Model on translating legal text from Deustch to Spanish. It was first released in this repository. This model is first pretrained all the translation data over some unsupervised task. Then the model is trained on three parallel corpus from jrc-acquis, europarl and dcep.
legal_t5_small_trans_de_es_small_finetuned is initially pretrained on unsupervised task with the all of the data of the training set. The unsupervised task was "masked language modelling". legal_t5_small_trans_de_es_small_finetuned is based on the
t5-small model and was trained on a large corpus of parallel text. This is a smaller model, which scales the baseline model of t5 down by using
dmodel = 512,
dff = 2,048, 8-headed attention, and only 6 layers each in the encoder and decoder. This variant has about 60 million parameters.
The model could be used for translation of legal texts from Deustch to Spanish.
Here is how to use this model to translate legal text from Deustch to Spanish in PyTorch:
from transformers import AutoTokenizer, AutoModelWithLMHead, TranslationPipeline pipeline = TranslationPipeline( model=AutoModelWithLMHead.from_pretrained("SEBIS/legal_t5_small_trans_de_es_small_finetuned"), tokenizer=AutoTokenizer.from_pretrained(pretrained_model_name_or_path = "SEBIS/legal_t5_small_trans_de_es", do_lower_case=False, skip_special_tokens=True), device=0 ) de_text = "Bei einer Kombination von Artikel 124 Absatz 14 mit Artikel 136 AEUV scheint die in den Artikeln 121 und 126 AEUV" pipeline([de_text], max_length=512)
The legal_t5_small_trans_de_es_small_finetuned (the supervised task which involved only the corresponding langauge pair and as well as unsupervised task where all of the data of all language pairs were available) model was trained on JRC-ACQUIS, EUROPARL, and DCEP dataset consisting of 8 Million parallel texts.
The model was trained on a single TPU Pod V3-8 for 250K steps in total, using sequence length 512 (batch size 4096). It has a total of approximately 220M parameters and was trained using the encoder-decoder architecture. The optimizer used is AdaFactor with inverse square root learning rate schedule for pre-training.
An unigram model trained with 88M lines of text from the parallel corpus (of all possible language pairs) to get the vocabulary (with byte pair encoding), which is used with this model.
The pre-training data was the combined data from all the 42 language pairs. The task for the model was to predict the portions of a sentence which were masked randomly.
When the model is used for translation test dataset, achieves the following results:
Test results :
- Downloads last month