Marco Lodola, Monument to Umberto Eco, Alessandria 2019

UmBERTo is a Roberta-based Language Model trained on large Italian Corpora. This implementation is based on Facebook Research AI code (https://github.com/pytorch/fairseq)

UmBERTo inherits from RoBERTa base model architecture which improves the initial BERT by identifying key hyperparameters for better results. Umberto extends Roberta and uses two innovative approaches: SentencePiece and Whole Word Masking. SentencePiece Model (SPM) is a language-independent subword tokenizer and detokenizer designed for Neural-based text processing and creates sub-word units specifically to the size of the chosen vocabulary and the language of the corpus. Whole Word Masking (WWM) applies mask to an entire word, if at least one of all tokens created by SentencePiece Tokenizer was originally chosen as mask. So only entire word are masked, not subwords.

Two models are released:

• umberto-wikipedia-uncased-v1, an uncased model trained on a relative small corpus (~7GB) extracted from Wikipedia-ITA.
• umberto-commoncrawl-cased-v1, a cased model trained on Commoncrawl ITA exploiting OSCAR (Open Super-large Crawled ALMAnaCH coRpus) Italian large corpus ( ~69GB)

Both models have 12-layer, 768-hidden, 12-heads, 110M parameters (BASE).

We trained both the models on 8 Nvidia V100 GPUs (p2.8xlarge P2 EC2 instance) during 4 days on AWS Sagemaker.

torch >= 1.3.1
sentencepiece
transformers
fairseq

pip install transformers

To install transformers from original repo (TESTED):

git clone https://github.com/huggingface/transformers.git
cd transformers
pip install .

To use a version of fairseq with UmBERTo support, build from source doing these steps:

git clone https://github.com/musixmatchresearch/fairseq
cd fairseq
pip install .

From official HuggingFace code.

from transformers import pipeline

fill_mask = pipeline(
	"fill-mask",
	model="Musixmatch/umberto-commoncrawl-cased-v1",
	tokenizer="Musixmatch/umberto-commoncrawl-cased-v1"
)

result = fill_mask("Umberto Eco è <mask> un grande scrittore")

#[{'sequence': '<s> Umberto Eco è considerato un grande scrittore</s>', 'score': 0.1859988570213318, 'token': 5032}, 
#{'sequence': '<s> Umberto Eco è stato un grande scrittore</s>', 'score': 0.1781671643257141, 'token': 471}, 
#{'sequence': '<s> Umberto Eco è sicuramente un grande scrittore</s>', 'score': 0.16565577685832977, 'token': 2654}, 
#{'sequence': '<s> UmbertoEco è indubbiamente un grande scrittore</s>', 'score': 0.09328985959291458, 'token': 17908}, 
#{'sequence': '<s> Umberto Eco è certamente un grande scrittore</s>', 'score': 0.05470150709152222, 'token': 5269}]

import torch
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("Musixmatch/umberto-commoncrawl-cased-v1")
umberto = AutoModel.from_pretrained("Musixmatch/umberto-commoncrawl-cased-v1")

encoded_input = tokenizer.encode("Umberto Eco è stato un grande scrittore")
input_ids = torch.tensor(encoded_input).unsqueeze(0)  # Batch size 1
outputs = umberto(input_ids)
last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output

import torch

umberto = torch.hub.load('musixmatchresearch/umberto', 'umberto_commoncrawl_cased')

assert isinstance(umberto.model, torch.nn.Module)
umberto.eval()  # disable dropout (or leave in train mode to finetune)

# Masked LM Inference
masked_line = 'Umberto Eco è <mask> un grande scrittore'
result = umberto.fill_mask(masked_line, topk=20)
# Output:
#('Umberto Eco è considerato un grande scrittore', 0.19939924776554108, ' considerato'), 
#('Umberto Eco è sicuramente un grande scrittore', 0.1669664829969406, ' sicuramente'), 
#('Umberto Eco è stato un grande scrittore', 0.16225320100784302, ' stato'), 
#('Umberto Eco è indubbiamente un grande scrittore', 0.09528309106826782, ' indubbiamente')
...

We obtained state-of-the-art results for POS tagging, confirming that cased models trained with WWM perform better than uncased ones. Our model Umberto-Wikipedia-Uncased trained with WWM on a smaller dataset and uncased, produces important results comparable to the cased results.

These results refers to umberto-wikipedia-uncased model.

These results refers to umberto-commoncrawl-cased model.

• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding Paper, Github
• CamemBERT: a Tasty French Language Model Paper, Page
• GilBERTo: An Italian pretrained language model based on RoBERTa Github
• RoBERTa: A Robustly Optimized BERT Pretraining Approach Paper, Github
• Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing Paper, Github
• Asynchronous Pipeline for Processing Huge Corpora on Medium to Low Resource Infrastructures Paper, Page
• Italy goes to Stanford: a collection of CoreNLP modules for Italian (TINT) Paper, Github, Page

All of the original datasets are publicly available or were released with the owners' grant. The datasets are all released under a CC0 or CCBY license.

• UD Italian-ISDT Dataset Github
• UD Italian-ParTUT Dataset Github
• WIKINER Page , Paper
• I-CAB (Italian Content Annotation Bank), EvalITA Page

@inproceedings {magnini2006annotazione,
	title = {Annotazione di contenuti concettuali in un corpus italiano: I - CAB},
	author = {Magnini,Bernardo and Cappelli,Amedeo and Pianta,Emanuele and Speranza,Manuela and Bartalesi Lenzi,V and Sprugnoli,Rachele and Romano,Lorenza and Girardi,Christian and Negri,Matteo},
	booktitle = {Proc.of SILFI 2006},
	year = {2006}
}
@inproceedings {magnini2006cab,
	title = {I - CAB: the Italian Content Annotation Bank.},
	author = {Magnini,Bernardo and Pianta,Emanuele and Girardi,Christian and Negri,Matteo and Romano,Lorenza and Speranza,Manuela and Lenzi,Valentina Bartalesi and Sprugnoli,Rachele},
	booktitle = {LREC},
	pages = {963--968},
	year = {2006},
	organization = {Citeseer}
}

Special thanks to I-CAB (Italian Content Annotation Bank) and EvalITA authors to provide the datasets as part of Master Thesis Research project with School of Engineering, University of Bologna.

Loreto Parisi: loreto at musixmatch dot com, loretoparisi
Simone Francia: simone.francia at musixmatch dot com, simonefrancia
Paolo Magnani: paul.magnani95 at gmail dot com, paulthemagno

Cite this work with:

@misc{musixmatch-2020-umberto,
  author = {Loreto Parisi and Simone Francia and Paolo Magnani},
  title = {UmBERTo: an Italian Language Model trained with Whole Word Masking},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/musixmatchresearch/umberto}}
}

We do Machine Learning and Artificial Intelligence @musixmatch
Follow us on Twitter Github