Reproduce “General2Specialized LLMs Translation for E-commerce” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kaidi Chen, Ben Chen, Dehong Gao, Huangyu Dai, Wen Jiang, Wei Ning, Shanqing Yu, Libin Yang and 1 more
Researchers from Alibaba Group, Northwest Polytechnical University, and Zhejiang University of Technology developed General2Specialized Translation (G2ST), a two-step fine-tuning paradigm with self-contrastive semantic enhancement, to adapt large language models for e-commerce translation. The approach, implemented with Qwen-14B, achieved the highest scores on an e-commerce parallel corpus, including a SacreBLEU of 39.85, outperforming baseline models by up to 1.63 points.
Key insights
- Strategic vocabulary expansion and initial fine-tuning with domain-specific term pairs are crucial for addressing out-of-vocabulary issues and learning specialized terminology in e-commerce.
- Subsequent fine-tuning with a carefully constructed parallel corpus enables models to effectively learn and apply the unique writing formulas and stylistic nuances of e-commerce texts.
- Incorporating self-contrastive semantic enhancement significantly improves model robustness and consistency in predictions across various NMT and LLM architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.