Developing a Complete Dialogue System Using Long Short-Term Memory

As technologies of natural language understanding and generation improve, the human interest towards human-computer interaction increases. The technologies can be applied for various applications of customer services. Most works related to this field are emphasizing on single sentence and speaker turn. Meanwhile, a conversation sometimes has its own context according to the previous one. Designing this kind of conversational system is challenging. Most conversational agents are built based on knowledge-based and rule based systems. This paper discusses a development of a complete dialogue system to understand the intent of a text and give response based on the dialogue state. The dialogue model is implemented using the combination of rule-based and data-driven approach by utilizing a long short-term memory (LSTM). Some experiments show that the developed system give a high performance. A detail observation informs that some errors come from the intent classifier that fails to classify some sentences not in the corpus. This system can be improved by increasing the performance of the intent classifier and incorporating an additional named entity recognition module.