Skip to content

Latest commit

 

History

History
109 lines (97 loc) · 3.78 KB

README.md

File metadata and controls

109 lines (97 loc) · 3.78 KB
Raw

Bezhta morphology

The current work is dedicated to building a morphological analyzer for Bezhta language (< Tsezic < Avar-Andic-Tsezic < Nakh-Dagestan; Glottolog: bezh1248). This repository contains a prototype for a Bezhta morphological analyzer. It is a part of a larger project by the students of the School of Linguistics and the Linguistic Convergence Laboratory at the NRU HSE that aims to provide digital tools for endangered languages.

The project is distributed under the GNU General Public License v3.0.

Sources

Grammar & dictionary

The parser follows (Comri et al., 2015) and (Madieva, 1965) descriptions of Bezhta Proper with the lexicon gathered from (Khalilov, 2015) dictionary. The digitized version of the dictionary is available at bezhta_dict.

Texts

For evaluation, I use Bezhta translation of The Gospel of Luke and The Book of Proverbs, a text from Madieva's grammar (1964) and two annotated texts. The texts are available in the corpora directory

Usage

The project requires lexd and hfst. You can get them by the following command:

curl -sS https://apertium.projectjj.com/apt/install-nightly.sh | sudo bash
apt install lexd
apt install hfst

Making the analyzer

make

Analyze a word:

echo 'соралила' | hfst-lookup bezhta.analyzer.hfst

Making the transliterator

Transliterator allows to transliterate Bezhta words from Cyrillic to Latin script.

make cy2lat.transliterator.disam.hfst

Transliterate a word:

echo 'соралила' | hfst-lookup cy2lat.transliterator.disam.hfst

Build transliterated analyzer:

make bezhta.tr.analyzer.hfst

Look up a word in Latin script:

echo 'soralila' | hfst-lookup bezhta.tr.analyzer.hfst

Making the segmenter

The segmenter identifies the morpheme boundaries in the input word.

make bezhta.segm.hfst

Segmenting a word:

echo 'нисойо' | hfst-lookup bezhta.segm.hfst

Result:

 нисойо        нисо>йо

Evaluating coverage

Analyzer:

make bezhta.analyzer.hfstol
mv bezhta.analyzer.hfstol coverage
cd coverage
make check-coverage

Additionally, make-check-unrecog can be used to get a list of unrecognized tokens. Note that all text files should start with text-

Current performance: ~75% naive coverage

Transliterator:

make bezhta.tr.analyzer.hfst
mv bezhta.tr.analyzer.hfst transliterator
make check-coverage

Note: some symbols may be recognized incorrectly, I recommend using transliterator_coverage.ipynb instead.

Evaluating accuracy

make bezhta.analyzer.hfstol
mv bezhta.analyzer.hfstol accuracy
cd accuracy

To analyze texts with the parser, use

hfst-proc bezhta.analyzer.hfstol text-annotated-1.txt > FILENAME-1.txt
hfst-proc bezhta.analyzer.hfstol text-annotated-1.txt > FILENAME-2.txt

Then compute accuracy:

python3 accuracy.py FILENAME-1.txt text-1-gold.txt
python3 accuracy.py FILENAME-2.txt text-2-gold.txt

Making the guesser

cd guesser
make bezhta.guesser.hfst

Guessing a token:

echo 'войъис' bezhta.guesser.hfst

For evaluation, see guesser_evaluation.ipynb