JamSpell

JamSpell is a spell checking library with following features:

• accurate - it considers words surroundings (context) for better correction
• fast - near 5K words per second
• multi-language - it's written in C++ and available for many languages with swig bindings

Colab example

JamSpellPro

jamspell.com - check out a new jamspell version with following features

• Improved accuracy (catboost gradient boosted decision trees candidates ranking model)
• Splits merged words
• Pre-trained models for many languages (small, medium, large) for:
  en, ru, de, fr, it, es, tr, uk, pl, nl, pt, hi, no
• Ability to add words / sentences at runtime
• Fine-tuning / additional training
• Memory optimization for training large models
• Static dictionary support
• Built-in Java, C#, Ruby support
• Windows support

Content

• Benchmarks
• Usage
  • Python
  • C++
  • Other languages
  • HTTP API
• Train

Benchmarks

	Errors	Top 7 Errors	Fix Rate	Top 7 Fix Rate	Broken	Speed (words/second)
JamSpell	3.25%	1.27%	79.53%	84.10%	0.64%	4854
Norvig	7.62%	5.00%	46.58%	66.51%	0.69%	395
Hunspell	13.10%	10.33%	47.52%	68.56%	7.14%	163
Dummy	13.14%	13.14%	0.00%	0.00%	0.00%	-

Model was trained on 300K wikipedia sentences + 300K news sentences (english). 95% was used for train, 5% was used for evaluation. Errors model was used to generate errored text from the original one. JamSpell corrector was compared with Norvig's one, Hunspell and a dummy one (no corrections).

We used following metrics:

• Errors - percent of words with errors after spell checker processed
• Top 7 Errors - percent of words missing in top7 candidated
• Fix Rate - percent of errored words fixed by spell checker
• Top 7 Fix Rate - percent of errored words fixed by one of top7 candidates
• Broken - percent of non-errored words broken by spell checker
• Speed - number of words per second

To ensure that our model is not too overfitted for wikipedia+news we checked it on "The Adventures of Sherlock Holmes" text:

	Errors	Top 7 Errors	Fix Rate	Top 7 Fix Rate	Broken	Speed (words per second)
JamSpell	3.56%	1.27%	72.03%	79.73%	0.50%	5524
Norvig	7.60%	5.30%	35.43%	56.06%	0.45%	647
Hunspell	9.36%	6.44%	39.61%	65.77%	2.95%	284
Dummy	11.16%	11.16%	0.00%	0.00%	0.00%	-

More details about reproducing available in "Train" section.

Usage

Python

1. Install swig3 (usually it is in your distro package manager)

2. Install jamspell:

pip install jamspell

3. Download or train language model

4. Use it:

import jamspell

corrector = jamspell.TSpellCorrector()
corrector.LoadLangModel('en.bin')

corrector.FixFragment('I am the begt spell cherken!')
# u'I am the best spell checker!'

corrector.GetCandidates(['i', 'am', 'the', 'begt', 'spell', 'cherken'], 3)
# (u'best', u'beat', u'belt', u'bet', u'bent', ... )

corrector.GetCandidates(['i', 'am', 'the', 'begt', 'spell', 'cherken'], 5)
# (u'checker', u'chicken', u'checked', u'wherein', u'coherent', ...)

C++

1. Add jamspell and contrib dirs to your project

2. Use it:

#include <jamspell/spell_corrector.hpp>

int main(int argc, const char** argv) {

    NJamSpell::TSpellCorrector corrector;
    corrector.LoadLangModel("model.bin");

    corrector.FixFragment(L"I am the begt spell cherken!");
    // "I am the best spell checker!"

    corrector.GetCandidates({L"i", L"am", L"the", L"begt", L"spell", L"cherken"}, 3);
    // "best", "beat", "belt", "bet", "bent", ... )

    corrector.GetCandidates({L"i", L"am", L"the", L"begt", L"spell", L"cherken"}, 3);
    // "checker", "chicken", "checked", "wherein", "coherent", ... )
    return 0;
}

Other languages

You can generate extensions for other languages using swig tutorial. The swig interface file is jamspell.i. Pull requests with build scripts are welcome.

HTTP API

• Install cmake

• Clone and build jamspell (it includes http server):

git clone https://github.com/bakwc/JamSpell.git
cd JamSpell
mkdir build
cd build
cmake ..
make

• Download or train language model
• Run http server:

./web_server/web_server en.bin localhost 8080

• GET Request example:

$ curl "http://localhost:8080/fix?text=I am the begt spell cherken"
I am the best spell checker

• POST Request example

$ curl -d "I am the begt spell cherken" http://localhost:8080/fix
I am the best spell checker

• Candidate example

curl "http://localhost:8080/candidates?text=I am the begt spell cherken"
# or
curl -d "I am the begt spell cherken" http://localhost:8080/candidates

{
    "results": [
        {
            "candidates": [
                "best",
                "beat",
                "belt",
                "bet",
                "bent",
                "beet",
                "beit"
            ],
            "len": 4,
            "pos_from": 9
        },
        {
            "candidates": [
                "checker",
                "chicken",
                "checked",
                "wherein",
                "coherent",
                "cheered",
                "cherokee"
            ],
            "len": 7,
            "pos_from": 20
        }
    ]
}

Here pos_from - misspelled word first letter position, len - misspelled word len

Train

To train custom model you need:

1. Install cmake

2. Clone and build jamspell:

git clone https://github.com/bakwc/JamSpell.git
cd JamSpell
mkdir build
cd build
cmake ..
make

3. Prepare a utf-8 text file with sentences to train at (eg. sherlockholmes.txt) and another file with language alphabet (eg. alphabet_en.txt)

4. Train model:

./main/jamspell train ../test_data/alphabet_en.txt ../test_data/sherlockholmes.txt model_sherlock.bin

5. To evaluate spellchecker you can use evaluate/evaluate.py script:

python evaluate/evaluate.py -a alphabet_file.txt -jsp your_model.bin -mx 50000 your_test_data.txt

6. You can use evaluate/generate_dataset.py to generate you train/test data. It supports txt files, Leipzig Corpora Collection format and fb2 books.

Download models

Here is a few simple models. They trained on 300K news + 300k wikipedia sentences. We strongly recommend to train your own model, at least on a few million sentences to achieve better quality. See Train section above.

• en.tar.gz (35Mb)
• fr.tar.gz (31Mb)
• ru.tar.gz (38Mb)