class: center, middle, inverse, title-slide # Decoding the Alt-Right ## Analyzing Online Toxicity on Social Media <br> <br> <font size='6'>Simon Roth<a href="#fn1" class="footnote-ref" id="fnref1">¹</a> & Fabio Votta<a href="#fn2" class="footnote-ref" id="fnref2">²</a> <br> 2018-20-08</font> ### ¹ E-Mail: <a href="mailto:nomis.roth@gmx.com">nomis.roth@gmx.com</a> <br> ² E-Mail: <a href="mailto:fabio.votta@gmail.com">fabio.votta@gmail.com</a> ### <br> Slides: <a href='decoding-the-altright.netlify.com'>https://decoding-the-altright.netlify.com/ </a> --- class: center, middle ## What is the Alt-Right? <style> .onehundredtwenty { font-size: 120%; } <style> .ninety { font-size: 90%; } .eightyfive { font-size: 85%; } .eighty { font-size: 80%; } .seventyfive { font-size: 75%; } .seventy { font-size: 70%; } </style> --- class: center, middle .onehundredtwenty[ **Non-ironic Nazism masquerading as ironic Nazism.** ~Andrew Anglin, Founder of the *The Daily Stormer* ] --- ## What is the Alt-Right? A group of loosely aligned individuals and subgroups that predominately consist of white, young & male frequent internet user. Although adherence to these beliefs varies, these are some common themes within the Alt-Right: -- + Antipathy towards both liberals (*“social justice warriors”*) and conservatives (*“cuckservatives”*) -- + Fighting a *"culture war"* against perceived liberal consensus within the establishment -- + Belief that freedom of speech is being threatened by *"political correctness"* -- + Strong animosity towards feminism, multiculturalism and immigration (*"cultural marxism"*) -- + Belief in biological differences in between races and genders that justifies separate treatment -- + Ambiguous and troll-like behavior as to ensure *plausible deniability* --- <!-- ## What is the Alt-Right? --> <!-- The Alt-Right consists of... --> <!-- -- --> <!-- > an amalgam of conspiracy theorists, techno-libertarians, white nationalists, Men’s Rights advocates, trolls, anti-feminists, anti-immigration activists, and bored young people [...]. [S]ome participants in this ecosystem are organized by their beliefs (like *"Men’s Rights Activists"*), while others are organized by particular media or platforms (such as individual blogs and podcasts). Thus, the lines between these groups are always blurry and uneven (Marwick and Lewis 2017). --> <!-- --- --> ## What is the Alt-Right? <img src="images/flower1.png"> --- ## What is the Alt-Right? <img src="images/flower2.png"> --- ## What is the Alt-Right? <img src="images/flower3.png"> --- ## What is the Alt-Right? <img src="images/flower4.png"> --- ## What is the Alt-Right? <img src="images/flower5.png"> --- ## What is the Alt-Right? <img src="images/flower6.png"> --- ## What is the Alt-Right? <img src="images/flower7.png"> --- class: inverse, center, middle ## Data Collection --- ## What data did we collect? -- + Text data from three Platforms: Facebook, Twitter and YouTube <center> <img src="images/socialmedia.png" width="80%"> </center> -- + Comments from 36 Alt-Right connected profiles/YouTube channels -- + Comments from 6 Media Outlets (CNN, MSNBC, ABC, Fox News, Washington Post, New York Times) <center> <img src="images/media.png" width="80%"> </center> -- + Comments from the entire year 2017 -- + In total over 30 million comments/tweets --- ## Who did we scrape? Alt-Right <img src="index_files/figure-html/tally1-1.png" style="display: block; margin: auto;" /> --- ## Who did we scrape? - Media <img src="index_files/figure-html/tally2-1.png" style="display: block; margin: auto;" /> --- class: inverse, center, middle ## Language Modeling: Toxicity --- ### Language Predictions ** Data** * Cooperation between Wikipedia and Jigsaw * Five thousand crowd-workers were involved * Approx. 160k comments annotated with **toxicity** scores * The dataset is documented [here](https://meta.wikimedia.org/wiki/Research:Detox/Data_Release). <center> <img src = "../images/jigsaw.png" width = 200> </center> **Machine Learning** * Balanced data to 30.000 * Text Normalization * Different Keras Models (Accuracy ~90%) * Prediction scores (*toxicity*) for each comment in the entire corpus --- class: inverse, center, middle ## User Activity Types ### How do Media and Alt-Right interact online? --- ## Differentiating between User Activity Types <!-- --> --- ### Shared Active Users: Alt-Right Profiles and Media Outlets <center><img src="../images/audience_graph.png" width = 940></center> --- ### Multilevel Logit Regression:  --- background-image: url("../images/toxic_graph.png") background-position: 50% 50% background-size: 110% class: center, top ## The Flow of Toxicity --- ## Average Toxicity by Day <br> <img src="index_files/figure-html/unnamed-chunk-1-1.png" style="display: block; margin: auto;" /> --- class: inverse, center, middle ## Content Analysis --- ### Multilevel Logit Regression:  --- ### Most Frequent Bigram Network - Muslims <center> <img src="../images/muslim_bigram_network.png" width = "95%"> </center> --- class: inverse, center, middle ## Conclusions --- ## Conclusions * Alt-Right pages exhibit much greater levels of toxicity than Media equivalents -- + Alt-Right toxicity spikes on days of major Terror attacks and events like "March Against Sharia" or Charlottesville -- + There is a good share of users that comment both on Alt-Right and Media pages -- + *Alex Jones* and *Paul Joseph Watson* share a considerable user base with each other as well as with *Fox News* and *CNN* -- + Highest levels of toxicity on Media pages are from polarized Alt-Right users -- + On Alt-Right pages, polarized Media users are almost as toxic as Alt-Right users -- + Polarized Alt-Right users use much more toxic language when talking about Immigrants, Black people and Muslims --- class: inverse, center, middle ## Thanks for Listening! Slides: <a href='decoding-the-altright.netlify.com'>https://decoding-the-altright.netlify.com/ </a> --- ## References Hawley, G. (2017). Making Sense of the Alt-right. Columbia University Press. Marwick, A., & Lewis, R. (2017). Media manipulation and disinformation online. New York: Data & Society Research Institute. --- class: inverse, center, middle ## Appendix --- .eightyfive[ **Milo Yiannopolous' characterization of the Alt-Right:** `\(^1\)` > *The alt-right is a movement born out of the youthful, subversive, **underground edges of the internet**. 4chan and 8chan are hubs of alt-right activity. For years, members of these forums – political and non-political – have delighted in **attention-grabbing, juvenile pranks**. Long before the alt-right, 4channers turned **trolling** the national media into an in-house sport.* **Marwick and Lewis 2017** > *Term coined by Richard Spencer to describe a version of **white nationalism** that positions itself as comprised of **younger, wealthier, and better-educated individuals** than traditional white supremacist groups like the Klu Klux Klan. Deeply entwined with social media and staples of **internet culture, like memes**. Believes that **“political correctness” threatens individual liberty**. Example: Jared Taylor of the American Renaissance, Mike Cernovich of Danger and Play.* **Hawley 2017** > *Using the loosest definition, we could say the Alt-Right includes anyone with right-wing sensibilities that **rejects the mainstream conservative movement**. But there are certain common, perhaps universal attitudes within the Alt-Right. The Alt-Right is **fundamentally concerned with race**. At its core, the Alt-Right is a white-nationalist movement, even if many (perhaps most) of the people who identify with the Alt-Right do not care for that term. The most energetic and significant figures of the movement want to see the creation of a **white ethnostate** in North America.* ] .seventy[ `\(^1\)`Allum Bokhari and Milo Yiannopoulos, “An Establishment Conservative’s Guide to the Alt-Right,” Breitbart, March 29, 2016, http://www.breitbart.com/tech/2016/03/29/an-establishment-conservatives-guide-to-the-alt-right/. ] --- ### Appendix: Total Number of Posts and Comments: Alt-Right <br> <img src="index_files/figure-html/unnamed-chunk-2-1.png" style="display: block; margin: auto;" /> --- ### Appendix: Text Predictions <!-- * different keras models: very complex, millions of parameters and several deep non-linear layers. --> <!-- * global metrics help to understand system performance -->  --- ### Appendix: Explainability I <!-- * Trusting your algorithemn is crusical for adoption, documentation, regulatory oversight and human acceptance. --> <!-- * How to? Zoom into individual instances of your data or your predictions and derive local explanations. --> <!-- * open black box algorithms --> <!-- * assessing local interpretation of predictions. --> <!-- * our examples --> <!-- * compare classifier --> <p> <span style="background-color:#150E3780; color: #1A1A1A">Communist </span> <span style="background-color:#150E3780; color: #1A1A1A">control </span> <span style="background-color:#150E3780; color: #1A1A1A">freak </span> <span style="background-color:#51127C80; color: #141414">engineered </span> <span style="background-color:#51127C80; color: #141414">scum </span> <span style="background-color:white; color: #000000">situation </span> <span style="background-color:white; color: #000000">this </span> <span style="background-color:white; color: #000000">attack </span> <span style="background-color:white; color: #000000">proves </span> <span style="background-color:white; color: #000000">it's </span> <span style="background-color:white; color: #000000">all </span> <span style="background-color:white; color: #000000">organized </span> <span style="background-color:white; color: #000000">ahead </span> <span style="background-color:white; color: #000000">of </span> <span style="background-color:white; color: #000000">time </span> </p><p> <span style="background-color:#150E3780; color: #1A1A1A">Shoot </span> <span style="background-color:#150E3780; color: #1A1A1A">dat </span> <span style="background-color:#150E3780; color: #1A1A1A">nigga </span> <span style="background-color:#150E3780; color: #1A1A1A">amp </span> <span style="background-color:#150E3780; color: #1A1A1A">his </span> <span style="background-color:white; color: #000000">shorty </span> <span style="background-color:#150E3780; color: #1A1A1A">bitch </span> <span style="background-color:white; color: #000000">shorty </span> <span style="background-color:#150E3780; color: #1A1A1A">bitch </span> </p><p> <span style="background-color:white; color: #000000">VICESports </span> <span style="background-color:white; color: #000000">it </span> <span style="background-color:white; color: #000000">past </span> <span style="background-color:white; color: #000000">your </span> <span style="background-color:white; color: #000000">bed </span> <span style="background-color:; color: ">time </span> <span style="background-color:; color: ">No </span> <span style="background-color:; color: ">wonder </span> <span style="background-color:#150E3780; color: #1A1A1A">no </span> <span style="background-color:#150E3780; color: #1A1A1A">one </span> <span style="background-color:#150E3780; color: #1A1A1A">reads </span> <span style="background-color:#150E3780; color: #1A1A1A">that </span> <span style="background-color:#150E3780; color: #1A1A1A">garbage </span> <span style="background-color:#150E3780; color: #1A1A1A">on </span> <span style="background-color:#150E3780; color: #1A1A1A">vice </span> <span style="background-color:#150E3780; color: #1A1A1A">because </span> <span style="background-color:#150E3780; color: #1A1A1A">faggots </span> <span style="background-color:#150E3780; color: #1A1A1A">like </span> <span style="background-color:#150E3780; color: #1A1A1A">you </span> <span style="background-color:white; color: #000000">write </span> <span style="background-color:#150E3780; color: #1A1A1A">shit </span> <span style="background-color:#150E3780; color: #1A1A1A">fuck </span> <span style="background-color:#CB3E7280; color: #080808">off </span> </p><p> <span style="background-color:#150E3780; color: #1A1A1A">Get </span> <span style="background-color:#150E3780; color: #1A1A1A">the </span> <span style="background-color:#150E3780; color: #1A1A1A">fuck </span> <span style="background-color:white; color: #000000">outta </span> <span style="background-color:; color: ">here </span> <span style="background-color:white; color: #000000">talking </span> <span style="background-color:white; color: #000000">bout </span> <span style="background-color:white; color: #000000">the </span> <span style="background-color:white; color: #000000">response </span> <span style="background-color:#150E3780; color: #1A1A1A">will </span> <span style="background-color:#150E3780; color: #1A1A1A">be </span> <span style="background-color:#150E3780; color: #1A1A1A">better </span> <span style="background-color:#150E3780; color: #1A1A1A">You </span> <span style="background-color:#150E3780; color: #1A1A1A">fucking </span> <span style="background-color:#150E3780; color: #1A1A1A">ignorant </span> <span style="background-color:#150E3780; color: #1A1A1A">stupid </span> <span style="background-color:#150E3780; color: #1A1A1A">asshole </span> </p> --- ### Appendix: Explainability II <!-- * open black box algorithms --> <!-- * assessing local interpretation of predictions. --> <!-- * our examples --> <!-- * compare classifier --> <center> <table id="tab1" style="font-size: 14px" class="ui compact table"> <thead> <tr> <th>word</th> <th>glove</th> <th>lstm</th> <th>gru</th> <th>multi</th> <th>stacked</th> <th>grucnn</th> <th>ngram</th> </tr> </thead> <tbody> <tr> <th>Communist</th> <th> <div style="background-color:#F76F5C80">0.607</div> </th> <th> <div style="background-color:#150E3780">0.98</div> </th> <th> <div style="background-color:#51127C80">0.926</div> </th> <th> <div style="background-color:#150E3780">0.969</div> </th> <th> <div style="background-color:#51127C80">0.948</div> </th> <th> <div style="background-color:#51127C80">0.928</div> </th> <th>Communist control freak</th> </tr> <tr> <th>control</th> <th> <div style="background-color:#F76F5C80">0.584</div> </th> <th> <div style="background-color:#51127C80">0.89</div> </th> <th> <div style="background-color:#51127C80">0.925</div> </th> <th> <div style="background-color:#51127C80">0.911</div> </th> <th> <div style="background-color:#51127C80">0.912</div> </th> <th> <div style="background-color:#F76F5C80">0.633</div> </th> <th>control freak engineered</th> </tr> <tr> <th>freak</th> <th> <div style="background-color:#F76F5C80">0.61</div> </th> <th> <div style="background-color:#150E3780">0.994</div> </th> <th> <div style="background-color:#150E3780">0.98</div> </th> <th> <div style="background-color:#150E3780">0.999</div> </th> <th> <div style="background-color:#150E3780">0.989</div> </th> <th> <div style="background-color:#150E3780">0.995</div> </th> <th>freak engineered scum</th> </tr> <tr> <th>engineered</th> <th> <div style="background-color:#F76F5C80">0.554</div> </th> <th> <div style="background-color:#150E3780">0.955</div> </th> <th> <div style="background-color:#150E3780">0.978</div> </th> <th> <div style="background-color:#150E3780">0.997</div> </th> <th> <div style="background-color:#150E3780">0.974</div> </th> <th> <div style="background-color:#150E3780">0.973</div> </th> <th>engineered scum situation</th> </tr> <tr> <th>scum</th> <th> <div style="background-color:white">0.537</div> </th> <th> <div style="background-color:#51127C80">0.945</div> </th> <th> <div style="background-color:#150E3780">0.973</div> </th> <th> <div style="background-color:#150E3780">0.995</div> </th> <th> <div style="background-color:#150E3780">0.974</div> </th> <th> <div style="background-color:#150E3780">0.972</div> </th> <th>scum situation this</th> </tr> <tr> <th>situation</th> <th> <div style="background-color:white">0.519</div> </th> <th> <div style="background-color:white">0.159</div> </th> <th> <div style="background-color:white">0.066</div> </th> <th> <div style="background-color:white">0.311</div> </th> <th> <div style="background-color:white">0.253</div> </th> <th> <div style="background-color:white">0.27</div> </th> <th>situation this attack</th> </tr> <tr> <th>this</th> <th> <div style="background-color:white">0.536</div> </th> <th> <div style="background-color:white">0.107</div> </th> <th> <div style="background-color:white">0.071</div> </th> <th> <div style="background-color:white">0.179</div> </th> <th> <div style="background-color:white">0.339</div> </th> <th> <div style="background-color:white">0.267</div> </th> <th>this attack proves</th> </tr> <tr> <th>attack</th> <th> <div style="background-color:#F76F5C80">0.557</div> </th> <th> <div style="background-color:white">0.144</div> </th> <th> <div style="background-color:white">0.073</div> </th> <th> <div style="background-color:white">0.113</div> </th> <th> <div style="background-color:white">0.348</div> </th> <th> <div style="background-color:white">0.294</div> </th> <th>attack proves it's</th> </tr> <tr> <th>proves</th> <th> <div style="background-color:#F76F5C80">0.572</div> </th> <th> <div style="background-color:white">0.38</div> </th> <th> <div style="background-color:white">0.111</div> </th> <th> <div style="background-color:white">0.067</div> </th> <th> <div style="background-color:white">0.33</div> </th> <th> <div style="background-color:#CB3E7280">0.707</div> </th> <th>proves it's all</th> </tr> <tr> <th>it's</th> <th> <div style="background-color:white">0.546</div> </th> <th> <div style="background-color:white">0.309</div> </th> <th> <div style="background-color:white">0.068</div> </th> <th> <div style="background-color:white">0.126</div> </th> <th> <div style="background-color:white">0.204</div> </th> <th> <div style="background-color:white">0.07</div> </th> <th>it's all organized</th> </tr> <tr> <th>all</th> <th> <div style="background-color:#F76F5C80">0.558</div> </th> <th> <div style="background-color:white">0.142</div> </th> <th> <div style="background-color:white">0.07</div> </th> <th> <div style="background-color:white">0.105</div> </th> <th> <div style="background-color:white">0.187</div> </th> <th> <div style="background-color:">0.027</div> </th> <th>all organized ahead</th> </tr> <tr> <th>organized</th> <th> <div style="background-color:white">0.512</div> </th> <th> <div style="background-color:">0.039</div> </th> <th> <div style="background-color:white">0.062</div> </th> <th> <div style="background-color:">0.031</div> </th> <th> <div style="background-color:white">0.169</div> </th> <th> <div style="background-color:">0.009</div> </th> <th>organized ahead of</th> </tr> <tr> <th>ahead</th> <th> <div style="background-color:white">0.54</div> </th> <th> <div style="background-color:white">0.055</div> </th> <th> <div style="background-color:white">0.068</div> </th> <th> <div style="background-color:">0.036</div> </th> <th> <div style="background-color:white">0.264</div> </th> <th> <div style="background-color:white">0.14</div> </th> <th>ahead of time</th> </tr> <tr> <th>of</th> <th> <div style="background-color:white">0.524</div> </th> <th> <div style="background-color:white">0.189</div> </th> <th> <div style="background-color:white">0.068</div> </th> <th> <div style="background-color:white">0.094</div> </th> <th> <div style="background-color:white">0.35</div> </th> <th> <div style="background-color:white">0.193</div> </th> <th>of</th> </tr> <tr> <th>time</th> <th> <div style="background-color:white">0.528</div> </th> <th> <div style="background-color:white">0.17</div> </th> <th> <div style="background-color:white">0.073</div> </th> <th> <div style="background-color:">0.03</div> </th> <th> <div style="background-color:white">0.314</div> </th> <th> <div style="background-color:white">0.159</div> </th> <th>time</th> </tr> </tbody> </table> </center> <!-- <center> --> <!-- <img src = "images/mbo_search.png" width = "80%"> --> <!-- </center> --> --- ### Appendix: Nullmodel  --- #### Appendix: Most Frequent Bigram Network - Jews <center> <img src="../images/jew_bigram_network.png" width = "95%"> </center> --- #### Appendix: Most Frequent Bigram Network - Migrants <center> <img src="../images/migrant_bigram_network.png" width = "95%"> </center>