Non-Monotonic Latent Alignments for CTC-Based Non-Autoregressive Machine Translation

Authors: Chenze Shao, Yang Feng

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on major WMT benchmarks show that our method substantially improves the translation performance of CTC-based models. Our best model achieves 30.06 BLEU on WMT14 En-De with only one-iteration decoding, closing the gap between non-autoregressive and autoregressive models.2
Researcher Affiliation Academia Chenze Shao1,2, Yang Feng1,2 1Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences 2University of Chinese Academy of Sciences
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Source code: https://github.com/ictnlp/NMLA-NAT.
Open Datasets Yes Datasets We evaluate our methods on the most widely used public benchmarks in previous NAT studies: WMT14 English$German (En$De, 4.5M sentence pairs) [5] and WMT16 English$Romanian (En$Ro, 0.6M sentence pairs) [6].
Dataset Splits Yes For WMT14 En$De, the validation set is newstest2013 and the test set is newstest2014. For WMT16 En$Ro, the validation set is newsdev-2016 and the test set is newstest-2016.
Hardware Specification Yes We use the Ge Force RTX 3090 GPU to train models and measure the translation latency.
Software Dependencies No We implement our models based on the open-source framework of fairseq [MIT License, 31].
Experiment Setup Yes On WMT14 En$De, we use a dropout rate of 0.2 to train NAT models and use a dropout rate of 0.1 for finetuning. On WMT16 En$Ro, the dropout rate is 0.3 for both the pretraining and finetuning. We use the batch size 64K and train NAT models for 300K steps on WMT14 En$De and 150K steps on WMT16 En$Ro. During the finetuning, we train NAT models for 6K steps with the batch size 256K. All models are optimized with Adam [27] with β = (0.9, 0.98) and = 10 8. The learning rate warms up to 5 10 4 within 10K steps in the pretraining and warms up to e 10 4 within 500 steps in the finetuning.