The model scores the whole message. To see why, it hides one word at a time and checks how much the score changes.
- highlighted words changed the score the most
- The model relied on those words more than others
- Unmarked words had little effect on the result