Cynthia Yee Ting Ng
Livia van Vliet
Not long after May 25, 2020 and the insurgent Black Lives Matter protests against police brutality and racism, politicians chipped in on social media; often tapping into trending hashtags in visual ways (especially poignant was the #BlackOutTuesday tag that was accompanied by black squares). How do politicians go about their digital visual strategies in this highly emotionally charged issue space? And how do people receive such displays of solidarity by politicians: as ‘virtue-signalling’...or as authentic? What determines their responses and what are the diverging visual discourses within issue groups and in the realm of politicians on Twitter, Facebook and Instagram?
Even though the project began with the question of authenticity regarding public displays of authenticity and solidarity regarding the Black Lives Matter Movement on Twitter, Facebook and Instagram after the killing of George Floyd, in practice this seemed difficult to pursue. That is why this research has taken a more comparative approach in questioning the differences between the information shared by political parties on national, European and international levels through Twitter, and comparing them to the grassroot movements on Instagram and Facebook.
Social media platforms have developed their image uploading and sharing capacities, some of them to the extent that they transformed themselves into primarily image-based media. In this sense images ‘living and dying’ on such platforms are often more communicative rather than commemorative (van Dijck, 2007), meaning that we conceptualize them as meaningful ‘texts’. Approached in this way images conveying solidarity with a certain cause, such as Black Lives Matter imagery, are in this project conceptualized as visual communication of self-disclosure and potentially virtue-signalling.
Our leading research question was:
What are the differences in the information that is shared by grassroot BLM movements (Insta and FB) versus top-down information diffused by political elites (Twitter)?
Some sub questions for the projects were:
Do politicians/grassroots movements talk about BLM?
What are the main words used when talking about BLM?
What is the sentiments of tweets related to BLM? And:
What are the images shared in this issue space?
Researchers undertaking visual research in the realm of social media are methodologically challenged, often having to tackle issues of selection due to the sheer amount and unstable stream of images uploaded to social platforms every day. Through casting the net wide -- by incorporating multiple platforms and a lot of data -- we stayed in the exploratory mode for a significant part of the Summer School. Through intense exploration we found multiple interesting directions for further analyses and other political and public visual strategies relating to Black Lives Matter. This resulted in a ‘birdseye view’, showing a variety of methodologies and subprojects and possible outcomes.
A variety of methodologies was implemented, including: hashtag collocation networks, tweet frequency over time, topic modelling, sentiment analysis, image tagging (computer vision), ImageSorter, image-label networks and Crowdtangle.
The initial datasets included:
Datasets of politicians (UK, US, EU) (via: www.twitterpoliticians.org)
#GeorgeFloyd tagged tweets subset
Instagram #blackouttuesday, #GeorgeFloyd, #AllLivesMatter and #BlackLivesMatter
For the Twitter politician datasets, we used the hashtag frequencies and hashtag co-occurrence network to determine what hashtags were used when discussing black lives matter in the US (between the house and senate), the UK and the European Parliament. We cleaned all the text, removing stopwords, converting it to lower case, and treated the hashtags as regular words (since we discovered through word frequency that politicians discuss black lives matter without necessarily using the hashtags. We then filtered each country based on the related hashtags. In the case of the European Parliament, we had to translate all the tweets first using the google translate python library before filtering them.
Following this, we used used the python library Matplotlib to plot bar graphs showing which party (or in the case of the European Parliament, political group) discussed BLM related issues the most, followed by generating wordclouds using the wordcloud library to see the most frequently used terms by politicians across countries.
We then created topic models using LDA modelling from the python libraries pyLDAvis and gensim to examine models based on unigrams, bigrams and trigrams for each country. We examined the number of topics and picked the number per dataset that would generate the most distinct topics.
Lastly, we used the Sentistrength from the NLTK Python library to conduct sentiment analysis on the US house and UK parliament tweets.
For the Twitter George Floyd hashtag based dataset, we followed a similar procedure as with the politician dataset when it came to cleaning the tweets. We did not filter them as they were already filtered by all tweets that contained ‘George Floyd’.
We ran a LDA topic model on the Tweets. We also did qualitative analysis on subsets of the data, specifically in the tagged MAGA and Antifa spaces.
We used both Instagram-Scraper and InstaLoader on Github to scrape issue-relevant Instagram posts associated with key hashtags.
We scraped #blackouttuesday on Github’s instagram-scraper, acquiring a set of over 13,000 images. #blackouttuesday is considered a “Social Media Moment,” which took place on June 2, 2020 and reverberated in the days following. The movement began in the music industry, drawing on #TheShowMustBePaused to show economic solidarity by refraining from advertising on social media. Some Instagram users responded by posting a single black box in the feed. In other cases, black artists and business were featured and promoted. For some, the day was used as a "day to disconnect from work and reconnect with our community." Unfortunately, due to time constraints, we were unable to acquire the full #blackouttuesday image set along with the metadata. However, this data set enabled exploration of dominant visual discourses using Image Sorter.
Quali quanti visual analysis with Vision APIWe also engaged in automated content analysis of images that was used as a starting point of a typically quali-quanti approach in assessing image content. We used the Memespector (Git) tool that taps into the Google Cloud Vision API to label images. This tool also generates a networked file in which images-as-nodes are organized by their annotated content labels (essentially organizing images by similar content images and not so much by their formal modalities which sets this direction apart from ImageSorter that organizes mainly by color). Although highly problematic to use as a stand-alone method for analysis of social media images and their expressive meaning, we demonstrate that combining machine vision with certain platform data—hashtag data in this case —presents a way forward, especially in the selection of visuals one should study, see also https://journals.sagepub.com/doi/full/10.1177/2056305120928485. Our fabulous designers added an interepretative layer on top of the image label networks so we could identify interesting clusters to read closely. Facebook For Facebook we determined the three most salient Reactions buttons and qualitatively read their typical visual vernacular. We did that in two ways: ordering such images by their Reactions scores (engagement) and through running a sample of images that co-occur with particular Reactions through the labelling Vision API which allowed us to qualitatively read the image label networks garnered by the tool Memespector tapping into Google Vision. Furthermore we ran a LDA topic modelling algorithm on the comments of the global BLM page on Facebook which led us to a significant topic surrounding the fake use of an image, initially posted to Instagram.
The ways politicians discuss Black Lives Matter
This first subproject focused on the top-down engagement of politicians with the BLM movement on twitter, where we focused mainly on the UK, the European Parliament and the US House and Senate. First, we explored Political Hashtag co-occurence networks, exemplified below by showing the United Kingdom and the European Parliament. Taking the data from the 25th of May, the day of the killing of George Floyd, for about three weeks. From the networks you can see that (of course) the corona virus comes across all the networks, but also that there are distinct topics on Black Lives Matter in both the UK and the European Parliament.
We also examined the hashtag co-occurences from the United States from the same dates. From the results for the United States House, we can see that they had two interesting topics running surrounding Black Lives Matter, namely “Black Lives Matter” on its own, and “George Floyd.”. Whereas in the senate, there is little to no discussion on Black Lives Matter, and George Floyd was a single ‘node’ that got mixed into a discussion on other topics such as “Iran”.
Subsequently, we went on to explore how these parties then specifically discussed Black lives Matter. As you can see (below) there are very different words used to discuss the topic. The European Parliament focuses on protest, racism and the police a lot, while the UK parliament stands on its own with the use of the word ‘solidarity’. The US House focuses more on George Floyd, police brutality, gun violence and United States nationalism.
Another way we explored the data, was by looking at tweet frequency over time. Juxtapositioning the timelines for the European Parliament and the UK parliament showed an interesting dissimilarity in Twitter traffic. Surprisingly, after the killing of George Floyd and the Minneapolis Protests there was a general increase on the amount of tweets for Black Lives Matter in the European Parliament. Then a huge spike followed when Trump threatened to deploy the military. Contrastingly, the UK activity showed that the discussion on Black Lives Matter was rather stagnant until Trump threatened to deploy the military, which was followed by a huge spike in activity. In the US we see something similar. On the day of the killing of George Floyd there was a big spike in activity, followed by a more stagnant period, followed by an even bigger spike after the threat by Trump. The big drop on the 6th of june represents the biggest Black Lives Matter protests in Washington. It is not completely clear why this caused a drop in the amount of tweets.
Next, we went on to explore the data through topic modeling; this showed the differences in the topics that would emerge from the tweets, similar to the word frequencies. In the US the biggest topic was indeed police brutality. Other topics were legislation, covid and Black Communities. The most outlying topic seems to be more specifically about George Floyd.
The UK parliament showed a different set of topics. The main topic here was racism, solidarity and injustice. Moreover, there were quite a lot of calls to action, such as calls to action against racism, and the call to action to not stay silent when you witnessed racism. There were also topics focusing on the protests in London and protests in general. Last, a small topic was police actions and social distancing.
Finally, when it came to the European parliament, we had less success since the topics were in many different languages. Two main topics in English were protests and police brutality. Other subtopics were in French and Spanish. From this we had learned a lesson; first translate the tweet, then filter it and process last.
Next, we performed a sentiment analysis to explore the sentiment of politicians for Black Lives Matter to find out whether political parties showed generally more positive or negative sentiments towards the topic. For the US House, it is clear that the democrats are talking a lot more about Black Lives Matter in general. A similar schism is seen in the distinction between the left wing and right wing politicians in the UK Parliament. Interestingly, a lot of ‘negative’ tweets came up. When looking at the singular tweets, it becomes clear that the classifier defined a lot of tweets as negative that were actually in solidarity with Black Lives Matter, due to the used vernacular (e.g. “tragic”, “shameful”, “crime”.). This shows that ‘out of the box’ sentiment analysis is not useful with this sort of topic. Hence, we abandoned this route.
We then went on to look at the images that were shared. We ran them through memespector and looked at the tags that were applied to the images. There appeared to be a big difference between the images that were shared by the US House and by the EU parliament. Below, you can see that the US House-images focused mainly on the specific case of George Floyd (“years old”, “American”, “George Floyd”), and the EU parliament focused more on the movement as a whole (“black lives”, “black lives matter”, “police”).
To conclude; politicians are quite segmented when talking about Black Lives Matter both within parties and across countries. Moreover, the data needs to be filtered well in order to capture tweets that ‘skirt’ the issue. For example, words like ‘justice system’ and ‘antifa’ are used more by republicans although they are still discussing issues related to Black Lives Matter. Last, most politicians tweets were classified as ‘neutral’, and negative tweets were not necessarily negative, showing that the classifier was not working well.
How this compares to Twitter Data of the grassroots movement
This subgroup focused on “George Floyd” related tweets by the bottom-up grassroots movement. We have results from a random sample, and again filtered that sample down to the MAGA/Antifa tweets. From these two samples it becomes clear that there are very different conversations going on on this topic.
Similarly, if we look at the hashtag frequencies comparison between Allivesmatter and MAGA, there are very different hashtags used within the same dataset by different groups.
Within the dataset, BarackObama and realDonaldTrump were mentioned a lot. This was due to the statement that Obama posted on the 29th of May on the death of George Floyd, which attracted more than 110.000 retweets within two days. Most of the tweets that were related to Obama also shared images of Obama and George Floyd related footage. In contrast, when looking at the tweets related to Donald Trump, the most tweeted images were not of Donald Trump at all, resulting in very different content connected to the two politicians.
Next, the topic models on MAGA and Antifa tweets showed interesting things as well. For example, Antifa and Trump appeared to be a separate Spanish bubble. Other topics were protesting, George Floyd’s physical conditions during the event that lead to his death and famous people that tweet about George Floyd.
Whereas if we look to the sample of George Floyd tweets, we first of all see similar topics, such as the George Floyd event, famous people talking on the topic and Trump. However, also topics here are antifa, fascists, MAGA and black violence.
To conclude: how do these bottom up tweets differ to the politician’s tweets? First of all the discussion on specific politicians stands out; politicians don not generally discuss Trump. Moreover, the hashtagged data seems to be mainly American focused and discusses Trump a lot (it should be noted that the MAGA space is not necessarily pro-Trump). Third, it becomes clear that politicians tend to discuss local issues, including Covid19, that are entangled with the global BLM issues, while the George Floyd hashtagged data does not seem to have much entanglement with other issues than BLM at all. Last, politician’s tweets seem to contain more ‘call to action’ content, whereas the George Floyd tweets tend to be more an expression of outrage.
Visual Displays of Solidarity on Instagram and Facebook
This last subproject incorporated other platforms into the research by comparing the previous discourse (both visual and textual) to the Instagram and Facebook grassroots spaces.
First, we focus on Instagram. Here we compared the visual vernaculars in #GeorgeFloyd, #BLM and #AllLivesMatter. The expectation was that for All Lives Matter, there would be a visually very different colloquialism. This appeared not to be the case. The visual repertoires we came across were very diverse; especially interesting were the reworkings of news events and images. The artistic interpretation of for example the death of George Floyd, combined with the statue of liberty. Also the visuals included a lot of protests and kneeling.
Especially evident from processing the images in imagesorter was also the historic visual connections that users made to past expressions of racism and oppression. Moreover, we could see a lot of motivational text based images and user generated content.
We also had data from #blackouttuesday, however there we missed the metadata. Still there were also some interesting findings, one of which was the limited number of black squares by politicians. The use of the hashtag #blackouttuesday seems to be limited to the grassroots movement. It could be that the black squares were pushed out by other forms of imagery. A possible critique could be that black squares make black bodies invisible. There seems to be a visual campaign going on that pushes out the black squares; this could be fleshed out further.
The vast proportion of the #blackouttuesday network features photos of people, typically headshots (potentially selfies) of black folx looking fabulous (see below, right). Further research into the text associated with these posts would help us better understand these images in relation to the research question. There is potential that black folx are indicating pride and self-love in resistance to mainstream media representations and politician framing of BLM, as well as how black squares make black bodies invisible. There is also potential that some posts are promoting black businesses (images of products, particularly clothing items, are also very prevalent in the data set - potentially associated with the promotion of black businesses as part of the #blackouttuesday campaign). Finally, platform vernaculars may be at play - selfies and polished forms of self-representation are part and parcel of Instagram as a platform.
Furthermore, we made image-label networks using google vision API for the hashtag #blacklivesmatter. This technology clustered the images based on crowd. Here there are always interesting discourses between the captions and the images themselves (or even the comments).
From the hashtag-image network of George Floyd it became clear that the black is beautiful vernacular has become entangled with the Black Lives Matter protests and that this was again entangled with Covid-19 and Arabic captions. Prominent memes are especially interesting since these were completely absent from the political discourse. It is also typical for instagram to bring items into the discourse that are not necessarily related to the issue.
Subsequently, we went into the hashtag #allivesmatter. Here we merely travelled the imagesorter, which gave us some surprising findings. This seemed to be a space where a surprising amount of content was pro-blacklivesmatter (apart from the amount of expected hateful content). Interestingly, there were quite some image clusters of hijacking hashtags and targeted clogging of the feed with non-related content. These images or memes were all posted within the same short time frame, with no username or full name associated and featured random images of memes, K-pop stars, the Joker, etc.
This last subproject focused on emoji button dynamics and visual vernaculars. Here we demarcated the data of a #bluelivesmatter query through facebook crowdtangle. We demarcated images through their most used Reactions buttons. In this case, we focused on most loved, most sad and most angry. Interestingly, analysis of the reactions showed that the images connected to ‘sad’ and ‘angry’ largely correlated, while the images connected to ‘love’ were more related to ‘like’. This links to findings in earlier DMI-based work (see also https://ijoc.org/index.php/ijoc/article/view/11657) that found that two buttons strongly co-occur signifying more ambiguous emotions at play. When displaying solidarity, the ‘sad’ button is least used in Blue Lives (way more Angry folks..). Emphasis in the Sad space is on the harm done by black people, as opposed to the killing of George Floyd by a white policeman. In Angry we see, quite unsurprisingly, the ‘back the blue’ demonstration and the injustice toward the police.
In both the Love and Sad spaces we see an interesting re-occurrence in both highly engaged as well as frequently used images: that of visuals depicting people of color that are conveying a testimonial for upholding either all or blue lives matter and at times bluntly denying racism altogether. See figure below for an example.
Topic modelling on the official Black Lives Matter page on Facebook showed that topic modelling lead to interesting content. In this case, one topic was relating the killing of George Floyd to the harassment of a pregnant woman, which appeared to be flagged content of fake news. It holds a terrible image of an American tourist who was depicted after being (clearly physically abused) and raped while on holidays in Madrid in 2018. The false caption enclosed told a story of her being abused by GeorgeFloyd at Walmart.
To conclude, some of politician’s visual displays are connected to grassroot movement displays (black squares, protest, historical claims…). However, in their visual displays we mostly find headshots and institutional photos, texts and user created content and calls for action, which is very different from the visual displays by the grasroots movement.
Moreover, there are some significant differences between the vernaculars of Instagram, Twitter and Facebook. For example, Twitter’s visual displays are self-referential and performative. Grassroots visual displays in Instagram are more focused in motivational texts, symbols, photos of people & protest. Also in Instagram we find selfies and polished forms of self-representation as well as commodities and branding.
Studying networked images and their meta-data garnered through platform-specific features, means working with messy data (incomplete and unstable dynamic metadata such as retweet metrics) and unstable objects of research: images that are constantly recontextualized as they travel within and across platform(s).
This project addresses challenges posed to researchers undertaking visual research in the realm of social media, one of which being the issue of selection. Hand (2016, p.8) summarizes this aptly as: “social media enable access to billions of images that are difficult to classify. For qualitative researchers, the identification, selection and organization of visual materials always involves negotiating complex issues of sample, representation, authenticity and exhaustiveness.”"
We took two main routes for selecting images to study social media; topic modelling and machine clustering images. We believe that the main way forward is to integrate these methodologies further. In order to know the direct responses to political messages, one has to go into the comments sphere further. The ambition is also to start working more with emoji’s; so not merely Reactions buttons, but in-text emojis. Hence; we would welcome a new Gephi plugin that would allow for the display of emojis
Hill, M., & Royal, C. (2018). “Thank You, Black Twitter”: State Violence, Digital Counterpublics, and Pedagogies of Resistance. Urban Education, 53(2), 286-302.
Online community empowerment, emotional connection and armed love in the BLM movement:
Book reviews of #1960now (historical context)
Visual methods for social media
'Photography meets Social Media: image making and sharing in a
continually networked present', in Pasternak, G. (ed.) The Handbook of Photography Studies. London: Bloomsbury.
Schreiber, M. (2017). Audiences, aesthetics and affordances, analysing practices of visual communication on social media. Digital Culture & Society, 3(2), 143–164
Hand, M. (2017). Visuality in Social Media: researching images, circulations and practices. In: Sloan, L. Quan-Haase, A. (eds.). The Sage Handbook of Social Media Research Methods. London: Sage
Pearce, W., Özuka, S.M., Greene, A.K., Telling, L., Bansard J.S., Omena J.J. & Rabello, E.T. (2018). Visual cross-platform analysis: Digital methods to research social media images. Information, Communication & Society: 1(20) doi:10.1080/1369118X.2018.1486871
Highfield T Leaver T (2016) Instagrammatics and digital methods: studying visual social media, from selfies and GIFs to memes and emoji. Communication Research and Practice 2(1): 47-62
Automated vision/machine vision APIs
Mintz, A. & Silva, T. (2019) Interrogating Vision APIs. SMART Data Sprint: Beyond Visible Engagement. Digital Media Winter Institute 2019. Retrieved from: https://smart.inovamedialab.org/smart-2019/project-reports/interrogating-vision-apis/
Bechmann, A. & Bowker G.C. (2019). Unsupervised by any other name: Hidden layers of knowledge production in artificial intelligence on social media. Big Data & Society, Vol. 6, No. 1
Bechmann, A. (2017). Keeping It Real: From Faces and Features to Social Values in Deep Learning Algorithms on Social Media Images. Proceedings of the 50th Hawaii International Conference on System Sciences.
|png||Screenshot 2019-07-22 at 15.22.51.png||manage||429 K||21 Oct 2019 - 13:20||EmilieDeKeulenaar|
|png||pasted image 0.png||manage||1 MB||21 Oct 2019 - 13:23||EmilieDeKeulenaar|
|png||Screenshot 2019-07-23 at 12.25.46.png||manage||60 K||21 Oct 2019 - 13:24||EmilieDeKeulenaar|
|jpg||Crusade_WikiTimeline-02.jpg||manage||70 K||21 Oct 2019 - 13:27||EmilieDeKeulenaar|
|jpg||Atlantis_WikiTimeline_Tavola disegno 1.jpg||manage||86 K||21 Oct 2019 - 13:28||EmilieDeKeulenaar|
|jpg||WW2_WikiTimeline-03.jpg||manage||66 K||21 Oct 2019 - 13:28||EmilieDeKeulenaar|
|jpg||Areagraph epoche.jpg||manage||202 K||21 Oct 2019 - 13:30||EmilieDeKeulenaar|
|png||Screenshot 2019-07-23 at 16.10.01.png||manage||327 K||21 Oct 2019 - 13:31||EmilieDeKeulenaar|
|png||pasted image 2.png||manage||1 MB||21 Oct 2019 - 13:32||EmilieDeKeulenaar|
|png||image-wall-e3b55f6d8e296e95f13bd18fc943dd55.png||manage||934 K||21 Oct 2019 - 13:33||EmilieDeKeulenaar|
|png||unnamed-2.png||manage||12 K||21 Oct 2019 - 13:34||EmilieDeKeulenaar|
|png||unnamed-3.png||manage||11 K||21 Oct 2019 - 13:34||EmilieDeKeulenaar|
|jpg||Areagraph_03_Tavola disegno 1.jpg||manage||302 K||21 Oct 2019 - 13:36||EmilieDeKeulenaar|
|png||Screenshot 2019-07-22 at 16.42.17.png||manage||527 K||21 Oct 2019 - 13:37||EmilieDeKeulenaar|
|png||unnamed-4.png||manage||54 K||21 Oct 2019 - 13:37||EmilieDeKeulenaar|
|png||cluster 2.png||manage||1 MB||21 Oct 2019 - 13:44||EmilieDeKeulenaar|