punkt

Build Status Crates.io Docs.rs

Overview

Implementation of Tibor Kiss' and Jan Strunk's Punkt algorithm for sentence tokenization. Results have been compared with small and large texts that have been tokenized using NLTK.

Training

Training data can be provided to a SentenceTokenizer for better results. Data can be acquired manually by training with a Trainer, or using already compiled data from NLTK (example: TrainingData::english()).

Typical Usage

The punkt algorithm allows you to derive all the necessary data to perform sentence tokenization from the document itself.

```rust # let trainer: Trainer = Trainer::new(); let mut data = TrainingData::new();

trainer.train(doc, &mut data);

for s in SentenceTokenizer::::new(doc, &data) { println!("{:?}", s); } ```

rust-punkt also provides pretrained data that can be loaded for certain languages.

rust # # let data = TrainingData::english();

rust-punkt also allows training data to be incrementally gathered.

```rust # let trainer: Trainer = Trainer::new(); let mut data = TrainingData::new();

for d in docs.iter() { trainer.train(d, &mut data);

for s in SentenceTokenizer::::new(d, &data) { println!("{:?}", s); } } ```

Customization

rust-punkt exposes a number of traits to customize how the trainer, sentence tokenizer, and internal tokenizers work. The default settings, which are nearly identical, to the ones available in the Python library are available in punkt::params::Standard.

To modify only how the trainer works:

```rust # struct MyParams;

impl DefinesInternalPunctuation for MyParams {} impl DefinesNonPrefixCharacters for MyParams {} impl DefinesNonWordCharacters for MyParams {} impl DefinesPunctuation for MyParams {} impl DefinesSentenceEndings for MyParams {}

impl TrainerParameters for MyParams { const ABBREVLOWERBOUND: f64 = 0.3; const ABBREVUPPERBOUND: f64 = 8f64; const IGNOREABBREVPENALTY: bool = false; const COLLOCATIONLOWERBOUND: f64 = 7.88; const SENTENCESTARTERLOWERBOUND: f64 = 35f64; const INCLUDEALLCOLLOCATIONS: bool = false; const INCLUDEABBREVCOLLOCATIONS: bool = true; const COLLOCATIONFREQUENCYLOWERBOUND: f64 = 0.8f64; } ```

To fully modify how everything works:

```rust # struct MyParams;

impl DefinesSentenceEndings for MyParams { // const SENTENCEENDINGS: &'static Set = &phfset![...]; }

impl DefinesInternalPunctuation for MyParams { // const INTERNALPUNCTUATION: &'static Set = &phfset![...]; }

impl DefinesNonWordCharacters for MyParams { // const NONWORDCHARS: &'static Set = &phfset![...]; }

impl DefinesPunctuation for MyParams { // const PUNCTUATION: &'static Set = &phf_set![...]; }

impl DefinesNonPrefixCharacters for MyParams { // const NONPREFIXCHARS: &'static Set = &phfset![...]; }

impl TrainerParameters for MyParams { // const ABBREVLOWERBOUND: f64 = ...; // const ABBREVUPPERBOUND: f64 = ...; // const IGNOREABBREVPENALTY: bool = ...; // const COLLOCATIONLOWERBOUND: f64 = ...; // const SENTENCESTARTERLOWERBOUND: f64 = ...; // const INCLUDEALLCOLLOCATIONS: bool = ...; // const INCLUDEABBREVCOLLOCATIONS: bool = true; // const COLLOCATIONFREQUENCYLOWERBOUND: f64 = ...; } ```

Benchmarks

Specs of my machine:

``` test tokenizer::benchsentencetokenizertrainondocumentlong ... bench: 129,877,668 ns/iter (+/- 6,935,294) test tokenizer::benchsentencetokenizertrainondocumentmedium ... bench: 901,867 ns/iter (+/- 12,984) test tokenizer::benchsentencetokenizertrainondocumentshort ... bench: 702,976 ns/iter (+/- 13,554) test tokenizer::wordtokenizerbenchlong ... bench: 14,897,528 ns/iter (+/- 689,138) test tokenizer::wordtokenizerbenchmedium ... bench: 339,535 ns/iter (+/- 21,692) test tokenizer::wordtokenizerbenchshort ... bench: 281,293 ns/iter (+/- 3,256) test tokenizer::wordtokenizerbenchverylong ... bench: 54,256,241 ns/iter (+/- 1,210,575) test trainer::benchtrainerlong ... bench: 27,674,731 ns/iter (+/- 550,338) test trainer::benchtrainermedium ... bench: 681,222 ns/iter (+/- 31,713) test trainer::benchtrainershort ... bench: 527,203 ns/iter (+/- 11,354) test trainer::benchtrainerverylong ... bench: 98,221,585 ns/iter (+/- 5,297,733)

```

Python results for sentence tokenization, and training on the document (the first 3 tests mirrored from above):

The following script was used to benchmark NLTK.

False is being passed into tokenize to prevent NLTK from aligning sentence boundaries. This functionality is currently unimplemented.

python timed = timeit.timeit('s.train(f0); [s for s in s.tokenize(f0, False)]', 'from bench import s, f0', number=tests) print(timed) print(timed / tests)

long - 1.3414202709775418 s = 1.34142 x 10^9 ns ~ 10.3283365927x improvement medium - 0.007250561956316233 s = 7.25056 x 10^6 ns ~ 8.03950245027x improvement short - 0.005532620595768094 s = 5.53262 x 10^6 ns ~ 7.870283759x improvement

License

Licensed under either of

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.