Knowledge Distillation: A Method for Making Neural Machine Translation More Efficient
Neural machine translation (NMT) systems have greatly improved the quality available from machine translation (MT) compared to statistical machine translation (SMT) systems. However, these state-of-the-art NMT models need much more computing power and data than SMT models, a requirement that is unsu...
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2022-02-01
|
| Series: | Information |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2078-2489/13/2/88 |
| _version_ | 1797479258143064064 |
|---|---|
| author | Wandri Jooste Rejwanul Haque Andy Way |
| author_facet | Wandri Jooste Rejwanul Haque Andy Way |
| author_sort | Wandri Jooste |
| collection | DOAJ |
| description | Neural machine translation (NMT) systems have greatly improved the quality available from machine translation (MT) compared to statistical machine translation (SMT) systems. However, these state-of-the-art NMT models need much more computing power and data than SMT models, a requirement that is unsustainable in the long run and of very limited benefit in low-resource scenarios. To some extent, model compression—more specifically state-of-the-art knowledge distillation techniques—can remedy this. In this work, we investigate knowledge distillation on a simulated low-resource German-to-English translation task. We show that sequence-level knowledge distillation can be used to train small student models on knowledge distilled from large teacher models. Part of this work examines the influence of hyperparameter tuning on model performance when lowering the number of Transformer heads or limiting the vocabulary size. Interestingly, the accuracy of these student models is higher than that of the teachers in some cases even though the student model training times are shorter in some cases. In a novel contribution, we demonstrate for a specific MT service provider that in the post-deployment phase, distilled student models can reduce emissions, as well as cost purely in monetary terms, by almost 50%. |
| first_indexed | 2024-03-09T21:43:17Z |
| format | Article |
| id | doaj.art-9f0b589c1a6c4230a4b6836bec89160d |
| institution | Directory Open Access Journal |
| issn | 2078-2489 |
| language | English |
| last_indexed | 2024-03-09T21:43:17Z |
| publishDate | 2022-02-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Information |
| spelling | doaj.art-9f0b589c1a6c4230a4b6836bec89160d2023-11-23T20:25:38ZengMDPI AGInformation2078-24892022-02-011328810.3390/info13020088Knowledge Distillation: A Method for Making Neural Machine Translation More EfficientWandri Jooste0Rejwanul Haque1Andy Way2ML-Labs, ADAPT Centre, Dublin City University, D09 Y074 Dublin, IrelandSchool of Computing, National College of Ireland, D01 Y300 Dublin, IrelandML-Labs, ADAPT Centre, Dublin City University, D09 Y074 Dublin, IrelandNeural machine translation (NMT) systems have greatly improved the quality available from machine translation (MT) compared to statistical machine translation (SMT) systems. However, these state-of-the-art NMT models need much more computing power and data than SMT models, a requirement that is unsustainable in the long run and of very limited benefit in low-resource scenarios. To some extent, model compression—more specifically state-of-the-art knowledge distillation techniques—can remedy this. In this work, we investigate knowledge distillation on a simulated low-resource German-to-English translation task. We show that sequence-level knowledge distillation can be used to train small student models on knowledge distilled from large teacher models. Part of this work examines the influence of hyperparameter tuning on model performance when lowering the number of Transformer heads or limiting the vocabulary size. Interestingly, the accuracy of these student models is higher than that of the teachers in some cases even though the student model training times are shorter in some cases. In a novel contribution, we demonstrate for a specific MT service provider that in the post-deployment phase, distilled student models can reduce emissions, as well as cost purely in monetary terms, by almost 50%.https://www.mdpi.com/2078-2489/13/2/88NMTGreen AIknowledge distillationCO<sub>2</sub> savings |
| spellingShingle | Wandri Jooste Rejwanul Haque Andy Way Knowledge Distillation: A Method for Making Neural Machine Translation More Efficient Information NMT Green AI knowledge distillation CO<sub>2</sub> savings |
| title | Knowledge Distillation: A Method for Making Neural Machine Translation More Efficient |
| title_full | Knowledge Distillation: A Method for Making Neural Machine Translation More Efficient |
| title_fullStr | Knowledge Distillation: A Method for Making Neural Machine Translation More Efficient |
| title_full_unstemmed | Knowledge Distillation: A Method for Making Neural Machine Translation More Efficient |
| title_short | Knowledge Distillation: A Method for Making Neural Machine Translation More Efficient |
| title_sort | knowledge distillation a method for making neural machine translation more efficient |
| topic | NMT Green AI knowledge distillation CO<sub>2</sub> savings |
| url | https://www.mdpi.com/2078-2489/13/2/88 |
| work_keys_str_mv | AT wandrijooste knowledgedistillationamethodformakingneuralmachinetranslationmoreefficient AT rejwanulhaque knowledgedistillationamethodformakingneuralmachinetranslationmoreefficient AT andyway knowledgedistillationamethodformakingneuralmachinetranslationmoreefficient |