Enhancing Text Classification with Modular Deep Encoder-Decoder Networks for Multi-modal Data

Our study centers on a Modular Deep Encoder-Decoder network (MDED) about text classification. The basic idea behind MDED is to extend a pre-trained model, which consists of a set of deep encoder-decoders, with an additional source of data modality by adding a new encoder to the network. MDED has been applied to text classification on Saudi Newspapers Arabic Corpus (SaudiNewsNet). A modular LSTM is built. As per the outcomes of our experiment, we found that algorithms can efficiently combine new data sources and pre-trained models. When adding a new data source, the training process can converge into a local optimum in a few epochs. With an excellent initial encoder-decoder, the modular LSTM networks can improve classification accuracy by up to 0.51%.