Ælæctra - A Step Towards More Efficient Danish Natural Language Processing

Ælæctra is a Danish Transformer-based language model created to enhance the variety of Danish NLP resources with a more efficient model compared to previous state-of-the-art (SOTA) models. Initially a cased and an uncased model are released. It was created as part of a Cognitive Science bachelor's thesis.

Ælæctra was pretrained with the ELECTRA-Small (Clark et al., 2020) pretraining approach by using the Danish Gigaword Corpus (Strømberg-Derczynski et al., 2020) and evaluated on Named Entity Recognition (NER) tasks. Since NER only presents a limited picture of Ælæctra's capabilities I am very interested in further evaluations. Therefore, if you employ it for any task, feel free to hit me up your findings!

Ælæctra was, as mentioned, created to enhance the Danish NLP capabilties and please do note how this GitHub still does not support the Danish characters " Æ, Ø and Å " as the title of this repository becomes " -l-ctra ". How ironic.?

Here is an example on how to load both the cased and the uncased Ælæctra model in PyTorch using the ?Transformers library:

from transformers import AutoTokenizer, AutoModelForPreTraining

tokenizer = AutoTokenizer.from_pretrained("Maltehb/-l-ctra-danish-electra-small-cased")
model = AutoModelForPreTraining.from_pretrained("Maltehb/-l-ctra-danish-electra-small-cased")

from transformers import AutoTokenizer, AutoModelForPreTraining

tokenizer = AutoTokenizer.from_pretrained("Maltehb/-l-ctra-danish-electra-small-uncased")
model = AutoModelForPreTraining.from_pretrained("Maltehb/-l-ctra-danish-electra-small-uncased")

Evaluation of current Danish Language Models

Ælæctra, Danish BERT (DaBERT) and multilingual BERT (mBERT) were evaluated:

On DaNE (Hvingelby et al., 2020), Ælæctra scores slightly worse than both cased and uncased Multilingual BERT (Devlin et al., 2019) and Danish BERT (Danish BERT, 2019/2020), however, Ælæctra is less than one third the size, and uses significantly fewer computational resources to pretrain and instantiate. For a full description of the evaluation and specification of the model read the thesis: 'Ælæctra - A Step Towards More Efficient Danish Natural Language Processing'.

Pretraining

To pretrain Ælæctra it is recommended to build a Docker Container from the Dockerfile . Next, simply follow the pretraining notebooks

The pretraining was done by utilizing a single NVIDIA Tesla V100 GPU with 16 GiB, endowed by the Danish data company KMD . The pretraining took approximately 4 days and 9.5 hours for both the cased and uncased model

Fine-tuning

To fine-tune any Ælæctra model follow the fine-tuning notebooks

References

Clark, K., Luong, M.-T., Le, Q. V., & Manning, C. D. (2020). ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. ArXiv:2003.10555 [Cs]. http://arxiv.org/abs/2003.10555

Danish BERT. (2020). BotXO. https://github.com/botxo/nordic_bert (Original work published 2019)

Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. ArXiv:1810.04805 [Cs]. http://arxiv.org/abs/1810.04805

Hvingelby, R., Pauli, A. B., Barrett, M., Rosted, C., Lidegaard, L. M., & Søgaard, A. (2020). DaNE: A Named Entity Resource for Danish. Proceedings of the 12th Language Resources and Evaluation Conference, 4597–4604. https://www.aclweb.org/anthology/2020.lrec-1.565

Strømberg-Derczynski, L., Baglini, R., Christiansen, M. H., Ciosici, M. R., Dalsgaard, J. A., Fusaroli, R., Henrichsen, P. J., Hvingelby, R., Kirkedal, A., Kjeldsen, A. S., Ladefoged, C., Nielsen, F. Å., Petersen, M. L., Rystrøm, J. H., & Varab, D. (2020). The Danish Gigaword Project. ArXiv:2005.03521 [Cs]. http://arxiv.org/abs/2005.03521

As the majority of this repository is build upon the works by the team at Google who created ELECTRA, a HUGE thanks to them is in order.

A Giga thanks also goes out to the incredible people who collected The Danish Gigaword Corpus (Strømberg-Derczynski et al., 2020).

Furthermore, I would like to thank my supervisor Riccardo Fusaroli for the support with the thesis, and a special thanks goes out to Kenneth Enevoldsen for his continuous feedback.

Lastly, i would like to thank KMD, my colleagues from KMD, and my peers and co-students from Cognitive Science for encouriging me to keep on working hard and holding my head up high!

For help or further information feel free to connect with the author Malte Højmark-Bertelsen on hjb@kmd.dk or any of the following platforms: