OpenNMT Tokenizer Versions Save

Fixes and improvements

• Consider escaped characters as single characters in BPE
• Ignore undefined scripts when resolving inherited or common scripts

New features

• Add tokenization option allow_isolated_marks to allow combining marks to appear isolated in the tokenization output in specific conditions

Fixes and improvements

• Fix infinite loop when the text contains an invalid Unicode character
• Fix segmentation fault when the BPELearner does not not find any pairs of characters in the tokenized data
• [Python] Update ICU to 72.1

New features

• [Python] Add argument vocabulary in the Tokenizer constructor to set the vocabulary with a list of tokens instead of using a file
• [Python] Add function pyonmttok.is_valid_language to check if a language code is valid and can be passed to the Tokenizer constructor

New features

• [Python] Add pickling support to pyonmttok.Vocab

Fixes and improvements

• Update pybind11 to 2.10.1
• Update cibuildwheel to 2.11.2

Changes

• [Python] Wheels are now built under manylinux2014 and requires pip >= 19.3 for installation

New features

• [Python] Build wheels for Python 3.11

Fixes and improvements

• Improve error handling when reading token frequencies in the vocabulary file
• [Python] Fix possible crash when pyonmttok is imported before torch
• [Python] Update ICU to 71.1
• [C++] Fix static compilation with -DBUILD_SHARED_LIBS=OFF
• [C++] Fix CMake warning when compiling the tests

New features

• [Python] Build ARM64 wheels for macOS

Fixes and improvements

• [CLI] Fix error when the option --segment_alphabet is not set
• Fix SentencePiece build warning when compiling with Clang

New features

• Add property pyonmttok.Vocab.counters to retrieve the number of occurrences of each token

Fixes and improvements

• Update pybind11 to 2.10.0
• Update cxxopts to 3.0.0

New features

• Add utilities to build and use vocabularies:
  • pyonmttok.Vocab
  • pyonmttok.build_vocab_from_tokens
  • pyonmttok.build_vocab_from_lines
• Define the method Tokenizer.__call__ to simplify the tokenizer usage when additional features are unused:

tokens = tokenizer(text)

Fixes and improvements

• Update pybind11 to 2.9.1

Fixes and improvements

• Fix deprecated languages codes in ICU that are incorrectly considered as invalid (e.g. "tl" for Tagalog)

New features

• [Python] Build wheels for AArch64 Linux

Fixes and improvements

• [Python] Update ICU to 70.1