IEEE Trans Neural Netw Learn Syst. 2021 Sep;32(9):3819-3830. doi: 10.1109/TNNLS.2020.3015912. Epub 2021 Aug 31.
Named entity recognition (NER) aims to recognize mentions of rigid designators from text belonging to predefined semantic types, such as person, location, and organization. In this article, we focus on a fundamental subtask of NER, named entity boundary detection, which aims at detecting the start and end boundaries of an entity mention in the text, without predicting its semantic type. The entity boundary detection is essentially a sequence labeling problem. Existing sequence labeling methods either suffer from sparse boundary tags (i.e., entities are rare and nonentities are common) or they cannot well handle the issue of variable size output vocabulary (i.e., need to retrain models with respect to different vocabularies). To address these two issues, we propose a novel entity boundary labeling model that leverages pointer networks to effectively infer boundaries depending on the input sequence. On the other hand, training models on source domains that generalize to new target domains at the test time are a challenging problem because of the performance degradation. To alleviate this issue, we propose Metabdry, a novel domain generalization approach for entity boundary detection without requiring any access to target domain information. Especially, adversarial learning is adopted to encourage domain-invariant representations. Meanwhile, metalearning is used to explicitly simulate a domain shift during training so that metaknowledge from multiple resource domains can be effectively aggregated. As such, Metabdry explicitly optimizes the capability of "learning to generalize," resulting in a more general and robust model to reduce the domain discrepancy. We first conduct experiments to demonstrate the effectiveness of our novel boundary labeling model. We then extensively evaluate Metabdry on eight data sets under domain generalization settings. The experimental results show that Metabdry achieves state-of-the-art results against the recent seven baselines.
命名实体识别 (NER) 的目标是从属于预定义语义类型（例如，人、地点和组织）的文本中识别刚性指示词的提及。在本文中，我们专注于 NER 的一个基本子任务，即实体边界检测，其旨在检测文本中实体提及的开始和结束边界，而无需预测其语义类型。实体边界检测本质上是一个序列标记问题。现有的序列标记方法要么受到稀疏边界标签的影响（即实体很少，非实体很常见），要么无法很好地处理可变大小输出词汇的问题（即需要针对不同的词汇重新训练模型）。为了解决这两个问题，我们提出了一种新的实体边界标记模型，该模型利用指针网络根据输入序列有效地推断边界。另一方面，在测试时在源域上训练模型并推广到新的目标域是一个具有挑战性的问题，因为性能会下降。为了解决这个问题，我们提出了 Metabdry，这是一种用于实体边界检测的新的域泛化方法，无需访问目标域信息。特别是，采用对抗学习来鼓励域不变表示。同时，元学习用于在训练期间显式模拟域转移，以便可以有效地聚合来自多个资源域的元知识。因此，Metabdry 显式优化了“学习泛化”的能力，从而使模型更具通用性和鲁棒性，从而减少了域差异。我们首先进行实验以证明我们的新型边界标记模型的有效性。然后，我们在域泛化设置下在八个数据集上广泛评估 Metabdry。实验结果表明，Metabdry 在最近的七个基线中取得了最先进的结果。
