TLDR: ViRanker is a new cross-encoder reranking model for Vietnamese, combining the BGE-M3 encoder with Rotary Position Encoding (RoPE) and Blockwise Parallel Transformer (BPT). It was trained on an 8 GB Vietnamese corpus using a hybrid hard-negative sampling strategy. ViRanker achieves strong early-rank accuracy on the MMARCO-VI benchmark, outperforming multilingual baselines and competing with PhoRanker, and is openly released on Hugging Face to foster wider adoption and research.
In the world of artificial intelligence, particularly in areas like search engines and question-answering systems, a crucial step after retrieving initial results is ‘reranking’. This process refines the order of results to present the most relevant information first. While significant progress has been made for high-resource languages like English, languages with complex syntax and unique characteristics, such as Vietnamese, often lag behind.
Introducing ViRanker: A Specialized Solution for Vietnamese
A new research paper introduces ViRanker, a cross-encoder reranking model specifically designed to tackle the challenges of the Vietnamese language. Developed by Phuong-Nam Dang, Kieu-Linh Nguyen, and Thanh-Hieu Pham, ViRanker aims to fill the gap for competitive rerankers in Vietnamese, which is considered a low-resource language with flexible word order, rich diacritics, and syntactic ambiguities.
ViRanker builds upon the robust BGE-M3 encoder and incorporates two key architectural enhancements to optimize its performance for Vietnamese: Rotary Position Encoding (RoPE) and the Blockwise Parallel Transformer (BPT).
Key Innovations Behind ViRanker’s Performance
The model’s effectiveness stems from several thoughtful design choices:
-
Rotary Position Encoding (RoPE): This feature is crucial for Vietnamese, as it helps the model better understand word order and the subtle nuances introduced by diacritics. This leads to improved accuracy, especially for the top-ranked results.
-
Blockwise Parallel Transformer (BPT): To efficiently handle longer documents without excessive memory usage, ViRanker utilizes BPT. This innovation allows the model to process more extensive contexts while maintaining accuracy, making it scalable across different types of GPUs.
-
Multi-Layer Perceptron (MLP) Head: An additional scoring head was integrated to adapt the BGE-M3 encoder, originally an embedding model, for direct query-document relevance scoring.
-
Hybrid Hard-Negative Mining: For robust training, ViRanker employs a sophisticated strategy to identify and learn from ‘hard negatives’ – documents that are similar to relevant ones but are actually irrelevant. This process involves using BM25 retrieval, dense similarity reranking, and Maximal Marginal Relevance (MMR) filtering to create a high-quality training dataset.
Training and Evaluation
ViRanker was trained on an extensive 8 GB curated corpus of Vietnamese text, gathered from Wikipedia, open-source GitHub repositories, and various books and reports. This dataset was carefully processed and used to generate 3.5 million query-document triplets for training.
When evaluated on the MMARCO-VI benchmark, ViRanker demonstrated strong early-rank accuracy, outperforming several multilingual baselines and showing highly competitive results against other Vietnamese-specific models like PhoRanker. For instance, it achieved an NDCG@3 of 0.6815 and an MRR@3 of 0.6641, highlighting its ability to deliver precise results at the very top of the ranking, which is critical for user satisfaction.
The model also proved to be efficient in terms of inference speed and memory usage across various GPU setups, from high-end A100s to mid-tier T4s, making it practical for real-world deployment.
Also Read:
- Navigating the Dialogue: A Deep Dive into Conversational Question Answering Systems
- Bridging Morphological Gaps: MoVoC for Ge’ez Script NLP
Looking Ahead
While ViRanker marks a significant advancement, the researchers acknowledge areas for future improvement. These include expanding the training corpus further, exploring parameter-efficient fine-tuning (PEFT) techniques, and investigating newer architectural innovations to enhance long-context handling and robustness against noisy inputs.
ViRanker is openly released on Hugging Face, encouraging reproducibility and broader adoption within the Vietnamese Natural Language Processing community. This work not only provides a powerful tool for Vietnamese information retrieval but also serves as a valuable case study for adapting advanced AI models to other underrepresented languages. To learn more about ViRanker and its technical details, you can refer to the full research paper here.
