File size: 3,765 Bytes
96e9002
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
# CrisisTransformers
CrisisTransformers is a family of pre-trained language models and sentence encoders introduced in the paper "[CrisisTransformers: Pre-trained language models and sentence encoders for crisis-related social media texts](https://arxiv.org/abs/2309.05494)". The models were trained based on the RoBERTa pre-training procedure on a massive corpus of over 15 billion word tokens sourced from tweets associated with 30+ crisis events such as disease outbreaks, natural disasters, conflicts, etc. Please refer to the associated paper for more details. 

CrisisTransformers were evaluated on 18 public crisis-specific datasets against strong baselines such as BERT, RoBERTa, BERTweet, etc. Our pre-trained models outperform the baselines across all 18 datasets in classification tasks, and our best-performing sentence-encoder outperforms the state-of-the-art by more than 17\% in sentence encoding tasks.

## Uses
CrisisTransformers has 8 pre-trained models and a sentence encoder. The pre-trained models should be finetuned for downstream tasks just like [BERT](https://huggingface.co/bert-base-cased) and [RoBERTa](https://huggingface.co/roberta-base). The sentence encoder can be used out-of-the-box just like [Sentence-Transformers](https://huggingface.co/sentence-transformers/all-mpnet-base-v2) for sentence encoding to facilitate tasks such as semantic search, clustering, topic modelling.

## Models and naming conventions
*CT-M1* models were trained from scratch up to 40 epochs, while *CT-M2* models were initialized with pre-trained RoBERTa's weights and *CT-M3* models were initialized with pre-trained BERTweet's weights and both trained for up to 20 epochs. *OneLook* represents the checkpoint after 1 epoch, *BestLoss* represents the checkpoint with the lowest loss during training, and *Complete* represents the checkpoint after completing all epochs. SE represents sentence encoder.

| pre-trained model | source |
|--|--|
|CT-M1-BestLoss|[crisistransformers/CT-M1-BestLoss](https://huggingface.co/crisistransformers/CT-M1-BestLoss)|
|CT-M1-Complete|[crisistransformers/CT-M1-Complete](https://huggingface.co/crisistransformers/CT-M1-Complete)|
|CT-M2-OneLook|[crisistransformers/CT-M2-OneLook](https://huggingface.co/crisistransformers/CT-M2-OneLook)|
|CT-M2-BestLoss|[crisistransformers/CT-M2-BestLoss](https://huggingface.co/crisistransformers/CT-M2-BestLoss)|
|CT-M2-Complete|[crisistransformers/CT-M2-Complete](https://huggingface.co/crisistransformers/CT-M2-Complete)|
|CT-M3-OneLook|[crisistransformers/CT-M3-OneLook](https://huggingface.co/crisistransformers/CT-M3-OneLook)|
|CT-M3-BestLoss|[crisistransformers/CT-M3-BestLoss](https://huggingface.co/crisistransformers/CT-M3-BestLoss)|
|CT-M3-Complete|[crisistransformers/CT-M3-Complete](https://huggingface.co/crisistransformers/CT-M3-Complete)|


| sentence encoder | source |
|--|--|
|CT-M1-Complete-SE|[crisistransformers/CT-M1-Complete-SE](https://huggingface.co/crisistransformers/CT-M1-Complete-SE)|


## Results
Here are the main results from the associated paper.

<p float="left">
<img width="80%" alt="classification" src="https://raw.githubusercontent.com/rabindralamsal/images/main/cls.png" />
<img width="55%" alt="sentence encoding" src="https://raw.githubusercontent.com/rabindralamsal/images/main/se.png" />
</p>

## Citation
If you use CrisisTransformers, please cite the following paper:
```
@misc{lamsal2023crisistransformers,
      title={CrisisTransformers: Pre-trained language models and sentence encoders for crisis-related social media texts}, 
      author={Rabindra Lamsal and
		      Maria Rodriguez Read and
		      Shanika Karunasekera},
      year={2023},
      eprint={2309.05494},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```