The paper describing it is well over 7000 words(!), but basically through natural language processing and machine learning.