.amu.edu.pl Abstract This paper presents Adam Mickiewicz Univer- sity’s (AMU) submissions to the constrained track of the WMT 2022 General MT Task. We participated in the Ukrainian ↔ Czech trans- lation directions. The systems are a weighted ensemble of four models based on the Trans- former (big) architecture. The models use source factors to utilize the information about named entities present in the input. Each of the models in the ensemble was trained using only the data provided by the shared task or- ganizers. A noisy back-translation technique was used to augment the training corpora. One of the models in the ensemble is a document- level model, trained on parallel and synthetic longer sequences. During the sentence-level decoding process, the ensemble generated the n-best list. The n-best list was merged with the n-best list generated by a single document-level model which translated multiple sentences at a time. Finally, existing quality estimation mod- els and minimum Bayes risk decoding were used to rerank the n-best list so that the best hy- pothesis was chosen according to the COMET evaluation metric. According to the automatic evaluation results, our systems rank first in both translation directions. 1 Introduction We describe Adam Mickiewicz University’s sub- missions to the constrained track of the WMT 2022 General MT Task. We participated in the Ukrainian ↔ Czech translation directions – a low-resource translation scenario between closely related lan- guages. The data provided by the shared task organizers was thoroughly cleaned and filtered, as described in section 2. The approach described in section 3 is based on combining various MT enhancement methods, *AN and GP contributed equally. †KG and MP contributed equally. including transfer learning from a high-resource language pair (Aji et al., 2020; Zoph et al., 2016), noisy back-translation (Edunov et al., 2018), NER- assisted translation (Modrzejewski et al., 2020), document-level translation, model ensembling, quality-aware decoding (Fernandes et al., 2022), and on-the-fly domain adaptation (Farajian et al., 2017). The results leading to the final submissions are presented in section 4. Additionally, we performed a statistical significance test with paired bootstrap resampling (Koehn, 2004), comparing the baseline solution with the final submission on the test set reference translations released by the shared task organizers. According to the automatic evaluation results based on COMET (Rei et al., 2020) scores, our systems rank first in both translation directions. 2 Data In the initial stage of system preparation, the sentence-level data was cleaned and filtered us- ing the OpusFilter (Aulamo et al., 2020) toolkit. With the use of the toolkit, language detection filtering based on fastText (Joulin et al., 2016) was performed, duplicates were …

Content truncated to 500 words