Shallow NLP: processing with minimal understanding of context.

Contextual information

• Spatial: "find the nearest X."
• Temporal: "meet at lunch time."

other implied contexts like "took it home" can mean a prize (because other references to taking a prize home exist)

mixtures of when to generalize or be specific. may have to infer "Taking the prize home" meaning 'prize' is an 'it' which means this response *could* answer "take 'it' home"

Watson "questions"

A question is a query to Watson.

A question goes through a set of steps.

• Feature extraction
• Generate hypotheses: find responses in a corpus with a suspected match
• Reasoning election: library of "reasoning algorithms" scores each hypothesis to the question
• Gradient boosting: scores adjusted based on their historic accuracy during training