Skip to content

Latest commit

 

History

History
165 lines (116 loc) · 4.74 KB

README.md

File metadata and controls

165 lines (116 loc) · 4.74 KB
Raw

divvunspell

An implementation of hfst-ospell in Rust, with added features like tokenization, case handling, and parallelisation.

CI

Building and installing commandline tools

# For the `divvunspell` binary:
cargo install divvunspell-bin

# For `thfst-tools` binary (most people can skip this one):
cargo install thfst-tools

# To build the development version from this source, cd into the relevant directory and:
cargo install --path .

Building with gpt2 support on macOS aarch64

(Skip this if you are not experimenting with gpt2 support. So skip. Now.)

Clone this repo then:

brew install libtorch
LIBTORCH=/opt/homebrew/opt/libtorch cargo build --features gpt2 --bin divvunspell

No Rust?

curl https://sh.rustup.rs -sSf | sh
source $HOME/.cargo/env
rustup default stable
cargo build --release

divvunspell

Usage:

Usage: divvunspell SUBCOMMAND [OPTIONS]

Optional arguments:
  -h, --help  print help message

Available subcommands:
  suggest   get suggestions for provided input
  tokenize  print input in word-separated tokenized form
  predict   predict next words using GPT2 model

$ divvunspell suggest -h
Usage: divvunspell suggest [OPTIONS]

Positional arguments:
  inputs                 words to be processed

Optional arguments:
  -h, --help             print help message
  -a, --archive ARCHIVE  BHFST or ZHFST archive to be used
  -S, --always-suggest   always show suggestions even if word is correct
  -w, --weight WEIGHT    maximum weight limit for suggestions
  -n, --nbest NBEST      maximum number of results
  --no-case-handling     disables case-handling algorithm (makes results more like hfst-ospell)
  --json                 output in JSON format

If you want to debug divvunspell behaviour, simply enable rust's logging features by setting RUST_LOG=trace on your commandline's environment variables.

accuracy

Building:

cd accuracy/
cargo install --path .

The resulting binary accuracy is placed in $HOME/.cargo/bin/, make sure it is on the path.

Usage:

divvunspell-accuracy 1.0.0-beta.1
Accuracy testing for DivvunSpell.

USAGE:
    accuracy [OPTIONS] [ARGS]

FLAGS:
    -h, --help       Prints help information
    -V, --version    Prints version information

OPTIONS:
    -c <config>             Provide JSON config file to override test defaults
    -o <JSON-OUTPUT>        The file path for the JSON report output
    -w <max-words>          Truncate typos list to max number of words specified
    -t <TSV-OUTPUT>         The file path for the TSV line append

ARGS:
    <WORDS>    The 'input -> expected' list in tab-delimited value file (TSV)
    <ZHFST>    Use the given ZHFST file

thfst-tools

Convert hfst and zhfst files to thfst and bhfst formats.

  • thfst: byte-aligned hfst for fast and efficient loading and memory mapping, required to run divvunspell on ARM processors
  • bhfst: thfst files wrapped in a box container; in the case of zhfst files converted to bhfst, the metadata file (index.xml in the zhfst archive) is converted to a json file for faster and leaner processing by the divvunspell library.

Usage:

thfst-tools 1.0.0-alpha.5
Tromsø-Helsinki Finite State Transducer toolkit.

USAGE:
    thfst-tools <SUBCOMMAND>

FLAGS:
    -h, --help       Prints help information
    -V, --version    Prints version information

SUBCOMMANDS:
    bhfst-info         Print metadata for BHFST
    help               Prints this message or the help of the given subcommand(s)
    hfst-to-thfst      Convert an HFST file to THFST
    thfsts-to-bhfst    Convert a THFST acceptor/errmodel pair to BHFST
    zhfst-to-bhfst     Convert a ZHFST file to BHFST

Speller testing

There's a prototype-level testing tool in support/accuracy-viewer. Use it like:

accuracy -o support/accuracy-viewer/public/report.json typos.txt sma.zhfst
cd support/accuracy-viewer
npm i && npm run dev

View in http://localhost:5000.

typos.txt is a TSV file with typos in the first column and expected correction in the second. More info by accuracy --help.

License

The crate divvunspell is licensed under either of

at your option.

The divvunspell, thfst-tools and accuracy binaries are licensed under the GPL version 3 license.

More docs?

We have GitHub pages site for divvunspell with some more tech docs and stuff (WIP).