Information extraction and NER

NER marks spans up in text: people, organisations, places, dates, amounts. The standard format is BIO:

Anna     B-PER
Svensson I-PER
works    O
at       O
Volvo    B-ORG
in       O
Göteborg B-LOC

B = the beginning of an entity, I = inside, O = outside.

Two approaches today:

	A fine-tuned encoder	An LLM with a schema
The data required	500+ labelled sentences	0–20 examples
The cost per document	very low	high
New entity types	retraining	change the prompt
Precision on the trained types	higher	somewhat lower

For large volumes and fixed types the encoder wins. For few documents or changing types the LLM wins.

Evaluate at the entity level, not the token level. A span counts as correct only if both the boundaries and the type are right. That is harder than a token-wise measurement and is what actually matters.

An error like «Anna Svensson» → «Svensson» therefore counts as both a miss (FN) and a false one (FP) — not as half right. seqeval does this correctly; an ordinary classification_report at the token level systematically overestimates.

Relation extraction goes a step further: find that «Anna Svensson» works at «Volvo». The common approaches are pairwise classification of entity pairs, or an LLM with a schema that requires both entities and relations in JSON.

Pitfalls specific to Swedish: compounds («Volvochefen» contains an organisation), genitive forms, and Swedish NER corpora being considerably smaller than English ones. KB-lab's models and the SUC corpus are the starting points.

from transformers import pipeline
from seqeval.metrics import classification_report

# A fine-tuned Swedish model
ner = pipeline("ner", model="KBLab/bert-base-swedish-cased-ner", aggregation_strategy="simple")
for e in ner("Anna Svensson arbetar på Volvo i Göteborg sedan 2019."):
    print(f"{e['entity_group']:6s} {e['word']:20s} {e['score']:.2f}")
# PER    Anna Svensson        0.99
# ORG    Volvo                0.98
# LOC    Göteborg             0.99

# Evaluation at the ENTITY level
true = [["B-PER", "I-PER", "O", "O", "B-ORG", "O", "B-LOC"]]
pred = [["B-PER", "I-PER", "O", "O", "B-ORG", "O", "O"]]
print(classification_report(true, pred, digits=3))

# The LLM alternative with a schema
from pydantic import BaseModel
class Entity(BaseModel):
    text: str
    type: str           # PER|ORG|LOC|TIME|AMOUNT
    start: int
class Extraction(BaseModel):
    entities: list[Entity]

Always compare the two on your material before choosing — the difference in precision and cost varies greatly with the domain.