Natural Language Understanding (NLU)

Natural Language Understanding (NLU) aims to transform natural language texts (sentences or questions) into their formal meaning representations as machine executable programs. Currently, we are working on two NLU engines for the question answering task, based on knowledge graph and web tables respectively.

Knowledge Graph-based NLU engine parses a natural language question into a lambda-calculus query based on a given knowledge graph (such as Freebased or Satori). Table-based NLU engine parses a natural language question into a SQL query based on a given table. We focus on table because it is one of the most common and valuable data sources across organizations in multiple domains. Our NLU engines are designed to handle questions with different types, such as single-relation questions, multi-hop questions, multi-constraint questions, Yes/No questions, multi-choice questions, common-sense questions, domain-specific questions, and etc. Both Bing and Xiaoice have already used our NLU engines in their QA and chat scenarios.

Recently, we are working on building a comprehensive & large-scale semantic parsing dataset, and will release it soon!