细粒度实体分类是一项多类别多标签任务,能协助广泛的下游任务(关系抽取、共指消解、问答 系统等)提高工作效率、优化准确率,已成为自然语言处理领域的一个研究热点。针对传统的细粒度实体分类方法人工标注大型语料库难度大,准确率偏低等问题,研究人员提出了基于神经网络的细粒度实体分类方 法,不仅能够解决人工标注费时费力的问题,而且可以提高分类的准确率。然而现有的神经网络模型大多需 要远程监督的参与,在此过程中会引入噪声标签等问题,通过噪声标签处理方法能够有效抑制噪声标签对分类结果的影响,进一步提升分类性能。 在相同评测数据集下,根据相同评价指标对比各类细粒度实体分类方 法的性能,可以发现在细粒度实体分类领域中采用 BiLSTM 处理实体指称上下文,并通过注意力机制提取更为重要的特征,有助于提高细粒度实体分类方法的准确率、Macro F1值和 Micro F1值。
Fine-grained entity typing is a multi-class and multi-label task, which can help a wide range of downstream tasks (relationship extraction, co-reference resolution, question answering system, etc. ) to enhance productivity and optimize accuracy. It has become a research hotspot in natural language processing field. In view of the difficulty and low accuracy of the traditional fine-grained entity typing method to annotate large corpus, researchers proposed the fine-grained entity typing method based on neural network, which can not only solve the time-consuming and laborious problem of manual annotation, but also improve the accuracy of classification. However, most of the existing neural network models require the participation of distant supervision, which will introduce noise labels and other problems in the process. The noise labels processing method can effectively suppress the impact of noise labels on the classification results and further improve the classification performance. Under the same evaluation datasets, we compared the performance of various fine-grained entity typing methods according to the same evaluation metrics. It can be found that in the field of fine-grained entity typing, using BiLSTM to process the context of entity mention and extracting more important features through the attention mechanism are helpful to improve the accuracy, Macro F1 value and Micro F1 value of fine-grained entity typing method.
ZHOU Qi, TAO Wan. A Comparative Study Based on Fine-grained Entity Classification[J]. Journal of Chongqing Technology and Business University(Natural Science Edition）,2022,39(4):9-18