Deep Graph-Long Short-Term Memory: A Deep Learning Based Approach for Text Classification

Abstract

Multi-label text classification is a challenging task in many real applications. Mostly, in all the traditional techniques, word2vec is used to show the sequential information among text. However, use of word2vec ignores logic and context relationship among text, and we treat each label as an individual unit. Therefore, the existing techniques failed to reflect the real scenarios and to gain the semantic information regarding the relationship among texts. In this paper, we propose a model Deep Graph-Long Short-Term Memory (DG-LSTM) for multi-label text classification. In the proposed model, we store the documents using the graph database. Initially, the documents are pre-processed using standard dictionaries, and afterwards it generates the classified dictionaries. These classified dictionaries are used to generate the subgraphs. The model maintains a lookup table to reduce the search space for the new documents. For classification, the model uses the deep learning technique DG-LSTM. DG-LSTM is using Deep Graph_Rectified Linear Unit activation function to avoid blow-up and dying neuron problem of Rectified Linear Unit activation function. We verify the proposed model on the legal case of Indian judiciary. The results show that the proposed model has achieved 99% accuracy to classify the fresh case into its corresponding category.