Simorgh at SemEval-2026 task 7: Region-Aware Hybrid Retrieval for Low-Resource Cultural Reasoning in Multilingual Question Answering

In three linesSimorgh proposes a region-aware hybrid retrieval approach combining BM25 lexical matching and dense semantic similarity for culturally grounded multilingual QA on BLEnD benchmark (30 languages). Uses quantized Qwen3-14B with logit-based answer selection. Improves cross-lingual stability but reveals performance gaps tied to training data imbalance.

Summary generated by Claude — human-verified