Skip to content

Latest commit

 

History

History
86 lines (60 loc) · 3.63 KB

README.md

File metadata and controls

86 lines (60 loc) · 3.63 KB
Raw

KENLM-RS

This crate is an experimental Rust wrapper around the kenlm C++ library. It uses autocxx to generate the glue-code. It uses a modified version of kenlm located in src/cxx/lm and src/cxx/util. Most of the modifications add constructor functions or expose config fields.

Both binary and arpa files are supported. Although you may want to avoid loading .arpa files, KenLM will scream at you from C++.

There is also a initial support for reading files within rust in src/reader/arpa.rs.

Maxorder

Kenlm's build-flag -DKENLM_MAX_ORDER governs the maximal ngram order you'll be able to load with this library. Loading a model with larger order than the library was built with will cause a runtime exception originating in C++. -DKENLM_MAX_ORDER also governs the size of state, you may set it via the env var KENLM_MAX_ORDER or by changing the default value in build.rs. The current default is 3. Increasing it comes at the cost of increased state-sizes.

The a state is a plain-old-data struct with:

[c_uint, KENLM_MAX_ORDER-1]
[f32, KENLM_MAX_ORDER-1]
u8

Usage

Score a sample sentence using a model

$ cargo run --example score_sentence -- --model-path carol.bin "the register of his burial was signed by the clergyman the clerk the undertaker and the chief mourner"
> Total score of the sentence "the register of his burial was signed by the clergyman the clerk the undertaker and the chief mourner" is: -23.76212

Library

use kenlm_rs::Model;

let model = Model::new("test_data/test.bin", false).unwrap();

let mut mem1 = model.new_state();
let mut mem2 = model.new_state();
let bos = true;

if bos {
    model.fill_state_with_bos_context(&mut mem1);
} else {
    model.fill_state_with_null_context(&mut mem1);
}
let mut score = 0f32;
for w in &["what", "a", "lovely", "sentence"] {
    let out = model.score_word_given_state(&mut mem1, &mut mem2, w);
    std::mem::swap(&mut mem1, &mut mem2);
    score += out;
}
eprintln!("{score:?}");

Check out examples/score_sentence.rs, examples/inspect_vocab.rs and Model in src/model/mod.rs.

Modifications to kenlm

Besides formatting, a few functions were added, mostly to provide easier access from Rust or to help autocxx. An incomplete list can be found below:

config.cc

In src/cxx/lm/config.cc, there are 3 added functions.

namespace lm
{
  namespace ngram
  {
    std::unique_ptr<Config> Config_Create();
    void Config_set_load_method(Config &config, util::LoadMethod load_method);
    void Config_set_enumerate_callback(Config &config, EnumerateVocab &enumerateCallback);
  }
}
  • Config_Create is a constructor
  • Config_set_load_method sets the load_method
  • Config_set_enumerate_callback sets the enumerate callback that gets executed for each vocab entry, see VocabCallback in src/cxx/bridge.rs for an example callback.

virtual_interface.cc

In src/cxx/lm/virtual_interface.cc there is a single added function LoadVirtualPtr, it is essentially LoadVirtual but returns a unique pointer.