VMSMO

Official code and dataset link for ''VMSMO: Learning to Generate Multimodal Summary for Video-based News Articles''

About the corpus

VMSMO corpus consists of 184,920 document-summary pairs, with 180,000 training pairs, 2,460 validation and test pairs.

We first publish the link (https://drive.google.com/drive/folders/1MpVv9naDaLINIo4ZKjGoZZHqp7v3_b-A?usp=sharing) to download each case in the dataset. The dataset consists of train.json, valid.json, and test.json. In each item in the json file, there are:

- ID: the ID number of the news
- content: the content of news
- original_pictures: whether the original microblog has pictures
- video_url: video URL
- image_url: video cover image URL
- publish_place: the place of publication
- publish_time: the release time of microblog
- publish_tool: microblog publishing method
- Up_num: number of likes
- retweet_num: number of forwarding
- comment_num: number of comments
- title: title of the weibo

Only the entries 'content', 'title', 'video_url' and 'image_url' are needed in our experiment. However, we keep all information in the json files for possible future uses.

About the code

Requirements

• python = 3.6
• tensorflow = 1.9
• numpy = 4.2
• opencv python = 1.16

### Commands

In the preprocess folder, we have videoprocess.pyto split the videos into frames, and dataprocess.py to read images, and find the image label for the video. Finally, by resnet152_img.py in sim folder, we use resnet to extract image features.

Train:

python run_summarization.py --mode=train --data_path=* --test_path=* --vocab_path=* --log_root=logs --exp_name=vmsmo --max_enc_steps=100 --max_dec_stpes=30 --vocab_size=50000 --lr=0.001

Test:

python run_summarization.py --mode=decode --data_path=* --test_path=* --vocab_path=* --log_root=logs --exp_name=vmsmo --max_enc_steps=100 --max_dec_stpes=30 --vocab_size=50000 --lr=0.001

We also give the crawler code used to crawl videos and text from weibo website, as shown in crawler-weibo folder.

Citation

We appreciate your citation if you find our dataset and code beneficial.

@inproceedings{Li2020VMSMO,
  title={VMSMO: Learning to Generate Multimodal Summary for Video-based News Articles},
  author={Mingzhe Li, Xiuying Chen, Shen Gao, Zhangming Chan, Dongyan Zhao, and Rui Yan},
  booktitle = {EMNLP},
  year = {2020}
}