Application of adversarial transfer learning in domain adaptive English machine translation

Abstract Machine translation systems have demonstrated a significant level of success, especially in rich language pairs endowed with heavy parallel corpora. When the model is used in domain-specific settings where in-domain data is scarce or even unavailable, however, its performance suffers greatly. To address this challenge, this research presents an adversarial transfer learning framework aimed at improving English machine translation across diverse domains such as medical, legal, and IT, where in-domain parallel data is limited. Instead of bilingual translation, domain adaptive adversarial transfer learning (DAATL) is used to domain shift between high-resource source domains and low-resource target domains using monolingual paraphrase pairs. This model is used to enhance translation quality in specialized or underrepresented English domains, such as medical or legal texts, by transferring knowledge from general-domain translation models trained on abundant parallel corpora. Tokenization is used as a data preprocessing technique to separate text into meaningful units, allowing for improved input representation. Bidirectional encoder representations from transformers is used for feature extraction, capturing deep contextualized embeddings to improve the model’s awareness of language subtleties across domains. The architecture comprises private encoder-decoder pairs for each domain to capture domain-specific linguistic features and a shared encoder-decoder pair to learn domain-invariant representations. A domain discriminator is integrated and trained adversarial to predict the domain of the encoded input. Simultaneously, the shared encoder is trained to confuse the discriminator, encouraging the learning of domain-invariant features through feature disentanglement and promoting robust cross-domain generalization. The experimental findings reveal that the DAATL technique greatly outperforms competing baselines in BLEU, achieving 66.2, indicating successful adversarial feature learning. These findings demonstrate the framework’s effectiveness for domain-adaptive monolingual style transfer in low-resource settings.