Skip to content

A Danish-speaking language model with entity-aware self-attention

License

Notifications You must be signed in to change notification settings

peleiden/daluke

Repository files navigation

DaLUKE: The Entity-aware, Danish Language Model

pytest

Implementation of the knowledge-enhanced transformer LUKE pretrained on the Danish Wikipedia and evaluated on named entity recognition (NER).

Installation

pip install daluke

For including optional requirements that are necessary for training and general analysis:

pip install daluke[full]

Python 3.8 or newer is required.

Explanation

For an explanation of the model, see our bachelor's thesis or the original LUKE paper.

Usage

Inference on simple NER or masked language modeling (MLM) examples

Python

For performing NER predictions

from daluke import AutoNERDaLUKE, predict_ner

daluke = AutoNERDaLUKE()

document = "Det Kgl. Bibliotek forvalter Danmarks største tekstsamling, der strækker sig fra middelalderen til det nyeste litteratur."
iob_list = predict_ner(document, daluke)

For testing MLM predictions

from daluke import AutoMLMDaLUKE, predict_mlm

daluke = AutoMLMDaLUKE()
# Empty list => No entity annotations in the string
document = "Professor i astrofysik, [MASK] [MASK], udtaler til avisen, at den nye måling sandsynligvis ikke er en fejl."
best_prediction, table = predict_mlm(document, list(), daluke)

CLI

daluke ner --text "Thomas Delaney fører Danmark til sejr ved EM i fodbold."
daluke masked --text "Slutresultatet af kampen mellem Danmark og Rusland bliver [MASK]-[MASK]."

For Windows, or systems where #!/usr/bin/env python3 is not linked to the correct Python interpreter, the command python -m daluke.api.cli can be used instead of daluke.

Training DaLUKE yourself

This part shows how to recreate the entire DaLUKE training pipeline from dataset preparation to fine-tuning. This guide is designed to be run in a bash shell. If you use Windows, you will probably have to make some modifications to the shell scripts used.

# Download forked luke submodule
git submodule update --init --recursive
# Install requirements
pip install -r requirements.txt
pip install -r optional-requirements.txt
pip install -r luke/requirements.txt

# Build dataset
# The script performs all the steps of building the dataset, including downloading the Danish Wikipedia
# You only need to modify DATA_PATH to where you want the data to be saved
# Be aware that this takes several hours
dev/build_data.sh

# Start pretraining using default hyperparameters
python daluke/pretrain/run.py <INSERT DATA_PATH HERE> -c configs/pretrain-main.ini --name $NAME --save-every 5 --parameter-updates 10000 --name daluke --fp16
# Optional: Make plots of pretraining
python daluke/plot/plot_pretraining.py <DATA_PATH>/daluke

# Fine-tune on DaNE
python daluke/collect_modelfile.py <DATA_PATH>/daluke <DATA_PATH>/ner/daluke.tar.gz
python daluke/ner/run.py <DATA_PATH>/ner/daluke -c configs/main-finetune.ini --model <DATA_PATH>/ner/daluke.tar.gz --name finetune --eval
# Evaluate on DaNE test set
python daluke/ner/run_eval.py <DATA_PATH>/ner/daluke/finetune --model <DATA_PATH>/ner/daluke/finetune/daluke_ner_best.tar.gz
# Optional: Fine-tuning plots
python daluke/plot/plot_finetune_ner.py <DATA_PATH>/ner/daluke/finetune/train-results