9524143564

Meme learn

Introduction

Meme (noun) - An image, video, piece of text, etc, typically hu.moro.us in naturę, that is .co.pi.ed and spread rapidly by Internet users, often with slight variations.    -

In our work, we acąuired and structured image memes from various so.ur.ces, effectively .creating a first ever (to the best of our knowledge) meme network. Having done that, we performed analyses and created a generative model.

Embeddings

10

Graph emhndding

Embeddings visualization with UMAP decomposition (color = meme templote)

10

-10

20-

10-

5

C- *rUv -b

ic

Comploto embedding

Meme generation

82.5%    66.5%

accuracy

We created a classification model, which selects the best matching image tern piąte fo r a given text. On top of this mod el, we bui lt a m eme gene rato r, whose sa m ple output is presented below.

Meme recommendation

Our simple recommendation system delivers the most similar memes to the one provided. It is based on cosine similarity and takes into cosideration multiple aspects of memes: image, text and users' soda I interactions.

JEŚLI NAUKA 0 DANYCH TO DANOLOCIA

TO CZY NAUKA 0 ZIEMI TO aEMOŁOGIA

^c=K-

t *

* MA fi *A« 7>    U\U

•A,

AUTHOR

l!n-«» w.illnho.t    KainnMK-jhiUir^irMI-

- COMMENTS

USERS

' )

Fig 1. Meme as .co.mp.lex data source

Meme Generation

Meme Recommendation

Model

"Zrobił zadanie, me zabrał do szkoły"

;ags

embeddings

Tag mbeddings

Piotr

Bielak

L

Politechnika

Wrocławska

\ /

Michał

Bieroński

Michał

jóźwiak

Processing pipeline

Fig 3, Applications

Finał thoughts

•    Memes' short texts were enough to perform pretty interesting tasks.

•    Based on embedding dustering, well defined communities do not seem to emerge from I mgfli p's user base.

•    Generative model works much better for Polish memes - higher nu.mb.er of parts of speech seems to h el p.