Ranking suggestions in divvunspell
Background
GiellaLT facilitates the use of two epeller engines:
We have recently (2024) moved to using divvunspell for the GiellaLT spellers.
- The Speller Error Model page documents how to rank correction suggestions based on letter substitutions.
Speller testing with divvunspell
There’s a prototype-level testing tool in the divvunspell directory. In order to use it, clone divvunspell (see the README file for details. Note that you will need rust to use divvunspell.
Use divvunspell like this (here with sma as an example).:
accuracy -o support/accuracy-viewer/public/report.json typos.txt sma.zhfst
cd support/accuracy-viewer
npm i && npm run dev
View in http://localhost:5000 (where the 5-digit number is given in the feedback.
More info by accuracy --help.
Using the results
The penalty points are explained on the Speller Error Model page. The goal is to get values for corrections as high as possible, this may be done by tweaking the penalty points.