Samsung Research Philippines - Datasaur AI’s Submission for the WMT22 Large Scale Multilingual Translation Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1034 - 1038 December 7-8, 2022 ©2022 Association for Computational Linguistics Samsung Research Philippines - Datasaur AI’s Submission for the WMT22 Large Scale Multilingual Translation Task Jan Christian Blaise Cruz Samsung Research Philippines Manila, Philippines jcb.cruz@samsung.com Lintang Sutawika∗ Datasaur AI San Francisco, California, USA lintang@datasaur.ai Abstract This paper describes the submission of the joint Samsung Research Philippines - Datasaur AI team for the WMT22 Large Scale Multilingual African Translation shared task. We approach the contest as a way to explore task compo- sition as a solution for low-resource multilin- gual translation, using adapter fusion to com- bine multiple task adapters that learn subsets of the total translation pairs. Our final model shows performance improvements in 32 out of the 44 translation directions that we participate in when compared to a single model system trained on multiple directions at once. 1 Introduction In this paper, we describe two systems that we submit to the WMT22 Large Scale Multilin- gual African Translation shared task: a base- line finetuned MT5 (Xue et al., 2020) model trained on multiple directions at once (referred to as SRPH-DAI-Baseline), and an MT5 model successively finetuned with task composition us- ing multiple pair-specific adapters (referred to as SRPH-DAI-Fusion). We first outline the preprocessing steps and fil- tering heuristics used to clean the contest dataset, then we show the training setup and experimental design used for constructing our submitted systems. We then report our results on the hidden test set via BLEU, spBLEU, and CHRF2++ automatic evalua- tion metrics. 2 Preprocessing In this section, we detail the preprocessing steps used to filter the contest dataset to ensure that data quality is as high as possible. Given that the contest dataset contains sentence pairs that were artificially aligned from crawled ∗ Work done while at Konvergen AI. data, we use a number of filters to reduce the pos- sibility of mismatched pairs in the final training dataset: • We filter out pairs where one or both sentences have too few (<= 3) or too many (>= 150) tokens post-sentencepiece tokenization. • We remove pairs if one or both sentences have too many repeated (>= 5) punctuations or symbols of the same type (e.g. “/////”), or con- tiguous punctuations/symbols of considerable (>= 3) length (e.g. “word $&**$”). • We also remove sentence pairs where one sen- tence has punctuation that is missing from the other (e.g. “word!!” → “word?”). • If a pair has a sentence where a large percent- age of the total characters (total >= 70%) are numbers or punctuations (e.g. “word ??! +22 8456 8967”), the pair is dropped. • An average word length filter is also used to re- move pairs where one sentence has words that are disproportionately longer than the words in the corresponding sentence. We get a ratio r by taking the sum of the lengths of each to- ken in a sentence, then dividing it by the num- ber of tokens. We only …

Content truncated to 500 words