Using the fine-grained entity typing method of distant supervision, when assigning type labels to entity mention, since the knowledge base contains all type labels of the entity, noisy labels will be introduced. This paper proposed a Fine-grained Entity Typing model combined with Features (FETF) to reduce the negative impact of noisy labels. It is different from the previous methods such as manual annotation and heuristic rule pruning. The model not only improves the classification efficiency, but also does not need to reduce the size of the training set, which can improve the overall performance of the classification model. FETF divides the training set into clean dataset and noisy dataset according to the type numbers of entity mention in candidate type set, and constructs different objective functions for them to achieve the purpose of reducing the impact of noisy labels. At the same time, FETF can use the feature generator to jointly learn the relational features of entity mention - type label, as well as the similarity and hierarchical features of type label - type label. The feature generator extracts semantic features other than context, so as to help the fine-grained entity typing model to assign type labels to entity mention. In addition, we introduce adversarial training in the context processor, which can effectively alleviate the model overfitting noisy labels, and improve the robustness and generalization ability of the model. Experimental results on the public datasets show that the method proposed in this paper can effectively alleviate the negative impact of noisy labels on the fine-grained entity typing model, and outperforms previous methods in accuracy, Macro F1 value and Micro F1 value.

使用远程监督的细粒度实体类型方法，在为实体mention分配类型标签时，由于知识库包含实体的所有类型标签，因此会引入噪声标签。本文提出了一种结合特征（FETF）的细粒度实体类型模型，以减少噪声标签的负面影响。它不同于以往的人工标注、启发式规则剪枝等方法。该模型不仅提高了分类效率，而且不需要减小训练集的大小，可以提高分类模型的整体性能。FETF根据候选类型集中实体提及的类型数将训练集分为干净数据集和噪声数据集，并为它们构造不同的目标函数，以达到降低噪声标签影响的目的。同时，FETF可以利用特征生成器联合学习实体mention-type label的关系特征，以及type label-type label的相似度和层次特征。特征生成器提取上下文以外的语义特征，以帮助细粒度的实体类型模型将类型标签分配给实体提及。此外，我们在上下文处理器中引入对抗性训练，可以有效缓解模型过拟合噪声标签，提高模型的鲁棒性和泛化能力。