Delayed Combination of Feature Embedding in Bidirectional LSTM CRF for NER

Named Entity Recognition (NER) plays a vital role in natural language processing (NLP). Currently, deep neural network models have achieved significant success in NER. Recent advances in NER systems have introduced various feature selections to identify appropriate representations and handle Out-Of-the-Vocabulary (OOV) words. After selecting the features, they are all concatenated at the embedding layer before being fed into a model to label the input sequences. However, when concatenating the features, information collisions may occur and this would cause the limitation or degradation of the performance. To overcome the information collisions, some works tried to directly connect some features to latter layers, which we call the delayed combination and show its effectiveness by comparing it to the early combination. As feature encodings for input, we selected the character-level Convolutional Neural Network (CNN) or Long Short-Term Memory (LSTM) word encoding, the pre-trained word embedding, and the contextual word embedding and additionally designed CNN-based sentence encoding using a dictionary. These feature encodings are combined at early or delayed position of the bidirectional LSTM Conditional Random Field (CRF) model according to each feature's characteristics. We evaluated the performance of this model on the CoNLL 2003 and OntoNotes 5.0 datasets using the F1 score and compared the delayed combination model with our own implementation of the early combination as well as the previous works. This comparison convinces us that our delayed combination is more effective than the early one and also highly competitive.