By looking at rejected addresses with scores just under 90, we might find tweaks to fault penalty values that reduce the number of false negatives. The impact of such changes should be studied before changing production.
One example is changing penalty of STREET_TYPE.missing from 6 to 4. This will allow the common input pattern:
civicNumber streetName locality province
to tolerate minor errors and stay a good match (score > 89). One instance of the pattern is as follows:
4000 seymour victoria bc => 4000 seymour pl saanich bc
This currently scores a 90 so if you spell the street name or locality wrong, or leave out a unitDesignator, the score will descend below 90 and become a bad match. Lowering STREET_TYPE.missing penalty by two points gives some room for these minor errors.
By looking at rejected addresses with scores just under 90, we might find tweaks to fault penalty values that reduce the number of false negatives. The impact of such changes should be studied before changing production.
One example is changing penalty of STREET_TYPE.missing from 6 to 4. This will allow the common input pattern:
civicNumber streetName locality province
to tolerate minor errors and stay a good match (score > 89). One instance of the pattern is as follows:
4000 seymour victoria bc => 4000 seymour pl saanich bc
This currently scores a 90 so if you spell the street name or locality wrong, or leave out a unitDesignator, the score will descend below 90 and become a bad match. Lowering STREET_TYPE.missing penalty by two points gives some room for these minor errors.