Arts and Culture
This paper shows how computer vision, machine learning and traditional critical discourse analysis can be combined to extract and study cultural representations from a large corpus of historical visual data. Based on Leiden University Library’s annotated database of North Korean posters, it explores what the automated processing of visual representations of people, politics and society can tell us about the historical and artistic evolution of the Democratic People’s Republic of Korea over the past few decades. From its origins in distributional semantics to the recent increase in corpus-based approaches, the field of discourse analysis has a long history of engaging with computational methods. Yet this engagement has by and large been dominated by a text-based perspective, with visual materials mobilized merely as representative illustrations of arguments built upon the analysis of textual data rather as elements of a larger visual discursive corpus with its own lexicon and grammar. This paper seeks to highlight both the prevalence and the shortcomings of such an approach. It argues that studies of visual representations too often rest upon the synechdochal assumption that an individual part (a single visual) can express the essence of the whole (an ideological structure), an assumption whose truth value is in fine only bound to the author’s authority. Such an argument, in turn, raises methodological questions about the representativeness of representations and stresses the need to use quantitative corpus studies to develop more robust approaches to the critical analysis of visual discourse.