Testing speller suggestions
We test speller suggestion with divvunspell.
Test procedures
Testing typos.tsv on the command line
Stand in any language repository, and do:
./autogen.sh
./configure --enable-spellers
make -j
make check -j
These commands build the speller for the selected language, and then run (among other things) a test that generates a speller accuracy report at docs/typosreport/report.json (plus docs/typosreport/report-<variant>.json for each speller variant, if the language has any). The report is written only if divvunspell (with the accuracy subcommand) is installed; otherwise that test is skipped.
Viewing the test report
On the web
The accuracy viewer is served by the shared Jekyll theme (giellalt/jekyll-theme-giellalt), and the report data is published automatically on every CI build — you do not commit report.json. After your speller changes are merged and the build has run, the latest report is at:
https://giellalt.github.io/lang-XXX/typosreport/
View locally
The viewer is a WebAssembly app that lives in the theme, not in the language repo, so a bare python3 -m http.server in docs/ no longer shows anything. Instead, run the helper script — it fetches the viewer from the theme, puts your local report.json next to it, and serves the two together:
docs/typosreport/preview.sh
Then open http://localhost:8000/ (the script tries to open it for you).
Options: --port N to change the port, --refresh to re-download the viewer, --no-open to skip opening a browser. The script needs curl and python3; everything it downloads goes in gitignored scratch dirs under docs/typosreport/.
Re-run make check -j to regenerate report.json, then re-run the script (or just reload the page after re-running the script) to see the new numbers.
Testing arbitrary tsv files on the command line
To try a typos.tsv that isn’t in a language repo, generate a report with divvunspell directly and view it with the same script. Build the speller first (make -j in the language repo), then:
divvunspell accuracy --verbose \
-o /path/to/lang-fit/docs/typosreport/report.json \
my-typos.tsv \
/path/to/lang-fit/tools/spellcheckers/fit.zhfst
/path/to/lang-fit/docs/typosreport/preview.sh
my-typos.tsv is a tab-separated input<TAB>expected list; rows with an empty expected column are treated as correct words (to measure false positives). --verbose adds the per-suggestion weight breakdown (lexicon / mutator / reweight) to the report.
Running a wordlist through divvunspell
The following command takes a wordlist (here, for fit), spits it out in json format with 5 suggestions, and shows only the words that according to divvunspell are spelled incorrectly.
cat wordlist | divvunspell suggest --json -n 5 -a tools/spellcheckers/fit.zhfst |jq -c|tr '{' '\n'|grep -v ":true"
Evaluate test results
In order to evaluate the effect of the suggestion mechanism, one must for each target word (each suggestion) take the corpus weight into consideration. The corpus weight of each target word we get as follows:
hfst-lookup tools/spellcheckers/analyser-desktopspeller-gt-norm.hfst