Využití hlubokých neuronových sítí při tvorbě chatovacího robota

The thesis is about usage of deep learning generative model for natural language processing in order to create a chatting robot (chatbot). Chatbot will to some extent help the worker of customer support answer questions from customers.In the first theoretical part of the thesis, the reasons behind a tremendous popularity growth of deep learning methods for natural language processing are being mentioned and subsequently also the most commonly used architectures are described, such as convolutional neural networks or autoencoders.Further on, the thesis concerns a topic of recurrent neural architectures and their extensions such LSTM or GRU cells.Next, the approaches for statistical text modelling in order to effectively represent semantical meaning of text are described. In this section the thesis delves into describing the most commonly used statistical model families, such as bag of words models, n-gram techniques and most importantly also neural vector models such as word2vec or Glove algorithms.In the following section, the thesis concerns generative models for chatbot creation as are for example seq2seq models and compares these approaches with older, rule based chatbots. It also describes problems with such generative models and how to overcome them by using for example an attention mechanism or beam search decoder.The practical part of the work is concerned towards a practical implementation of deep learning generative seq2seq model for a chatbot creation. The implementation uses also neural vector models and methods used for data preprocessing are also described in this section. Also an empirical evaluation of the obtained results is carried out.At last, the thesis also mentions some ethical questions which arise with the usage of deep learning architectures in commercial applications.