Reproduce “Multi-CPR: A Multi Domain Chinese Dataset for Passage Retrieval” 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
Dingkun Long, Qiong Gao, Kuan Zou, Guangwei Xu, Pengjun Xie, Ruijie Guo, Jian Xu, Guanjun Jiang and 2 more
Researchers from Alibaba Group introduced Multi-CPR, the first large-scale, multi-domain Chinese dataset for passage retrieval, featuring human-annotated query-passage pairs from E-commerce, Entertainment video, and Medical domains. Experiments on Multi-CPR demonstrated that dense retrieval models, especially with in-domain continuous pre-training, significantly outperform sparse models and general-domain trained models.
Key insights
- High-quality in-domain human-annotated data is crucial for effectively training dense retrieval models, enabling them to significantly outperform sparse models for domain-specific Chinese passage retrieval.
- Continuous pre-training of the BERT backbone on in-domain raw text further enhances the performance of dense retrieval models by adapting to domain-specific linguistic characteristics.
- Dense models exhibit greater optimization potential and better semantic generalization capabilities compared to sparse models, though sparse methods still provide competitive baselines for explicit queries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.