Skip to content

Suhanip/Multimodal-Meme-Classification

Repository files navigation

Multimodal-Meme-Classification

Identifying offensive content in image and text

This repository contains the code for identifying offensive content in image and text. Stacked LSTM + VGG16 is one of the way to identifying whether the given meme is offensive or not as mentioned in [1].

Technique used: Early Stopping to combine the image and the text modality and compare it with text- and an image-only baseline.

Prerequisite : Download the file "http://nlp.stanford.edu/data/glove.6B.zip" for before the execution of Glove Embedding of 50d.

Multi-OFF Dataset was leveraged by the authors to the 2016 U.S.Presidential election. Dataset can be downloaded using the link https://drive.google.com/drive/folders/1hKLOtpVmF45IoBmJPwojgq6XraLtHmV6?usp=sharing

References

[1] Suryawanshi, Shardul and Chakravarthi, Bharathi Raja and Arcan, Mihael and Buitelaar, Paul (2020). Multimodal Meme Dataset (MultiOFF) for Identifying Offensive Content in Image and Text. Association for Computational Linguistics.