UNLP 2026

Ukrainian document question answering with retrieval and reranking

When Local Wins Fail to Generalize: UNLP 2026 Ukrainian Document QA

PFW shared-task project, ACL 2026

Built a retrieval-augmented question-answering pipeline for Ukrainian documents, combining BGE-M3 dense retrieval, a Qwen3-0.6B page reranker, and MamayLM-Gemma-3-12B-IT for multiple-choice answering.

Results

  • Achieved a private score of 0.8722, ranking 7/15 teams
  • Reached 0.933 document retrieval at rank 1 on the development evaluation
  • Found that conservative retrieval baselines generalized better than locally optimized variants under hidden domain shift