Abstract:

Fine-grained entity classification is a task that requires a fine-grained type label to be assigned to a given entity mention. Most of the existing fine-grained entity classification uses distant supervision method. All of the type labels corresponding to the entities in the knowledge base are assigned to the entity mention, which will introduce irrelevant or specific noise labels. In distant supervision, type labels that are not related to the entity mention context are classified as out-of-context noise labels, and type labels whose assignment of fine-grained labels leads to inacc-urate entity meaning in the context are classified as overly-specific noise labels. In order to reduce the impact of noise, manual labeling and heuristic pruning methods have been used in the past, but there are some problems such as low efficiency and reducing the size of the training set, which leads to the deterioration of the overall performance of the classification model. By introducing the memory network, the classification model can deeply learn the correlation between the entity mention context and the type label, enhance the memory representation of the type label corresponding to the similar entity mention context, and effectively reduce the influence of out-of-context noise labels. At the same time, transformative hierarchical loss function is used to effectively learn the hierarchical relationship between type labels, so as to alleviate the negative impact of overly-specific noise labels. In addition, using the L2 regularization function can prevent the model from overfitting noise labels. Experimental results on public datasets show that the proposed method can effectively alleviate the negative effects of out-of-context noise labels and overly-specific noise labels on the classification model, and its performance in accuracy, Macro F1 value and Micro F1 value is superior to previous methods for processing noise labels.

Key words: fine-grained entity classification, noise processing, memory network, type label

摘要：

细粒度实体分类是在给定实体指称后要求为其分配细粒度类型标签的任务。大多数细粒度实体分类采用远程监督的方法，为实体指称分配知识库中实体所对应的全部类型标签，这会引入无关或具体的噪声标签。在远程监督中，将分配与指称上下文无关的类型标签归为无关噪声标签，分配细粒度标签导致在上下文中实体含义不准确的类型标签归为具体噪声标签。为减轻噪声影响，以往采用人工标注、启发式规则剪枝等方法，但存在效率低、缩减训练集规模导致分类模型整体性能变差等问题。通过引入记忆网络，分类模型能深入学习实体指称上下文与类型标签之间的关联性，增强对相似的指称上下文所对应类型标签的记忆表示，有效减轻无关噪声标签的影响。与此同时，利用变形的层次损失函数有效学习类型标签之间的层次关系，从而缓解具体噪声标签的负面影响。此外，结合L2正则化函数防止训练模型对噪声标签的过拟合。在公开数据集上的实验结果表明，提出的方法能够有效缓解无关噪声标签和具体噪声标签对分类模型的消极影响，在准确率、Macro F1值、Micro F1值上表现均优于以往处理标签噪声的方法。

关键词: 细粒度实体分类, 噪声处理, 记忆网络, 类型标签