Multi-Figurative Language Generation (COLING 2022)

Overview

Quick Start

How to use

git clone git@github.com:laihuiyuan/mFLAG.git
cd mFLAG

from model import MultiFigurativeGeneration
from tokenization_mflag import MFlagTokenizerFast
tokenizer = MFlagTokenizerFast.from_pretrained('laihuiyuan/mFLAG')
model = MultiFigurativeGeneration.from_pretrained('laihuiyuan/mFLAG')


# an example for hyperbole-to-sarcasm generation
# a token (<hyperbole>) is added at the beginning of the source sentence to indicate its figure of speech
inp_ids = tokenizer.encode("<hyperbole> I am not happy that he urged me to finish all the hardest tasks in the world", return_tensors="pt")
# the target figurative form (<sarcasm>)
fig_ids = tokenizer.encode("<sarcasm>", add_special_tokens=False, return_tensors="pt")
outs = model.generate(input_ids=inp_ids[:, 1:], fig_ids=fig_ids, forced_bos_token_id=fig_ids.item(), num_beams=5, max_length=60,)
text = tokenizer.decode(outs[0, 2:].tolist(), skip_special_tokens=True, clean_up_tokenization_spaces=False)
# special tokens: <literal>, <hyperbole>, <idiom>, <sarcasm>, <metaphor>, or <simile>

Training

Step 1: Pre-training

python train_pt.py -dataset ParapFG -figs hyperbole idiom metaphor sarcasm simile

Step 2: Fine-tuning

# parallel paraphrase pretraining data
python train_ft.py -dataset ParapFG -figs hyperbole idiom metaphor sarcasm simile

# literal-figurative parallel data
python train_ft.py -dataset MultiFG -figs hyperbole idiom metaphor sarcasm simile

Step 3: Figurative Generation

# Generating idioms form hyperbolic text
python inference.py -src_form hyperbole -tgt_form idiom

Model and Outputs

• Our model mFLAG can be found in Hugging Face, the corresponding outputs are in the /data/outputs/ directory

Citation

@inproceedings{lai-etal-2022-multi,
    title = "Multi-Figurative Language Generation",
    author = "Lai, Huiyuan and Nissim, Malvina",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = October,
    year = "2022",
    address = "Gyeongju, Republic of korea",
}