The Journal for Transdisciplinary Research in Southern Africa 
ISSN: (Online) 2415-2005, (Print) 1817-4434
Page 1 of 8 Original Research
Not just a language with white faces: 
Analysing #taalmonument on Instagram  
using machine learning
Authors: From the late 19th century, and especially during apartheid (1948–1994), Afrikaans became 
Eduan Kotze1 
1 inextricably tied with white people, white domination and apartheid. This association has persisted Burgert Senekal  
after 1994, and calls to preserve Afrikaans are often derided with claims that the preference for 
Affiliations: Afrikaans is also a preference for racial segregation. In such anti-Afrikaans views, Afrikaans is 
1Department of Computer seen as synonymous with white people and apartheid despite the fact that Afrikaans was never 
Science and Informatics, exclusively spoken by white people. This prejudice towards Afrikaans is also shown towards the 
Faculty of Natural and 
Agricultural Sciences, Afrikaanse Taalmonument, which was unveiled in 1975 to commemorate this language.
University of the Free State, 
Bloemfontein, South Africa Using machine learning and a large dataset of social media posts from Instagram, we show 
that not only white people visit this monument to Afrikaans, but also take pictures here and 
Corresponding author: post about it on one of the largest social media platforms. As such, we show that the interest 
Burgert Senekal, in this monument – just like the language itself – is not exclusively tied to one race. We also 
burgertsenekal@yahoo.co.uk
make suggestions for further research, such as using machine learning for image recognition 
Dates: using social media datasets that could illuminate how other South African monuments are 
Received: 29 Apr. 2020 seen in the contemporary world.
Accepted: 18 Sept. 2020
Published: 15 Dec. 2020 Keywords: Afrikaans; Taalmonument; machine learning; convolutional neural networks; 
Instagram; apartheid.
How to cite this article:
Kotze, E. & Senekal, B., 2020, 
‘Not just a language with 
white faces: Analysing Introduction
#taalmonument on Instagram 
using machine learning’, The Since the late 19th century, Afrikaans ‘was constructed as a “white language,” with a “white 
Journal for Transdisciplinary history” and “white faces”’ (Willemse 2017). Because the Afrikaner-dominated National Party 
Research in Southern Africa (NP) carried out its policy of racial segregation (apartheid) in South Africa from 1948 to 1994, 
16(1), a871. https://doi. Afrikaans also became associated with apartheid. In particular, the 1976 Soweto riots, which was 
org/10.4102/td.v16i1.871
to a large extent opposition towards Afrikaans as a medium of education, turned the focus of 
Copyright: anti-apartheid resistance towards Afrikaans, ‘This rebellion stigmatized or “further stigmatized” 
© 2020. The Authors. Afrikaans, because the apartheid policy and its application caused injustice and increased a 
Licensee: AOSIS. This work negative attitude towards Afrikaners and Standard Afrikaans’ (Steyn 2014:418).1
is licensed under the 
Creative Commons 
Attribution License. Afrikaans is still associated with the apartheid government and related concepts such as 
oppression and the restriction of freedom, which has led to a resentment towards the language by 
a large proportion of the South African population (Van Zyl & Rossouw 2016:310). Recent protests 
at South African university campuses (e.g. #AfrikaansMustFall and #OpenStellenbosch) saw 
black students mobilising to remove Afrikaans as a language of tertiary education, arguing that 
it remained a barrier to education, offered an unfair advantage to white students, perpetuated 
racial segregation and alienated black students. This hostility towards Afrikaans can also be 
seen in the conduct of African National Congress (ANC) officials, in particular Gauteng MEC for 
Education, Panyaza Lesufi, and Minister of Higher Education, Blade Nzimande, who have made 
numerous statements against Afrikaans (Friedman 2019; Steward 2014). Nzimande, for instance, 
called the private Afrikaans-only tertiary education institution, Akademia, ‘racist’ because of its 
language policy (Steward 2014), whilest Lesufi made similar comments about Sol-Tech (Friedman 
2019). Such views ignore the fact that the majority of Afrikaans speakers (60%) are not white 
people (Willemse 2017) but have become commonplace in South Africa nevertheless.
Read online: The monument to the Afrikaans language, the Afrikaanse Taalmonument (Afrikaans Language 
Scan this QR 
code with your Monument), is likewise associated with apartheid by some, and there have been calls to dismantle 
smart phone or 
mobile device 1.Author’s own translation from the original Afrikaans, ‘Hierdie opstand het Afrikaans gestigmatiseer of “verder gestigmatiseer,” want die 
to read online. apartheidsbeleid en die toepassing daarvan het onreg veroorsaak en ’n negatiewe gesindheid teenoor Afrikaners en Standaardafrikaans 
laat toeneem’.
http://www.td-sa.net Open Access
Page 2 of 8 Original Research
the Taalmonument in the interest of nation-building (Smith speak Afrikaans as a first language (Smith 2013:133). 
2013:124; Van Zyl & Rossouw 2016:310). Groenewald Nevertheless, the Afrikaner is generally seen as a white 
(2018:230) calls the ANC-regime ‘antagonistic to the language nation (Senekal 2019) (note that Afrikaner and Afrikaans-
that the monument valorises’, and hence hostility towards speaking are two different labels, the former generally 
the monument itself can be expected. denotating an ethnic group and the latter a linguistic group). 
In light of this association between white people and 
The current study investigates posts made with the hashtag, Afrikaans, calls for the preservation of Afrikaans are often 
#taalmonument, on the social media platform, Instagram. As seen as an attempt to maintain segregation and ‘white 
Instagram posts constitute a voluntary association with this privilege’ (see e.g. Pilane 2015).
monument in the public sphere, the objective of the current 
study is to determine whether only white people – the race The Taalmonument symbolises Afrikaans’s diverse roots, 
associated with Afrikaans – voluntarily associate themselves including Western European (Dutch, French, German and 
with this monument or whether people of other races do the Portuguese), Malaysian and African languages, including 
same and to what extent. To this end, we develop, train and those of the Khoi-Khoi, San and other black Africans 
evaluate our own machine learning image recognition (Smith 2013:144; Van Wijk 2014:76; Van Zyl & Rossouw 
classifier after constructing our own annotated corpus of 2016:300–301). Nevertheless, there has been fierce criticism of 
images, which is also benchmarked against an internationally this monument, including that it is an ‘apartheidmonument’ 
recognised dataset. We also make suggestions for future (Van Zyl & Rossouw 2016:309). However, Van Wijk (2014:21) 
research. states that he did not design the monument for white 
Afrikaners but rather for the language itself (see also Van Zyl 
Background to the Taalmonument & Rossouw 2016:310). Moreover, an effort was made to 
The first proposal to erect a monument to Afrikaans was secure the attendance of coloured Afrikaans speakers and 
made at a commemoration of the founding of the Genootskap authors at the inauguration of the Taalmonument in 1975 
van Regte Afrikaners (Association of Real Afrikaners) in 1942 (Smith 2013:146; Van Zyl & Rossouw 2016:309). A poem by 
(Groenewald 2018:227; Van Zyl & Rossouw 2016:299). Adam Small (one of the most prominent coloured Afrikaans 
Following this proposal, the Afrikaanse Taalmonumentkomitee authors), ‘Nkosi sikelel’ iAfrika’, was also recited at the 
(Afrikaans Language Monument Committee) was founded to opening (Smith 2013:146). From the beginning, then, the 
raise funds for this purpose (Groenewald 2018:227; Van Zyl & Taalmonument has aimed at shedding the stigma of 
Rossouw 2016:299). More than 20 years later, in 1964, a Afrikaans being a language reserved for white people. 
competition was held to select an architect to design the However, with the Soweto riots occurring just the year after 
monument, and the architect Jan Van Wijk was chosen to the opening of the Taalmonument, this attempt at making 
design the monument (De Vaal-Senekal, De Kock & Putter Afrikaans more inclusive seems to have had little effect.
2018:198; Van Zyl & Rossouw 2016:300). The monument was 
unveiled by Prime Minister BJ Vorster on 10th October 1975, Today, the Taalmonument still aims at inclusivity, ‘The ATM 
and the accompanying Taalmuseum (Language Museum) strives for all South Africans to appreciate Afrikaans. In this 
was inaugurated on 14th August 1975 (De Vaal-Senekal et al. spirit, the ATM works hard to encourage and support 
2018:197; Van Zyl & Rossouw 2016:298). Afrikaans among the youth and non-mother-tongue 
speakers’ (De Vaal-Senekal et al. 2018:198, see also Van Zyl & 
A monument to Afrikaans will inevitably be placed in the 
Rossouw 2016:311; Smith 2013:138).
racialised discourse that is associated with this language. 
Although Afrikaans is currently associated with white 
people and apartheid, this was not always the case: When This effort to broaden the appeal of the Taalmonument and 
the Genootskap van Regte Afrikaners was founded in 1875, the museum should lead to a diverse collection of visitors. In 
most Afrikaans speakers were not white people, and the contemporary world, visitors to monuments and 
Afrikaans was often referred to as a hotnotstaal2 (Groenewald museums often share their visits with others on social media 
2018:228). As Willemse reminds us, ‘Afrikaans also has a platforms, such as Instagram, which provides the opportunity 
“black history” rather than just the known hegemonic to analyse social media posts to obtain a better understanding 
apartheid history inculcated by white Afrikaner Christian of who visits monuments and why. The following section 
national education, propaganda and the media’. Throughout provides a short background on Instagram.
the apartheid years (1948–1994), however, the Afrikaner 
was depicted as a white nation, with Afrikaans-speaking Instagram
coloured people marginalised by the apartheid state. Today, 
the majority of Afrikaans speakers are not white people, and Being founded in 2010, Instagram quickly became a major 
whilest 60% of South African white people have Afrikaans role player as a social media platform. Currently, Instagram 
as a first language, over 90% of the coloured population has around a billion worldwide users each month and 
500 million users each day, with over 50 billion photos 
2.A term for a person of colour from the Cape area in South Africa. The use of this 
term is now deprecated and considered offensive. The term ‘hotnot’ was historically uploaded to date (Aslam 2020). In South Africa, Facebook is 
used to refer to the non-Bantu indigenous nomadic pastoralist people of the the most popular social media platform, followed by 
Western Cape Province of South Africa. The preferred name for the non-Bantu 
indigenous people is currently Khoi, Khoikhoi or Khoisan. YouTube, WhatsApp, Facebook Messenger, LinkedIn, Twitter 
http://www.td-sa.net Open Access
Page 3 of 8 Original Research
and Instagram (Qwerty 2017:12). Instagram is a photo-based the South African census and in public discourses. Most 
platform that allows only photo and video posts, that is, no university staff have experienced being obligated to indicate 
text-only posts similar to Facebook and Twitter. their race on administrative forms as well, with racial 
categories reminiscent of the Population Registration Act 
Instagram is, however, not representative of the entire (Union of South Africa 1950) (white-, black-, coloured-, 
population of a country as Instagram users tend to be Indian people and other). We would therefore like to 
younger (Anderson & Jiang 2018; Aslam 2020; Duncan 2016). emphasise that we trained a model to conduct racial 
This is particularly relevant in the current study, as people classification because the discourse on Afrikaans and the 
who visit the Taalmonument and post pictures of their visits Taalmonument is already racialised; the irony of deracialising 
later will probably be from a younger generation that is less this discourse is that we first need to be able to distinguish 
tied to a first-hand experience of apartheid and the NP. Note, between races to ascertain whether visitors to the 
however, that we do not have access to users’ ages. Taalmonument who post about their visits afterwards on 
Instagram belong to one or various races.
To investigate whether only white people or people of 
different races associate themselves with the Taalmonument A variety of racial classification methods using machine 
on Instagram, we first had to train a model to distinguish learning have been proposed. Fu et al. (2014:2487) note 
between different races. The following section provides a ‘statistically significant variances in facial anthropometric 
background to machine learning for racial classification, after dimensions between all race groups’, which ‘pave the way of 
which we discuss the specific methods we used. anthropometry-based automatic race recognition’. The 
question is what to measure. There is a common misconception 
Machine learning for image classification that race is defined by a skin colour (as exemplified by 
referring to people as ‘white’ or ‘black’), and numerous 
Machine learning is a subfield of artificial intelligence (AI) efforts have been made to use skin colour to differentiate 
and was developed from the 1960s onwards (Kononenko between races, but Fu et al. (2014:2485) argue ‘skin color is 
2001; Michie 1968), in particular through the works of such a variable visual feature within any given race that it is 
Rosenblatt (1962), Nilsson (1965) and Hunt, Martin and Stone actually one of the least important factors in distinguishing 
(1966). The field gained ground in the most recent two between races’. A second view holds that ‘physical 
decades because of the big data revolution (Jordan & Mitchell characteristics such as hairshaft morphologic characteristics 
2015:256), leading Jordan and Mitchell (2015:260) to claim, and craniofacial measurements are viewed as significant 
‘machine learning is likely to be one of the most transformative indicators of race belongings’ (2014:2485), whilest another 
technologies of the 21st century’. method compares the eyes of subjects; Fu et al. (2014:2490) 
note ‘Statistically significant race differences in retinal 
A large amount of recent research has been directed towards geometric characteristics’, which have been reported in 
identifying race in images using machine learning (Fu, He & several studies. We opted for a more holistic approach by 
Hou 2014; Trivedi & Amali 2017; Vo, Nguyen & Le 2018). extracting whole faces and teaching a model to which race 
Although the concept of race is a contentious issue, faces belong, as discussed below.
particularly as the term is often used interchangeably or 
confused with ethnicity (see, e.g. Bartlett 2001; Collins 2004; Depending on the criteria and level of analysis, there are 
Markus 2008), Fu et al. (2014:2483) define the difference between three and 200 races (Coon 1962). Fu et al. (2014:2485) 
between race and ethnicity simply, ‘race refers to a person’s distinguish between seven races, which cover about 95% of 
physical appearance or characteristics, while ethnicity is the world population: African/African American, caucasian, 
more viewed as a culture concept, relating to nationality, East Asian, Native American/American Indian, Pacific 
rituals and cultural heritages, or even ideology’. We prefer Islander, Asian Indian and Hispanic/Latino. These seven 
this simple distinction between race and ethnicity and focus races, of course, exclude coloured people. In adapting racial 
the rest of our discussion on race. classifications for the South African context, we initially used 
the classifications suggested by Jan Raats, whose classification 
Racial classification is in one sense a highly controversial was used by the NP government through the Population 
topic, because it carries the baggage of the Population Registration Act (James 2012; Union of South Africa 1950) and 
Registration Act (Union of South Africa 1950) that, can still be found on administrative forms in South Africa 
along with other apartheid-era legislation, led to racial today. These categories distinguish between four races: 
discrimination and human rights abuses in South Africa white-, black-, coloured- and Asiatic people (we substitute his 
before 1994. In contrast, racial classification is not controversial classification of ‘bantu’ for the more politically acceptable 
in contemporary South Africa: Broad-Based Black Economic term ‘black’). However, the difference between Indian- and 
Empowerment (BBBEE), as well as the discourse around Asian people is so striking that we decided to split the 
white monopoly capital, transformation, white privilege and land Asiatic category into Asian- and Indian people.
expropriation, assumes racial categories. Despite the 
abolishment of racial categories in South Africa during the People of a mixed-race origin pose a significant challenge 
final years of apartheid, racial categories have persisted in to existing facial recognition models (Fu et al. 2014:2502). 
http://www.td-sa.net Open Access
Page 4 of 8 Original Research
This predicts that there will be difficulty in classifying a wrapper package for OpenCV Python bindings to perform 
South Africans who have been mixing for the past 350 years, image processing. OpenCV is a modern implementation of 
especially for the coloured population. Afrikaners, although the novel Classifier Cascade face detection algorithm (Viola & 
generally considered white people, are also not exclusively Jones 2001) and provides the CascadeClassifier class that 
caucasian in their genetic makeup (Erasmus, Klingenberg & allowed us to create a cascade classifier for face detection. A 
Greeff 2015; Greeff 2007; H. Heese 1979, 1984; J. Heese 1971). cascade, in machine learning terms, is an approach where a 
function is trained from numerous positive and negative 
Our experiments confirmed Fu et al.’s (2014:2502) assertion images. This will allow the image classifier to detect objects 
and encountered substantial difficulty in distinguishing (such as faces) in images. The result is that OpenCV allows us 
between white-, coloured- and black faces. When all five to extract faces from images, regardless of how many faces 
categories were included, we failed to move beyond an there are in a single image.
accuracy level of 70%, regardless of how we refined our 
model. We, therefore, simplified our racial categories to a Using the face detection classifier, we were able to successfully 
binary classification, white or black, as the objective of the detect and extract 3534 faces (2129 that were annotated as 
current study is in any case to determine whether only white people and 1405 that were annotated as black people) 
white people associate themselves with the Taalmonument from our training dataset, using a confidence factor of 0.98 
and whether other races do the same, regardless of which (in other words, we only allowed OpenCV to extract faces if 
race those people belong to. it was 0.98% certain that it was a face it had identified). We 
then randomly selected from each image dataset to create the 
The following section describes how the model was training and testing datasets. Table 1 shows the number of 
constructed and trained. images we used for the training and validation of the model.
Methods UTKFACE dataset
Model training We also wanted to compare our annotator’s classification 
by benchmarking our classifier with an internationally 
#modelsofinstagram dataset recognised dataset. The UTKFace dataset by Zhang, Song and 
A random sample of images was downloaded from Instagram Qi (2017) is a large face dataset with a long age span (subjects 
to collect sufficient training data that could be used in the of between 0 and 116 years old) and consists of over 20 000 
construction of a classifier. Images placed on Instagram are images with annotations in terms of age, gender and race. We 
already annotated to some degree by placing them with a applied an age filter (18–65 years old) on the dataset resulting 
hashtag, but the hashtag indicates to which discourse the in 17 655 images, as the #modelsofinstagram facial images 
image belongs and not necessarily what the content of the were extracted from Instagram users who will most likely fall 
image is. A picture with the hashtag #europeans could, for within this age range, as will the images posted with 
instance, show the picture of an African slave, as Europeans #taalmonument. We then randomly selected from these 
are known for slavery, but the hashtag does not indicate images selecting only white people (race = 0) and black people 
that the content of the image is a black African. We (race = 1) images to create the training and testing datasets. 
experimented with various possible hashtags that could be We did not filter on gender, because we wanted our classifier 
used to construct a labelled dataset, but possible hashtags to function across genders. Table 2 shows the training and 
differed considerably across races: Whilest #blackmodels and validation sets we used from the UTKFace dataset.
#indianmodels collected images of people belonging to these 
races, #whitemodels had a very limited selection of images Data augmentation
and #colouredmodels created problems with the different 
meanings associated with the term. Hashtags such as #san, As both datasets consist of a relatively small number of 
#european and #sotho did not return a meaningful number of training examples, one can inadvertently introduce 
relevant images. The hashtag #afrikaner delivered a overfitting into a model. Overfitting occurs when a model 
considerable number of irrelevant images, again partly because learns the noise instead of the signal of the training data and 
the term carries different meanings in different languages. We consequently will not generalise well from training data 
eventually decided to use a single hashtag, #modelsofinstagram, 
TABLE 1: Number of training and validation images for dataset 1.
and used an annotator to classify people according to race. The Variables White people Black people Total
annotator is in his late thirties and thoroughly familiar with Training 1000 1000 2000
racial categories in a South African context. Validation or testing 400 400 800
Total 1400 1400 2800
After the annotator had labelled the images, we began work 
on developing an image classifier. To classify an image TABLE 2: Number of training and validation images for dataset 2.
according to race, we first needed to perform face detection Variable White people Black people Total
and extract a face from an Instagram image, because this Training 3000 3000 6000
reduced the amount of noise in an image. For automatic face Validation or testing 500 500 1000
detection, we used opencv-python 4.2.0.34 (Heinisuo 2020), Total 3500 3500 7000
http://www.td-sa.net Open Access
Page 5 of 8 Original Research
to unseen data. In predictive modelling, the signal is the As the name convolutional neural network indicates, the neural 
underlying pattern that the machine learning model should network model employs a mathematical operation called a 
learn from the data. In other words, overfitting refers to when convolution. A convolution is a specialised kind of linear 
the model does not accurately learn what it is supposed to operation and enables a CNN to use convolution instead of a 
evaluate because of a small dataset. Suppose a large number general matrix multiplication in at least of one its layers 
of images of dogs also contain cars, the model can mistakenly (Goodfellow et al. 2016:327). After a convolution operation, 
identify cars with dogs and classify a cat as a dog because of the network will perform pooling to reduce the dimensionality. 
the presence of a car in the image. This enables the network to reduce the number of training 
parameters and as a result also shortens the training time. 
One way to overcome overfitting is to introduce data The  most common type of pooling is max pooling, which is 
augmentation by generating more training examples from the the  same type of pooling we use in our classifiers. This enabled 
existing training dataset. Data augmentation may include the models to reduce the input to the pooling layer (e.g. 32 × 32 
flipping, rotating or blurring images. The goal is to use × 10 dimensionality) to a 16 × 16 × 10 feature map as illustrated 
random transformations that create believable-looking in Figure 2 (adapted from Dertat (2017).
images and consequently artificially increase the number of 
training examples. For our experiment, we made use of the For our study, we constructed three CNNs: A CNN model 
ImageDataGenerator class that is bundled with Keras (Chollet consisting of three convolution blocks (Model1), a CNN 
2017), a Python deep learning library, to create batches of model consisting of four convolution blocks (Model2) and 
images with real-time data augmentation to both training a CNN model based on the VGG16 model proposed by 
datasets (#modelsofinstagram and UTKFace). These included Simonyan and Zisserman (2014) (Model3).
flipping images horizontally, rotating images by 45 degrees 
and zooming images up to 50% randomly. Finally, we also The first CNN model consists of three convolution blocks 
applied width shift and height shift by factors of 0.15.
(3 × 3 filter) with the same padding and a max pool layer 
(2 × 2 filter) in each of them resulting in eight layers. For the 
Deep learning models
For our machine learning classifiers, we made use of 
convolutional neural networks (CNNs) (Goodfellow, Bengio 
& Courville 2016:326). These classifiers are a specialised kind 
of neural network that process vector space (grid-like topology) 
datasets. These datasets can be a one-dimensional grid (1-D) 
such as time-series data or a two-dimensional grid (2-D) of 32 16Pooling
pixels such as image data. Convolutional Neural Networks 
have been used successfully in applications such as facial 16
recognition and more recently in natural language processing. 10
Examples of CNN image recognition models are MobileNet 
by Howard et al. (2017), Levi and Hassner’s (2015) age and 
gender recognition model and Campos, Jou and Giró-i-Nieto’s 32
(2017) image sentiment recognition model.
Convolutional neural networks consists of a series of 10
convolutional and pooling layers, and all CNN models have a Source: Adapted from Dertat, A., 2017, Applied deep learning – Part 4: Convolutional neural 
networks, viewed 22 April 2020, from https://towardsdatascience.com/applied-deep-
similar architecture. The architecture of CNN models is learning-part-4-convolutional-neural-networks-584bc134c1e2 
shown in Figure 1, which is adapted from Dertat (2017). FIGURE 2: Convolutional Neural Network pooling.
Input Conv Pool Conv Pool FC FC Somax
Source: Adapted from Dertat, A., 2017, Applied deep learning – Part 4: Convolutional neural networks, viewed 22 April 2020, from https://towardsdatascience.com/applied-deep-learning-part-4-
convolutional-neural-networks-584bc134c1e2 
Conv, convolution; Pool, pooling operations; FC, fully connected.
FIGURE 1: Convolutional neural network architecture.
http://www.td-sa.net Open Access
Page 6 of 8 Original Research
classification block, there were two fully connected layers classifier not to label a sample as positive if it was negative. 
with 512 units on top of the convolution blocks that were Recall is the ability of the classifier to find all the positive 
activated by a relu activation function. In deep learning samples. Accuracy returns the number of correctly classified 
neural networks, the activation function is responsible for samples whilest F1 is the weighted average of precision and 
transforming the summed weight input from a node into recall. As the training of the models took a substantial amount 
the activation of the node or output for that node of time, we did not train using n-fold cross validation. Cross-
(Brownlee 2019). Popular activation functions include sigmoid validation is a resampling technique to evaluate machine 
(or logistic), tanh (hyperbolic tangent) or relu (rectified linear learning models on a limited dataset whilest n (in n-fold) 
units). We opted for relu as it allows for backpropagation of refers to the number of groups that a given dataset is split 
errors to train our deep learning models (Goodfellow et al. into. Instead, we made use of Model Checkpoint and Early 
2016:226). In total, there were 10 904 097 trainable parameters. Stopping. Model Checkpoint monitors a specific parameter 
of the model (we used val_loss or validation loss) and Early 
The second CNN model consists of four convolution blocks Stopping will stop the training process of the model if there 
(3 × 3 filter) with the same padding and a max pool layer was no improvement in validation loss after a number of 
(2 × 2 filter) in each of them resulting in 12 layers. For the epochs. An epoch refers to the number of times that a learning 
classification block, there was a single fully connected layer algorithm with work through a training dataset. We set the 
with 512 units on top of the convolution blocks that were maximum number of epochs at 100 and allowed the model to 
activated by a relu activation function. In total, there were stop after 10 epochs if there was no improvement in validation 
7 595 809 trainable parameters. loss. After the training was completed, we tested each 
model with both #modelsofinstagram (n = 800) and UTFFace 
The third CNN model was a scaled-down version of the (n = 1000) testing datasets. Table 3 provides the test evaluation 
original VGG16 model proposed by Simonyan and Zisserman metrics of each model and dataset.
(2014). The original VGG16 model consists of five convolution 
blocks (3 × 3 filter) with a max pool layer (2 × 2 filter) in each From the testing of the models, CNN Model2 with our own 
of them where the ‘16’ refers to 16 layers that have weights. #models of instagram dataset performed the best. When 
Our model consists of four convolution blocks (3 × 3 filter) examining the model during testing, we noted a test loss of 
with the same padding and a max pool layer (2 × 2 filter) in 0.106758 with an accuracy of 0.97. The model reached the 
each of them resulting in 14 layers that have weights. For the optimal training validation loss value at n = 33 epochs. In 
classification block, there were two fully connected layers other words, our model is capable of predicting a person’s 
with 512 units on top of the convolution blocks that were race with 97% accuracy. With the model created, trained and 
activated by a relu activation function. In total, there were evaluated, we could now apply it to a dataset of images 
12 790 433 trainable parameters. downloaded with the hashtag #taalmonument, as discussed 
in the following section.
All three models output class probabilities based on a binary 
classification by the sigmoid activation function for output. Data gathering
We made use of the ADAM optimiser and a binary cross 
Before we could investigate the race of people that posted 
entropy loss function. Adam is an adaptive learning rate 
with the hashtag #taalmonument, we first had to download 
optimisation algorithm specifically designed for deep 
all posts tagged with this hashtag. Posts were downloaded 
learning (Kingma & Ba 2017). As we are using a binary 
using the application, InstaBro, on 14 February 2020. The first 
classifier, our loss function will also be binary and use cross 
post was made on 01 July 2012, meaning that the dataset 
entropy to measure how far from the true value (0 or 1) our 
spans over 7 years. There were 2988 photos posted with this 
prediction of each image was. The loss function will then 
hashtag (#taalmonument) during this period. Note that we 
average these class-wise errors to obtain the final loss 
could not gather any data about users, including their names, 
(Peltarion 2020). We also experimented with dropout, a 
age, location or gender. Importantly, we could only download 
regularization technique used to reduce the overfitting of a 
network. Dropout takes a fractional number as its input posts from public profiles, that is, we were not required to 
value, in the form such as 0.1, 0.2, 0.4, etc. This means follow users in order to include their content in the analysis 
dropping out 10%, 20% or 40% of the output units randomly below. In other words, these posts were made openly, in 
from the applied layer. For CNN model 1 and CNN model 3, front of an audience numbering around a billion, which 
we applied dropout to the last max pool layer (0.3 for Model means that the dataset constitutes posts made by people who 
1 and 0.2 for Model 3). TABLE 3: The test evaluation metrics of each model and dataset.
Dataset CNN Accuracy Precision Recall F1
Testing the models #modelsofinstagram Model1 0.95 0.95 0.95 0.95
Model2 0.97 0.97 0.97 0.97
We trained the three CNN models on both datasets. As the Model3 0.97 0.97 0.97 0.97
classifier is a binary classifier (only two labels, i.e. white UTKFace Model1 0.94 0.94 0.94 0.94
people or black people), we report the precision, recall, F1 Model2 0.95 0.95 0.95 0.95
and accuracy as the evaluation metrics used to assess the Model3 0.94 0.94 0.94 0.94
performance of the CNN models. Precision is the ability of a CNN, Convolutional Neural Network; F1, weighted average of precision and recall.
http://www.td-sa.net Open Access
Page 7 of 8 Original Research
TABLE 4: Results. interest in the monument. It may, for instance, be that a smaller 
Variable Black people (label = 0) White people (label = 1) Total proportion of coloured people show an interest in this 
CNN Model 2 139 529 688 monument than is the case for white people, but we have no 
CNN, Convolutional Neural Network. data to explain this skewed distribution and other factors 
could be at play.
openly chose to associate themselves with the Taalmonument. 
Furthermore, not including any information about users Of course, although the above shows a diverse association of 
prevents violating user privacy. For the same reason, we people with the Taalmonument on Instagram, this study did 
cannot provide examples of the racial classifications of users not conduct a representative investigation into attitudes 
and rather use the results in aggregate. towards the Taalmonument. Such a study of attitudes can 
better be conducted using a large sample of questionnaires or 
Results interviews. However, the above does show that, contrary to 
To perform predictions on the unlabeled facial image claims that it is an ‘apartheidsmonument’, users on Instagram 
dataset (#taalmonument), we deployed the best-performing take the time and effort to publicly associate themselves with 
classification model (CNN Model 2). First, the unlabelled this monument even if they are not white people.
dataset from Instagram (#taalmonument) was preprocessed, 
which included scaling and extracting human faces. From the Conclusion
2988 unlabelled photos, 668 human faces were identified This article showed that people who visit the Taalmonument 
and extracted using OpenCV (the rest of the photos were of and post about their visits later on Instagram are from various 
the monument or the landscape around the Taalmonument). racial backgrounds. Contrary to the racialised discourse on 
We then passed these facial images to our model as input and Afrikaans in South Africa, our study shows that not only 
received a label as output. Table 4 summarizes the results.
white people take the time and effort to travel to this 
monument, take pictures and post about it afterwards on 
The following section discusses these results.
Instagram – in other words, voluntarily associate with this 
Discussion monument on a highly public global platform. Our study therefore does not suggest that the Afrikaanse Taalmonument 
Census data show that most Afrikaans speakers are not is considered to be a ‘white people only’ or ‘apartheid’ 
white people, but as noted in the section discussing the monument but rather a monument that has enough 
background, the Taalmonument and Afrikaans are both significance for people of other races to also take the time and 
accused of being exclusively white phenomena. The results effort to take photos here and post about it on social media.
in the previous section, however, show that this is not 
entirely the case in our study. Of 688 faces identified We only investigated one monument and one factor, namely 
from Instagram posts made with the hashtag #taalmonument, race. Future studies could apply a similar method to 
529 (76.89%) were white people and 139 (20.2%) were black investigate the demographics of visitors who post on social 
people. As we showed our classifier to predict people’s race media with relation to other museums and monuments in 
with 97% accuracy, this shows that 20% of people who chose South Africa, including, for example, considering visitors’ age 
to associate themselves with the Taalmonument are not 
and gender. Social media provides a wealth of data with which 
white people. The key issue here is voluntary association: 
whilest people may attend an Afrikaans university based on to investigate how museums and monuments function in the 
the geographical location, the availability of transport, contemporary world and in people’s lives, and much of this 
limited course options or for other reasons, people who take opportunity has not been realised in academic research yet.
a photo at a monument and post it to Instagram do so 
willingly and intentionally. Moreover, taking the time to Acknowledgements
travel to the monument, taking a picture and posting it on Authors’ contributions
Instagram constitute a significant effort on the part of the 
user. The fact that a substantial number of people who All authors contributed equally to this work.
associate themselves with the Taalmonument are not white 
shows that this monument does not only garner attention Funding information
from the white population but rather functions in an 
inclusive capacity, as intended by Van Wijk. University of the Free State Interdisciplinary Research Grant.
However, it is unclear why only 20% of the faces we identified Ethical consideration
are not white people, whilest white people are a minority both This article followed all ethical standards for carrying out 
in the national population of South Africa and amongst research.
Afrikaans speakers. This over-representation of white people 
may reflect Instagram user demographics (no data are 
available on the distribution of Instagram use by race in SA), Data availability statement
cultural differences or it may indicate a smaller proportional The data are not publicly available due to privacy restrictions. 
http://www.td-sa.net Open Access
Page 8 of 8 Original Research
Disclaimer Hunt, E., Martin, J. & Stone, P., 1966, Experiments in induction, Academic Press, New York, NY.
The views and opinions expressed in this article are those of James, W., 2012, The strange career of race classification in South Africa, viewed 
04 April 2019, from http://politicsweb.co.za/news-and-analysis/the-strange-
the authors and do not necessarily reflect the official policy or career-of-race-classification-in-south.
position of any affiliated agency of the authors. Jordan, M.I. & Mitchell, T.M., 2015, ‘Machine learning: Trends, perspectives, and 
prospects’, Science 349(6245), 255–260. https://doi.org/10.1126/science.aaa8415
References Kingma, D.P. & Ba, J., 2017, ‘Adam: A method for stochastic optimization’, arXiv Preprint:arXiv:1412.6980v9.
Kononenko, I., 2001, ‘Machine learning for medical diagnosis: History, state of the art 
Anderson, M. & Jiang, J., 2018, Teens, social media and technology 2018, viewed and perspective’, Artificial Intelligence in Medicine 23(1), 89–109. https://doi.
23 March 2020, from https://www.pewresearch.org/internet/2018/05/31/teens- org/10.1016/S0933-3657(01)00077-X
social-media-technology-2018/.
Levi, G. & Hassner, T., 2015, Age and gender classification using convolutional neural 
Aslam, S., 2020, Instagram by the numbers: Stats, demographics and fun facts, viewed 
23 March 2020, from https://www.omnicoreagency.com/instagram-statistics/. networks, s.n., Boston, MA.
Bartlett, R., 2001, ‘Medieval and modern concepts of race and ethnicity’, Journal of Markus, H.R., 2008, ‘Pride, prejudice, and ambivalence: Toward a unified theory of 
Medieval and Early Modern Studies 31(1), 39–56. https://doi.org/10.1215/ race and ethnicity’, American Psychologist 63(8), 651–670. https://doi.org/ 
10829636-31-1-39 10.1037/0003-066X.63.8.651
Brownlee, J., 2019, A gentle introduction to the Rectified Linear Unit (ReLU), viewed 22 Michie, D., 1968, ‘“Memo” functions and machine learning’, Nature 218(5136), 19–22. 
April 2020, from https://machinelearningmastery.com/rectified-linear-activation- https://doi.org/10.1038/218019a0
function-for-deep-learning-neural-networks/. Nilsson, N., 1965, Learning machines, McGraw-Hill, New York, NY.
Campos, V., Jou, B. & Giró-i-Nieto, X., 2017, ‘From pixels to sentiment: Fine-tuning 
CNNs for visual sentiment prediction’, Image and Vision Computing 65, 15–22. Peltarion, 2020, Binary crossentropy, viewed 22 April 2020, from https://peltarion.
https://doi.org/10.1016/j.imavis.2017.01.011 com/knowledge-center/documentation/modeling-view/build-an-ai-model/loss-functions/binary-crossentropy.
Chollet, F., 2017, Deep learning with python, Manning Publications, Shelter Island, NY. Pilane, P., 2015, Black students’ dissent on afrikaans campuses is about more than just 
Collins, F.S., 2004, ‘What we do and don’t know about “race,” “ethnicity,” genetics and language policy, viewed 20 April 2020, from https://www.thedailyvox.co.za/
health at the dawn of the genome era’, Nature 36(11), 13–15. https://doi. black-students-dissent-on-afrikaans-campuses-is-about-more-than-just-
org/10.1038/ng1436 language-policy/.
Coon, C., 1962, The origins of races, Knopf, New York, NY. Qwerty, 2017, The digital landscape in South Africa 2017. A data driven look at South 
De Vaal-Senekal, P., De Kock, C. & Putter, M., 2018, ‘Challenges in the archives of Africa’s relationship with digital, viewed 19 March 2020, from http://qwertydigital.
the Afrikaans Language Museum, Paarl, Western Cape, South Africa: A case co.za/wp-content/uploads/2017/08/Digital-Statistics-in-South-Africa-2017-
study’, Atlanti 28(1), 195–205. https://doi.org/10.33700/2670-451X.28.1. 195- Report.pdf.
205(2018) Rosenblatt, F., 1962, Principles of neurodynamics, Spartan Books, Washington, DC.
Dertat, A., 2017, Applied deep learning – Part 4: Convolutional neural networks, 
viewed 22 April 2020, from https://towardsdatascience.com/applied-deep- Senekal, B.A., 2019, ‘Ras en afrikaneretnisiteit: ʼn kwantitatiewe ondersoek na huidige 
learning-part-4-convolutional-neural-networks-584bc134c1e2. opvattinge’, Ensovoort 40(8), 1. 
Duncan, F., 2016, So long social media: The kids are opting out of the online public Simonyan, K. & Zisserman, A., 2014, Very deep convolutional networks for large-scale 
sphere, viewed 17 September 2018, from http://theconversation.com/so-long- image recognition, s.n., s.l.
social-media-the-kids-are-opting-out-of-the-online-public-square-53274. Smith, S.J., 2013, ‘Monumentalising language: Visitor experience and meaning 
Erasmus, J.C., Klingenberg, A. & Greeff, J.M., 2015, ‘Allele frequencies of AVPR1A and making at the Afrikaanse Taalmonument’, Unpublished PhD-thesis, Southern 
MAOA in the Afrikaner population’, South African Journal of Science 111(7), 1–6. Cross University.
https://doi.org/10.17159/sajs.2015/20150074 Steward, D., 2014, Nzimande and the Afrikaners, viewed 20 April 2020, from https://
Friedman, B., 2019, Panyaza Lesufi takes a swipe at private Afrikaans university www.politicsweb.co.za/opinion/nzimande-and-the-afrikaners.
under construction, viewed 20 April 2020, from http://www.702.co.za/articles/ 
360944/panyaza-lesufi-takes-a-swipe-at-private-afrikaans-university-under- Steyn, J., 2014, ‘Ons gaan ’n taal maak’ Afrikaans sedert die Patriot-jare, Kraal 
construction. Uitgewers, Pretoria.
Fu, S., He, H. & Hou, Z.-G., 2014, ‘Learning race from face: A survey’, IEEE Transactions Trivedi, A. & Amali, D.G.B., 2017, ‘A comparative study of machine learning models for 
on Pattern Analysis and Machine Intelligence 36(12), 2483–2509. https://doi. ethnicity classification’, IOP Conference Series: Materials Science and Engineering 
org/10.1109/TPAMI.2014.2321570 263(4), 042091. https://10.1088/1757-899X/2 63/4/042091
Goodfellow, I., Bengio, Y. & Courville, A., 2016, Deep learning: Machine learning book, Union of South Africa, 1950, Population registration act, 1950, Union of South Africa, 
MIT Press Ltd., Cambridge. Cape Town.
Greeff, J.M., 2007, ‘Deconstructing Jaco: Genetic Heritage of an Afrikaner’, Annals Van Wijk, J., 2014, Taalmonument, Historical Media, Tokai.
of Human Genetics 71(5), 674–688. https://doi.org/10.1111/j.1469-1809.2 007. 
00363.x Van Zyl, A. & Rossouw, J., 2016, ‘Die Afrikaanse Taalmuseum en -monument in die 
Paarl: 40 jaar later’, Tydskrif vir Geestewetenskappe 56(2), 295–313. https://doi.
Groenewald, M., 2018, ‘An interrogation of the visual rhetoric of South African graphic org/10.17159/2224-7912/2016/v56n2-1a2
designer Ernst de Jong (1934–2016)’, Unpublished PhD-thesis, University of Pretoria.
Viola, P. & Jones, M., 2001, Rapid object detection using a boosted cascade of simple 
Heese, H., 1979, ‘Identiteitsprobleme gedurende die 17de eeu’, Kronos 1, 27–33. features, s.n., I-I, Kauai.
Heese, H., 1984, Groep sonder grense. Die rol en status van die gemengde bevolking Vo, T., Nguyen, T. & Le, C.T., 2018, ‘Race recognition using deep convolutional neural 
aan die Kaap, 1652–1795, Universiteit van Wes-Kaapland, Bellville. networks’, Symmetry 10(564), 1–15. https://doi.org/10.3390/sym10110564
Heese, J., 1971, Die herkoms van die Afrikaner, 1657–1867, A.A. Balkema, Kaapstad. Willemse, H., 2017, More than an oppressor’s language: Reclaiming the hidden 
Heinisuo, O.-P., 2020, Opencv-python 4.2.0.34, viewed 22 April 2020, from https:// history of Afrikaans, viewed 20 April 2020, from https://theconversation.com/
pypi.org/project/opencv-python/. more-than-an-oppressors-language-reclaiming-the-hidden-history-of-
afrikaans-71838.
Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, 
M. & Adam, H., 2017, ‘MobileNets: Efficient convolutional neural networks for Zhang, Z., Song, Y. & Qi, H., 2017, Age progression/regression by conditional adversarial 
mobile vision applications’, viewed n.d., from https://arxiv.org/pdf/1704.04861.pdf. autoencoder, viewed n.d., from https://arxiv.org/pdf/1702. 08423.pdf.
http://www.td-sa.net Open Access