Non-Fluent Synthetic Target-Language Data Improve Neural Machine Translation.

Area of Science:

• Natural Language Processing
• Machine Translation
• Artificial Intelligence

Background:

• Neural machine translation (NMT) models require large parallel corpora for effective training.
• Data augmentation techniques, like generating synthetic parallel sentences, are crucial when parallel data is scarce.
• Current methods often assume synthetic data must mimic in-domain parallel corpora, potentially limiting performance.

Purpose of the Study:

• To investigate the impact of non-fluent synthetic target sentences on NMT performance.
• To propose a novel approach using non-fluent synthetic data within a multilingual NMT framework.
• To evaluate the effectiveness of this method across various resource scenarios.

Main Methods:

• Generating synthetic parallel sentences with non-fluent target sides.
• Integrating these synthetic sentences into a multilingual NMT system, treating them as data from another language.
• Conducting comparative experiments against state-of-the-art synthetic data generation methods.

Main Results:

• The proposed method consistently improved translation performance across ten low-resource and four high-resource tasks.
• Performance gains were observed compared to existing state-of-the-art synthetic data generation techniques.
• The improvements were independent of the original training corpus size.

Conclusions:

• Non-fluent synthetic training data can enhance NMT performance when utilized in a multilingual setting.
• This approach leads to more robust NMT systems, less susceptible to domain shift and hallucination.
• The findings challenge the assumption that synthetic data must strictly adhere to target-side fluency for optimal results.