@inproceedings{yang-etal-2026-dutir,
    title = "{DUTIR} at {S}em{E}val-2026 Task 8: A Hybrid Retrieval and Faithfulness-Guarded Framework for Multi-Turn {RAG}",
    author = "Yang, Jin  and
      Chen, Yichong  and
      Yang, Liang",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.48/",
    pages = "328--332",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper describes the system submittedby DUTIRtaskC for SemEval-2026 Task 8:MTRAGEval (Task C). Multi-turn RetrievalAugmented Generation (RAG) poses significant challenges in context tracking, retrievalprecision, and hallucination mitigation. Ourproposed system addresses these by employinga multi-stage pipeline consisting of: (1) LLMbased query rewriting (powered by GPT-5.2) toresolve conversational dependencies; (2) a hybrid retrieval module combining dense embeddings (BGE-M3) and sparse retrieval (BM25)with Reciprocal Rank Fusion (RRF); (3) aconfidence-based answerability gating mechanism; and (4) a post-generation faithfulnessguard. Experimental results on the blind test setshow that our approach achieves a CompositeScore of 0.5576, ranking 4th out of 29 participating teams. Detailed analysis reveals that oursystem significantly outperforms strong baselines in faithfulness and successfully handlesunderspecified queries."
}