UoM&MMU at TSAR-2022 Shared Task - Prompt Learning for Lexical Simplification: prompt-ls-en-2
We present PromptLS, a method for fine-tuning large pre-trained masked language models to perform the task of Lexical Simplification. This model is part of a series of models presented at the TSAR-2022 Shared Task by the University of Manchester and Manchester Metropolitan University (UoM&MMU) Team in English, Spanish and Portuguese. You can find more details about the project in our GitHub.
Models
Our models were fine-tuned using prompt-learning for Lexical Simplification. These are the available models you can use (current model page in bold):
Model Name | Run # | Language | Setting |
---|---|---|---|
prompt-ls-en-1 | 1 | English | fine-tune |
prompt-ls-en-2 | 2 | English | fine-tune |
roberta-large | 3 | English | zero-shot |
prompt-ls-es-1 | 1 | Spanish | fine-tune |
prompt-ls-es-2 | 2 | Spanish | fine-tune |
prompt-ls-es-3 | 3 | Spanish | fine-tune |
prompt-ls-pt-1 | 1 | Portuguese | fine-tune |
prompt-ls-pt-2 | 2 | Portuguese | fine-tune |
prompt-ls-pt-3 | 3 | Portuguese | fine-tune |
For the zero-shot setting, we used the original models with no further training. Links to these models are also updated in the table above.
Results
We include the official results from the competition test set as a reference. However, we encourage the users to also check our results in the development set, which show an increased performance for Spanish and Portuguese. You can find more details in our paper.
Language | # | Model | Setting | Prompt1 | Prompt2 | w | k | Acc@1 | A@3 | M@3 | P@3 |
---|---|---|---|---|---|---|---|---|---|---|---|
English | 1 | RoBERTa-L | fine-tune | simple | word | 5 | 5 | 0.6353 | 0.5308 | 0.4244 | 0.8739 |
English | 2 | mBERT | multilingual | easier | word | 10 | 10 | 0.4959 | 0.4235 | 0.3273 | 0.7560 |
English | 3 | RoBERTa-L | zero-shot | easier | word | 5 | - | 0.2654 | 0.268 | 0.1820 | 0.4906 |
Spanish | 1 | BERTIN | fine-tune | sinónimo | fácil | - | 3 | 0.3451 | 0.2907 | 0.2238 | 0.5543 |
Spanish | 2 | BERTIN | fine-tune | palabra | simple | - | 10 | 0.3614 | 0.2907 | 0.2225 | 0.538 |
Spanish | 3 | BERTIN | fine-tune | sinónimo | fácil | 10 | 10 | 0.3668 | 0.269 | 0.2128 | 0.5326 |
Portuguese | 1 | BR_BERTo | fine-tune | palavra | simples | - | 8 | 0.1711 | 0.1096 | 0.1011 | 0.2486 |
Portuguese | 2 | BR_BERTo | fine-tune | sinônimo | fácil | - | 10 | 0.1363 | 0.0962 | 0.0944 | 0.2379 |
Portuguese | 3 | BR_BERTo | fine-tune | sinônimo | simples | 5 | 10 | 0.1577 | 0.1283 | 0.1071 | 0.2834 |
Citation
If you use our results and scripts in your research, please cite our work: "UoM&MMU at TSAR-2022 Shared Task: Prompt Learning for Lexical Simplification".
@inproceedings{vasquez-rodriguez-etal-2022-prompt-ls,
title = "UoM\&MMU at TSAR-2022 Shared Task: Prompt Learning for Lexical Simplification",
author = "V{\'a}squez-Rodr{\'\i}guez, Laura and
Nguyen, Nhung T. H. and
Shardlow, Matthew and
Ananiadou, Sophia",
booktitle = "Shared Task on Text Simplification, Accessibility, and Readability (TSAR-2022), EMNLP 2022",
month = dec,
year = "2022",
}
- Downloads last month
- 4