MET-Meme: A Multi-modal Meme Dataset Rich in Metaphors

MET-Meme spans across two languages, amounting to 10039 text-image pairs with manual annotations.
Cimages.rar is a compressed file containing 6045 Chinese images.
Eimages.rar is a compressed file containing 3994 English images.
images can be downloaded here

Introduction to Met-meme：

(1): metaphor annotation:

1.metaphor occurrence : literal meme(0); metaphorical meme(1).
2.metaphor category: complementary;text dominant;image dominant.
3.source domain and target domain are annotated with the structure of "adj + n" or "verb + n".
4.source modality and target modality :complementary;text dominant;image dominant.

(2):sentiment analysis: happiness(1);love(2);anger(3);sorrow(4);fear(5);hate(6);surprise(7).

(3):intention category: interactive(1);expressive(2);purely entertaining(3);offensive(4);other(5).

(4):offensiveness degree: non-offensive(0);slightly offensive(1);moderately offensive(2);very offensive(3).

MET-Meme's code

The code we provide can be run directly, but the pre-training process of multilingual Bert, vgg16 and resnet50 requires the reader to complete beforehand. Detailed hyperparameters are shown in our paper.