Deep Learning Digital Humanities Propaganda in Context

Propaganda in Context

The goal of this project fund­ed by DI4DH is twofold: First, it is to be con­sid­ered a con­tri­bu­tion to the emerg­ing field of “dis­tant view­ing”, which uses quan­ti­ta­tive meth­ods to assess a cor­pus con­sist­ing of a large num­ber of visu­al media. Cur­rent­ly, deep learn­ing meth­ods play a minor role in dis­tant view­ing, as most of the projects use pre­trained net­works. This is under­stand­able, as train­ing is not triv­ial. How­ev­er, using pre­trained net­works sig­nif­i­cant­ly reduces the amount of pos­si­ble research ques­tions. More­over, a bet­ter under­stand­ing of the train­ing process allows us to con­tribute to the field of “crit­i­cal machine learn­ing” as well; specif­i­cal­ly we try to point out some of the ben­e­fits and pit­falls of train­ing an arti­fi­cial neur­al net­work for a human­i­ties research project.

We select­ed YouTube as an exam­ple, which has become the most impor­tant online media out­let in Rus­sia. In 2020, 82% of those aged 14–64 years use it dai­ly, mak­ing in the most suc­cess­ful exam­ple of Social Media in Rus­sia. There­fore, it is of vital impor­tance for Slav­ic cul­tur­al and media stud­ies to devel­op ana­lyt­ic tools for this platform. 

Three test cas­es are used in our study: Ukrain­ian nation­al­ist Stepan Ban­dera (1909–1959), who was instru­men­tal­ized by both sides of the Ukraine con­flict start­ing in 2013; promi­nent Russ­ian oppo­si­tion leader Alek­sei Naval’nyi; and Belaru­sian pres­i­dent Ali­ak­san­dr Lukashen­ka, who was re-elect­ed in fall 2020. In the case of Naval’nyi and Lukashen­ka, YouTube clips with sev­er­al mil­lion views helped to bring the protest to the streets. Nat­u­ral­ly, demon­stra­tions rely a lot on visu­al sym­bols such as flags and thus, allow us to test our the­o­ries. Pro­test­ers in Belarus, for exam­ple, do not use the offi­cial flag and coat of arms, which stem from Sovi­et times, but rather those of the first Belaru­sian repub­lic found­ed in 1918.

For these test cas­es, Deep Learn­ing is used to train an arti­fi­cial neur­al net­work (Resnet1010) to auto­mat­i­cal­ly detect 45 pre­de­fined nationalist(ic) sym­bols and 40 politi­cians from East­ern Europe, which in turn allow to ana­lyze the sym­bol­ic lan­guage of both state-run and oppo­si­tion­al pro­pa­gan­da dis­cours­es in East­ern Europe. This research ques­tion is not only inter­est­ing for Slav­ic media stud­ies, but also for East­ern Euro­pean His­to­ry and polit­i­cal sci­ences; the method­ol­o­gy, on the oth­er hand, pos­es impor­tant impuls­es for open­ing Dig­i­tal Human­i­ties for non-text media.

First exam­ples of our trained net­works are avail­able on Github.