sv-ngram-corpus

I'm building this corpus while tuning confusion-pair rules for Swedish in LanguageTool, on the sv-improvements branch of katalystnord/languagetool.

LanguageTool already ships n-gram based confusion detection (de/dem, den/det, and similar pairs) for several languages; Swedish never had a corpus or index behind it, hence this corpus and work. I'm publishing it here so anyone else working on Swedish NLP can use the same material as a reference rather than rebuilding it from scratch.

Contents

File	Source	Size	Lines
corpus/wikipedia-svwiki-2026-09-09.txt	Swedish Wikipedia	581 MB	3,326,770
corpus/c4-sv-shards000-175.txt	c4-sv (allenai/c4), shards 0-175 of 1024	2.2 GB	12,376,021

Each file is added once and never edited afterward: a new batch means a new file, not a change to an existing one (see "Extending this corpus" below).

To get one combined file, for example for org.languagetool.dev.bigdata.CommonCrawlToNgram3:

cat corpus/wikipedia-svwiki-2026-09-09.txt corpus/c4-sv-shards000-175.txt > combined.txt

How it was built

Wikipedia. I downloaded the "latest" Swedish Wikipedia dump from dumps.wikimedia.org on 2026-09-09 (1.5 GB, pages-articles.xml.bz2). A first attempt at extracting all roughly 2.7M articles with 20 parallel WikiExtractor processes drowned my computer, so I had to cut it down to a smaller, well-formed subset instead: 377,994 articles, extracted single-threaded with 3 processes. I checked the result for Lsjbot contamination specifically (a bot that once wrote close to half of Swedish Wikipedia as short templated species and geography stubs): only 356 of the 3,326,770 lines, 0.01%, match its boilerplate phrasing, and a manual read of a sample came back as solid, varied, human-written prose.

Because I used the "latest" URL rather than a dated snapshot, this file isn't reproducible byte-for-byte against a specific archived dump after the fact. Anyone extending this who needs exact reproducibility should pin a dated dump instead; I didn't need that for tuning n-gram statistics.

c4-sv. Shards 0-175 of the 1024 multilingual/c4-sv.tfrecord-*-of-01024.json.gz shards on allenai/c4, downloaded via huggingface_hub. A document is kept only if langdetect calls it Swedish and ftfy finds no encoding damage; paragraphs under 20 characters are dropped. The Hugging Face shards are fixed and versioned, so this part is exactly reproducible: I confirmed that by redownloading shard 0 and comparing it byte-for-byte, matching SHA-256, against the original download.

Full script: scripts/build_corpus.sh.

License

Swedish Wikipedia text is CC BY-SA 4.0 (or GFDL, reuser's choice); Wikimedia's own terms of use invite bulk reuse and redistribution, with attribution satisfied by a link back to the source article or its history page. c4-sv is ODC-BY; AllenAI released it specifically to permit reuse and redistribution with attribution.

Concatenating the two into one corpus file is a compilation, not a derivative of either source. That combination doesn't trigger CC BY-SA's ShareAlike clause, which covers transforming a work, not aggregating it next to other material. Use this corpus, in whole or in part, under the same two licenses.

Extending this corpus

More c4-sv shards get downloaded and processed over time (176 of 1024 so far). To add a batch:

1. Run scripts/build_corpus.sh c4-sv <start> <end> for the next shard range, or scripts/build_corpus.sh wikipedia [YYYYMMDD] for a newer Wikipedia snapshot.
2. Add the resulting file to corpus/ and a row to the table above.
3. Rebuild sv-ngram-index from the full set of files and tag the commit. That index needs a full rebuild each time; it can't be extended incrementally the way this corpus can.

Related

• katalystnord/sv-ngram-index: the Lucene n-gram index built from this corpus.
• katalystnord/languagetool (sv-improvements branch): the rule work this corpus supports.
• katalystnord/swedish-pos-dict: the POS/morphology dictionary pipeline.
• katalystnord/LT-Swedish-grammer: project notes.