Effective integration of morphological analysis and named entity recognition based on a recurrent neural network

Abstract

Morphological analysis (MA) and named entity recognition (NER) are essential steps in natural language processing. Because NER is generally considered the next step after MA, many previous studies have adopted a pipeline architecture in which results of MA are used as inputs of NER. However, under this kind of pipeline architecture, MA errors lead to decreasing performance in NER models. To alleviate this error propagation problem, we propose an integrated neural network model that performs MA and NER simultaneously. The proposed model consists of two layers of bidirectional gated recurrent unit models with conditional random field layers: a lower layer for MA and an upper layer for NER. To optimize weighting parameters of the proposed model, we use a two-phase training scheme. The first phase trains all layers for NER, and the second trains the lower layer for MA. In our experiments, the proposed model outperforms both an independent MA model and independent NER model. Based on the experimental results, we conclude that the proposed model can effectively alleviate the error propagation problem that frequently occurs in the pipeline architecture.