greynirseq 0.4

Natural language processing for Icelandic

---

GreynirSeq

GreynirSeq is a natural language parsing toolkit for Icelandic focused on sequence modeling with neural networks. It is under active development and is in its early stages.

The modeling part (nicenlp) of GreynirSeq is built on top of the excellent fairseq from Facebook (which is built on top of pytorch).

GreynirSeq is licensed under the GNU AFFERO GPLv3 license unless otherwise stated at the top of a file.

What's new?

• An Icelandic RoBERTa model, IceBERT finetuned for NER and POS tagging.
• Icelandic - English translation.

What's on the horizon?

• More fine tuning tasks for Icelandic, constituency parsing and grammatical error detection

---

Be aware that usage of the CLI or otherwise downloading model files will result in downloading of gigabytes of data. Simply installing greynirseq will not download any models, they are automatically downloaded on-demand.

Installation

In a suitable virtual environment

# From PyPI
$ pip install greynirseq
# or from git main branch
$ pip install git+https://github.com/mideind/greynirseq@main

Features

TL;DR give me the CLI

The greynirseq CLI interface can be used to run pretrained models for various tasks. Run pip install greynirseq && greynirseq -h to see what options are available.

POS

Input is accepted from file containing a single tokenized sentence per line, or from stdin.

$ echo "Systurnar Guðrún og Monique átu einar um jólin á McDonalds ." | greynirseq pos --input -

nvfng nven-s c n---s sfg3fþ lvfnsf af nhfog af n----s pl

NER

Input is accepted from file containing a single tokenized sentence per line, or from stdin.

$ echo "Systurnar Guðrún og Monique átu einar um jólin á McDonalds ." | greynirseq ner --input -

O B-Person O B-Person O O O O O B-Organization O

Translation

Input is accepted from file containing a single untokenized sentence per line, or from stdin.

# For en->is translation
$ echo "This is an awesome test that shows how to use a pretrained translation model." | greynirseq translate --source-lang en --target-lang is

Þetta er æðislegt próf sem sýnir hvernig nota má forprófað þýðingarlíkan.

# For is->en translation
$ echo "Þetta er æðislegt próf sem sýnir hvernig nota má forprófað þýðingarlíkan." | greynirseq translate --source-lang is --target-lang en

This is an awesome test that shows how a pre-tested translation model can be used.

Neural Icelandic Language Processing - NIceNLP

IceBERT is an Icelandic BERT-based (RoBERTa) language model that is suitable for fine tuning on downstream tasks.

The following fine tuning tasks are available both through the greynirseq CLI and for loading programmatically.

1. POS tagging
2. NER tagging

There are also a some translation models available. They are Transformer models trained from scratch or finetuned based on mBART25.

1. Translation

Development

To install GreynirSeq in development mode we recommend using poetry as shown below

pip install poetry && poetry install

Linting

All code is checked with Super-Linter in a GitHub Action, we recommend running it locally before pushing

./run_linter.sh

Type annotation

Type annotation will soon be checked with mypy and should be included.