rucrf contains a trainer and an estimator for Conditional Random Fields (CRFs). This library supports: - [x] lattices with variable length edges, - [x] L1 and L2 regularization, and - [x] multi-threading during training.
```rust use std::num::NonZeroU32;
use rucrf::{Edge, FeatureProvider, FeatureSet, Lattice, Model, Trainer};
// Train: // 京(kyo) 都(to) // 東(to) 京(kyo) // 京(kei) 浜(hin) // 京(kyo) の(no) 都(miyako) // // Test: // 水(mizu) の(no) 都(miyako) // // 1-gram features: // 京: 1, 都: 2, 東: 3, 浜: 4, の: 5, 水: 6 // 2-gram features: // kyo: 1, to: 2, kei: 3, hin: 4, no: 5, miyako: 6, mizu: 7
let mut provider = FeatureProvider::new(); let label京kyo = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(1).unwrap()], &[NonZeroU32::new(1)], &[NonZeroU32::new(1)], ))?; let label都to = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(2).unwrap()], &[NonZeroU32::new(2)], &[NonZeroU32::new(2)], ))?; let label東to = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(3).unwrap()], &[NonZeroU32::new(2)], &[NonZeroU32::new(2)], ))?; let label京kei = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(1).unwrap()], &[NonZeroU32::new(3)], &[NonZeroU32::new(3)], ))?; let label浜hin = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(4).unwrap()], &[NonZeroU32::new(4)], &[NonZeroU32::new(4)], ))?; let labelのno = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(5).unwrap()], &[NonZeroU32::new(5)], &[NonZeroU32::new(5)], ))?; let label都miyako = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(2).unwrap()], &[NonZeroU32::new(6)], &[NonZeroU32::new(6)], ))?; let label水mizu = provider.addfeatureset(FeatureSet::new( &[NonZeroU32::new(6).unwrap()], &[NonZeroU32::new(7)], &[NonZeroU32::new(7)], ))?;
let mut lattices = vec![];
// 京都 (kyo to) let mut lattice = Lattice::new(2)?; lattice.addedge(0, Edge::new(1, label京kyo))?; lattice.addedge(1, Edge::new(2, label都to))?;
lattice.addedge(0, Edge::new(1, label京kei))?; lattice.addedge(1, Edge::new(2, label都miyako))?;
lattices.push(lattice);
// 東京 (to kyo) let mut lattice = Lattice::new(2)?; lattice.addedge(0, Edge::new(1, label東to))?; lattice.addedge(1, Edge::new(2, label京kyo))?;
lattice.addedge(1, Edge::new(2, label京kei))?;
lattices.push(lattice);
// 京浜 (kei hin) let mut lattice = Lattice::new(2)?; lattice.addedge(0, Edge::new(1, label京kei))?; lattice.addedge(1, Edge::new(2, label浜hin))?;
lattice.addedge(0, Edge::new(1, label京kyo))?;
lattices.push(lattice);
// 京の都 (kyo no miyako) let mut lattice = Lattice::new(3)?; lattice.addedge(0, Edge::new(1, label京kyo))?; lattice.addedge(1, Edge::new(2, labelのno))?; lattice.addedge(2, Edge::new(3, label都miyako))?;
lattice.addedge(0, Edge::new(1, label京kei))?; lattice.addedge(2, Edge::new(3, label都to))?;
lattices.push(lattice);
// Generates a model let trainer = Trainer::new(); let model = trainer.train(&lattices, provider);
// 水の都 (mizu no miyako) let mut lattice = Lattice::new(3)?; lattice.addedge(0, Edge::new(1, label水mizu))?; lattice.addedge(1, Edge::new(2, labelのno))?; lattice.addedge(2, Edge::new(3, label都to))?; lattice.addedge(2, Edge::new(3, label都miyako))?;
let (path, ) = model.searchbest_path(&lattice);
asserteq!(vec![ Edge::new(1, label水mizu), Edge::new(2, labelのno), Edge::new(3, label都miyako), ], path); ```
Licensed under either of
at your option.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.