nerrs 0.1.0

CRF-based Named Entity Recognition for Persian text
Documentation

nerrs

CRF-based Named Entity Recognition for Persian text, written in Rust.

Uses the averaged structured perceptron from crfrs with Viterbi decoding. You supply a BIO-tagged corpus; the crate handles feature extraction, training, and inference.

Features

  • BIO-encoded NER with 9 entity types (PER, ORG, LOC, DAT, TIM, MON, PCT, EVE)
  • Persian-specific feature extraction (character n-grams, orthographic flags, context window)
  • Train from any IOB-tagged corpus
  • Save/load models as JSON
  • Token-level evaluation

Usage

[dependencies]
nerrs = "0.1"

use nerrs::{NerTagger, crfrs::TrainConfig};

// Train
let corpus: Vec<Vec<(String, String)>> = vec![/* (word, BIO-tag) pairs */];
NerTagger::train_and_save(&corpus, "ner.model", TrainConfig::default()).unwrap();

// Tag
let mut tagger = NerTagger::new();
tagger.load_model("ner.model").unwrap();
let tags = tagger.tag(&["علی", "به", "تهران", "رفت", "."]).unwrap();
// → [("علی","B-PER"), ("به","O"), ("تهران","B-LOC"), ("رفت","O"), (".","O")]

API

Item Description
NerTagger::new() Create a tagger with no loaded model
NerTagger::train_and_save(corpus, path, config) Train and save model to disk
NerTagger::fit(&mut self, corpus, path, config) Train, save, and load in one step
NerTagger::load_model(path) Load a previously saved model
NerTagger::tag(words) Tag a single sentence
NerTagger::tag_sents(sents) Tag multiple sentences
NerTagger::evaluate(test) Token-level accuracy on a labelled corpus

Entity types

Tag Meaning
O Outside any entity
B-PER / I-PER Person name
B-ORG / I-ORG Organisation
B-LOC / I-LOC Location
B-DAT / I-DAT Date
B-TIM / I-TIM Time
B-MON / I-MON Money / currency amount
B-PCT / I-PCT Percentage
B-EVE / I-EVE Event

License

MIT