"Shumor: A Benchmark and Multi-Agent Framework for Decoding Humor in Game Reviews"

"Humor understanding remains a distinct challenge for large language models (LLMs), particularly in vertical domains rich in cultural nuances, such as video games. In this paper, we introduce Shumor, a large-scale, bilingual humor dataset constructed from extensive Steam game reviews. By benchmarking state-of-the-art (SOTA) LLMs on Shumor for humor intensity prediction, we identified a significant disparity between model predictions and human perception regarding humor intensity in gaming contexts. Facing the tasks of humor classification and explanation, we have found that existing models lack domain knowledge and deep reasoning in the gaming field. Therefore, we propose SOPA-SC (Script-Opposition-Purpose Architecture via Synergistic Collaboration), a novel framework based on the incongruity theory of humor, which combines chain-of-thought (CoT) reasoning with multi-agent collaboration. By simulating the cognitive synergy of specialized agents (Gamer, Linguist, and Psychologist), SOPA-SC effectively decomposes complex domain-specific humor. Experimental results demonstrate that SOPA-SC achieves substantial improvements in both F1-score and Jaccard index for humor classification. Notably, it also significantly boosts recall, effectively mitigating the common issue of \"missed detections\" prevalent in general-purpose models. Furthermore, qualitative evaluations confirm that SOPA-SC significantly surpasses existing foundation models in interpretive depth, accurately decoding implicit community memes and psychological intent."

幽默理解仍然是大语言模型（Large Language Models，LLMs）面临的一项独特挑战，尤其在诸如电子游戏这类蕴含丰富文化意蕴的垂直领域中。本研究提出了Shumor——一个源自海量Steam游戏评测的大规模双语幽默数据集。我们通过在Shumor数据集上针对幽默强度预测任务对当前最优（State-of-the-Art，SOTA）大语言模型开展基准测试，发现游戏场景下的幽默强度预测中，模型预测结果与人类感知存在显著偏差。针对幽默分类与幽默解释两类任务，我们发现现有模型缺乏游戏领域的专业知识与深度推理能力。为此，我们提出了SOPA-SC（协同协作式脚本-对立-意图架构，Script-Opposition-Purpose Architecture via Synergistic Collaboration）——一种基于幽默不和谐理论的新型框架，该框架将思维链（Chain-of-Thought，CoT）推理与多智能体协作相结合。通过模拟玩家、语言学家与心理学家三类专业智能体的认知协同过程，SOPA-SC可有效拆解复杂的领域专属幽默。实验结果表明，在幽默分类任务中，SOPA-SC在F1值与杰卡德指数两项指标上均实现了大幅提升；值得注意的是，该框架还显著提升了召回率，有效缓解了通用模型普遍存在的“漏检”问题。此外，定性评估结果证实，SOPA-SC在解释深度上显著优于现有基础模型，能够精准解码隐含的社区梗与心理意图。