Text Classification With Deep Neural Networks

The thesis explores different extensions of Deep Neural Networks in learning underlying natural language representations and how to apply them in Natural Language Processing tasks. Novel methods of learning lower or higher level features of natural languages are given in which word and phrase dense representations are derived from unlabelled corpora. Word representations are learned by training Deep Neural Networks to predict context from each sentence while phrase representations are learned by unsupervised learning with Convolutional Restricted Boltzmann Machine. It is shown that word representations learned from architectures which preserve text input as sequences have better word similarity and relatedness than bag-of-word approaches. Additionally phrase representations learned with Convolutional Restricted Boltzmann Machine when combined with bag-of-word features improve results of text classification tasks over only bag-of-word features. Beside learning word and phrase representations, to the best of my knowledge, the work in the thesis is first to explore Deep Neural Networks in Adverse Drug Reaction detection task where my architectures when used with pre-trained word representations significantly outperform the state-of-the-art models. In addition, outputs from my proposed attentional architecture can be used to highlight important word spans without explicit training labels. In the future I propose the learned representations to be used with the discussed Deep Neural Networks in different NLP tasks such as Dialog Systems, Machine Translation or Natural Language Inference.