Many researchers jointly model multiple linguistic tasks (e.g., joint modeling of named entity recognition and named entity classification and joint modeling of syntactic parsing and semantic parsing) with an implicit assumption that these individual tasks can enhance each other via the joint modeling. Before conducting research on jointly modeling multiple tasks, however, such researchers hardly examine whether such assumption is true or not. In this paper, we empirically examine whether named entity classification improves the performance of named entity recognition as an empirical case of examining whether semantics improves the performance of a syntactic task. To this end, we firstly specify the way to determine whether a linguistic task is a syntactic task or a semantic task according to both syntactic theory and semantic theory. After that, we design and conduct extensive experiments on two well-known benchmark datasets using three representative yet diverse state-of-the-art models. Experimental results demonstrate that named entity recognition does not lie at the semantic level and is not a semantic task; instead, it is a syntactic task and that the joint modeling of named entity recognition and classification does not improve the performance of named entity recognition. Experimental results also demonstrate that traditional handcrafted feature models can achieve state-of-the-art performance in comparison with the auto-learned feature model on named entity recognition.

许多研究人员在隐式假设这些单独的任务可以通过联合建模而彼此增强的前提下，共同对多个语言任务进行建模（例如，命名实体识别和命名实体分类的联合建模以及句法解析和语义解析的联合建模）。但是，在对多个任务进行联合建模进行研究之前，此类研究人员几乎不会检查这种假设是否正确。在本文中，我们通过实证研究命名实体分类是否可以提高命名实体识别的性能，作为检查语义是否可以提高句法任务性能的实证案例。为此，我们首先根据句法理论和语义理论确定一种确定语言任务是句法任务还是语义任务的方法。之后，我们使用三个具有代表性但又多样化的最新模型，对两个著名的基准数据集进行了设计和进行了广泛的实验。实验结果表明，命名实体识别不属于语义层次，也不是语义任务。相反，这是一个语法任务，并且命名实体识别和分类的联合建模不会提高命名实体识别的性能。实验结果还表明，与在命名实体识别上自动学习的特征模型相比，传统的手工特征模型可以实现最先进的性能。实验结果表明，命名实体识别不属于语义层次，也不是语义任务。相反，这是一个语法任务，并且命名实体识别和分类的联合建模不会提高命名实体识别的性能。实验结果还表明，与在命名实体识别上自动学习的特征模型相比，传统的手工特征模型可以实现最先进的性能。实验结果表明，命名实体识别不属于语义层次，也不是语义任务。相反，这是一个语法任务，并且命名实体识别和分类的联合建模不会提高命名实体识别的性能。实验结果还表明，与在命名实体识别上自动学习的特征模型相比，传统的手工特征模型可以实现最先进的性能。