Statistics, and the fluent join#
SLOP8217 · Week 4
A rule-based system breaks in a way that looks broken.
Where the training text came from, why the errors moved from the word to the sentence, and how to date an artefact by the shape of its failure
A statistical one breaks in a way that looks like English for six words and then turns.
The reason is where the text came from.
Outline#
- parallel corpora: what was available, and what that made the systems good at
- the phrase table as a suspect with a memory but no understanding
Outline#
- locating a join, and what its position tells you about unit length
- dating an artefact from the shape of its errors
At the bench on Wednesday
Find the seam#
Wednesday, 17 March
- Material
- signage from before 2016
- Operation
- locating the join
- Leaves
- a dated estimate and the reasoning behind it