Steps towards question answering with deep neural memory networks

Since the rise of neural networks in science and industry much progress has been made in the field of Artificial Intelligence. In this thesis, steps towards automatic Question Answering (QA) on text passages are examined by applying deep neural networks with memory components to large-scale datasets for the greater aim of developing the human-like ability of reading some piece of text and answering arbitrary questions on it. We explore the neural models of LSTM and Memory Networks by using three major text datasets of SQuAD with SNLI for QA, and MSRP for Paraphrase Detection. We empirically show that the LSTM achieve high performance for the binary step towards QA and are in addition expandable to Paraphrase Detection with a performance close to the state-of-the-art. Besides, we demonstrate that in contrast to their success on the bAbI tasks Memory Networks are not suitable for QA on the realistic dataset of SQuAD. Finally, we discover the usability of the measure of cosine similarity on the internal hidden vectors of LSTM. The thesis is closed by presenting a semantic search algorithm for text pairs and its further development to use cases for the landscape of business activities