End-to-end Parser for Eastern Armenian

This repository contains necessary tools to parse raw Eastern Armenian text. It has a script, run.sh, which takes raw text as an input and produces a CoNLL-U file with lemmas, morphological features, part-of-speech tags and dependency trees.

• The parser segments the text into sentences and tokenizes them using ArmTreeBank's Tokenizer module.
• Lemmatization, POS tagging and dependency parsing is performed by a neural network called COMBO, which is developed and open-sourced by Piotr Rybak and Alina Wroblewska from Institute of Computer Science, Polish Academy of Sciences. If you use this network, please cite their paper.
• We have trained COMBO on the training set of the ArmTDP treebank from UD v2.3.
• The accuracy of the parser is far from perfect. It has been trained only on ~500 sentences. The table below shows the accuracy on the test set of the same treebank.

Metric	Accuracy
Lemmatization	88.05%
Part-of-speech tagging	85.07%
Morphological features	70.21%
Dependency parsing (Labelled attachment score)	55.25%

Visualization of the current parser

The model is hosted on DigitalOcean: https://parser.yerevann.com/

Instructions (for End to end parsing)

• Make sure you have all the requirements installed

pip install -r requirements.txt

• Clone the repo (to get the submodules don't forget to include the --recursive flag)

git clone --recursive https://github.com/Armtreebank/End-to-end-Parser.git

• Run the following command to get the .conllu file with predictions for every sentence of the input

python3 predict.py --model_path path_to_model.pkl --input_path sample.txt --output_path sample.conllu

Instructions (for COMBO training)

cd COMBO
python3 -m src.main --mode autotrain --train train_data_path.conllu --valid valid_data_path.conllu --model model.pkl --force_trees

Acknowledgements

This project is supported by ANSEF grant Lingu-5008 and ISTC Research Grant.