Due to the recent strong growth of speech recognition and synthesis technologies, smart devices leveraging natural language
interfaces (NLI) are heavily promoted to the consumer market. With Google and Amazon taking the lead, major global
IT and telecom companies are entering fierce competition to enable NLI on all kinds of devices.
In fact, natural language processing (NLP) technology still faces many challenges when it comes to modeling dialogs accurately, and often it is difficult to use NLI in its desired form. In order to understand natural language accurately, it is important to understand human knowledge. This task is not a simple text processing technique. Advanced NLP and NLU require enormous general knowledge, understanding the human reasoning system and generic patterns in human conversation. Well-labeled data is required to process these types of complex information and patterns correctly, which takes much effort and time to produce such data, and also, the information itself is constantly changing. Finding new ways for synthesizing these massive amount of information is the field of artificial intelligence.
Our focus on conversational AI is to gradually develop technologies that can enable human-like conversations. Starting with recognizing the user’s intentions and slots from small labeled data, we are further developing technologies for modeling the dialog flow, and progressive learning in a semi-supervised manner from unlabeled data.
- Common patterns are pretrained from generic datasets, making it possible to learn intentions and slots with a small amount of domain-specific data.
- Dialog flow modeling
- For Korean language, higher intent classification precision (max over 20%) and recall (max over 40%) compared to IBM Watson®, MS LUIS, and Google DialogFlow.
- *Based on Korean customer service dataset.