NorskHelsenett/LOS_Document_classification_ETI

LOS Document Classification (ETI)

Norwegian public-sector documents labelled with their top-level LOS category (Felles vokabular / common vocabulary, level 1). Built to train a small, self-hostable text classifier by distilling LLM-generated labels.

Splits

split	rows	labels	use
train	~1982	silver — LLM-generated (Claude Haiku 4.5, few-shot)	training
test	551	human gold — manually annotated	honest evaluation

The train labels are weak supervision (macro-F1 ≈ 0.70 vs. the human set); the test split is the manually-annotated ground truth. URLs overlapping between the two were removed from train to prevent leakage. Primary metric: macro-F1 on test.

Columns

column	type	description
level	int	LOS level (1 for this release)
url	string	source document URL
text	string	full document text (reconstructed from the RAG chunk store)
summary	string	concise LLM summary (<512 tokens; for classification-on-summary + RAG)
class	string	LOS level-1 category (one of six)

Labels (LOS level 1)

Arbeid, Helse og omsorg, Innbygger, Kultur-Idrett-Fritid, Opplæring og utdanning, Sosial og økonomisk trygghet

The distribution is heavily imbalanced (~64% Helse og omsorg; Arbeid/Innbygger/ Kultur-Idrett-Fritid each <5%), so macro-F1 is the appropriate metric.

How it was built

1. Text — Norwegian public-sector pages, reconstructed from the project's Elasticsearch chunk index (chunks ordered by sequence and de-duplicated).
2. Labels — one LOS level-1 category per document via few-shot LLM prompting (Claude Haiku 4.5, temperature 0).
3. Summaries — one standalone <512-token summary per document (Azure OpenAI), capturing topic, audience, service type and key entities — usable both as a compact classifier input and as a contextual-RAG representation.
4. Cleaning — documents with <300 chars of source text (unreliable summary/label) and non-Norwegian encyclopedic content were removed (~10 docs).

Intended use

Training/evaluating a sequence-classification model (e.g. XLM-RoBERTa) to tag Norwegian public documents with LOS level-1 categories, deployable via Text Embeddings Inference (TEI) or Infinity. Recommended: evaluate on the separate human-labelled set for honest numbers.

Limitations

• Silver labels (LLM, not human) — bounded by the labelling model's accuracy.
• Severe class imbalance; rare classes have few examples.
• text is a chunk reconstruction, not a fresh scrape (may differ slightly from the live page).

Provenance

Produced for the ETI project (Norsk helsenett SF). LOS taxonomy by Digdir: https://www.digdir.no/informasjonsforvaltning/los-felles-vokabular-klassifisering-av-offentlige-tjenester-og-ressurser/2434