PLVCG Model

This is the implementation of PLVCG model for Live Video Comment Generation based on PyTorch and Huggingface's Transformers.

This model is a seq-to-seq model to generate live video comments by their context comments and visual frames.

Requirment

Install Transformers:

pip install git+https://github.com/huggingface/transformers.git@7b75aa9fa55bee577e2c7403301ed31103125a35

Dataset

You can download our dataset at here and put them into the LiveBot folder.

This dataset is based on Livebot and the raw data can be found at livebot.

Config

All the parameters can be set at: LiveBot\MyPLVCG\config.

Train

Praining step:

python pretrain.py

Generate fine-tuning step:

python fine_tune_generate.py

Classificatin fine-tuning step:

python fine_tune_classification.py

Test:

Candidate comments ranking:

python test_rank.py

Generate comments:

python test_generate.py

Citation

If this is helpful for your work, please cite our paper as:

@inproceedings{DBLP:conf/pakdd/ZengGXT21,
  author    = {Zehua Zeng and
               Neng Gao and
               Cong Xue and
               Chenyang Tu},
  title     = {{PLVCG:} {A} Pretraining Based Model for Live Video Comment Generation},
  booktitle = {Advances in Knowledge Discovery and Data Mining - 25th Pacific-Asia
               Conference, {PAKDD} 2021, Virtual Event, May 11-14, 2021, Proceedings,
               Part {II}},
  year      = {2021}
}