通过生成预测对话脱轨
Forecasting Conversation Derailments Through Generation
AI总结：
针对现有语言模型难预测未来对话脱轨的问题，提出基于微调LLM生成多条未来对话轨迹并结合共识预测的方法，引入社会语言属性引导生成，在英文基准上取得SOTA结果。
摘要：
预测对话脱轨在在线内容审核、冲突解决和商务谈判等现实场景中具有实用价值。然而，尽管语言模型在识别对话中存在的冒犯性言论方面表现出色，但它们在预测未来对话脱轨方面仍存在困难。与仅基于过往对话历史预测对话结果的已有研究不同，我们的方法使用经过微调的大语言模型（LLM），以现有对话历史为条件采样多条未来对话轨迹，并基于这些轨迹的共识预测对话结果。我们还尝试利用反映话轮级对话动态的社会语言属性，作为生成未来对话时的引导。我们的未来对话轨迹方法在英文对话脱轨预测基准上超越了当前最优结果，且在消融实验中展现出显著的准确率提升。
英文摘要：
Forecasting conversation derailment can be useful in real-world settings such as online content moderation, conflict resolution, and business negotiations. However, despite language models' success at identifying offensive speech present in conversations, they struggle to forecast future conversation derailments. In contrast to prior work that predicts conversation outcomes solely based on the past conversation history, our approach samples multiple future conversation trajectories conditioned on existing conversation history using a fine-tuned LLM. It predicts the conversation outcome based on the consensus of these trajectories. We also experimented with leveraging socio-linguistic attributes, which reflect turn-level conversation dynamics, as guidance when generating future conversations. Our method of future conversation trajectories surpasses state-of-the-art results on English conversation derailment prediction benchmarks and demonstrates significant accuracy gains in ablation studies.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
