Why neural networks arent fit for natural language understanding