Model on translating legal text from English to Deustch. It was first released in
this repository. The model is parallely trained on the three parallel corpus with 42 language pair
from jrc-acquis, europarl and dcep along with the unsupervised task where the model followed the task of prediction in a masked language model.
No pretraining is involved in case of legal_t5_small_multitask_en_de model, rather the unsupervised task is added with all the translation task
to realize the multitask learning scenario.
The model could be used for translation of legal texts from English to Deustch.
Here is how to use this model to translate legal text from English to Deustch in PyTorch:
1from transformers import AutoTokenizer, AutoModelWithLMHead, TranslationPipeline
2
3pipeline = TranslationPipeline(
4model=AutoModelWithLMHead.from_pretrained("SEBIS/legal_t5_small_multitask_en_de"),
5tokenizer=AutoTokenizer.from_pretrained(pretrained_model_name_or_path = "SEBIS/legal_t5_small_multitask_en_de", do_lower_case=False,
6 skip_special_tokens=True),
7 device=0
8)
9
10en_text = "Reiterates its call on the Commission to submit a proposal to the Parliament and Council as soon as possible in order to ensure that bunker oil for engine fuel in new ships is stored in safer, double-hull tanks since freight or container ships often contain heavy fuel as engine fuel in their bunkers the quantity of which may considerably exceed the cargoes of smaller oil tankers; considers that, before submitting such a proposal, the Commission should ascertain whether or not the existing IMO rules laid down in Resolution MEPC.141(54) are sufficient to guarantee the safe transport of bunker oil used as fuel;"
11
12pipeline([en_text], max_length=512)
The legal_t5_small_multitask_en_de model (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 9 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.
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.
When the model is used for translation test dataset, achieves the following results:
