search

fesvhtr / DocMSU

[AAAI 2024] Official repository of the paper "DocMSU: A Comprehensive Benchmark for Document-level Multimodal Sarcasm Understanding"
7 stars 0 forks source link

Data sample size #2

Open ChoongwonKang opened 2 weeks ago

ChoongwonKang commented 2 weeks ago

The img.zip and anno.zip files contain only 71,828 data points. The dataset is stated to have 102,588 entries. Could you please clarify where the remaining 30,760 entries are located?

fesvhtr commented 2 weeks ago

As we mentioned in the paper, the dataset has been augmented by GPT. And, the current release of the dataset is the pre-augmentation version, we recommend that you use this original version. The complete data still needs to be organised.

Also, if you need more data, you can augment the text and images by yourself using LLM, etc. The data in this paper was augmented two years ago by an earlier version of the GPT, and the current technology allows for significantly higher quality data augmentation.

Thank you.