sv-ngram-index

A Lucene n-gram frequency index (1/2/3-grams) for Swedish, built from sv-ngram-corpus to evaluate confusion-pair rules for Swedish in LanguageTool (katalystnord/languagetool, sv-improvements branch). LanguageTool already ships this kind of index for English, German, French, Spanish, Italian, Russian, Chinese, and Hebrew; This is the first for Swedish.

This holds aggregate n-gram counts, not the underlying text. The text itself is in sv-ngram-corpus, kept in a separate repo.

Structure

sv/
  1grams/   124 MB
  2grams/   1.3 GB
  3grams/   3.7 GB

Built from the combined corpus (2.84 GB, 15,702,791 lines: Wikipedia plus c4-sv shards 0-175 of 1024), 44.1 GB raw, force-merged down to about 5.1 GB total. Drop-in compatible with org.languagetool.dev.bigdata.ConfusionRuleEvaluator's languageModelTopDir argument, the same way as LanguageTool's own official per-language n-gram directories.

Building it

cat corpus/wikipedia-svwiki-2026-09-09.txt corpus/c4-sv-shards000-175.txt > combined.txt
xz -k combined.txt
java -cp <languagetool-dev classpath> \
  org.languagetool.dev.bigdata.CommonCrawlToNgram3 \
  sv combined.txt.xz ./ngram-index

(from a checkout of sv-ngram-corpus.)

Reproducing a confusion pair's numbers

Every pair in confusion_sets.txt on sv-improvements carries its precision and recall as an inline comment, for example:

att; ett; 10;    # p=0.995, r=0.964, f0.5=0.989, s=0.995, 1000+1000, 3grams, 2026-09-13, ...

To reproduce one:

java -cp <classpath> org.languagetool.dev.bigdata.ConfusionRuleEvaluator \
  att ett sv /path/to/this/repo combined.txt

using a checkout of sv-ngram-corpus from the same point in time as the index revision being checked. Tagged commits (for example v26-index) exist for this reason: a pair's recorded numbers are only guaranteed to reproduce against the index that was current when it was evaluated, not necessarily main.

This index will keep growing as more of c4-sv's shards get mined. Each new batch means a full rebuild, not an incremental update; that's how Lucene index construction works here.

License

Aggregate counts derived from Swedish Wikipedia (CC BY-SA 4.0) and c4-sv (ODC-BY). See sv-ngram-corpus's README for the attribution statement. Released under the same two licenses as the source text.

Related

• katalystnord/sv-ngram-corpus: the corpus text and build recipe behind this index.
• katalystnord/languagetool (sv-improvements branch): the rule work this index supports.
• katalystnord/LT-Swedish-grammer: project notes.