Combining Knowledge-driven & Data-driven AI in Real-World NLP Solutions

Knowledge-driven expert systems have been considered "too brittle/fragile" for real-world applications.
Well-engineered
brute-force end-2-end machine-learning solutions are often considered a
practical approach to overcome expert systems limitations.
But the
experience in real world projects shows that this is not true, and both
approach "alone" are often doomed to failure when a model needs to be
put into work in a real-world scenario.

In this presentation I will try to:Promote a culture of critical evaluation of ML solutions (no myths, no magic, no free lunch, complexity requires hard work :) )Present
Expert.ai current approach (which is not perfect and could suffer from a
number of the issues I will be presenting) and possible future
directions.