Exploring Discriminative Word-Level Domain Contexts for Multi-Domain Neural Machine Translation

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Multi-domain Neural Machine Translation (NMT) models often use unified architectures with mixed-domain corpora.
• Existing models struggle to differentiate the varying relevance of words within a sentence to its specific domain.
• This limitation hinders optimal performance in translating across diverse subject areas.

Purpose of the Study:

• To develop a novel multi-domain NMT model that effectively distinguishes and utilizes word-level domain contexts.
• To improve translation accuracy by accounting for the differential impact of words based on their domain relevance.
• To enhance the adaptability and performance of NMT systems across various specialized domains.

Main Methods:

• Employed multi-task learning to jointly train NMT with attention-based monolingual domain classification.
• Introduced a domain classifier and an adversarial domain classifier to generate domain-specific and shared annotations via gating vectors.
• Integrated an attentional domain classifier into the decoder to refine training using word-level cost weighting based on domain relevance.

Main Results:

• Experimental results demonstrated the effectiveness of the proposed model in multi-domain translation tasks.
• The model successfully discriminated the impact of target words based on their relevance to the sentence domain.
• Improved translation quality was observed across several multi-domain test cases.

Conclusions:

• The proposed method effectively distinguishes and exploits word-level domain contexts in multi-domain NMT.
• Jointly modeling NMT with domain classification tasks significantly enhances translation performance.
• This approach offers a promising direction for developing more robust and accurate multi-domain translation systems.