BERTweet: A pre-trained language model for English Tweets

BERTweet is the first public large-scale language model pre-trained for English Tweets. BERTweet is trained based on the RoBERTa pre-training procedure. The corpus used to pre-train BERTweet consists of 850M English Tweets (16B word tokens ~ 80GB), containing 845M Tweets streamed from 01/2012 to 08/2019 and 5M Tweets related to the COVID-19 pandemic. The general architecture and experimental results of BERTweet can be found in our paper:

@inproceedings{bertweet,
title     = {{BERTweet: A pre-trained language model for English Tweets}},
author    = {Dat Quoc Nguyen and Thanh Vu and Anh Tuan Nguyen},
booktitle = {Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
pages     = {9--14},
year      = {2020}
}

Please CITE our paper when BERTweet is used to help produce published results or is incorporated into other software.

For further information or requests, please go to BERTweet's homepage!

Main results

postagging ner

sentiment irony

Pre-trained models

Model	#params	Arch.	Pre-training data
`vinai/bertweet-base`	135M	base	850M English Tweets (cased)
`vinai/bertweet-covid19-base-cased`	135M	base	23M COVID-19 English Tweets (cased)
`vinai/bertweet-covid19-base-uncased`	135M	base	23M COVID-19 English Tweets (uncased)
`vinai/bertweet-large`	355M	large	873M English Tweets (cased)

Example usage

import torch
from transformers import AutoModel, AutoTokenizer 

bertweet = AutoModel.from_pretrained("vinai/bertweet-large")
tokenizer = AutoTokenizer.from_pretrained("vinai/bertweet-large")

# INPUT TWEET IS ALREADY NORMALIZED!
line = "SC has first two presumptive cases of coronavirus , DHEC confirms HTTPURL via @USER :cry:"

input_ids = torch.tensor([tokenizer.encode(line)])

with torch.no_grad():
    features = bertweet(input_ids)  # Models outputs are now tuples
    
## With TensorFlow 2.0+:
# from transformers import TFAutoModel
# bertweet = TFAutoModel.from_pretrained("vinai/bertweet-large")

Normalize raw input Tweets

Before applying BPE to the pre-training corpus of English Tweets, we tokenized these Tweets using TweetTokenizer from the NLTK toolkit and used the emoji package to translate emotion icons into text strings (here, each icon is referred to as a word token). We also normalized the Tweets by converting user mentions and web/url links into special tokens @USER and HTTPURL, respectively. Thus it is recommended to also apply the same pre-processing step for BERTweet-based downstream applications w.r.t. the raw input Tweets.

For vinai/bertweet-large, given the raw input Tweets, to obtain the same pre-processing output, users could employ our TweetNormalizer module.

Installation: pip3 install nltk emoji==0.6.0
The emoji version must be either 0.5.4 or 0.6.0. Newer emoji versions have been updated to newer versions of the Emoji Charts, thus not consistent with the one used for pre-processing our pre-training Tweet corpus.

import torch
from transformers import AutoTokenizer
from TweetNormalizer import normalizeTweet

tokenizer = AutoTokenizer.from_pretrained("vinai/bertweet-large")

line = normalizeTweet("DHEC confirms https://postandcourier.com/health/covid19/sc-has-first-two-presumptive-cases-of-coronavirus-dhec-confirms/article_bddfe4ae-5fd3-11ea-9ce4-5f495366cee6.html?utm_medium=social&utm_source=twitter&utm_campaign=user-share… via @postandcourier 😢")

input_ids = torch.tensor([tokenizer.encode(line)])