When the machine called sarcasm a compliment

When the machine called sarcasm a compliment
It was a Friday when an automated sentiment model told a sneaker brand the launch was 78% positive. The social lead felt the floor was on fire. She pulled 200 mentions and hand-coded them. True positive: 51%. The model had read "oh great, another drop I can't afford" as praise.

Sarcasm, slang, and a regional in-joke had fooled it on roughly one in four posts. On a 9,000-mention launch, that's a 2,200-post error in the wrong direction.

The fix wasn't to abandon the model. They kept it for tracking week-over-week trend, where consistent bias cancels out, and hand-coded a 200-post sample whenever they needed an absolute number for the boardroom.

The takeaway: trust automated sentiment for direction, never for the decimal. The model is a compass, not a scale.