Authors:
Chhavi Sharma
and
Viswanath Pulabaigari
Affiliation:
Department of Computer Science Engineering, Indian Institute of Information Technology, Sri City, India
Keyword(s):
Meme, Dataset, Multi-modality.
Abstract:
In recent times internet ”memes” have led the social media-based communications from the front. Specifically, the more viral memes tend to be, higher is the likelihood of them leading to a social movement, that has significant polarizing potential. Online hate-speeches are typically studied from a textual perspective, whereas memes being a combination of images and texts have been a very recent challenge that is beginning to be acknowledged. Our paper primarily focuses on the meme vs. non-meme classification, to address the crucial primary step towards studying memes. To characterize a meme, metric based empirical analysis is performed, and a system is built for classifying images as meme/non-meme using visual and textual features. An exhaustive set of experimentation to evaluate conventional image processing techniques towards extracting low-level descriptors from an image is performed, which suggests the effectiveness of Haar wavelet transform based feature extraction. Further stud
y establishes the importance of both graphic and linguistic content within a meme, towards their characterization and detection. Along-with the deduction of an optimal F-1 score for meme/non-meme classification, we also highlight the efficiency induced by our proposed approach, in comparison with other popular techniques. The insights gained in understanding the nature of memes through our systematic approach, could possibly help detect memes and flag the ones that are potentially disruptive in nature.
(More)