In this thesis neural network based model for unsupervised modeling of task-oriented dialogue data is proposed. Task-oriented dialogue system recently gained a lot of interest from both academic and industrial areas for its potential for being natural and intuitive interface for applications based on artificial intelligence. However, traditional approaches for modeling task-oriented dialogue had some restrictions concerning with expressiveness of model and scalable training with large amount of data. We propose Slot Entity Memory Network model which is a generative model based on neural networks. Being sufficiently expressive to cope with complex and diverse natural language utterances using neural networks, it can be trained without any human annotated labels on task-oriented dialogues. By incorporating task-oriented dialogue structure in neural network model, Slot Entity Memory Network can help system developers to understand what the model had learned from data. We train the proposed model with CamRest676 dataset and report quantitative and qualitative results with analysis of dialogue structure that model has learned.