Conversational stance detection aims to identify a speaker’s attitude toward a target in multi-turn dialogues, while the zero-shot setting requires generalization to unseen targets. Existing methods often rely on coarse-grained semantic representations or static knowledge injection, limiting their ability to capture subtle stance cues and utilize commonsense knowledge. To address this, we propose FIRE, a reasoning-enhanced framework that integrates fine-grained stance cues and commonsense knowledge for zero-shot conversational stance detection. FIRE adaptively mines sentiment, rhetorical, referential, cross-turn stance dependency cues and commonsense knowledge, and fuses them into explicit reasoning chains for target-aware stance inference. By combining supervised instruction tuning with reinforcement learning, FIRE further improves reasoning-chain quality and stance prediction. Experiments on benchmark datasets show that FIRE consistently outperforms strong baselines across different target types and LLM backbones, demonstrating the effectiveness of fine-grained cue modeling and knowledge-driven reasoning.

“中文语音实体关系三元组抽取任务(Chinese Speech Entity-Relation Triple Extraction Task, CSRTE)是第二十四届中国计算语言学大会中的一项技术评测,旨在从中文语音数据中自动识别并提取实体及其相互关系,构建结构化的语音关系三元组(头实体、关系、尾实体)。本任务的目标是提升中文语音关系三元组抽取的准确性与效率,增强模型在不同语境和复杂语音场景下的鲁棒性,实现从语音输入到文本三元组输出的全流程自动化处理。通过本次评测,有助于推动中文语音信息抽取技术的发展,促进语音与自然语言处理技术的深度融合,为智能应用提供更加丰富且精准的基础数据支持。此次评测共有257支队伍报名参赛,其中59支队伍提交了A榜成绩。成绩排名前15的队伍晋级A榜,并且表现突出的前朷支队伍提交了技术报告。”