Robust Optimization for Multilingual Translation with Imbalanced Data

Authors: Xian Li, Hongyu Gong

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We ran experiments on common benchmarks (TED, WMT and OPUS-100) with varying degrees of data imbalance. CATS effectively improved multilingual optimization and as a result demonstrated consistent gains on low resources (+0.8 to +2.2 BLEU) without hurting high resources.
Researcher Affiliation Industry Facebook AI {xianl, hygong}@fb.com
Pseudocode Yes Algorithm 1 Curvature Aware Task Scaling (CATS).
Open Source Code No The paper does not provide an explicit statement about the release of its own source code, nor does it include a link to a code repository for its methodology.
Open Datasets Yes Datasets. We experiment on three public benchmarks of multilingual machine translation with varying characteristics of imbalanced data as is shown in Table 1. ...TED [53] WMT[34] OPUS-100[62]...
Dataset Splits Yes We choose the best checkpoint by validation perplexity and only use the single best model without ensembling. We use the same preprocessed data by the Multi DDS baseline authors [53], and followed the same procedure to preprocess OPUS-100 data released by the baseline [62].
Hardware Specification No The paper mentions training 'with the same compute budget' but does not provide specific hardware details such as GPU/CPU models or memory.
Software Dependencies No The paper mentions using the 'Transformer architecture' and cites 'fairseq' as a toolkit, but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We provide detailed training hyperparameters in Appendix B.