Emotional Clicktivism: Facebook Reactions across pages and publics

Team Members

Anna Scuttari, Nathan Stolero, Arran Ridley, Livia Teernstra, Denise van de Wetering, Michele Invernizzi, Jasper Schelling, Marloes Geboers

Contents

Summary of Key Findings

In week 1 we explored Reactions usage in the Syrian Revolution Network page (https://www.facebook.com/Syrian.Revolution/) which is used by a closed community. In this follow-up project, we explored Reaction usage across pages and publics, using pages that relate to different topics (Women's March and Black Lives Matter) and using pages that relate to Syria as well, but are serving a more diverse public i.e. not a closed community (news outlets pages, BBC, Al Jazeera English and USA Today) and a public that is geographically or socially more distant to the Syrian conflict. We encountered different intensities in the use of buttons, however, we also saw similar correlations between the use of buttons (similar to the Syrian Revolution Network page). An interesting finding is how Reactions is used differently among news outlets pages: while all three explored news outlets (BBC, Al Jazeera English and USA Today) were similar in their use of the sad button in time-slots of disaster, the difference in the use of Reactions appears in the aftermath of disaster, in this case that is the day after the chemical attack on April 7th, 2017. That day showed a significant use of the love button in the USA Today public. This might indicate a communicative practice sending love towards the US army that acted upon this situation with air strikes on the Assad regime. However, this could also point in the direction of different cultural practices leading to different usages of the Reactions feature in news publics. This should be further studied.

From looking at the top news posts in sad spikes (Sad Reactions) it is interesting to see that all three news outlets use similar generic images (breaking news sign and a dark picture of a missile launch). However, while the Al Jazeera English public reacted with the highest number of sad and angry reactions, BBC news's public reacted with anger, making it the post that got the highest number of angry reponses. It could be because of the cultural closeness of the Al Jazeera public to the Syrian people, compared to the distance of the BBC news public.

An interesting question that could be further studied is about differences in the images used by closed activist communities such as the Syrian Revolution Network community and by non activist communities (maybe somewhat more diverse communities). In the images on the pages of the three studied news outlets (non activist), for example, we hardly see any signholders and it seems that images in news outlets reporting immediate disasters are more often generic whereas communities 'on the conflict ground' seem to be more willing to publish explicit images.

As was already mentioned in the project that preceded this follow up (see Emotional Clicktivism week 1), the use of the Reactions buttons seems to reflect the notion of a structural classification of emotions as constructed by Laros and Steenkamp (2005) where there are different hierarchical levels of emotions. According to this interpretation of emotional structures, there is a superordinate level of emotion (positive vs. negative affect), a basic emotion level (four basic positive and four negative emotional states), and finally a subordinate level (42 classified emotions). The combining of sad and angry buttons and the more distinct usage of positive buttons seem to point in the direction of this hierarchical perspective.

Figure 1 shows the network of icorrelations in the Women's March page

Then, the visual objects found in the content analysis (see network below) and their clustering around certain emotions (sometimes in correlation with two buttons, which results in images falling in between buttons) offer a very interesting starting point to asses how people respond affectively with symbols of disaster. In this research children are representing victimhood and adult men represent rebels/freedom fighters in the Syrian conflict. In both Women's March page and the Syrian Revolution Network page signholders represent activists. In Women's March page women represent similar symbols as man do in the Syrian Revolution Network page, i.e. freedom fighters.

Figure 2 shows the network in which the highlighted part represents images with children, clustering in the correlation zone of Sad and Angry

1. Introduction

In an earlier account of the workings of the like economy, Gerlitz and Helmond (2013) show how Facebook enables particular forms of social engagement and affective responses through its protocol, collapsing the social with the traceable and marketable and filtering it for positive affects. This project aims to build on this in providing insight into how Reactions buttons, implemented in 2016, relate to this filtering of content that evokes certain emotional reactions.

The introduction of Facebook’s Reactions in 2016 is exemplary for how a platform affords for emotional engagement with content. While still refraining from implementing the Dislike button, which cannot be collapsed in the composite Like counter (Gerlitz and Helmond, 2013), Facebook decided to add buttons, including buttons that portray negativity, mainly angry and sad. This project zooms in on the use of these (negative) buttons in the Syrian news space containing public Facebook pages.

In the commercial endeavour of Facebook, Gerlitz and Helmond (2013) argue, it is only the traceable social that matters to Facebook, as the still intensive, non-measurable, non-visible social is of no actual value for the company: it can neither enter data mining processes nor be scaled up further. Following Gerlitz and Helmond, we should understand Reactions as an attempt to further ‘metrify’ a part of what might have been conceived as non-measurable: emotions evoked in users.

Social buttons both pre-structure and enable the possibilities of expressing affective responses to or engaging with web content, while at the same time measuring and aggregating them (Gerlitz and Helmond, 2013). With implementing the Reactions buttons Facebook is adding several layers to the original like button, extending the ways in which content on Facebook can be qualified in an affective way. The question arises whether the collective use of Reactions, can be 'repurposed' for pinpointng certain publics within a page and for understanding the visual content they engage with.

2. Initial Data Sets

Figure 3: Schematic overview procedure

Scraped and used in week 1 (link), continued in week 2 for visual content analysis with this set:

Facebook Syrian Revolution Network https://www.facebook.com/Syrian.Revolution/ with 1.937.954 followers was scraped using Netvizz, scraping a time span beginning just before the implementation of Reactions (16 February 2016 - 27 June 2017). Implementation was February 24 2016.

Scraped and analyzed in week 2

Demarcated timeslots to scrape, using time slots derived from three spikes in distinct Reactions usage in the Syrian Revolution Network, see below:

Figure 4 shows the bumpchart of Reactions usage

Across publics within topic:

BBC: https://www.facebook.com/search/top/?q=bbc%20news (42,242,883 followers)

Al Jazeera English: https://www.facebook.com/aljazeera/ (9,984,178 followers)

USA Today: https://www.facebook.com/usatoday/ (8,295,826 followers)

Across pages (other topic, closed communities)

Women's march: https://www.facebook.com/womensmarchonwash/ (783,112 followers)

Black Lives Matter: https://www.facebook.com/BlackLivesMatter/ (282,367 followers)

For scraping we used the page data module, scraping for full stats and posts by page only

3. Research Questions

How are topics, content and publics related to Reactions use on Facebook?

4. Methodology

This project is constituted of several 'sub projects':

1 Visual content analysis of the Syrian Revolution Network page (community).

In the first week, a small content analysis was performed. However, we wanted to have a better understanding of what exactly was going on in the dataset. Therefore, we decided to perform a bigger and more precise content analysis, in which we coded for both emotions shown in pictures and the presence of objects. Because the whole Syrian revolution dataset would have been too big to perform a content analysis on (N = 6.409), we decided to perform the content analysis on the sample we drew in week one (N = 363). This sample is assumed to be representative for the whole dataset, since the sample mean (M = 2526.16, SD = 2686.06) did not significantly differ (t = 0.675, p = .500) from the full dataset (M = 2631.72, SD = 2905.34).

First, we printed the scatterplot we designed with the use of Gephi in week one. This scatterplot contained all 363 images from the sample, clustered around the reaction buttons. By printing this Gephi plot, we were able to analyse the pictures in the clusters by close reading them. We realized that pictures were clustering based on content, for example in the Love/Sad cluster. It was decided to look at emotions expressed by people in the pictures to get a better understanding of the relation between the used reaction buttons and the presence of emotions. Based on literature, (Laros and Steenkamp, 2005) we decided to code for the following variables and their indicators:

• Anger: angry, frustrated, irritated, hostility, unfulfilled, discontented
• Fear: scared, afraid, panicky, nervous, worried, tense
• Sadness: depressed, sad, miserable, helpless, nostalgia, guilty
• Shame: embarrassed, ashamed, humiliated
• Contentment: contented, fulfilled, peaceful
• Happiness: optimistic, encouraged, hopeful, happy, pleased, joyful, relieved, thrilled, enthusiastic

Besides identifying emotions, the codebook contained variables to indicate objects present in the sample. We decided to also look at those, since the results of the API were not always covering the full content of the photograph correctly. We coded for presence of people in the picture, whom we quantified as male, female and child. If the gender was not clear, the person was coded unknown. If no person was present in the picture, it was coded none. It was also decided to include text as a variable for the cases in which only a logo or a text was present. Images which contained an image of a person were coded as text. Examples of this are infographics, cartoons, memes and posters. We did not look at the gender of the person depicted in the text category because it is was decided that this was not a real reflection of the person. The content of the text is thus manipulated or constructed by a person and therefore is decided not to include the subject as a person in the dataset.

The first 30 images within the sample set were coded by two people, according to the codebook, to determine how they perceived the emotions and therein derive a consensus for the emotional reading and coding of images. It appeared that there were many signholders present in the Gephi plot. Therefore, we decided to include the presence of a signholder as a separate variable to be able to understand which kind of reaction buttons are used when signholders are published. After recoding the 30 pictures again, when including the new variable signholder, the file was split into 2 sections so the two coders could work on it concurrently.

2 Button use and visual exploration in Women's March and Black Lives Matters movement pages (community comparasion).

The Women's March and Black Lives matter were chosen as non-violent protest movements that could be used to compare with a violent conflict, for the purposes of traingulation of the results concerning emotional correlations and those related to symbols of rebellion. Data analysis of women's march was performed according to the same steps used in the project that preceded this follow up (see Emotional Clicktivism week 1) and according to the content analysis performed in this project on the same Facebook page. Following steps were taken:

• Scraping data through Netvizz;

• download of pictures through DownThemAll,

• attribution of content tags through Google Vision API

• descriptives

• visual content analysis

• T-test to compare means in relation to encoded objects

• network analysis with relation to the coded encoded object.

• The number of pictures scraped amounted to 552 units, due to the newness of the Facebook page, therefore no samplig was needed to go through all steps (N=552). Some additional steps were added in relation to the first results of the analysis. In particular, Google Vision API showed a significative amount of "text" tags, which reflected the relatively high amount of pictures with text and claims for activism and protest. To futher explore content and compare visual discourse with the Syrian Revolution page, an exploratory analysis of text content was performed usin tag clouds, but we decided not to explore text in depth because of time constraints and less relation to visual analysis.

Content analysis followed the same scheme as the one on Syrian Revolution page, in which we coded for both emotions shown in pictures and the presence of objects. All images scraped were coded (N = 552). A single difference compared to the Syrian Revolution page was introduced with regard to cartoons in pictures. Indeed, cartoons were not frequent in the Syrian Revolution page and therefore they were not treated as pictures showing individuals (e.g. women, men, children). On the Women's March page, cartoons were often used in visual discourse andd therefore their content was coded.

3 Making the networks (connected to 2,3, and 4)

This protocol is for creating a network of connections between images and the reactions that they generated on Facebook.
Tools used: Netvizz, Google sheet, OpenRefine, Downthemall, Gephi, BBedit, Illustrator

Steps taken:

Phase 01:

1. Open the full data .tab file in Google Sheet;

2. Keep only photos, filter out rows with errors.

Phase 02: Get the images

1. Copy the full_picture column to a new file and save it as image-list.txt;

2. Open dTa! Manager(part of DownThemAll!) and load the .txt file (Import from file -> make sure the format in the options is set to text files -> Open);

3. Select all the images -> Select a folder to place them -> in the mask input box type flatcurl (this will rename the images with part of their url);

4. Download all the images.

Phase 03: Prepare the final dataset

For Gephi to work we would need two separate datasets: a nodes table and an edges table. In order to do that we can use a few Google Sheet functions to make these tables starting from the Netvizz data.

1. Nodes table (using these header names will save a lot of pain later on)

2. Edges table (using these header names will save a lot of pain later)

Phase 04: Create the network

• Import datasets on Gephi;

• Spatialize the graph;

• Export as svg (it’s important to also export the labels).

Phase 05: Make the final svg

1. Open the svg in any text editor that has the Find and Replace function;

2. Remove all the <circle> tag for the images, remove all the <text> tag for the reactions;

3. Replace the <text> tag for the images with an <image> tag, that will have as attributes: the same x and y attributes of the text tag, an xlink:href attribute with the name of the image, a width and height attribute (1% each and you can always change this value later if the images get too small);

4. Save the svg and put it in the same folder where all the images are;

5. Open the svg with Illustrator and make the final touches.

5. Findings

Community - The Syrian Revolution page

Before diving into the statistical analysis we looked at the descriptives of the reaction buttons. It appreared that the Sad is the most frequent used button to express emotion on Facebook, followed by Love and Angry. Wow did not seem to play a big role in the dataset based on its low frequency.

Table 1: Descriptions sample (N = 363)

	Max	Mean	Std. dev	Count
Love	308	19.94	35.15	7237
Wow	31	1.23	2.76	445
Haha	193	5.90	21.18	2142
Sad	1278	56.63	141.1	20557
Angry	339	19.28	43.23	6998

Correlation

First, we explored the correlation between the reactions given in the sample we drew. This sample was assumed to be representative of the whole dataset. We drew this conclusion based on the results of an independent sample t-test, in which we compared the sample mean with the mean of the dataset. There was no significant difference between the sample and the dataset, and thus can we assume that the sample is representative.

The correlation matrix shows some interesting results. First of all, we found a strong and significant correlation between sad and angry (r = .703, p < .001), which is the same result we found when testing correlations for the whole dataset.

Table 2: Correlation reaction buttons in sample

	Love	Wow	Haha	Sad	Angry
Love	1
Wow	.349, p = .000	1
Haha	.179, p = .001	.350, p = .000	1
Sad	.059, p = .261	.044, p = .403	-.088, p = .095	1
Angry	-.159, p = .003	.017, p = .752	-.083, p = .114	.703, p = .000	1

Next, we correlated the emotions we coded in the sample. It appeared that fear and sadness, shame and anger, and shame and happiness were significant related in a positive direction. However, the strength of the relations were weak. Sadness and happiness were significant related but in a negative direction. The relation is again not very strong.

Table 3: Correlation coded emotions

	Anger	Fear	Sadness	Shame	Contentment	Happiness
Anger	1
Fear	.007, p = .896	1
Sadness	.073, p = .167	.275, p = .000	1
Shame	.108, p = .040	-.013, p = .812	.078, p = .138	1
Contentment	.013, p = .799	-.077, p = .146	-.097, p = .065	.065, p = .217	1
Happiness	. 044, p = .406	-.052, p = .328	-.123, p = .020	.110, p = .036	.097, p = .066	1

Independent sample t-test

First, we are interested in the different reactions of the presence of males in the pictures among the different reaction buttons. Interesting to see is the difference in group means when looking at the expressed love for pictures with or without males. It appeared that the presence of males get significant (t (361) = -3.026, p = .003) more love reactions (M = 24.92, SD = 31.45) compared to pictures without males (M = 13.82, SD = 31.45). Furthermore, the presence of a child (M = 111.70, SD = 212.69) evokes significantly (t(361) = -4.138, p < .001) more Sad reactions compare to images on which no child is present (M = 40.31, SD = 106.61). The same appeared for the use of the Angry button when a child was present: images containing children received significantly (t(361) = -3.450, p = .001) more Angry reactions (M = 33.45, SD = 66.84) compared to images without children (M = 15.08, SD = 32.16).

Interesting to see is that the presence of signholders influenced the amount of received Wow, Haha, Sad and Angry reactions in a significant but negative direction. For an extensive description of these results, please consult table 4.

Table 4: Independent sample t-test results objects and reactions

		Love	Wow	Haha	Sad	Angry
Male		T = -3.026, p = .003	T = .732, p = .465	T = -.956, p = .340	T = -1.310, p = .191	T = 1.663, p = .097
		M = 13.82, SD = 31.45	M = 1.34, SD = 3.09	M = 4.72, SD = 14.06	M = 45.90, SD = 101.46	M = 23.45, SD = 49.00
	1	M = 24.92, SD = 37.24	M = 1.13, SD = 2.47	M = 6.86, SD = 25.55	M = 65.38, SD = 166.32	M = 15.88, SD = 37.67
Female		T = -.254, p = .800	T = 1.282, p = .201	T = 1.460, p = .145	T = -.609, p = .543	T = -.211, p = .833
		M = 19.78, SD = 35.96	M = 1.29, SD = 2.89	M = 6.43, SD = 22.21	M = 55.16, SD = 143.37	M = 19.12, SD = 43.87
	1	M = 21.37, SD = 26.77	M = .66, SD = .94	M = .94, SD = 1.41	M = 70.46, SD = 118.55	M = 20.74, SD = 37.18
Child		T = 1.137, p = .256	T = 1.712, p = .088	T = 1.712, p = .088	T = -4.138, p = .000	T = -3.450, p = .001
		M = 21.08, SD = 38.03	M = 1.36, SD = 3.08	M = 1.36, SD = 3.08	M = 40.31, SD = 106.61	M = 15.08, SD = 32.16
	1	M = 16.08, SD = 22.67	M = .77, SD = 1.10	M = .77, SD = 1.10	M = 111.70, SD = 212.69	M = 33.45, SD = 66.84
Text		T = 2.744, p = .006	T = 1.512, p = .132	T = 1.648, p = .100	T = 2.517, p = .012	T = .970, p = .333
		M = 22.52, SD = 37.80	M = 1.34, SD = 3.01	M = 6.84, SD = 23.54	M = 66.27, SD = 156.02	M = 20.46, SD = 46.17
	1	M = 10.11, SD = 19.96	M = .8, SD = 1.47	M = 2.32, SD = 6.00	M = 20.49, SD = 39.72	M = 15.01, SD = 29.57
Signholder		T = 1.34, p = .182	T = 2.228, p = .027	T = 2.339, p = .020	T = 2.520, p = .012	T = 3.293, p = .001
		M = 21.15, SD = 37.46	M = 1.39, SD = 3.02	M = 7.16, SD = 23.39	M = 65.80, SD = 154.14	M = 22.92, SD = 47.03
	1	M = 14.86, SD = 22.79	M = .57, SD = .93	M = .57, SD = 1.14	M = 18.52, SD = 45.57	M = 4.10, SD = 12.37

Secondary and third parties publics - The three news outlets

In the analysis of the news outlets parties, representing the secondary and third parties which are more distant to the Syria conflict, than the public of the Syrian Revolution page, we scraped only the four spikes that were found in the first week's analysis.

"Sad Beats" and "Love Explosion" Spikes

In the Syrian revolution page, we found that a week in August and three days in October, had a spike of "Love-Explosion", when the most used reaction button was "love". April (2016), December (2016) and April (2017), the day of the deadly chemical attack, were all "sad-beats" spikes.

Charts i-iii shows how those spikes were represented in the use of reaction buttons in the three news outlets' pages. Chart i shows how the spikes were distributed in the Al Jazeera English page.

Figure 5: Chart i. Bump chart for Al Jazeera English spikes.

The chart shows that while the Syrian Revolution page showed two love-explosion spikes, in the Al Jazeera English page all the spikes were sad. The use of love reactions button was very minimal. The greatest spike of sad reactions was on the day of the chemical attack, with a small angry spike afterwards. A similar pattern was shown in the BBC News page. Chart ii shows the distribution of the use of reaction button on that page.

   

Figure 6:

 

Chart ii. Bump chart for BBC News spikes.

The chart shows that BBC also had only sad spikes, and specifically on that dates that the Syrian Revolution page had love-explosion spikes. An interesting change in comparison to the Al Jazeera page, is that while on the day of the chemical attack there was a big sad spike, after that spike, there was a very big spike of use of angry reaction buttons. This spike was even bigger than the sad spike. And when looking at the chart of the USA Today page (Chart iii), a difference can be found also there.

 

Figure 7:

   

Chart iii. Bump chart for USA Today spikes.

Chart iii shows that a very big love-explosion spike followed the sad spike of the chemical attack. So while all news outlets were similar in their sad spikes around the time-slots that we examined, the difference is in the type of spike that followed the day of the chemical attack on April 7th, 2017. Figure A shows the posts that were shared on the three pages on that date. These were also the posts and images that received the highest number of sad and angry, angry and love reactions in Al Jazeera, BBC News and USA Today, accordingly.

_ Figure 8: The stories which were shared on April 7th by the three outlets._

This shows that the three posts actually dealt with the same subject - the American response to the chemical attack. They even used similar generic images (breaking news sign and a dark picture of a missile launch). However, while the Al Jazeera English public reacted with the highest number of sad and angry reactions, BBC news's public reacted with anger, making it the post that got the highest number of angry reponses. It could be because of the cultural closeness of the Al Jazeera public to the Syrian people, compared to the distance of the BBC news public. Compared to those publics, the USA Today american public reacted to this post with love - making it the post that had the highest number of love reactions. This could signal empathy to the situation of the Syrian people or, another possible explanation, love towards the US army that acted upon this situation.

Looking at the correlations in button usage

While examining the statistics of the news pages, a few similarities and a few differences can be noticed, compared to the Syrian Revolution page. Table 1 shows the descriptive statistics of the use of reaction buttons accross all the pages.

Table 1: Shares and reactions across news pages

	Maximum	Mean	Std. Deviation	Total
Share_Count	209718	2462.64	17536.676
Love	684	105.44	140.434	2040
Wow	3553	106.66	354.280	2517
Haha	583	33.53	62.154	1566
Sad	66413	1229.74	5742.020	110990
Angry	32559	645.56	2828.217	56195
Valid N (listwise)	144

Table 1 shows that negative reactions were very frequent among the news outlets, with sad (M=1229.74; SD=5742.020) being the top reaction and angry (M=645.56; SD=2828.217) afterwards. Positive emotions were far behind. However, tables 2 - 4 shows these statistics on each page separately.

Table 2: Al Jazeera English Reactions

	Maximum	Mean	Std. Deviation	Total
Share_Count	209718	3806.60	24483.370
Love	479	34.51	69.498	1180
Wow	687	37.47	110.764	2086
Haha	583	22.60	70.288	1150
Sad	66413	1593.63	7776.848	82554
Angry	32559	807.73	3805.112	47411
Valid N (listwise)	73

Table 3: BBC Reactions

	Maximum	Mean	Std. Deviation	Total
Share_Count	16721	1603.68	3262.358
Love	684	105.21	138.977	3998
Wow	3553	208.13	606.188	7909
Haha	264	48.21	62.655	1832
Sad	14057	1352.24	2927.905	51385
Angry	7581	819.13	1532.025	31127
Valid N (listwise)	38

Table 4: USA Today reactions

	Minimum	Maximum	Mean	Std. Deviation	Total
Share_Count	3	11883	478.73	2050.816
Love	2	622	262.61	134.089	8666
Wow	2	1700	142.88	289.666	4715
Haha	0	174	40.82	32.025	1347
Sad	14	1300	283.70	323.608	9362
Angry	10	1100	86.94	189.264	2869
Valid N (listwise)	33

These tables show that while Al Jazeera and BBC had similar descriptive statistics, in the case of USA Today the top reaction was sad (M=283.7; SD=323.608), while the second frequent reaction, and not very distance from sad reaction, was love (M=262.61; SD=134.089). Angry (M=86.94; SD=189.264) was only fourth, after Wow (M=142.88; SD=289.666) which was third.

The differences could also be found in the correlations tables. Tables 5 - 7 show the correlation between the different reaction buttons in the three pages.

Table 5: Al Jazeera English reaction correlations

	Love	Wow	Haha	Sad	Angry
Love	1
Wow	.526**	1
Haha	.264*	.280*	1
Sad	.384**	.675**	0.162	1
Angry	.394**	.716**	0.173	.994**	1
**. Correlation is significant at the 0.01 level (2-tailed).

Table 5 shows similar correlations to those which were found in the Syrian Revolution page. Two different findings that can be found are the very high correlation (higher than in the Syrian revolution page) between sad and angry, making this correlation almost perfect (r=.994, p<.01), and a new correlation beween the use of Wow and angry reaction buttons (r=.716, p>.01).

Table 6: BBC reaction correlations

	Love	Wow	Haha	Sad	Angry
Love	1
Wow	.434**	1
Haha	0.065	0.295	1
Sad	-0.082	0.069	0.039	1
Angry	0.179	.572**	0.183	.654**	1
**. Correlation is significant at the 0.01 level (2-tailed).

Table 6 shows that the correlation between Wow and angry that was found in the arab news community (Al Jazeera) page appears also in the western broadsheet outlet (r=.572, p<.01). Sad and angry were also correlated, but in a correlation which is similar to the one which was found in the Syrian revolution page (r=.654, p<.01).

Table 7: USA Today reaction correlations

	Love	Wow	Haha	Sad	Angry
Love	1
Wow	.632**	1
Haha	.636**	0.294	1
Sad	-0.178	0.016	-0.309	1
Angry	.434*	.922**	0.157	0.291	1

In the western tabloid, USA Today, suprisingly there was no correlation of angry and sad. Moreover, sad reactions were not correlated with any other reaction. However, the angry and Wow reactions were highly correlated, resulting in an almost perfect correlation (r=.922, p<.01). Love was correlated with all the other reactions, Wow (r=.632, p<.01), Haha (r=.636, p<.01) and Angry (r=.434, p<.01), but not with sad.

Content analysis: The existance of different objects

Table 8 shows the frequency of different types of content in the manual content analysis.

Table 8: Content analysis descriptives

	Count
Male	95
Female	29
Child	43
None	34
Text	6
Signholder	1

Similarly to the Syrian revolution page, there was a high representation of male characters (95) and children (25), along with female characters (29). However, while in the Syrian Revolution page there were lots of sign holders, in the news outlets they appeared only once. Table 9 shows an independent samples t-test for the differences between posts that included certain objects and those we did not.

Table 9: T-tests for the presence of certain image content and reactions

		Love	Wow	Haha	Sad	Angry
Male		T = 4.145, p < .001	T = 2.756, p = .007
	No	M = 170.21, SD = 184.94	M = 222.28, SD = 591.14
	Yes	M = 71.52, SD = 99.05	M = 50.92, SD = 100.05
Female		Not significant
	No
	Yes
Child					T = -2.407, p = .017
	No				M = 482.67, SD = 1554.53
	Yes				M = 2982.65, SD = 10106.13
Text			T = -5.411, p < .001
	No		M = 76.59, SD = 199.79
	Yes		M = 811.33, SD = 1374.63
Signholder		Not significant
	No
	Yes

Table 9 shows that posts that included male objects got less love and wow reactions compared to those to did not have male objects (t=-.331, p<.01; t=-.227, p<.01), while posts who included no human characteristics at all received more love and wow reactions compared to those who included human characteristics (t=.282, p<.01; t=.294, p<.01). Posts that included text received more Wow reactions (t=.416, p<.01) and posts which included children received more sad reactions (t=.199, p<.01).

Network Analysis

Figure 8 shows how the different images clustered around the different reaction buttons. The size of the reaction face represents its centrality in the network.

Figure 9. Network Analysis of pictures' clusters in Al Jazeera English, BBC News and USA Today.

The network shows that most images clustered between sad and angry, showing their large correlation. The purple circle in the figure shows unique pictures in Al Jazeera English, while the red circle shows unique pictures of USA Today. Despite the fact that those picture are on the opposite site of the network - a close examination of them show that they were similar in content.

Another interesting finding can be seen in Figure C, showing the network of each of the news outlets separately, showing that while Al Jazeera English and BBC News had many shared spaces, USA Today's pictures tended to cluster around unique reactions.

Figure 10. Network Analysis of pictures' clusters in Al Jazeera English, BBC News and USA Today - separated.

Different Spaces

The Women's March Facebook group

Since it's conception in December 2016, there has been a lot of engagement with the Women's March group through reactions and likes, but not that many comments. The highest average reaction (aside from 'like') was love. In fact, love reactions dominated this page, with the secondary most frequent reaction being sad, although love reactions occured 12x more on average than sad ones. Table 9 shows the full descriptive statistics for the page engagement, whereas table 10 shows the descriptive statistics for the reactions.

Table 10: Women's March Descriptive statistics

	Maximum	Mean	Std. Deviation
likes	91776	4181.57	8155.965
comments	5532	228.34	506.633
reactions	121004	5053.09	10072.007
shares	35859	1002.30	2560.885
engagement	160521	6283.72	12569.034

Table 11: Women's March reactions

	Maximum	Mean	Std. Deviation
Like	90727	4126.14	8055.014
Love	27416	679.57	1774.395
Wow	1239	20.24	102.806
Haha	2973	11.46	129.627
Sad	9496	70.44	496.503
Angry	16839	51.44	727.820
Thankful	930	2.24	39.942

Similar to the Syrian datasets, likes way outnumber all the other reactions. Hence once more we consider it to be too ambigious to be an emotional reaction. Another interesting point was is the enormous deviation of angry reactions from the mean, indicating that there is an outlying post that caused a lot of angry reactions, although these posts do not seem to be the norm.

Figure 10: Correlations between emotions over time

Three combinations of emotions were found to be significant, which can be seen in Figure D, and was also confirmed by Pearson's R correlation tests. Similar to the Syrian Revolution page, there is a very strong positive correlation between sad and angry (r = .815, p =.000). There is also a very strong positive correlation between love and wow (r = .7, p = .000). Lastly, there is a weak positive correlation between love and haha (r = .117, p = .006). For instance, the correlation between love and wow is apparent at the end of january, and the correlation between sadness and anger occurs in the middle of January. Therefore, similar combinations of emotions can be observed across issues, publics and cultures. Lastly, there is a big spike of the thankful reaction which can be observed in mid-may, although this button is not always available, which could explain why it has no use in other time frames.

Figure 12: Image tags from the Women's March

Unlike the Syrian Revolution facebook group, the image tags for text and font were very relevant and accurate for the Women's March group (see Fig. E). Tag frequencies were counted using the tool on writewords.org.uk. Most images in the group were just text promoting marches and ideas of solidary amongst women. This shows the different way that people use images on a platform like facebook to spread short messages. Images grab more attention than the text description under it, especially if the message is short and wants to be disseminated easily. Sharing an image is easier than copy-pasting a status. Therefore, the tags of font, text and poster were very accurate to describe the images on the Women's March facebook group.

However, examining the emotional reactions ot the tags did not make sense, because of the different use of images on this page. Hence, a visual analysis not really a useful method of analysis. Therefore, we found it would tell us more if we examined the words present within these images and their association with the dominant emotional reaction. So we used the Optical Character Recognition (OCR) feature of Google’s Vision API to examine the text within the images, which could perhaps tell us more of the motivations of reactions to certain images.

Figure 13: Tag cloud for words within images receiving sad reactions

As an example, Figure F shows some of the most prominent words that recieved 'sad' reactions, extracted from images with OCR. At face value, this is already far more meaningful to exaplain the abundance of sad reactions than the tag 'text' or 'font'. Consequently, we recommend for groups which use images for spreading textual messages to extract the text from those images in order to draw meaningful conclusions about the content. Word frequencies as well as N-grams are useful for examining the text, however for the sake of brevity (and that this summer school has a visual theme) we have left the further analysis out.

Overall, we discovered that the Women's March page was used mainly as a call to action, and the images are simply reappropriated as ways to spread textual messages.

Content analysis Women's March

Coding was executed on the visual data in a similar way to the coding of the images in the Syrian Revolution Network. As the dataset was not toot big, we did not take a sample. We coded 553 images for Anger, Fear, Sadness, Shame, Contentment, Happiness, Male, Female, Child, Unknown, None, Text, Sign Holder and Other stuff (the latter was basically used to code for cartoon).

_Table 12 shows the count for the content analysis of Women's March (we left out emotions, as it turned out to be too unstable)._

	Count
Male	85
Female	255
Child	57
Unknown	58
None	247
Text	392
Sign Holder	87
Cartoons	123

Results of the content analysis were used to test if some objects or subjects were more associated with particular emotional reactions. We cross-checked this aspect using both t-test and network analysis. Both of them gave no evidence of particular association of sentiments with subjects. This was due to the relatively dominant presence of "love" reactions, which were able to attract the majority of relevant subjects: women, children and signholders.

The Black Lives Matter Facebook group

The Black Lives matter was scraped for the same time frame as the Women's march (December 2016 - July 2017), resulting in N = 134 of image only posts. Once again, there were more reactions (M = 1412, SD = 2757) than comments (M = 95, SD = 149) or shares (M = 341, SD = 1110). Therefore, people comment less than they react, and also do not share much. Also similarly to the Women's March page, on average most post recieved the love reaction (M = 169, SD = 383). The second most common reaction on average was sad (M = 66, SD = 380).

Examinig the correlations between emotions shows that results from the Syrian Revolution and Women's March pages are once again replicated. There is almost a perfect correlation between sad and angry (r = 958, p = .000), a moderate correlation between love and haha (r = .53, p = .000) and a weak to moderate correlation between love and wow (r = .405, p = .000), and love and thankful (r = .408, p = .000). Therefore we repeatedly saw a clear emergence of positive/negative emotions, but people express the dominant emotion different.y Again, we see the rise of more complex emotions that can be categorized on a positive - negative spectrum.

6. Discussion

On the methods

Mining Facebook reactions can show interesting emergent, complex emotions from groups. The emotions can be clearly polarized, where different expressions of emotional reactions used together (such as sad and angry) can show a negative overall mood, whereas happy and love can show an overall positive mood. The thankful button is not too useful to examine, because it is not always available and seems to be used infrequently. It might be interesting if you know the exact time span that Facebook makes it available, otherwise you don’t know if people aren’t replying thankful because they don’t feel that way, or if the button simply wasn’t available. Across the board, the "like" button is too ambiguous and does not necessarily represent an emotional reaction (although, it can). So it’s quite difficult to include it when examining emotional reactions, as it lends itself to misguided conclusions. Finally, many online tools are readily available and free to use in order to conduct this type of analysis.

On the emotional correlations

Across all pages the wow button was used, whether to express it together with something positive (haha) or something negative (anger), except in the case of the Syrian Revolution facebook group. Hence it is clear that the closer the group is to the violent event (in this case, the Syrian war), there is no expression of shock or surprise. This could be due to the constant exposure and geographic closeness of the group to the events, which therefore do not result in any surprise. That said, only images were examined, and hence there might be differences when taking into account textual messages also.

Interestingly when examining the Syrian news space, there was a very strong positive correlation between angry and wow that emerged. This correlation was especially strong in the USA Today audience, which interestingly was the only audience that did not experience a plurality of anger and sadness. Therefore, with this audience we see a completely different collective expression (or lack thereof). This may be due to either the specific audience or overall culture.

Both the Women’s March and Black lives matter pages showed the same correlations between emotions as the Syrian Revolution page and Syrian news space; Angry and sad (frustration), love and wow (enthrallment), love and haha (general happiness). There was also an additional correlation in the Black Lives Matter page between thankful and love, although the use of the thankful button is not always possible, and therefore its interpretation should be taken cautiously. In the future, one should consider regrouping the variables into groups of correlated emotions. This will increase accuracy of conclusions when examining reactions to images, since the pluarlity of emotions when using the reaction button is an important consideration.

Future questions relate to questions about:

Whether we can pinpoint certain communities in pages that serve a more or less diverse audience.

Does the use of Reactions within a cluster of 'objects' that appears, point to a certain group. For example within the child cluster, does the use of the sad button connect to a different public than the use of the angy button?

7. Conclusions

When comparing the overall news reactions to the Syrian revolution page, there are some similar emotional correlations to be found, including; Angry and sad, Love and wow as well as Love and haha. The key difference is that the news sites showed a strong positive correlation between wow and angry, as well as sad and wow. Compared to the Syrian revolution page, there was not a lot of use of the 'wow' button, hinting at the lack of surprise of the events in Syria in this community. Moreover, the lack of the use of the haha reaction in the news sites indicate that there was nothing funny being reported about Syria.

In addition, although there were some similarities between the news communities, it is important to examine the nuances based on the different news outlets. The lack of correlation between sad and angry, in favour of a strong positive correlation between wow and angry shows that there is a clear difference the way that news is presented and recieved amongst audiences.

Overall, there is a clear demarcation between positive and negative emotions when examining images across topics, pages and communities.

8. References

Berger, J., Milkman, K. (2012). What makes online content viral? Journal of Marketing Research, 192-205

Burgess, J., Green, J., (2009). YouTube and participatory culture. Cambridge: Polity Press

Couldry, N., Hepp, A. (2017). The mediatized construction of Reality. Cambridge: Polity Press

Gerlitz, C., Helmond, A. (2013). The Like Economy: social buttons and the data-intensive web. New Media and Society

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).

John, N. (2017). The age of Sharing. Cambridge: Polity Press

Laros, F. J., & Steenkamp, J. B. E. (2005). Emotions in consumer behavior: a hierarchical approach. Journal of business Research, 58(10), 1437-1445.

Latour, B. (2007). Reassembling the Social: An Introduction to Actor-Network-Theory. New York: Oxford University Press

Osofsky, J. (2016). Information about trending topics. Facebook Newsroom. https://newsroom.

fb.com/news/2016/05/information-about-trending-topics/

Rogers, R. (2013). Digital methods. Cambridge, Massachusetts; London: The MIT Press.

Stieglitz,S., Dang-Xuan, L. (2013). Emotions and Information Diffusion in Social Media - Sentiment of Microblogs

and Sharing Behavior. Journal of Management Information Systems, 29:4, pp. 217–247.

Tian, Y., Galery, T., Dulcinati, G., Molimpakis, E., Sun, C. (2017) Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media, Valencia, Spain, April 3-7, 2017, https://www.dropbox.com/s/7skeeu1hredct4m/7_Paper.pdf?dl=0