基于模式引导和焦点损失的对话状态追踪算法

Schema-guided dialog state tracking with focal loss

为解决以往对话状态追踪算法只能处理少数对话现象,且没有考虑到这些对话现象在数据集中分布不平衡的问题,提出一种基于模式引导和焦点损失的对话状态追踪算法。利用模式将槽位划分为可分类和不可分类,将用户表达归纳为4种现象(词汇多样性、未知值、值共享和推荐接受),以利于合理地填充对应槽位;使用焦点损失函数保证难样本得到充分训练。在大规模任务型对话数据集MultiWOZ2.2上进行大量消融实验,验证了模型的可行性和算法的有效性。

To solve the problem that previous dialog state tracking works can only handle a few dialog phenomena and do not consider the imbalance of these phenomena in the dataset,a schema-guided dialogue state tracking with focal loss was proposed.Schema was used to divide all slots into categorical and non-categorical,and user expressions were categorized into four phenomena(diverse value,unseen value,value sharing and recommendation acceptance),which facilitated filling the corresponding slots using a reasonable method.The focal loss was used to ensure that hard samples were adequately trained.A large number of ablation experiments were carried out on the large-scale task-oriented dialog dataset MultiWOZ 2.2 to verify the feasibility of the model and the effectiveness of the algorithm.