Italian Trulli

Advertising Gender - Using Computer Vision to Trace Gender Displays in Historical Advertisements, 1920-1990

This study applies computer vision techniques to examine the representation of gender in historical advertisements. Using information on the relative size, position, and gaze of men and women in thousands of images, we chart gender displays in Dutch newspaper adverts between 1920 and 1990.

In the 1925 edition of Psychology in Advertising psychologist Alfred Poffenberger encouraged ad makers to ‘short-circuit the consumer's mind through vivid, pictorial appeals to fundamental emotions’ (Marchand, 1985). The images and visual clichés tap into a representational system that produces meaning outside the realm of the advertised product (Goldman, 1992). In the late 1970s, sociologist Erving Goffman examined the semiotic content in advertisements printed in contemporary newspapers and glossy magazines. He noted that displays of gender are often ‘conveyed and perceived as if they were somehow natural, deriving, like temperature and pulse, from the way people are and needful, therefore, of no social or historical analysis’ (Goffman, 1976). Of course, almost the exact opposite is true. Goffman conceived of five different categories to study the depictions of, and the relations between men, women, and children in advertisements: relative size, feminine touch, function ranking, ritualization of subordination, and licensed withdrawal. Drawing from this approach, Jonathan Schroeder, contends that in advertisements the male regularly represents the ‘the active subject, the business-like, self-assured decision maker, while the female occupies the passive object, the observed sexual/sensual body, eroticized and inactive’ (Schroeder, 2004).

Studies like that of Goffman and Schroeder relied on a limited number of advertisements, and they also do not take the historicity of ads into account. For example, Goffman looked at around 400 different advertisements, which he more-or-less randomly selected from ‘newspapers and magazines easy to hand - at least to my hand.’ As Kang (1997) notes, he was often criticized for this method. Instead of relying on a random sample, he purposefully ‘selected images that mirrored gender differences.’ Goffman emphasized that he choose his categories partly on the basis of the fact that he could find almost no images that disproved his categories. For example, images in which women were relatively larger and had the executive role.

In an earlier project, we developed methods to extract visual material from historical newspapers (Wevers and Smits, 2019). Using a large-scale data set of digitized historical advertisements extracted these methods, we examine continuity and change in displays of gender using computational means. By applying state-of-the-art computer vision methods to a diachronic set of digitized Dutch newspapers advertisements published between 1920-1990, we can test existing hypotheses and contribute to existing replication studies about gender displays in advertisements (Kang, 1997; Belknap, 1991; Lidner, 2004). 1 We will start our research after the First World War, because Dutch advertising practice changed rapidly in the Interbellum. Technological advances made it cheaper to print images in newspapers. Influenced by American ad agencies, Dutch ad makers saw the potential of using visual material to increase sales and to convey particular brand identities (Schreurs, 2001).

In this short paper, we operationalize Erving Goffman's theory on gender displays in two ways. First, Goffman argues that ‘differences in size will correlate with differences in social weight’ (Kang, 1997). Using facial recognition software (Geitney, 2018; Zhang, 2016), we select adverts that include people, and then train a gender detection algorithm using a convolutional neural network to estimate whether men or woman were represented in the images (Levi, 2015). This allows us to visually represent the changing faces of men and women in advertisements. Part of the process also entails a reflection on the inherent bias in these algorithms.

Second, Goffman contends that body postures and the engagement in social settings show that men are often portrayed in ‘executive roles’ and that women are often depicted as being removed from the social situation gazing off into the distance. Using information extracted in the first step, we quantify whether how people were represented on image, e.g. how was position where on the image? In Figure 1, for example, we see a social gathering of four people with the central figure drinking a bottle of Coca-Cola. Using the described computer vision techniques, we can extract faces and determining where on the images these are located. This reveals a man as a central figure with two women and man surrounding the central man. By extracting these features from a large set of advertisements and applying clustering methods, we can detect patterns in advertisements over time that can be compared to findings by Goffman and others. (Impett, 2017).

Coca-Cola advertisements.  , August 14, 1959.
Figure 1. Coca-Cola advertisements. Nieuwsblad van het Noorden, August 14, 1959.

We argue that it is not only possible to test existing theories on displays of gender in a more extensive database, but also to shed light on their historicity: the fact that these displays seem natural and constant, but can, and will, change, sometimes rapidly, over time. In line with the conference theme of complexity, we are aware that it is problematic to reduce displays of gender to binary categories. In this study, we are operationalizing displays of gender to a limited set of features, thereby reducing the complexity of gender. However, the coarse-graining of displays of gender into features that are comprehensible and computable allows us to extract trends from collections of advertisements. These trends offer context to particularities and can reveal the use of visual clichés in advertising discourse. Moreover, in the paper, we evaluate and reflect on the use and performance of models trained on contemporary training data as well as relying on probabilities to define gender.

Appendix A

Bibliography
  1. Belknap, P. and Leonard, W.M. (1991 ). A Conceptual Replication and Extension of Erving Goffman’s Study of Gender Advertisements. Sex Roles , 25, 3: 103–118.
  2. Geitney, A. (2018). The world's simplest facial recognition api for Python and the command line. https://github.com/ageitgey/face_recognition (accessed 27 april, 2019).
  3. Goffman, E. (1976). Gender Advertisements . New York: Harper and Row.
  4. Goldman, R. (1992). Reading ads socially . London: Routledge.
  5. Impett, L and Moretti, F. (2017). Totentanz. Operationalizing Aby Warburgs Pathosformeln. Technical report Stanford Literary Lab .
  6. Kang, M.E. (1997). “The portrayal of women’s images in magazine advertisements: Goffmans gender analysis revisited”. Sex roles 37, 11-12: 979.
  7. Levi, G. and Hassner. T. (2015) “Age and gender classification using convolutional neural networks”. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops : 34–42.
  8. Lindner, K. (2004). “Images of Women in General Interest and Fashion Magazine Advertisements from 1955 to 2002”. Sex Roles , 51, 7: 409–421.
  9. Marchand, R. (1985). Advertising the American dream: making way for modernity, 1920-1940 . Berkeley: University of California Press.
  10. Schreurs, W. (2001). Geschiedenis van de reclame in Nederland . Utrecht: Het Spectrum.
  11. Schroeder, J.E. and Zwick, D. (2004). “Mirrors of Masculinity: Representation and Identity in Advertising Images”. Consumption Markets & Culture , 7, 1: 21-52.
  12. Wevers, M., & Smits, T. (2019). The visual digital turn: Using neural networks to study historical images. Digital Scholarship in the Humanities .
  13. Zhang, K., Zhang, Z., Li, Z., & Qiao, Y. (2016). Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters , 23 (10), 1499-1503.
Notes
1.

The data is kindly provided by the National Library of the Netherlands.

Melvin Wevers (melvin.wevers@dh.huc.knaw.nl), KNAW Humanities Cluster, The Netherlands and Thomas Smits (t.p.smits@uu.nl), Utrecht University, The Netherlands