Abstract: In recent years, controllable story generation has been a popular task in the field of natural language processing. Current research has been able to effectively integrate contextual and event features through cross attention mechanism. Despite these advancements, the existing research lacks efficient application of commonsense knowledge and still adopts post-training on public commonsense knowledge datasets. Although this method can improve the overall performance of the model to a certain extent, there remains ample room for improvement in terms of adaptability. To address the above challenges, a new model named adaptive knowledge enhancement (AKE) is proposed. The commonsense knowledge building module can adaptively create matching commonsense knowledge according to the training datasets during finetuning process, which provides more relevant additional information for the model. In addition, a multi-task learning component is trained using auxiliary function that encourage the model to learn more discriminative feature representations, leading to improved generalization capabilities. Extensive experiments show that the proposed AKE model significantly outperforms the state-of-the-art baseline models in terms of both automatic metrics and human evaluations, demonstrating the effectiveness of the proposed model in leveraging commonsense knowledge.

Key words: natural language processing, controllable story generation, cross attention mechanism, multi-task learning, knowledge enhancement

摘要： 可控故事生成是近年来自然语言处理领域内的热点方向。目前的研究通过交叉注意力机制已经能够有效地融合文本特征和事件特征，但其缺乏对于常识化知识的高效应用，仍然采用在公共的常识化知识语料库上后训练的方式，尽管这能够在一定程度上提升模型的性能，但在自适应性的方面仍然有很大的提升空间。为了解决此问题，提出自适应知识增强（adaptive knowledge enhancement，AKE）的可控故事生成模型，其中的常识化知识构建模块能够针对精调时的训练用数据集自适应地构建匹配的常识化知识语料库，确保为模型提供更相关的额外信息。此外，模型中加入了使用辅助函数进行训练的多任务学习组件，确保其能够学习到更具有判别性的特征表示，从而提升模型的泛化能力。实验结果表明，AKE在自动评测指标和人工评测指标上相较于其他基线模型均有显著提升，验证了此模型在利用常识化知识方面的优越性。

关键词: 自然语言处理, 可控故事生成, 交叉注意力机制, 多任务学习, 知识增强