A neural machine translation method based on split graph convolutional self-attention encoding.

Area of Science:

• Natural Language Processing
• Deep Learning
• Computational Linguistics

Background:

• Neural machine translation (NMT) is crucial for cross-language team communication.
• Current NMT methods struggle with non-Euclidean spaces and model complexity in dependency relation encoding.

Purpose of the Study:

• To propose a novel approach, split graph convolutional self-attention encoding (SGSE), for enhanced NMT.
• To improve utilization of syntactic dependency relations and reduce model complexity.

Main Methods:

• Extracting syntactic dependency relations and constructing a syntax dependency graph in non-Euclidean space.
• Developing split self-attention and syntactic semantic self-attention networks within a unified model.

Main Results:

• SGSE significantly enhances translation performance across multiple datasets.
• The proposed method effectively mitigates model complexity.
• Improved results observed in team collaboration and enterprise management scenarios.

Conclusions:

• SGSE offers a promising approach for advanced NMT.
• This method can enhance communication efficiency in cross-language collaborative teams.
• The approach effectively balances translation quality and model complexity.