Neural Network based Named Entity Recognition (NER) models often require large labelled datasets to be trained. In order to speed up the dataset construction process, a predefined set of named entities can be utilised for autonomous sentence labelling by tagging words found in the prededined set. However, NER models trained in such manner often overfit to the dataset and do not generalize to unseen data. In order to improve the generalization capability, we propose a regularisation technique that restrict the model’s input to part of speech tags of the sentence. We show that this technique allows the model to correctly predict a sequence even if it is incorrectly labelled in the training dataset. Poster