NAMED ENTITY RECOGNITION USING INDEPENDENTLY RECURRENT NEURAL NETWORKS

Abstract

The task of Named Entity Recognition is one the Natural Language Processing applications. The popular models to address the problem of understanding sequential data are LSTM and GRU. These models not only improve upon the long term dependencies but provide a good understanding of the context to predict tokens and their tags. However, gradient vanishing over deeper and longer layers has been an issue with these state-of-the-art models as well. Hence, IndRNN, which has recently been proposed as an alternative for sequential data processing has been introduced in our approach. Its application on NER has still not been discovered. This thesis work deals with the effectiveness of independently recurrent neural network on the task of Named Entity Recognition. IndRNN provides independence within the layers which helps improve the understanding of the functioning of the neural network. Also, each RNN is connected to each RNN from another layer which provides the correlation between text that we need.