The repository contains the code for Paper "Evolutionary Verbalizer Search for Prompt-based Few Shot Text Classification"
This repo includes various verbalziers in prompt-tuning, (based on OpenPrompt) and our Eoverb. We conduct experiments in five different datasets: AGNews, DBPedia, Yahoo Answers, IMDB, Amazon. More details can be seen in our paper.
# download datsets
sh download.sh
python train.pyYou can download Yahoo dataset and get the compressed file yahoo_answers_csv.tgz.
tar -zxvf yahoo_answers_csv.tgz -C ./datasets
cd datasets
mv yahoo_answers_csv yahoo| Arguments | Description |
|---|---|
| dataset | The input dataset: agnews, dbpedia, amazon, yahoo, imdb |
| model_type | Plms: bert, roberta |
| model_name_or_path | Path to pretrained model or shortcut name of the model |
| verb | verbalizer function: one, many, know, auto, soft, proto, evo |
| device_id | which gpu to be used |
| temp_id | which template to be used, you can input 0~3 |
| k | k-shot experiment setting |
| max_seq_length | The maximum total input sequence length after tokenization. |
| batch_size | Batch size per GPU/CPU for training. |
| learning_rate | The initial learning rate for AdamW. |
| num_train_epochs | Total number of training epochs to perform. |
| seed | random seed |
| Nl | label_word_num_per_class |
| Nc | the number of candidates |
| output_dir | the output directory. |