Natural Language Understanding for Conversational Agents

In recent years, intelligent assistants have gained great popularity since they provide a new way for people to interact with the Internet. These assistants are acting as conversational agents that can interpret and respond via human language automatically. An important and basic question here is how to correctly understand the input natural language. To answer this question, this research focuses on building an intelligent natural language understanding (NLU) system for conversational agents. On one hand, we need the system to understand the natural language exhaustively. Specifically, we need the system to understand the sentence in two levels, both token-level and sentence-level. Two different tasks are involved including named entity recognition and intent detection. On the other hand, we want to build a human-like system that can dynamically update itself to recognize upcoming new classes with a few examples. Different directions are explored for NLU tasks, including zero-shot learning, few-shot learning and incremental few-shot learning.