Social Media and Forced Displacement: Big Data Analytics and Machine Learning (White Paper)
1 like3,045 views
This white paper summarizes a project using social media data and machine learning to understand perspectives related to the Europe refugee emergency. The project conducted ten mini-studies analyzing Twitter data to monitor interactions between refugees and service providers, and understand host community sentiment toward refugees. Initial results were inconclusive for monitoring refugee interactions but revealed that a small number of tweets connected refugees to terrorist attacks in local Twitter communities. The paper outlines the methodology used and lessons learned to inform humanitarian decision-making and response through social media analysis.
Social Media and Forced Displacement: Big Data Analytics and Machine Learning (White Paper)
- 1. “Social Media and Forced Displacement: Big Data Analytics & Machine-Learning” White Paper September 2017 UN GLOBAL PULSE | UNHCR INNOVATION SERVICE
- 2. TABLE OF CONTENTS ACKNOWLEDGMENTS 4 SUMMARY BACKGROUND 5 PROJECT OVERVIEW A COMPENDIUM OF MINI-STUDIES USING SOCIAL MEDIA DATA 6 Queries and Taxonomies Classification Iteration 1 7 Table 1: Initial Monitors Overview Hypotheses Setup 8 Insights Iteration 2 9 Table 2: Situational Awareness Monitors Overview Setup Categories 10 Insights 10 LIMITATIONS AND LESSONS LEARNED THE WAY FORWARD 11 ANNEXES 13 Annex I: Data Query Taxonomies per Hypothesis Annex II: Tweets found and catalogued by AI 16 Annex III: Data Visualizations (Quantitative inputs) 19 Annex III: Data Visualizations (Qualitative inputs) 21 Annex IV: Interactive map 23
- 3. 4 5 SUMMARY This white paper summarizes the initial findings and lessons learned from a project conducted by UNHCR’s Innovation Service and UN Global Pulse1 to inform on the viability and value of social media an- alytics to complement understandings of the Europe Refugee Emergency. Ongoing conflicts and violence around the world2 led over 1.4 million people to seek refuge in Europe between 2015 and the first part of 2017. Data from social media offers a wealth of information that can be parsed to better understand what people think, and how people feel about things affecting their lives, such as the displacement and movement of large volumes of people. Researchers in turn, can use this data to inform topics of interest; decision makers can use such data as evidence on which to inform for example, programmatic responses and alterations. The paper outlines the process, questions and meth- odology used to develop the project and presents pre- liminary observations on how aspects of the Europe Refugee Emergency are related on Twitter. The paper describes ten quantitative social media mini-studies that were developed as part of the project. The project team initially set out to explore the value of social media both for monitoring Persons of Concern’s (PoC), sentiment towards the provision of services, and their interactions with service provid- ers3 . However, based on inconclusive initial results and anticipating an increase in negative public views towards PoC following the 2015-2016 terrorist attacks in Europe, the project refocused on the analysis of host communities’ sentiment towards PoC in reaction to incidents taking place in different European coun- tries. Findings revealed that within local active Twitter communities, a small number of people connected PoC and the different terrorist attacks. Being able to assess peoples’ views in real-time pro- vides a unique opportunity for UNHCR to counter non-conducive behaviour online. It also allows the Agency to better understand generalized perceptions vis-a-vis longer-term solutions for PoC. The processes detailed herein are intended to serve as examples and to inspire other agencies looking to use social media and data analytics to inform on de- cision-making processes, operational responses and policy development in emergency-related contexts. 1 UN Global Pulse is flagship innovation initiative of the United Nations Secretary-General on big data. Global Pulse functions as a network of innovation labs where research on Big Data for Development is conceived and coordinated. UNHCR Innovation is a service unit to UNHCR dedicated to facilitating innovation and experimentation through future-oriented approaches and organizational change processes to make UNHCR more efficient and impactful for PoC. 2 UNHCR, (2017). Emergencies: Europe Situation. 3 Including smugglers, NGOs, UN agencies, volunteers, Governments Keywords: social media monitoring, big data, big data analytics, machine-learning, artifi- cial intelligence, data parsing, forced displace- ment, refugees, asylum, migrants, Europe, sen- timent analysis, xenophobia, data science. BACKGROUND The Europe Refugee Emergency was a constantly, and rapidly-changing context. Ongoing conflicts and violence around the world4 led over 1.4 million people to seek refuge in Europe between 2015 and the first part of 2017. This included increasing numbers of families, women, and unaccompanied and separated children—some seeking to reunite with other family members already in Europe. This new movement was challenging for many organizations, including UNHCR; people moved quickly, often across several international boundaries in very short periods of time; sometimes encountering changing protection risks, particularly when legal practices evolved, when borders closed, or when alternative routes begin to develop. According to a report released by Social Media for Good5 , social media monitoring can provide signifi- cant value to decision makers in such dynamic con- texts, where humanitarian access is poor, the informa- tion landscape fragmented, and social media widely used. For example, UNHCR’s report “From a Refugee Perspective” portrays the discourse of refugees and migrants and the use of social media.6 Social media platforms are powerful communications tools for hu- manitarian organizations, both at a strategic corpo- rate level and an operations level to directly interact with affected communities. They also contain a wealth of information that can be parsed to measure and monitor conversations and emerging narratives. Further, sentiment analysis of social media content can be used to capture public perceptions of an orga- nization and its activities in a particular context to not only help develop new strategies, but also to ensure that existing programmes and projects are re-aligned and course-corrected in real-time. Several pilot and research projects have shown the feasibility of using social media data to crowdsource topics of relevance to sustainable development and humanitarian action. However, there has been little effort in extending the quantification of online sentiment to inform on interac- tions between PoC and services providers. Similarly, organizations would benefit from understanding how host communities view PoC on social media to inform their decision-making processes. 4 UNHCR, (2017). Emergencies: Europe Situation. 5 Luege, T. (2015). Social Media Monitoring in Humanitarian Crises: Lessons Learned from the Nepal Earthquake. Social media for Good. 6 UNHCR (2016). From a Refugee Perspective: Discourse of Arabic Speaking and Afghan refugees and migrants on social media UNHCR currently uses social media for two main pur- poses7 : 1) to publicly portray the Agency’s work, and digitally engage with public audiences; and 2) to com- municate with affected communities (CwC)8 . UNHCR has a strong influence among different audiences on platforms, such as Twitter, Facebook, and Instagram, and it has clear guidelines on the use of these plat- forms for communication purposes. While the use of social media for CwC-supported activities is rela- tively new to the Agency, there are many promising efforts underway, both at Headquarters, and in field operations. PROJECT OVERVIEW The work described in this paper was initiated by the UNHCR Winter Operations Cell9 and UNHCR’s Innovation Service in November 2015. UNHCR rec- ognized that big data analytics could provide ad- ditional insights into understanding the protection environment within the Europe Refugee Emergency. However, it did not have vast in-house knowledge, skills, or the necessary tools to conduct large-scale analyses. Therefore, it was limited in its ability to feed potentially valuable information contained in big data into operational responses. To validate the value of social media data in emergen- cy situations, UNHCR’s Innovation Service partnered with UN Global Pulse in January 2016. UN Global Pulse provided technical guidance, coaching and tools for the project. The joint-collaboration explored how alternative sources of data can and should play a role in pursuing humanitarian outcomes. The project team identified two opportunities in which social media could be harnessed to better un- derstand the Europe Refugee Emergency: O1 : Monitor interactions between PoC, and between service providers and PoC, in an aggregated form, and; O2 : Understand the sentiment of PoC, host commu- nities, and communities through which PoC have tran- sited, in aggregated form. The project envisioned a near-real-time monitoring system that could inform operational responses in support of the Europe Emergency Regional protec- tion strategy. This system would have a two-tier ar- chitecture, with a machine learning component in the 7 Note that these two uses of social media are distinct from those described within this white paper. 8 UNHCR Innovation (2016) Emergency Lab definition: Communication with Persons of Concern ensures they have access to the information they need through the most appropriate and trusted channels, enabling them to make informed decisions to protect themselves and each other. For UNHCR, communication implies continuous listening to and dialogue with and between Persons of Concern. This contributes to their sense of connectedness and dignity while facilitating channels for their voices to be heard and acted-on. 9 For more details on UNHCR Winter Cell, please see www.unhcr. org/news/.../big-chill-threatens-refugees-unhcrs-winter-cell-responds.html ACKNOWLEDGMENTS This paper was developed by colleagues from UNHCR Innovation Service and UN Global Pulse, an innovation initiative of the United Nations. UN Global Pulse would also like to thank the Government of the Netherlands for supporting its network of Pulse Labs and the activities under this project.
- 4. 6 7 backend that would process and classify social media posts according to predefined categories (e.g., posts related to abuse, or of xenophobic nature), and an information visualization interface that would enable UNHCR staff to routinely monitor, and analyze rele- vant social media feeds in six different languages: Arabic, Farsi, English, Greek, German and French. To inform the feasibility of this system, and to ensure the opportunities identified were substantial, the project team iteratively conducted a series of ten quantitative mini-studies using Twitter posts. A COMPENDIUM OF MINI-STUDIES USING SOCIAL MEDIA DATA The ten studies were divided into two main iterations. For each of the iterations, a methodology largely in- spired by the Harvard Data Science curriculum10 was used. The following sections detail the data and tools that were employed, discuss the main hypotheses, and share the general iterative procedure Data and Tools Twitter posts, or tweets, are mostly public expressions of ideas and opinions11 —as opposed to Facebook posts, which are mostly private. As of 2014, only 5.1% of Twitter accounts are protected12 . Therefore, given that the majority of Facebook posts are private, and potentially a PoC could be the one expressing an opinion, the project chose tweets as the main source of data, complying with UNHCR’s data protection policy13 . UN Global Pulse has a long-term research partnership with Crimson Hexagon that allowed the project team to use the company’s ForSight tool to access and analyze social media posts. Crimson Hexagon pro- vides an online social media monitoring platform that enables users to create monitors, which have built-in machine learning capabilities to semi-automatically classify and extract sentiment from posts. These ca- pabilities are based on algorithms that are iteratively improved using a training dataset, i.e., a curated col- lection of posts that helps train the monitor to correct- ly interpret any new incoming post14 . The three main steps in setting up and training a monitor are: 1) defining a taxonomy to identify the keywords, hashtags, and phrases that will help re- trieve the most relevant posts from the social media platform of interest (e.g., Twitter), 2) formulating a query using those terms to retrieve the posts; and 3) 10 Blitzstein, J. & Pfister, H. (2015). The Data Science Process. Harvard Data Science 11 Page, C. (2014). Twitter has almost 430 million inactive users. The inquirer. 12 Idem 13 UNHCR (2015). Policy on the Protection of Personal Data of Persons of Concern to UNHCR. RefWorld. 14 SAS (2016). Machine Learning. SAS Institute. manually classifying an initial subset of the retrieved posts to establish the training dataset. Once a monitor is trained, it provides different views of automatically classified posts, as well as of the sentiments extracted from them, which enable users to conduct a variety of quantitative analyses. While the project described in this paper was implemented with the ForSight tool, the methodology is generic and can be executed with other technological solutions. Queries and Taxonomies Formulating appropriate queries is not always straightforward, and a significant amount of effort can be put into training a monitor before an inadequate query is detected. This is an iterative process that in- volves a certain degree of trial and error. For example, selecting the appropriate vocabulary can be difficult. Tweets abound with colloquial language and “inter- net-speak”—Arabic slang typically varies across coun- tries and regions, and can be written in either Arabic abjad (which is dextrosinistral), or using the Roman al- phabet (e.g., ArabEasy, which is sinistrodextral). The 140-character restriction on Twitter also encourages word abbreviation. As a rule of thumb, a lack of rele- vant posts can be indicative of a poor query, or a very restrictive combination of keywords based on the use of logical operators (AND/OR/NOT). In addition, certain assumptions must be made re- garding tweeters’ specific knowledge of the topic of interest. For example, a premise of this project is that, in general, people who tweet have little to no knowl- edge of the legal and protection differences between migrants and refugees. Both terms were used as synonyms in the queries, even though they have dif- ferent implications for UNHCR. Contrary to migrants, refugees are specifically defined and protected by international law, particularly regarding refoulement15 . Finally, queries can be restrained in space and time. These two dimensions can be helpful for further bringing out the voice and opinions of for example, PoC vs. host communities. Geo-referencing of social media posts16 can be done based on a combination of the location declared by the user in his/her profile, and the latest location(s) from where s/he posted. Classification The classification process first requires determining a set of relevant categories, in which the queried posts will be filed. In initial explorations, the project team found that simple dichotomous categories are most effective, like racist–non-racist, or positive–negative. 15 Expel or return a refugee to the territories where her/his life or freedom would be threatened on the account of race, religion, nationality, membership of a particular social group or political opinion. UNHCR (1977). Note on non-refoulement. 16 Crimson Hexagon FAQ: How does Crimson base its geographical data. Categories for irrelevant and neutral posts are also useful, since all posts may not fit into the dichotomous pair, either because the content is inapplicable, or because it is incongruous17 . All the categories deter- mined in the mini-studies are presented in Annex I, along with their respective queries. An initial subset of posts must then be filed manual- ly to create the training dataset (or training tweets), which the monitor’s underlying algorithms will use to automatically classify new incoming posts. This involves personal judgment as to whether content is relevant or not, and can turn out to be a lengthy procedure. Typically, the more categories there are, the more posts there are that need to be read, and manually sorted. Iteration 1 The project team conducted six mini-studies in this it- eration. For each study, a unique monitor on Crimson Hexagon was created (see Table 1). Table 1: Initial Monitors Overview Monitor Unit of Analysis: Geography Unit of Analysis: Timeframe Language Number of Posts* Identified Opportunity 1. Interactions Arabic Greece February 1, 2015– April 18th , 2017 Arabic 6,341 O1 2. Interactions Farsi Greece February 1, 2016– April 18th , 2017 Farsi 1,483 O1 3. Xenophobia Greek Greece June 1, 2015– April 18th , 2017 Greek / Greeklish (latin chars.) 248,691 O2 4. Xenophobia English Greece June 1, 2015– April 18th , 2017 English 26,466 O2 5. Xenophobia Arabic Greece February 1, 2015– April 18th , 2017 Arabic 196 O2 6. Xenophobia Farsi Greece February 1, 2016– April 18th , 2017 Farsi 160 O2 Source: Crimson Hexagon ForSight tool Hypotheses The assumption for O1 was that analyzing social media posts could provide insights into, for example, altered routes, or the conversations PoC are having with ser- vices providers, including smugglers; and that this could provide better situational awareness for deci- sion making, and thereby better inform the orientation of resource allocations, and advocacy efforts. 17 See Annex II for a detailed classification of racist, non-racist, neutral, and irrelevant tweets. The hypothesis for O1 : O1 H: Monitoring interactions will reveal behavior- al patterns and intent of PoC with regard to service provision and access to territory and asylum. This can inform UNHCR programme design and plan- ning strategies for Strategy Objective 1 of the Europe Emergency Regional Protection Strategy: Access to territory and asylum is safe. The assumption for O2 was that understanding the sentiment that host communities express on social media could help identify pockets of, for example, xe- nophobic attitudes towards PoC; and that this could help UNHCR to improve the conditions in which durable solutions may occur, by better targeting corporate communications, and advocacy-related actions around legislation in specific countries. * Number of post analyzed by the machine to the date: April 18th, 2017. The hypothesis for O2 : O2 H: Understanding PoC and host communities’ mutual sentiments will reveal how both groups view and react to asylum conditions and protection. This can inform programme design and planning strate- gies for Strategy Objective 3 of the Europe Emergency Regional Protection Strategy: Access to protection systems and durable solutions are reinforced. This will also provide a baseline for responses, their ad- justment, and their possible improvement.
- 5. 8 9 Setup The flexibility of language use on social media re- quires native speakers to query and classify the train- ing tweets. Native speakers alone can understand the semantic nuances, colloquial, and even unusual uses of local language, as well as typical abbreviations discussed above. The project team relied on a group of English, Farsi, Arabic, and Greek native speakers, all of whom have basic knowledge in computer pro- gramming. This paper will refer to the team of native speakers as the monitor trainers. The monitor trainers were coordinated by UNHCR’s Innovation Service’s Data Scientist, and overseen by UN Global Pulse’s technical team. The project team concentrated on Greece for the first iteration, as the country has played many different roles throughout the Europe Refugee Emergency—it is now a host country for a static population of refu- gees and migrants. Based on UNHCR data from the Europe Regional emergency—which includes demo- graphic population data—it was assumed that PoC would largely be Arabic or Farsi speakers, and that posts from Greece in these languages would likely be those of PoC. To address O1 H, the trainers set up two specific mon- itors to track interactions between PoC, service pro- viders, and the general public regarding access to territory, asylum conditions, shelter conditions, trans- portation, and movement in Greece. The first monitor was set up for Arabic, and was called Interactions Arabic. The second monitor was set up for Farsi, and was called Interactions Farsi. A full description of the categories and queries used is provided in Annex I. Tweets were monitored from February 1st , 2015 or February 1st 2016 to April 18th , 2017 (end of the study). February 2015 corresponds to a period of major influx of PoC in Europe. February 1st , 2016 is approximately one month prior the EU-Turkey Agreement. To address O2 H, the trainers set up four monitors to track negative sentiment and perceptions, like xenophobic, discriminatory, or racist sentiments, of host communities towards PoC for Greek, English, Arabic, and Farsi. The monitors were respective- ly called Xenophobia Greek, Xenophobia English, Xenophobia Arabic, and Xenophobia Farsi (see Annex 1). Tweets were monitored for the period February 1st , 2015 or February 1st , 2016 to April 18th , 2017. To build the Xenophobia monitors (O2 H), the project used the following categories to classify the tweets: Xenophobic: tweets that express negative attitude, prejudice, or hostile sentiment that vilifies PoC; Non-Xenophobic: tweets that express explicit support, positive attitude, or friendly sentiment towards PoC; Neutral: tweets that describe facts about PoC (for example, news articles) but that do not express a strong sentiment or any sentiment at all; Irrelevant: tweets that are not related to PoC. The monitor trainers identified posts as belonging to the xenophobic category based on the UNESCO18 definition of xenophobia: “xenophobic behavior is hostility based on existing racial, ethnic, religious, cultural, or national prejudice”; and the UN Fund for Contemporary Forms of Slavery, OHCHR, declaration definition of xenophobia: “attitudes, prejudices, and behavior that reject, exclude and often vilify persons, based on the perception that they are outsiders or foreigners to the community, society or national identity”19 . The trainers also distinguished between factual, opin- ion-driven, rumor-driven, and breaking news tweets, in order to adequately train the machine for the neutral category. They further subtracted re-tweets (RT) from certain queries, following the findings of Mendoza et al20 , especially for the monitors related to O2 H, to avoid ‘inflating’ the number of xenophobic posts. Insights The Interactions Arabic monitor was successfully trained but did not retrieve a large number of rele- vant tweets (<7,000—see Table 1). Annex II portrays a subset of these posts. The Interactions Farsi monitor, however, could not be trained, due to an apparent lack of tweets in Farsi (<1,500) regarding access to territory, shelter conditions, and transportation. These results did not provide enough data to confirm or refute Q1 H, and could indicate that PoC—assumed to be either Arabic or Farsi speakers—a) simply do not use Twitter to inquire about, complain, or request ser- vices; b) do not have access to Twitter; or c) prefer other communications channels. The latter two pos- sibilities seem further supported by the Xenophobia Arabic, and Xenophobia Farsi monitors, which also retrieved a very low number of tweets (<200 and <160, respectively). The analysis also showed it is difficult to systemati- cally separate tweets coming from PoC, host com- munities, and the general public for further analy- sis. Only few tweets described access to territory in Europe—including closing borders and entry restric- tions—asylum conditions, and the economic chal- lenges encountered during, and at the end of their journey, while many expressed the sentiment of host 18 UNESCO (2016). Xenophobia. Learning to Live together. International Migration. 19 OHCHR (2011). Declaration on Racism, discrimination, Xenophobia and Related Intolerance against Migrants and Trafficked Persons. Asia-Pacific NGO Meeting for the World Conference Against Racism, Racial Discrimination, Xenophobia and Related Intolerance. Teheran, Iran 20 Mendoza, M., Poblete, B. & Castillo, C. (2010). Twitter Under Crisis: Can we trust the RT? 1st Workshop on Social Media Analytics (SOMA ’10), July 25, 2010, Washington, DC, USA. communities towards PoC. In hindsight, this could have been caused by improper querying and training of the monitors. Based on these early insights, the project decided to concentrate on Q2. The analysis of the O2 H monitors found few online signals for the Arabic and Farsi monitors. For English and Greek however, the number of posts was much bigger, in the order of thousands. Interestingly, only 5% of the tweets retrieved by the Xenophobia Greek monitor (12,423 out of 248,691—see Table 1) were clas- sified as xenophobic, compared to 15% (3,969 out of 26,466) in the Xenophobia English monitor. Although the monitors (queries) retrieved a larger number of posts in Greek, the analysis did not reveal the abso- lute number of tweets. However, with the sample re- trieved, there were more xenophobic posts in English than in Greek for this particular geographic location. See Annex 3 for a summary of the main topics dis- cussed in tweets retrieved by the Xenophobia Greek monitor. Iteration 2 Four follow up studies were conducted in the second iteration. The project created a unique monitor for each study using Crimson Hexagon (see Table 2). Table 2: Situational Awareness Monitors Overview Monitor Unit of Analysis: Geography Unit of Analysis: Timeframe Language Number of Posts* 7. Situation Awareness Nice Worldwide Date of event (14 July 2016) – April 18th , 2017 English, French, Greek, German 3,748,198 8. Situation Awareness Munich Worldwide Date of event (22 July 2016) – April 18th , 2017 English, French, Greek, German 58,815,918 9. Situation Awareness Saint-Étienne Worldwide Date of event (27 July 2016) – April 18th , 2017 English, French 28,884,522 10. Situation Awareness Berlin Worldwide Date of the event (18 December 2016) – April 18th , 2017 English, French, Greek, German 353,580,956 Source: Crimson Hexagon Forsight tool Hypotheses Based on the first iteration, and in reaction to incon- clusive results as well as in reaction to a number of terrorist attacks which occurred in Europe—resulting in refugees being mentioned in various media, includ- ing social media, in potentially concerning ways—, the project refocused to explore whether social media could provide a way to: O3 : Monitor the general public’s opinion on possible mislead relations between PoC and terrorist attacks, in aggregated form. The project team focused on measuring the volume of posts that either blamed, or defended PoC to gauge public opinion and understand whether opin- ions were generally in favour or against PoC. The hypothesis for O3 : O3 H: Host communities and the general public may make a link between PoC and terrorist attacks. Setup To address O3 H, the monitor trainers created four additional monitors that covered the unforeseen in- cidents in Nice (FR), Munich (DE), Saint-Étienne (FR), and Berlin (DE), which occurred on the 14th , 22nd , and 27th of July, and on the 18th of December 2016, respectively. Each was intended to gauge responses to the ter- rorist attacks, and how these might be related to PoC in the global Twittersphere. They were respec- tively called Situation Awareness Nice, Situation Awareness Munich, Situation Awareness Saint- Étienne, and Situation Awareness Berlin. All were trained in English, French, Greek, and German— except the Situation Awareness Saint-Étienne, which was only trained in English and French—using * Number of post analyzed by the machine to the date: April 18th, 2017. almost exactly the same query—only some local ref- erences, and particular hashtags specific to each incident varied. Particular attention was given to employing the same vocabulary for each language to enable a relative degree of comparison between monitors. For example, “attack” in English was trans- lated to “attaque” in French, “επίθεση” in Greek, and “Anschlag” in German. The monitors were not restrained to specific geo- graphic boundaries, but rather looked to under- stand global reactions and opinion. Nevertheless, the choice of language did concentrate the tweets
- 6. 10 11 retrieved to areas where those languages are spoken (see Annex I for details on the categories and queries that were used for each language). Tweets in each monitor were tracked onwards from the date of the incident covered, i.e., in the aftermath of the terrorist attack, until April 18th , 2017. Categories The following categories were used for the situation awareness monitors (O3 H). Blame: tweets that explicitly blame PoC for the incident; Don’t Blame: tweets that advocate for not blaming PoC for the incident, or at least that attempt to de- attach them; No reference to PoC: tweets that describe facts about the incident, but that do not mention PoC; Irrelevant: tweets that mention PoC, but that are not related to the incident; Off-topic: tweets that are neither related to PoC, or the incident. Insights While the Situation Awareness Saint-Étienne monitor gathered a significant number of tweets (>28M—see Table 2), the religious nature of the incident tended to skew the results: PoC were rather linked with funda- mental Islam, than with the event itself. As such, this monitor was discarded, and the project team further inspected only the incidents that did not specifically target religion. Only 6% of the tweets retrieved by the Situation Awareness Nice monitor, 11% by the Situation Awareness Munich monitor, and 5% by the Situation Awareness Berlin blamed PoC for the incident. There were also more don’t blame tweets in the Situation Awareness Berlin monitor than in the other two, with 7% of posts expressing explicit support for PoC in German, condemning racism and xenophobia, and stating that terrorism and violence are the main reasons why PoC flee their homes in the first place. It is also important to note that while the percentages of posts blaming refugees for the incidents are small, they still represent between hundreds of thousands to several millions of spontaneous messages of this direction: 0.2M tweets (Nice), 6.4 M tweets (Munich), and 17.6 M (Berlin). The Situation Awareness Berlin monitor retrieved a significantly higher absolute number of tweets con- necting PoC with the attack. These results could be attributed to several instances. First, because the police quickly identified and arrested a Pakistani asylum seeker as the perpetrator of the attack. Although he was later released when found inno- cent21, Twitter users pursued the discussion on a possible relationship between the incident and PoC. Some even continued to associate the incident with the Pakistani suspect, even after the police had clari- fied there was a Tunisian suspect. Secondly, this was the third in a series of recent incidents in Germany, after the Munich attack, and the Würzburg train in- cident, the latter carried out by a 17-year old Afghan asylum seeker. Third, on September 15th, 2016, Angela Merkel made some remarks regarding the in- tegration process of PoC in Germany, in an interview aired on RBB-Inforadio, one of Berlin’s main radio sta- tions. In one comment, she stated that “drivers are needed everywhere” in Germany22. This comment fueled negative posts on social media, which drew a link with the lorry truck attack, and blamed Germany’s Open Door Policy. LIMITATIONS AND LESSONS LEARNED While the results of the second iteration provide some interesting insights into the way people perceive issues related to the European Refugee Emergency, they should be considered cautiously, as social media alone can seldom provide a comprehensive overview of needs and opinions. For example, tweets are gen- erally not representative of socio-economic diversity and age. Only people with access to connectivity, and who have an account, can post, or respond on Twitter. In addition, although Crimson Hexagon’s machine learning and geo-referencing capabilities are ad- vanced, they may not always be entirely accurate. This means that the geo-based queries may have retrieved additional tweets posted from outside the determined geographic boundary (false positives) while omitting others posted from inside the geographic boundary (false negatives). Furthermore, machine classification is not always accurate. At the same time, the project assumed that the general public would not make the legal difference between refugees and migrants, and used both terms interchangeably in the queries. Also, the project did not establish a clear, system- atic distinction between host communities and the general public, in terms of language vis-à-vis geo- graphical location. This means that there could be, for example, a French person posting opinions in German language, currently residing in Germany; or a German person, posting in German language, currently resid- ing outside Germany. Either example could be cate- gorized as both host community, or general public, depending on different perspectives and proximity to the community. Furthermore, the automatic classification and senti- ment extraction from tweets may have missed some 21 The Guardian (2017). Berlin Truck Attack: first suspect released as drive thought to be still at large - as it happened. February 17, 2017. 22 Der Zeit (2016). Angela Merkel: Flüchtlinge sollen schnell in Arbeitsmarkt integriert werden. September 16, 2016. important contextual cues, as both procedures use only the posts’ textual content. Tweets may contain links, or allude to other information. They may also be part of a broader, ongoing conversation. Overlooking, or omitting these contextual cues may result in a misin- terpretation of certain uses of language, such as irony, satire, or metaphoric speech. This was observed in the Situation Awareness Berlin monitor, where there were many sarcastic references to the radio interview with Angela Merkel. The project initially imagined it would be possible to filter out the voices of PoC by collecting posts in Arabic and Farsi, and by geo-locating their point of origin. However, it turned out to be extremely com- plicated to determine whether a person tweeting in these languages was indeed a migrant or refugee, or simply a person from the host community or local diaspora—especially seeing that the Arabic and Farsi monitors in the first iteration did not retrieve many tweets. This would induce a high degree of uncer- tainty in any attempt to address hypotheses related to PoC vs. host communities, which is why a deeper analysis of O1 was not employed. However, extend- ing this research to other social media platforms like Facebook, which PoC use extensively23 , might facilitate the distinction, and help better understand interactions. The second iteration also showed the extent to which the comprehensiveness of the vocabulary used in a query could both increase the volume of retrieved posts, and their overall accuracy. Firstly, several iterations are needed to capture all the infor- mation. Secondly, the messages should be manual- ly scanned to be able to re-classify and re-train the machine to include relevant words. Although it was generally similar, the query used for the Situation Awareness Berlin monitor was more sophisticated, and better tailored, than those used for the other situ- ation awareness monitors (see Annex I). This was the result of immediate feedback received from end-us- ers within Germany on specific language nuances. The project used colloquial which were also used by local media and the general public when referring to PoC. UNHCR’s Innovation Service’s Community and Content Manager also assisted, pointing to specific hashtags and keywords that were being used. More generally, findings showed that working with social media requires a dynamic mindset. The project had to adapt and iterate rapidly. The hypotheses from the first iteration proved too broad, and for them to be of any tangible use to UNHCR, they had to be adjust- ed. The project also required more resources than ini- tially identified within UNHCR, as linking social media monitoring with operational responses and planning is a new concept for the Agency—the mini-studies were 23 European Commission (2016). Effective use of technology and social media for refugees’ labour market integration. typically inaccurately conflated with a range of other social-media-driven projects, including CwC efforts, Information Management work, Communication and Public Information (PI) activities, and UNHCR digital brand marketing. In addition, while this white paper refers to the current work as a series of “mini-studies,” it should be em- phasized this was a labor-intensive process. UNHCR’s Innovation Service had to resolve to recruit monitor trainers, all of whom had both relevant native language skills, and basic knowledge in computer program- ming. UN Global Pulse also invested the equivalent of a full-time staff member, in addition to providing the necessary partnership agreements, which enabled access to the data and tools. In the future, it will be important for UNHCR to have qualified and dedicat- ed personnel to develop similar approaches to col- lecting, processing, and analyzing big data sources. These efforts should be integrated across different units, to further the Agency’s understanding—as a whole—of the potential of social media and big data to inform operational decisions, advocacy activities, and strategic communications, as well as to improve listening to different affected communities, in order to demystify non-accurate information. The limitations of the initial interactions and the lessons learned throughout the course of the project helped reshape its initial scope. The project set out to use social media posts to build a better, more nuanced understanding of different complex aspects of the Europe Refugee Emergency, that are other- wise difficult to assess with traditional tools—such as surveys. However, it now sees added value in trying to use social media to detect unexpected signals of ongoing events that could put PoC at risk, and that UNHCR may need to quickly respond to, or act upon. The streaming nature of social media posts affords the detection of such signals in near-real-time, which could be useful in cases similar to the aftermath of the Berlin terrorist attack, where more than 17.6 million tweets linked the incident with PoC. There are few data sources that can facilitate such in-depth, rapid response mechanisms, and the project intends to continue exploring their potential. THE WAY FORWARD UNHCR routinely collects massive amounts of data, through, for example, registration and information management exercises, programme and project im- plementation, and financial activities. The main chal- lenge, and therefore an important opportunity for the Agency, is to find ways of accompanying the inte- gration of new data sources into this culture, and to bring more data-driven evidence into decision-mak- ing processes and advocacy efforts, particularly in developing an institutional policy against xenophobia,
- 7. 12 13 discrimination, and racism against PoC. The current project intends to continue exploring this integration with the development of a social media monitoring system (an early snapshot of which is presented in Annex IV), which will use streamed posts as a way to detect signals of ongoing events, which the Agency may need to act upon. Beyond this, there are several other opportunities for UNHCR in the future like: ● Defining clear, rigorous methodologies and protocols to distill relevant information ex- tracted from biased data sources like social media. Interpreting social media using quan- titative methods and machine intelligence is complex, particularly when the context of the composite data24 is nuanced and sometimes unclear from individual pieces of information; ● Integrating these new types of insights into operational workflows. Social media posts can typically feed into operations, policy or advocacy, and communications. This is another opportunity for UNHCR; ● Adopting the relevant ethical and privacy frameworks relating to data protection, privacy, anonymity, and security; ● Building internal data literacy and specialized capacities within the Agency. This last point should further help improve UNHCR’s capac- ity to make data-driven decisions. During the Committee on the Elimination of Racial Discrimination in March 2011, UNHCR’s Senior Legal Coordinator explained that “Combating racism, xeno- phobia and related forms of intolerance against refu- gees, asylum-seekers and stateless persons is one of the principle objectives of UNHCR, and these forms of discrimination are one of the greatest threats to the rights of refugees and asylum-seekers, in Europe and elsewhere”25 . From impacting the right to seek asylum, to better understanding how xenophobia is related to the primary root causes of persecution or negatively affecting integration opportunities, this is an area of work UNHCR must be more proactive in. In fact, not addressing xenophobia towards PoC in a strategic way would constitute a shortcoming of UNHCR’s overall protection mandate as an agency The 2009 “Combating Racism, Racial Discrimination, Xenophobia and Related Intolerance through a Strategic Approach” along with the 2015 evaluation of UNHCR’s Southern Africa Programmes “Protection from Xenophobia” layout specific guidelines on how the agency is addressing the issue. 24 Composite data or compound data is any data type which can be constructed in a program using the programming language’s primitive data types. In summary, is any language data type that isn’t a machine number 25 OHCHR (2011). Committee on the Elimination of Racial Discrimination. Thematic discussion: “Racial discrimination against People of African Descent”. UNHCR DIP However, confronting growing intolerance and xe- nophobia are just some of the many challenges that may lie ahead for UNHCR, in a world that is more con- nected, and where ideas and words can be shared across many channels, including digital channels. The European Network Against Racism (ENAR) published a study that highlights an increase in protests, political/ elections rhetoric, and formation of structured groups against refugees and asylum seekers in Europe26 . They mention that “social media is becoming increas- ingly crucial in forming opinions about migrants, and there has been a growing dissemination of fake eth- nicity-related news about migrants with alarming and sensationalist headlines.” 26 ENAR (2016). Racism and Discrimination in the Context of Migration in Europe. ENAR Shadow Report. Annex I: Data Query Taxonomies per Hypothesis 1.1 O1 H: Monitor Interactions ● Negative perception: bad conditions in access to services or to territory of asylum, police brutality, closed border, means of transportation ● Taxonomy for link: basic neutral, basic posi- tive, basic negative ● Geography: Greece, national level Machine learning query: untrained/discarded monitors PoC Farsi: پلیس OR شهربانی OR مرز OR پناهنده OR (پناه AND )جوی OR کوچگر OR اروپا OR یونان OR ثب OR بازداشت OR دستگیری OR اردوگاه OR (اردوگاه AND )پناهندگان OR وضع OR حال OR بهداشت OR قایق OR اتوبوس OR (اهربزرگ AND )بلوک OR اهر OR جاده OR قیمت OR ازوی OR خشونت OR (بد AND )رفتاری OR ربایی OR قاچاق Translation: Police Police OR border refugee OR (harbor AND barley) OR migrant Greece OR OR OR Europe Registration arrest OR OR OR arrested Camp OR (AND refugee camp) OR conditions are OR OR OR Health BOAT OR BUS OR (Highway AND Block) OR The price OR VISA OR OR OR roads violence OR (bad behavior AND) OR kidnapping OR trafficking General Public: Arabic OR وضع OR حدود OR الحدود OR أوروبا OR محاجرين OR الجئين OR الجئ OR محاجر اليونانيين OR يونانيين OR اليونان OR الوضع Translation: Mohajer (migrant), laje’, (refugee) OR Laj’een (refugees) Mohajeryn (migrants) Europe OR borders border OR situation OR the situation OR Greece OR Greeks OR the Greeks General Public: English “refugee” AND (“move” OR “movement” OR “move” OR “boat” OR “plane” OR “relocation” OR “resettle- ment” OR “removed” OR “returned” OR “reintegrat- ed” OR “walk” OR “road” OR “bus” OR “train” OR “money”) 1.2 O2 H: Understanding sentiment ● Negative perception: racists, extremist or xe- nophobic comments from host communities in their native language, negative sentiment and feelings towards refugees and migrants. ● Taxonomy: racist, non-racist, neutral, irrelevant ● Geography: Greece, national level ANNEXES
- 8. 14 15 Machine-learning query A) Xenophobia English ((migrant OR refugee OR refugees OR immigrants) AND (Greece OR Greeks OR fear OR hatred OR racism OR xenophobia OR foreigners OR arrivals OR Syrians)) AND ((migrant OR refugee OR refugees OR immigrants) AND -(RT OR US OR America OR UK OR Trump OR Brexit OR Merkel)) B) Xenophobia Greek ((μετανάστης OR πρόσφυγας OR πρόσφυγες OR μετανάστες) AND -(RT OR Βρυξέλλες OR Τσίπρας OR Μέρκελ OR Brexit OR Γερμανία)) OR ((μετανάστης OR πρόσφυγας OR πρόσφυγες OR μετανάστες) AND (φόβος OR μίσος OR ρατσισμός OR ξενοφοβία OR ξένοι OR αφίξεις OR Σύριοι)) OR ((metanastis OR metanasths OR metanastes OR prosfugas OR pros- fuges) AND -(RT1 OR Merkel OR Tsipras OR Brexit OR Germania)) OR((metanastis OR metanasths OR meta- nastes OR prosfugas OR prosfuges) AND (fovos OR fobos OR misos OR ratsismos OR xenofovia OR xeno- phobia OR afiksi OR xenoi OR afiskeis OR Syrioi OR Surioi)) C) Xenophobia Arabic محاجرين OR الجئين OR الجئ OR محاجر Translation: migrant OR migrants OR refugee OR refugees D) Xenophobia Farsi OR خارجی OR رسیدن OR یونان OR کوچگر OR خوشامد OR پناهندگان OR پناهنده خارجیها Translation: refugees, refugee, welcome, migrant, Greece, to arrive, foreign, foreigners 1.3 O3 : Incidents Linkage ● Linking incidents: blame refugees for attacks/incidents, terrorism activities in Europe, Munich, Nice, St. Etienne, #donot- blame refugees, #PrayforMunich, #offeneTür, Bastille, #BerlinAttack ● Taxonomy: blame refugees, do not blame refugee, neutral, irrelevant ● Geography: Worldwide Machine-learning query A) Situation Awareness Munich (Munich OR MunichAttack OR PrayForMunich OR offeneTür OR Beschuldige OR Flüchtlinge OR Flüchtlingen OR Schuld OR Attacke OR Tod OR Töten OR Opfer OR Schießen OR Schiessen OR Attentäter OR Gewehr OR Pistole ) OR (attack OR killer OR kill OR killed OR dead OR deadly OR death OR shoot- ing OR gun OR bullets OR victims OR killing) OR 1 Excluding Retweets (RT) (Μόναχο OR Μόναχο OR επίθεση OR PrayForMunich OR πρόσφυγες OR κατηγορούν OR πρόσφυγες OR ένοχος OR επιθέσεις OR θάνατοι OR θάνατο OR θυμάτων OR όπλο) OR (attaque OR attaques OR at- tentat OR attentats OR tué OR tueur OR assassin OR mort OR morts OR tournage OR fusillade OR pistolet OR fusil OR balles OR victimes) B) Situation Awareness Nice (Nice AND (terrorist OR attacks OR France OR dead OR (Bastille AND Day) OR terror OR deaths OR blame OR refugees OR refugee OR deaths OR attack OR victims OR assassins OR gun)) OR (Νίκαια AND (τρομοκράτης OR τρομοκρατική OR Γαλλία OR νεκρός OR νεκροί OR (Bastille AND Day) OR τρόμος OR θάνατοι OR επίθεση)) OR (Nizza AND (terroris- tischen OR Attacke OR Frankreich OR Tot OR (Bastille AND Tag) OR terror OR Tötten OR Beschuldige OR Flüchtlinge OR Flüchtlingen OR Schuld OR Töten OR Opfer OR Schießen OR Schiessen OR Attentäter OR Gewehr OR Pistole)) OR (Nice AND (terroriste OR attaque OR attaques OR attenat OR faute OR atten- tats OR France OR mort OR morts OR (Jour AND de AND la AND Bastille) OR terreur OR mortes OR blâme OR réfugiés OR réfugiés OR blâmer OR attaque OR mort OR victimes OR assassin OR pistolet OR (14 AND juillet) OR terreur)) C) Situation Awareness Saint-Etienne: otage OR armés OR (Saint AND Etienne AND du AND Rouvray) OR mort OR morts OR (prise AND d’otage) OR église OR prêtre OR assaillants OR tué OR blessé d) Situation Awareness Berlin (Berlin OR BerlinAttack OR BerlinTerrorAttack OR (Berlin AND Terrorist AND Anschlag) OR (Berlin AND Terroranschlag) OR Breitscheidplatz OR merkel- deutschland OR Weihnachtsmarkt OR (Weihnachts AND Markt) OR Anschlag OR offeneTür OR Beschuldige OR Flüchtlinge OR Flüchtlingen OR Schuld OR Attacke OR Tod OR Töten OR Opfer OR Weihnachten OR Attentäter OR Gewehr OR LKW OR Islam OR Pakistaner OR Pakistanisch OR Islamophobie OR Liberale OR Immigrant OR Asyl OR Lastwagen OR Asylant OR Asylanten OR Fluechtlingsbewerber OR Asylbewerber OR Lastkraftwagen OR Migranten OR Rassismus OR Fremdenfeindlichkeit OR (Beschuldige AND Flüchtlinge AND nicht) OR (Beschuldige AND Flüchtlingen AND nicht) OR Einwanderer OR vorw- erfen OR (scheiß AND Flüchtlinge) OR (scheiss AND Flüchtlinge) OR (scheiße AND Flüchtlinge) OR (sche- isse AND Flüchtlinge) OR anschuldigen OR anklagen OR Vorwürfemachen OR Muslime OR (Die AND Schuld AND den AND Flüchtlingen AND zuschieben)) OR (attack OR blamerefugees OR (blame AND refugees) OR terror OR terroristattack OR terrorist OR killer OR Merkel OR (open AND door) OR opendoor OR kill OR killed OR dead OR deadly OR death OR ISIS OR islam OR Pakistani OR Christmas OR christmasmarket OR truck OR victims OR killing OR RefugeesWelcome OR liberal OR immigrant OR migrant OR asylum OR lorry OR Afghan OR jihad OR islamophobia OR racism OR (don’t AND blame AND refugees) OR dontblamerefu- gees OR Asylmafia OR xenophobia OR thanksMerkel OR ThankyouMerkel) OR (Βερολίνο OR επίθεση OR τρόμος OR τρομοκράτης OR προσφύγων OR πρόσφυγας OR πρόσφυγες OR κατηγορούν OR πρόσφυγες OR ένοχος OR επιθέσεις OR θάνατοι OR θάνατο OR θυμάτων OR Χριστούγεννα OR φορτηγό OR (χριστουγεννιάτικος AND αγορά) OR Αφγανός OR Πακιστανός OR τζιχάντ OR ισλαμοφοβία OR ρατσισμός OR ξενοφοβία OR Μουσουλμάνος) OR (attaque OR attaques OR attentat OR attentats OR tué OR tueur OR terreur OR assassin OR mort OR morts OR tournage OR victimes OR camion OR Natale OR (marché AND de AND Noël) OR pakistanais OR asile OR (porte AND ouverte) OR refugie OR réfugié OR xénophobie OR MerciMerkel OR Musulman OR (ne AND blâmez AND pas AND les AND réfugiés))
- 9. 16 17 Non-Xenophobic Neutral Irrelevant Annex II: Tweets found and catalogued by AI O1 H: Monitor Interactions Translation: You are frustrated by all the refugees dying in the sea but words don’t do us much, open the borders Translation: The governor of Greek Central Macedonia: There are about 13,000 refugees are swarming to the Greek-Macedonian borders in miserable conditions Translation: Greece is currently facing a huge economic crisis.. and the circumstances for the refugees are even more difficult O2 H: Understanding sentiment Xenophobia English Monitor Category: Xenophobic
- 10. 18 19 Irrelevant Translation: On monday the first Syrian refugees will move from Turkey to Germany. Annex III: Data Visualizations (Quantitative inputs) O3 H: Incidents Linkage Total Number of tweets analyzed: 3,433,800 (Nice) + 297,506,445 (Munich) = 300,940,245 posts up to Jan 10th , 2017. Xenophobic: Munich (8%) in yellow and Nice (6%) in purple Geography: Worldwide Xenophobic Nice and Munich (January 10th, 2017) O3 H: Incidents Linkage Total Number of tweets analyzed: Munich (58,815,918 posts), Nice (3,748,198 posts) and Berlin (353,580,956 posts) = total 416,145,072 posts Xenophobic: Munich (8%) in yellow, Nice (7%) in green and Berlin (5%) in purple Geography: Worldwide Xenophobic O2 H: Understanding sentiment Xenophobia Greek Monitor Category: Xenophobic Translation: They treat Greek people Bad, to make space for ‘refugees’. Non-Xenophobic Translation: Humanitarian help for the refugees in Heraklio. Neutral Translation: More than 53.900 refugees and immigrants in the country.
- 11. 20 21 Including Berlin (April 18th, 2017) Not-Xenophobic O3 H: Incidents Linkage Total Number of tweets analyzed: Munich (58,815,918 posts), Nice (3,748,198 posts) and Berlin (353,580,956 posts) = total 416,145,072 posts Non-Xenophobic: Munich (<1%) in yellow, Nice (11%) in green and Berlin (7%) in purple Geography: Worldwide Annex III: Data Visualizations (Qualitative inputs) Data Visualization type: Word Cloud, Munich situation awareness monitor Data Visualization type: Cluster, Munich situation awareness monitor
- 12. 22 23 Annex IV: Interactive map (under-construction by UNGP) Tweets geo-located by route, interactive map (under construction), Python basedData Visualization type: Topic Wheel, Xenophobia Greece in Greek monitor
- 13. How to cite this document: UN Global Pulse, UNHCR Innovation Service, ‘Social Media and Forced Displacement: Big Data Analytics & Machine-Learning’, 2017 The opinions expressed in this paper are those of the authors and do not necessarily represent the position of UN Global Pulse or UNHCR.