pl-itn 0.1.0rc1

Polish FST Inverse Text Normalization

pl_itn

Inverse Text Normalization is an NLP task of changing the spoken form of a phrase to written form, for example:

one two three -> 1 2 3

pl_itn is an opensource Polish ITN Python library and REST API for practical applications.

This project is an implementation of NeMo Inverse Text Normalization for Polish.

Table of contents

Prerequisites
Setup
Docker
Usage
gRPC service
Building custom grammars
Documentation
Contributing
License
References

Prerequisites

For pynini

• A standards-compliant C++17 compiler (GCC >= 7 or Clang >= 700)
• The compatible recent version of OpenFst built with the grm extensions (see deps/install_openfst.md)

Setup

Make sure to first install prerequisites, especially OpenFST.

Install from PyPI

pip install pl_itn

Build from source

pip install .

Editable install for development

pip install -e .[dev]

Docker

To build docker image containing pl_itn library use pl_itn_lib.dockerfile file.
To build docker image with gRPC service use grpc_service.dockerfile file.

docker build -t <IMAGE:TAG> -f <DOCKERFILE> .

Usage

Console app

usage: pl_itn [-h] (-t TEXT | -i) [--tagger TAGGER] [--verbalizer VERBALIZER] [--config CONFIG]
              [--log-level {debug,info}]

Inverse Text Normalization based on Finite State Transducers

options:
  -h, --help            show this help message and exit
  -t TEXT, --text TEXT  Input text
  -i, --interactive     If used, demo will process phrases from stdin interactively.
  --tagger TAGGER
  --verbalizer VERBALIZER
  --config CONFIG       Optionally provide yaml config with tagger and verbalizer paths.
  --log-level {debug,info}
                        return a step back value.

pl_itn -t "jest za pięć druga"
jest 01:55

pl_itn -t "drugi listopada dwa tysiące osiemnastego roku"
2 listopada 2018 roku

Python

>>> from pl_itn import Normalizer
>>> normalizer = Normalizer()
>>> normalizer.normalize("za pięć dwunasta")
'11:55'

Docker

Existing docker image containing pl_itn library is required. For build command refer to Docker section.

docker run --rm -it <IMAGE:TAG> --help

gRPC Service

gRPC service methods are described in grpc_service/pl_itn_api/api.proto file. Docker container is suggested approach for running the service. For build command refer to Docker section. Service within container serves on port 10010.

Example of building the image and starting the service.

docker build -t pl_itn_service:test -f grpc_service.dockerfile .
docker run -p 10010:10010 pl_itn_service:test

Building custom grammars

Custom grammars can be built using build_grammar/build_grammar.py script.

There are three demo grammars available:

• not declined cardinal numbers (e.g. "jeden", "dwa", "trzy")
• declined cardinal numbers (e.g. "jednego", "dwóch", "trzech")
• ordinal numbers (e.g. "pierwszy", "druga", "trzecie")

Normalization types can be included and excluded from the grammar through the config file, which is set by default to build_grammar/grammar_config.yaml.

# cardinals_basic_forms: True
# cardinals_declined: True
# ordinals: True

$ python3 build_grammar/build_grammar.py --grammars-dir all

$ pl_itn \
  --tagger all/tagger.fst \
  --verbalizer all/verbalizer.fst \
  -t "Jeden trzech piąta"

1 3 5

# cardinals_basic_forms: True
# cardinals_declined: False
# ordinals: True

$ python3 build_grammar/build_grammar.py --grammars-dir cardinals_basic_ordinals

$ pl_itn \
  --tagger cardinals_basic_ordinals/tagger.fst \
  --verbalizer cardinals_basic_ordinals/verbalizer.fst \
  -t "Jeden trzech piąta"

1 trzech 5

# cardinals_basic_forms: True
# cardinals_declined: False
# ordinals: False

$ python3 build_grammar/build_grammar.py --grammars-dir only_basic_cardinals

$ pl_itn \
  --tagger only_basic_cardinals/tagger.fst \
  --verbalizer only_basic_cardinals/verbalizer.fst \
  -t "Jeden trzech piąta"

1 trzech piąta

See Documentation for more details.

Documentation

Contributing

License

Rerences

• K. Gorman. 2016. Pynini: A Python library for weighted finite-state grammar compilation. In Proc. ACL Workshop on Statistical NLP and Weighted Automata, 75-80.
• Y. Zhang, E. Bakhturina, K. Gorman, and B. Ginsburg. 2021. NeMo Inverse Text Normalization: From Development To Production.