Result filters

Metadata provider

  • DSpace

Language

Resource type

Availability

Active filters:

  • Metadata provider: DSpace
  • Keywords: model
8 record(s) found

Search results

  • Semi-supervised Icelandic-Polish Translation System (22.09)

    This Icelandic-Polish translation model (bi-directional) was trained using fairseq (https://github.com/facebookresearch/fairseq) by means of semi-supervised translation by starting with the mBART50 model. The model was then trained using a multi-task curriculum to first learn to denoise sentences. Then the model was trained to translate using aligned parallel texts. Finally the model was provided with monolingual texts in both Icelandic and Polish with which it iteratively creates back-translations. For the PL-IS direction the model achieves a BLEU score of 27.60 on held out true parallel training data and 15.30 on the out-of-domain Flores devset. For the IS-PL direction the model achieves a score of 27.70 on the true data and 13.30 on the Flores devset. -- Þetta íslensk-pólska þýðingarlíkan (tvíátta) var þjálfað með fairseq (https://github.com/facebookresearch/fairseq) með hálf-sjálfvirkum aðferðum frá mBART50 líkaninu. Líkanið var þjálfað á þremur verkefnum, afruglun, samhliða þýðingum og bakþýðingum sem voru myndaðar á þjálfunartíma. Fyrir PL-IS áttina fæst BLEU skor 27.60 á raun gögnum sem voru tekin til hliðar og 15.30 á Flores þróunargögnunum. Fyrir IS-PL áttina fæst skor 27.70 á raun gögnunum og 13.30 á Flores þróunargögnunum.

  • GreynirTranslate - mBART25 NMT (with layer drop) models for Translations between Icelandic and English (1.0)

    These are the models in http://hdl.handle.net/20.500.12537/125 trained with 40% layer drop. They are suitable for inference using every other layer for optimized inference speed with lower translation performance. We refer to the prior submission for usage and the documentation on layerdrop at https://github.com/pytorch/fairseq/blob/fcca32258c8e8bcc9f9890bf4714fa2f96b6b3e1/examples/layerdrop/README.md. Þessi líkön eru þjálfuð með 40% laga missi (e. layer drop) á líkönunum í http://hdl.handle.net/20.500.12537/125. Þau henta vel til þýðinga þar sem er búið að henda öðru hverju lagi í netinu og þannig er hægt að hraða á þýðingum á kostnað gæða. Leiðbeiningar um notkun netanna er að finna með upphaflegu líkönunum og í notkunarleiðbeiningum Fairseq í https://github.com/pytorch/fairseq/blob/fcca32258c8e8bcc9f9890bf4714fa2f96b6b3e1/examples/layerdrop/README.md.

  • GreynirTranslate - mBART25 NMT models for Translations between Icelandic and English (1.0)

    Provided are a general domain IS-EN and EN-IS translation models developed by Miðeind ehf. They are based on a multilingual BART model (https://arxiv.org/pdf/2001.08210.pdf) and finetuned for translation on parallel and backtranslated data. The model is trained using the Fairseq sequence modeling toolkit by PyTorch. Provided here are a model files, sentencepiece subword-tokenizing model and dictionary files for running the model locally. You can run the scripts infer-enis.sh and infer-isen.sh to test the model by translating sentences command-line. For translating documents and evaluating results you will need to binarize the data using fairseq-preprocess and use fairseq-generate for translating. Please refer to the Fairseq documentation for further information on running a pre-trained model: https://fairseq.readthedocs.io/en/latest/ - Pakkinn inniheldur almenn þýðingarlíkön fyrir áttirnar IS-EN og EN-IS þróuð af Miðeind ehf. Þau eru byggð á margmála BART líkani (https://arxiv.org/pdf/2001.08210.pdf) og fínþjálfuð fyrir þýðingar. Líkönin eru þjálfað með Fairseq og PyTorch. Líkönin sjálf og ásamt sentencepiece tilreiðingarlíkani eru gerð aðgengileg. Skripturnar infer-enis.sh og infer-isen.sh gefa dæmi um hvernig er hægt að keyra líkönin á skipanalínu. Til að þýða stór skjöl og meta niðurstöður þarf að nota fairseq-preprocess skipunina ásamt fairseq-generate. Frekari upplýsingar er að finna í Fairseq leiðbeiningunum: https://fairseq.readthedocs.io/en/latest/

  • Slovene Conformer CTC BPE E2E Automated Speech Recognition model PROTOVERB-ASR-E2E 1.0

    This Conformer CTC BPE E2E Automated Speech Recognition model was trained following the NVIDIA NeMo Conformer-CTC fine-tuning recipe (for details see the official NVIDIA NeMo NMT documentation, https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/intro.html, and NVIDIA NeMo GitHub repository https://github.com/NVIDIA/NeMo). It provides functionality for transcribing Slovene speech to text. The starting point was the Conformer CTC BPE E2E Automated Speech Recognition model RSDO-DS2-ASR-E2E 2.0, which was fine-tuned on the Protoverb closed dataset. The model was fine-tuned for 20 epochs, which improved the performance on the Protoverb test dataset for 9.8% relative WER, and for 3.3% relative WER on the Slobench dataset.

  • Neural Machine Translation model for Slovene-English language pair RSDO-DS4-NMT 1.2.6

    This Neural Machine Translation model for Slovene-English language pair was trained following the NVIDIA NeMo NMT AAYN recipe (for details see the official NVIDIA NeMo NMT documentation, https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/machine_translation/machine_translation.html, and NVIDIA NeMo GitHub repository https://github.com/NVIDIA/NeMo). It provides functionality for translating text written in Slovene language to English and vice versa. The training corpus was built from publicly available datasets, including Parallel corpus EN-SL RSDO4 1.0 (https://www.clarin.si/repository/xmlui/handle/11356/1457), as well as a small portion of proprietary data. In total the training corpus consisted of 32.638.758 translation pairs and the validation corpus consisted of 8.163 translation pairs. The model was trained on 64GPUs and on the validation corpus reached a SacreBleu score of 48.3191 (at epoch 37) for translation from Slovene to English and a SacreBleu score of 53.8191 (at epoch 47) for translation from English to Slovene.

  • Slovene Conformer CTC BPE E2E Automated Speech Recognition model RSDO-DS2-ASR-E2E 2.0

    This Conformer CTC BPE E2E Automated Speech Recognition model was trained following the NVIDIA NeMo Conformer-CTC recipe (for details see the official NVIDIA NeMo NMT documentation, https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/intro.html, and NVIDIA NeMo GitHub repository https://github.com/NVIDIA/NeMo). It provides functionality for transcribing Slovene speech to text. The training, development and test datasets were based on the Artur dataset and consisted of 630.38, 16.48 and 15.12 hours of transcribed speech in standardised form, respectively. The model was trained for 200 epochs and reached WER 0.0429 on the development and WER 0.0558 on the test dataset.

  • LitLat BERT

    Trilingual BERT-like (Bidirectional Encoder Representations from Transformers) model, trained on Lithuanian, Latvian, and English data. State of the art tool representing words/tokens as contextually dependent word embeddings, used for various NLP classification tasks by fine-tuning the model end-to-end. LitLat BERT are neural network weights and configuration files in pytorch format (i.e. to be used with pytorch library). The corpora used for training the model have 4.07 billion tokens in total, of which 2.32 billion are English, 1.21 billion are Lithuanian and 0.53 billion are Latvian. LitLat BERT is based on XLM-RoBERTa model and comes in two versions, one for usage with transformers library (https://github.com/huggingface/transformers), and one for usage with fairseq library (https://github.com/pytorch/fairseq). More information is in the readme.txt.

  • Slovene Punctuation and Capitalisation model RSDO-DS2-P&C 3.6

    This Punctuation and Capitalisation model was trained following the NVIDIA NeMo Punctuation and Capitalisation recipe (for details see the official NVIDIA NeMo P&C documentation, https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/punctuation_and_capitalization.html, and NVIDIA NeMo GitHub repository https://github.com/NVIDIA/NeMo). It provides functionality for restoring punctuation (,.!?) and capital letters in lowercased non-punctuated Slovene text. The training corpus was built from publicly available datasets, as well as a small portion of proprietary data. In total the training corpus consisted of 38.829.529 sentences and the validation corpus consisted of 2.092.497 sentences.