Human evaluation.

With the success of pre-trained language models, the performance of story ending generation has been dramatically improved while remaining challenging due to the lack of commonsense reasoning ability. Most previous works mainly focus on using commonsense knowledge to enhance the implicit correlations between words but ignore the hidden causality of sentences or events. In this paper, we propose Causal commonsense Enhanced joint model for story ending Generation (CEG), which incorporates causal commonsense events knowledge to generate a reasonable story ending. Specifically, we first develop a commonsense events inference model trained on GLUCOSE, which converts static knowledge into a dynamic generation model to discover unseen knowledge. It uses prompts to produce various commonsense events behind the stories as pseudo-labels of the dataset. Then, we propose a joint model for the causal events inference task and the story ending generation task to inject inference knowledge into the generation, which consists of a shared encoder, an inference decoder, and a generation decoder. In the causal events inference task, we use the shared encoder and the inference decoder to reason the causal events behind each sentence of the story context to help the model better understand the story and provide long-distance dependencies for the story ending generation. In story ending generation, we combine the hidden states of the causal events with the story context to generate the story ending by the shared encoder and the generation decoder. We jointly train the model on two tasks so that the generation decoder produces the story endings that better match the clues. Experimental results on the ROCStories dataset show that our model outperforms the previous works, demonstrating the effectiveness of the joint model and the generated causal events.伴随预训练语言模型的蓬勃发展，故事结尾生成任务的性能已得到大幅提升，但受限于常识推理能力的不足，该任务仍存在诸多挑战。现有多数研究主要聚焦于利用常识知识强化词汇间的隐式关联，却忽略了语句或事件背后潜藏的因果关系。本文提出因果常识增强型故事结尾生成联合模型（Causal commonsense Enhanced joint model for story ending Generation，CEG），该模型融合因果常识事件知识以生成合理的故事结尾。具体而言，本文首先基于GLUCOSE数据集训练得到常识事件推理模型，该模型将静态知识转化为动态生成模型以挖掘未被观测的知识；该模型通过提示词生成故事背后的各类常识事件，将其作为数据集的伪标签。随后，本文提出面向因果事件推理任务与故事结尾生成任务的联合模型，将推理知识注入生成流程，该模型包含共享编码器、推理解码器与生成解码器三个核心模块。在因果事件推理任务中，本文利用共享编码器与推理解码器对故事上下文各语句背后的因果事件进行推理，以帮助模型更好地理解故事内容，并为故事结尾生成提供长距离依赖支持；在故事结尾生成阶段，本文通过共享编码器与生成解码器，将因果事件的隐状态与故事上下文进行融合，最终生成故事结尾。本文对双任务进行联合训练，使得生成解码器能够生成更贴合故事线索的合理结尾。在ROCStories数据集上的实验结果表明，本文所提模型优于现有相关研究，验证了该联合模型与生成的因果事件的有效性。