README.md

Tokenizer

Tokenizer is a fast, generic, and customizable text tokenization library for C++ and Python with minimal dependencies.

Overview

By default, the Tokenizer applies a simple tokenization based on Unicode types. It can be customized in several ways:

• Reversible tokenization
  Marking joints or spaces by annotating tokens or injecting modifier characters.
• Subword tokenization
  Support for training and using BPE and SentencePiece models.
• Advanced text segmentation
  Split digits, segment on case or alphabet change, segment each character of selected alphabets, etc.
• Case management
  Lowercase text and return case information as a separate feature or inject case modifier tokens.
• Protected sequences
  Sequences can be protected against tokenization with the special characters ｟ and ｠.

See the available options for an overview of supported features.

Using

The Tokenizer can be used in Python, C++, or command line. Each mode exposes the same set of options.

Python API

pip install pyonmttok

>>> import pyonmttok
>>> tokenizer = pyonmttok.Tokenizer("conservative", joiner_annotate=True)
>>> tokens = tokenizer("Hello World!")
>>> tokens
['Hello', 'World', '￭!']
>>> tokenizer.detokenize(tokens)
'Hello World!'

See the Python API description for more details.

C++ API

#include <onmt/Tokenizer.h>

using namespace onmt;

int main() {
  Tokenizer tokenizer(Tokenizer::Mode::Conservative, Tokenizer::Flags::JoinerAnnotate);
  std::vector<std::string> tokens;
  tokenizer.tokenize("Hello World!", tokens);
}

See the Tokenizer class for more details.

Command line clients

$ echo "Hello World!" | cli/tokenize --mode conservative --joiner_annotate
Hello World ￭!
$ echo "Hello World!" | cli/tokenize --mode conservative --joiner_annotate | cli/detokenize
Hello World!

See the -h flag to list the available options.

Development

Dependencies

• ICU

Compiling

CMake and a compiler that supports the C++11 standard are required to compile the project.

git submodule update --init
mkdir build
cd build
cmake ..
make

It will produce the dynamic library libOpenNMTTokenizer and tokenization clients in cli/.

• To compile only the library, use the -DLIB_ONLY=ON flag.

Testing

The tests are using Google Test which is included as a Git submodule. Run the tests with:

mkdir build
cd build
cmake -DBUILD_TESTS=ON ..
make
test/onmt_tokenizer_test ../test/data