Struct MentionMatcher 

pub struct MentionMatcher { /* private fields */ }

A gazetteer compiled for one-pass matching.

Scanning a document once per phrase is O(document × phrases), and on a large corpus the gazetteer grows into the thousands, so projection was quadratic in corpus size — the reference implementation avoids this only by having a neural tagger read each document once, which the model-free build has no equivalent for. An Aho-Corasick automaton is that equivalent: built once from every phrase, it finds all of them in a single pass over the text, independent of how many phrases there are.

The automaton is a prefilter. It finds substring occurrences; word_spans then confirms word boundaries on the few phrases that occurred, so the result is exactly mentions.iter().filter(|m| contains_term(doc, m)) would give — a word-boundary match is necessarily a substring match — at a fraction of the cost.

Implementations

impl MentionMatcher

pub fn new(mentions: &[String]) -> MentionMatcher

pub fn present<'a>(&'a self, doc: &str) -> Vec<&'a String>

Which registered surfaces occur in doc as whole words, in registration order, deduplicated.