KoVRE: New Korean Visual Document Retrieval Model
A new model called KoVRE (Korean Visual Document Retrieval Embedding) has been developed to enhance the retrieval of Korean visual documents. This single-vector retriever overcomes the limitations of current VDR systems, which are often focused on English and depend on large backbones or complex multi-vector formats. KoVRE was trained on a dataset consisting of 708,729 Korean and English query-page pairs, utilizing positive-aware hard-negative mining techniques. The training process involved careful evaluations of the training data, hard-negative strategies, and knowledge distillation through rerankers. This 2B parameter model significantly outperforms its base backbone and exceeds the performance of an 8B model in Korean visual document retrieval tasks. The research can be found on arXiv under the identifier 2608.01389.
Key facts
- KoVRE is a single-vector retriever for Korean visual documents.
- Trained on 708,729 Korean and English query-page pairs.
- Uses positive-aware hard-negative mining.
- Includes controlled analyses of training-data composition, hard-negative treatment, and reranker-based knowledge distillation.
- The 2B model outperforms the 8B model on Korean benchmarks.
- Addresses English-centric nature of existing VDR models.
- Avoids massive backbones and storage-intensive multi-vector representations.
- Available on arXiv with identifier 2608.01389.
Entities
Institutions
- arXiv