Varja-Dominators at MedGenVidQA 2026: Hybrid Video and Document Retrieval using PubMedBERT, T5 Query Expansion, and Cross-Encoder Re-Ranking

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research · Depth: Expert, medium

Summary

The Varja-Dominators system, developed for Task A of the MedGenVidQA 2026 shared task, simultaneously retrieves relevant PubMed documents and medical videos for 60 consumer health topics. This unified multi-stage pipeline treats video and document retrieval as complementary. For video retrieval, it fine-tunes a PubMedBERT bi-encoder on 2,710 MedVidQA training samples, segments 833 video transcripts into 32,489 30-second chunks, and employs T5-based query expansion for BM25 sparse retrieval alongside FAISS dense retrieval. The system fuses these results using weighted Reciprocal Rank Fusion (RRF) with a 0.75 dense weight and 0.25 sparse weight, then re-ranks the top-200 candidates with a MiniLM-L-6-v2 cross-encoder to yield the final top-10 videos. Document retrieval utilizes the NCBI PubMed ESearch API with a progressive keyword fallback. The system achieved a MAP of 0.3898, Recall@10 of 0.8449, and NDCG@10 of 0.1079, covering all 60 topics.

Key takeaway

For Machine Learning Engineers developing medical information retrieval systems, integrating hybrid approaches is crucial. You should consider combining dense and sparse retrieval methods, like PubMedBERT with T5 query expansion and FAISS, fused by weighted RRF. This strategy improves performance metrics such as MAP (0.3898) and Recall@10 (0.8449), ensuring comprehensive topic coverage. Be mindful of relying solely on transcript text and external APIs, which can introduce limitations.

Key insights

Hybrid retrieval systems combining dense and sparse methods enhance medical information access.

Principles

• Treat video and document retrieval as complementary.
• Fuse diverse retrieval methods for improved ranking.
• Hard negative mining enhances bi-encoder performance.

Method

A multi-stage pipeline: PubMedBERT bi-encoder for video, T5 query expansion for sparse retrieval, FAISS for dense, RRF fusion, and MiniLM-L-6-v2 cross-encoder re-ranking. PubMed ESearch API for documents.

In practice

• Segment long videos into overlapping chunks.
• Use weighted RRF to combine dense and sparse scores.
• Implement keyword fallback for API robustness.

Topics

• Hybrid Retrieval
• Medical Video Retrieval
• PubMedBERT
• T5 Query Expansion
• Cross-Encoder Re-ranking
• Reciprocal Rank Fusion

Best for: NLP Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Benchmarking Retrieval Strategies for Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study — Cross-Encoder Reranking significantly improves biomedical RAG performance by enhancing contextual precision.
• Not your usual search system — feat. BM25, FAISS and XGBoost — Production-grade search systems combine lexical and semantic retrieval with learning-to-rank for optimal performance.
• DocRetriever: A Plug-and-Play Framework for Multimodal Document Retrieval with Comprehensive Benchmark — DocRetriever improves multimodal document retrieval via layout-aware sparse embeddings and a generalizable, reasoning-augmented reranker.
• HiKEY: Hierarchical Multimodal Retrieval for Open-Domain Document Question Answering — HiKEY uses document hierarchy and multimodal fusion for efficient, accurate RAG in ODQA.
• Article: Why Vector Search Alone Isn't Enough: Hybrid Retrieval for RAG — Hybrid retrieval combining vector search and BM25 via Reciprocal Rank Fusion is essential for robust RAG systems handling diverse query types.
• Lost in a Single Vector: Improving Long-Document Retrieval with Chunk Evidence Aggregation — Improving long-document retrieval by aggregating independently encoded chunk-level evidence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.