Module span_f1 

Span-based (entity-level) precision / recall / F1 — the seqeval metric.

Token-level accuracy is a poor measure for sequence labelling: a model that gets every O right but mangles entity boundaries can still score highly. The standard NER / chunking metric (CoNLL-2000/2003, implemented by the seqeval library) instead compares entity spans: a predicted span is a true positive only when an identical span — same entity type and same start/end boundaries — appears in the gold sequence.

Given gold tags y and predicted tags ŷ, spans are extracted with crate::tagging::bioes::extract_spans and matched exactly:

precision = |gold ∩ pred| / |pred|
recall    = |gold ∩ pred| / |gold|
f1        = 2 P R / (P + R)

Per-type breakdowns and a micro-averaged total are provided.

Structs

PrfScore

Precision / recall / F1 with the raw TP/FP/FN counts behind them.

SpanF1Report

Full span-F1 report: a micro-averaged overall score plus per-type scores.

Functions

span_f1

Compute span-F1 directly from gold and predicted tag sequences.

span_f1_from_spans

Compute the micro-averaged span-F1 from pre-extracted gold/predicted spans.