SloganGeneration

Accpeted in CIKM 2023 short paper.

[Title] Slogan Generation with Noise Perturbation
paper link

Evironment

1. Prepare Data

Data format should be source-target pair matching.

• data from paper is crawled from (https://sloganlist.com) based on crawling_slogan.ipynb.
• (your own data) prepare your own data in csv file with 2 column including source and target.

2. Download Cuda with adaptable version

My version of torch version.

torch                    1.13.0
torchaudio               0.12.1
torchmetrics             0.11.0
torchvision              0.14.0

3. Install necessary packages in requirements.txt

installation

pip install -r requirement.txt

파일 설명

├── infer_model # where pretrained models are saved for inference 
│   ├── mt5-kr0131-Noise1+ep1200  #: pretrained korean model
│   │    ├── pytorch_model.bin  
│   │    ├── pytorch_model700.bin / pytorch_model800.bin 
│   │    └── tokenizer_config.json / config.json 
│   ├── t5-ensg-Noise1+ep50  #: english t5 small pre-trained model
├── crawling_slogan.ipynb # code for crawling slogans in sloganlist.com
├── inference.py  # inference code
└── requirements.txt #env settings

※ The model you want to use should be named 'pytorch_model.bin'.

▶ How to Run

python inference.py --input "brand description~~" --num_sequences 5 --language "trans"

ㄴinput : brand description sentence
ㄴnum_sequences : number of slogans to generate
ㄴlanguage : slogan generation language currently supports korean and english. Korean->ko, English->en, KOEN Translation->trans