Team ENGsTF at PAN 2026: Hybrid Sparse-Dense Retrieval with Two-Stage Score Fusion for Generative Plagiarism Detection
الملخص
Generative plagiarism generated by large language models preserves semantic content while exhibiting low lexical overlap, which limits sparse retrieval methods and complicates dense retrieval over long documents. In this work, we present a hybrid source retrie...
Generative plagiarism generated by large language models preserves semantic content while exhibiting low lexical overlap, which limits sparse retrieval methods and complicates dense retrieval over long documents. In this work, we present a hybrid source retrieval system for the PAN 2026 Generative Plagiarism Detection task. Our approach combines BM25-based sparse retrieval with dense semantic retrieval using BGE-M3, integrating both signals via weighted score fusion. To better handle long scientific documents, source papers are segmented into overlapping token-based chunks, while queries are encoded as full-length inputs using the long-context capability of BGE-M3. Retrieved candidates are then reranked using BGE-reranker-v2-m3, with scores fused back into the retrieval scores via a second weighted combination. Experiments show that the hybrid approach outperforms sparse-only and dense-only baselines. Long-context embeddings improve retrieval quality, while score fusion is crucial for maintaining performance, as replacing retrieval scores with reranker outputs degrades effectiveness. On the PAN 2026 test set, the system achieves an nDCG@10 of 0.6422 and Recall@10 of 0.6708, the best performance among evaluated configurations.
الكلمات المفتاحية
الباحثون
- رياض سنبل مؤلف مشارك
ملفات المنشور
منشورات ذات صلة
Wathiq-SPU at AraSeg Shared Task 2026: Calibrated Transformer Ensembles for Arabic Sentence Segmentation across Punctuation and Paragraph Conditions
2026
WathiqTeam at AraGenre 2026: Class-Balanced Hybrid Retrieval with Selective LLM Reranking
2026
SPU_DeepSimplify at the CLEF 2026 SimpleText Track: A Planning-Based Approach to Scientific Text Simplification
2026