@inproceedings{nguyen-nguyen-2026-dangnt,
    title = "{DANGNT}@{SGU} at {S}em{E}val-2026 Task 1: A Two-Stage Mistral Generator with {D}istil{BERT} Reranking for {E}nglish Humor Generation",
    author = "Nguyen, Tan Loc  and
      Nguyen, Dang Tuan",
    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-1.375/",
    pages = "2986--2990",
    ISBN = "979-8-89176-414-9",
    abstract = "We describe DANGNT@SGU{'}s system for the English track of SemEval-2026 Task 1 (MWAHAHA), Subtask A (text-based humor generation). Our pipeline combines a two-stage QLoRA-adapted generator based on mistralai/Mistral-7B-Instruct-v0.2 with a DistilBERT reranker trained to distinguish jokes from non-jokes. The generator is first adapted on a raw joke corpus for general humor style, then further tuned on synthetic task-format instruction{--}response pairs for Word Inclusion and News Headlineprompts. At inference time, we generate five candidates per input, optionally enforce lexical constraints for Word Inclusion prompts, and rerank candidates with the classifier. In the official English Subtask A results, our team DANGNT@SGU obtained Elo 962 (95{\%} CI: 926{--}986), ranking 13th. The system is practical, reproducible, and based entirely on open models and public data."
}