NLP in Swedish: resources and pitfalls

Swedish resources to know about:

Resource	What	The source
KB-BERT, KB-Whisper	an encoder and speech recognition	the National Library of Sweden
GPT-SW3	generative models for the Nordic languages	AI Sweden
SUC	a balanced corpus with parts of speech	Språkbanken
Talbanken	a dependency treebank	Språkbanken
SweDN, SuperLim	benchmark suites for Swedish	Språkbanken
The Riksdag's open data	speeches and documents	the Riksdag
Wikipedia sv	~2.6 million articles	CC BY-SA

Three Swedish peculiarities that genuinely matter:

1. Compounds. «arbetsmarknadsutbildning» (labour market training) is one word. The tokenizer splits it into pieces and search systems miss the parts.
2. Rich morphology. The definite form, the genitive, the plural — character-based metrics (chrF) and lemmatisation help.
3. A small amount of data compared with English, both in pretraining and in evaluation suites.

Measure instead of assuming. Four measurements that decide whether a model will do for Swedish:

1. Tokens per word on representative Swedish text. An English-centric tokenizer gives 2.0–2.7; a Nordic one 1.3–1.6. That affects both the cost and the effective context length directly.
2. Factual knowledge about Sweden — your own questions with an answer key (public agencies, laws, geography, history).
3. Language quality — a blinded human judgement of fluency and idiom. Anglicisms and directly translated expressions are common.
4. Your actual task on Swedish data. That is the only thing that decides.

Common errors in Swedish NLP projects:

• Using English stop-word lists or lemmatisers.
• Comparing against English benchmark figures and assuming they hold.
• Forgetting that compounds make keyword search unreliable — here subword indexing or a hybrid with embeddings helps.
• Treating «å», «ä», «ö» as special characters in normalisation.

When Swedish is not enough: RAG against Swedish sources solves factual knowledge most cheaply; fine-tuning solves style and terminology; switching to a Nordic model solves the tokenisation.