vieenrose / tonalizer

Tonalizer - CRF-based Tone Reconstitution Tool

Tonalizer - CRF-based Tone Reconstitution Tool

Introduction

In many languages such as Bambara or Arabic, tone markers (diacritics) may be written but are actually often omitted. Lack of tone markers contribute to ambiguities and subsequent difficulties when reading texts. To circumvent this problem, tonalization may be used, as a word sense disambiguation task, relying on context to add diacritics that partially disambiguate words as well as senses.

Objective

Tonalizer we propose here models tonalizaion of a tonalized text in UTF-8, and generate a model file. With the model file leant, tonalizer can performs reconstitue tone markers of a new raw text without diacritics.

System Requirements

• python >= 2.4
• python-Levenshtein
• python-crfsuite
• nltk

Installation

• Unix-like Operating System (Linux, MacOS)

sudo pip install python-Levenshtein python-crfsuite nltk

General Functions

Undiacritizing

• Undiacritize a UTF-8 stream received by stdin and show result to stdout

PROGRAM | python tonalizer.py -u

• Undiacritize a UTF-8 raw text (.txt) and show result to stdout

python tonalizer.py -i infile -u

• Undiacritize a UTF-8 raw text (.txt) and save result to file

python tonalizer.py -i infile -u -o outfile

Training

• Generate a model file from a diacritized text, then store it in a file and show accuracy to screen

python tonalizer.py -i infile -l model -s report.csv

• Generate a model file from a UTF-8 stream received by stdin, then store it in a file and show accuracy to screen

PROGRAM | python tonalizer.py -l model -s report.csv

Diacritizing

• Use a model to recover diacrtize of a UTF-8 stream, then outputs to screen

PROGRAM | python tonalizer.py -d model 

• Use a model to recover diacrtize of a new text, then outputs to a file

python tonalizer.py -i infile -d model -o outfile

• Use a model to recover diacrtize of a UTF-8 stream, then outputs to a file

PROGRAM | python tonalizer.py -d model -o outfile 

Using pre-trained models

• French

PROGRAM | python tonalizer.py -d models/fr.mdl 

• Arabic

PROGRAM | python tonalizer.py -d model 

Experiment Results (Preliminary)

Ressource	File Size	File Name	Language	Accuracy	Train/Test Ratio
Project Gutenberg.	696 KB	pg17489.txt	French	0.975	90 : 10
T. Zerrouki et al.		معالم القربة في طلب الحسبة.txt	Arabic

Reference

• Project Gutenberg. (n.d.). Retrieved July 26, 2017, from www.gutenberg.org. link
• T. Zerrouki, A. Balla, Tashkeela: Novel corpus of Arabic vocalized texts, data for auto-diacritization systems, Data in Brief (2017) link