UNIVERSITEIT VAN AMSTERDAM
FACULTY OF SCIENCE
INFORMATION STUDIES
BUSINESS INFORMATION SYSTEMS TRACK
Mining Twitter to identify incidents of violence
against minorities
Maria Othon
11118709
Supervisor: Dr. Frank Nack
Second Reader: Dr. André Nusselder
1
Mining Twitter to identify incidents of violence
against minorities
University of Amsterdam
Information Studies
Institute for Informatics
Maria Othon +31650411148 othonmaria@yahoo.gr
ABSTRACT
Incidents of harassment against minorities are increasing dramatically in Europe. An application that intends to inform minority members about violent acts against them can be of help to users when travelling, commuting or establishing in a large city centers. However, there are problems on supporting such applications with data. In this research, Social Media are explored as an information source of identifying such incidents. We address this problem through mining from the Twitter Streaming API and try to identify: who the minority member was, where and when the violent incident happened and what kind of violence was performed. Through a survey and a Tweet analysis we gain insights and conclude that violent incidents against minorities cannot currently be harnessed from Social Media. In the end, recommendations and a discussion follow on establishing such an application.
Keywords
Violence, Minorities, Incidents, Social Media, Twitter, Twitter Streaming API, Tweet Analysis.
1. INTRODUCTION
Over the last years an increase of crime against specific groups of people can be observed. Europe has been affected by criminal acts of hate and discrimination. One quarter of people from minority groups claimed in 2009 that they were victims of such behavior in the last 12 months[1], this indicates that minorities encounter big problems in Europe. The current situation can be called hate crime, it is defined by the European Union Agency for Fundamental Rights (FRA) as "the criminal acts motivated by bias or prejudice towards particular groups of people"1. In 2014 the OECD reported 4.259 hate crime incidents in 46 European states regarding the main bias motivations. In this report the amount of incidents reported by official data sources was smaller than the civil society data sources2, this underlines that there is no cooperation or information exchange between those stakeholders. It can be also observed that civil society data collection mechanisms, due to the fact that they are closer to individuals, tend to acquire more information than the official data collection mechanisms. In the European Union Minorities and Discrimination Survey (EU‐MIDIS) a striking insight was that the majority of victims did not report anything to the police. 1 http://fra.europa.eu/en/theme/hate-crime 2 http://hatecrime.osce.org/infocus/2014-hate-crime-data-now-available
Among the reasons for non-reporting was the “doubt that the police could help”, “the incident was unimportant to report”, the “fear of the executioner” and the “negative attitude towards the police” [1]. This fact also impedes people from having insights in real data-oriented situations since most of them are not even recorded.
Regarding the official data collection, every year the European Union member countries are asked to submit their hate crime data. Among the problems that the EU has to respond are the change in legislation and data collection methods that occur in every country. Also, due to the fact that each country varies in the way that it reports the data (Appendix A, Appendix B) and the bias motivations, there is an increased complexity. These facts result in gaps in hate crime data collection in Europe (Table 1). This table categorizes the quality of data that are collected per European country as either limited, good or comprehensive. As a result, by considering Table 1 in page 2, Appendix A and Appendix B, the collected data cannot represent the reality, nor can data between different European countries be easily compared [2].
Limited Data Good Data Comprehensive Data Few incidents and a
narrow range of bias motivations are recorded.
Data are usually not published
A range of bias motivations are recorded.
Data are generally published
A range of bias motivations, types of crimes and characteristics of incidents are recorded
Data are always published Bulgaria Cyprus Estonia Greece Hungary Ireland Italy Latvia Luxemburg Malta Portugal Slovenia Spain Romania Austria Belgium Czech Republic Denmark France Germany Lithuania Poland Slovakia Finland Netherlands Sweden United Kingdom
Table 1. Classification of official data collection mechanisms pertaining to hate crime September 2012 by EU Member State
Source: FRA desk Research and FRA analysis of data provided by FRA’s research network
2 Τhe non-reporting or under-reporting of violent incidents as well as the concealment of hate crime data by the police and other official organizations [3], has led to the search for more accurate and up-to-date solutions. Nowadays, technology is in a respectful level that can meet minority members’ needs to be informed. Mobile applications can serve ample requirements and communicate precise and real time information by avoiding the involvement of different stakeholders. This results in the development of new technological solutions such as smart mobile applications that aim to inform users when travelling, commuting or establishing in large city centers. Such solutions seem fully applicable in the minority context because minority members appear more vulnerable to violence, and those acts occurring in specific areas of cities.
In the context only of violence some applications utilize crowd sourcing [4], [5], [6], [7] and/or others mining techniques 3 [5], [8], [9]. Specifically, in crowd sourcing users input the information in order to supply the application with data [4]. The previous mining applications extract a combination of social media points of interest (POI) and statistical data, e-news and blogs. However, none application deals with informing the minority members about violence against them. The social media is considered as an information source because their extensive use has made them concentrate valuable data. Facebook has 1.65 billion4 and Twitter has 310 million active mobile users5 by meaning that the proliferation of social media and the huge amount of data produced daily can skip the barrier of the various disconnected European data collection mechanisms. Thus, Social Media could be considered as a potential solution for extracting the violent incidents against specific minorities in Europe. The aim of this research is to get a better understanding of whether Social Media could solve the problem of acquiring incident data about violence against minorities. Another goal is to identify how Social media could be harvested for supplying a uniform European, location-based application that intends to inform minority members about violent incidents against them. The structure is as follows: first we conduct a literature review on the types of violence, minorities, social media topics (Section 2), second the problem statement and research question are presented (Section 3), third we conduct a survey (Section 6), fourth is the mining part with a Tweet Analysis on violent incidents against minorities (Section 7). In the end we discuss (Section 8) and make conclusions (Section 9).
2. LITERATURE REVIEW
In this part the different definitions of violence (Section 2.1) and minorities (Section 2.2) will be stated. Then a Social Media are presented.
2.1 Violence
2.1.1 Definitions
The definition of violence given by the Violence Prevention Alliance (VPA) in World Report on Violence and Health (WRVH) is considered as “the intentional use of physical force or 3 http://www.fearsquare.com 4 http://newsroom.fb.com/company-info 5 https://about.twitter.com/company
power, threatened or actual, against oneself, another person, or against a group or community, that either results in or has a high likelihood of resulting in injury, death, psychological harm, maldevelopment or deprivation” 6. However, the Government of Newfoundland and Labrador in Canada, define violence as “a pattern of behavior intended to establish and maintain control over an individual or a group”7.
2.1.2 Types of Violence
Violence has been categorized in different types according to the people who commit the violent act and the nature of the violent acts .Regarding the people who participate in a violent act, Krug [11] divides violence in three extensive categories:
Self-inflicted : Suicidal behavior Self-abuse Interpersonal: Family/Partner o Child o Partner o Elder Community o Acquaintance o Stranger Collective : Social Political Economic
Krug sorts the nature of violence in four distinct categories
Physical, Sexual, Psychological and Deprivation or Neglect
[11]. However, the Government of Canada7 divides the types of violence in a more analytical approach in Physical, Sexual,
Emotional, Psychological, Spiritual/Religious, Cultural, Financial Abuse and Neglect. In the following list you can
observe the definitions found from various sources:
Physical :
“Physical violence occurs when someone uses a part of their body or an object to control a person’s actions.”7
“The intentional use of physical force.” 8
Sexual :
“Sexual violence occurs when a person is forced to unwillingly take part in sexual activity.”7
“A sexual act committed against someone without that person’s freely given consent”8
Emotional :
“Say or do to make a person feel stupid or worthless” 7
Psychological :
“Threats to cause fear in order to gain control” 7
Spiritual/Religious :
“when someone uses an individual’s spiritual beliefs to manipulate, dominate or control that person.”7 6 http://www.who.int/violenceprevention/approach/definition/en/ 7 http://www.gov.nl.ca/VPI/types/index.html 8 http://www.cdc.gov/violenceprevention/intimatepartnerviolence/ definitions.html
3 “When religion is either the subject or object of violent behavior.” [12] (page 10)
Cultural :
“When an individual is harmed as a result of practices that are part of her or his culture, religion or tradition.” 7
Verbal Abuse:
“When someone uses language whether spoken or written to cause harm to an individual.” 7
Financial Abuse:
“When someone controls an individual’s financial resources without the person’s consent or misuses those resources.” 7
Neglect:
“When someone has the responsibility to provide care or assistance for an individual but does not..” 7 It has to be mentioned that there are more definitions and categorizations of violence but these were critically chosen because of their popularity and context relevance.
2.2 Minorities
2.2.1 Definitions
The definition of minorities is rather complex and experts argue that it should be broad and regularly updated [10]. According to a definition offered in 1977 by Francesco Capotorti, Special Rapporteur of the United Nations Sub-Commission on Prevention of Discrimination and Protection of Minorities, a minority is: "A group numerically inferior to the rest of the population of a State, in a non-dominant position, whose members - being nationals of the State - possess ethnic, religious or linguistic characteristics differing from those of the rest of the population and show, if only implicitly, a sense of solidarity, directed towards preserving their culture, traditions, religion or language"9. Also, Feagin [13]
(page 10) pinpoints that every minority has at least one of the following features:
(1) suffering discrimination and subordination
(2) physical and/or cultural traits that set them apart, and which are disapproved by the dominant group
(3) a shared sense of collective identity and common burdens (4) socially shared rules about who belongs and who does not determine minority status
(5) tendency to marry within the group.
2.2.2 Types of Minorities
Given the definitions above, the minority types are usually established after defining what is considered a minority. In 2014 until now, the OSCE2 recognizes the following categories European minorities as bias motivations:
Muslims
Racism and Xenophobia
Bias against Roma and Sinti
Anti-Semitism
Bias against Christians and other religions
9
www.ohchr.org/Documents/Publications/MinorityRights_en.pdf
Bias against LGBT people
Bias against people with disabilities
Another minority categorization, in a more general approach, encloses the Ethnicity, Race, Sexual, Religion, Disability and Other [7]. A new category that can be introduced in the minority part is illness10, this category can be used because of the stigma ill people might have and was restricted to cancer and aids.
Within these minority fields a wide variety of subcategories can be introduced. For example, in the sexuality section the LGBTQ (Lesbian, Gay, Bisexual, Transgender, Queers) group is not the only sexual minority set. A larger set such as the LGBTTTQQIAA (Lesbian, Gay, Bisexual, Transgender, Transsexual, Two-spirited, Queer, Questioning, Intersex, Asexual, Ally) 11 also exists. This indicates a need for providing a definition.
Apart from the different categories of each minority type, there are also various subcategories. For instance, in the race section the “White” category has subcategories of races such as the European southern race (Mediterranean), the European central race (Alpine) and the European notherrn race (Teutonic) [14]. Also, in the religious category Muslims have their own branches such as the Sunni, Shia, Alawi, Sufi, Wahabi12.Thus, it is important to clearly define those groups.
2.3 Overview
Given a plethora of the definitions and the types, of minorities and violence in Section 2.1 and 2.2, it is important for people to understand the various approaches about the topic. Also, individuals can gain insights on the importance and usefulness of Social Media. In the next part, the social media are discussed.
2.4 Social Media
2.4.1 Social Networking Sites
Social Media sites are in essence Web 2.0 online communities, they consist of groups around any type of interest where people can help one another with information, advice and personal networks. Social Media have changed the way societies communicate and allows interactions to occur in real time, and break the boundaries between segments and sections of society [17]. As society turns to social media as a primary method of communication and creative expression, social media is supplementing and in some cases supplanting letters, journals, serial publications and other sources routinely collected by research libraries13.Thus, they are a way of self- expression, as a surprisingly large number of people have had a strong desire for self-expression and desire for self-satisfaction that comes from helping others. Through posting personal thoughts many people appear to derive a sense of self-assurance and belonging. Social networking sites are a subcategory of social media, the main distinction between them is that social networking sites require mutual communication, while social media encompass and additional internet technologies such as blogs and video channels. 10 www.berkeleydailyplanet.com/issue/2014-08- 08/article/42355?headline=ON-MENTAL-ILLNESS-We-Should-be-Considered-a-Minority--Jack-Bragen 11 http://ok2bme.ca/resources/kids-teens/what-does-lgbtq-mean 12 www.wildolive.co.uk/islam_denominations.htm 13 https://thesocialweb2014.wordpress.com
4 A social network site is defined as “a networked communication platform in which participants have uniquely identifiable profiles that consist of user-supplied content, content provided by other users, and/or system-provided data, can publicly articulate connections that can be viewed and traversed by others and can consume, produce, and/or interact with streams of user generated content provided by their connections on the site” [18] (page 7). There are almost 200 social networking websites 14. Facebook is the worldwide leading social network with 1.65 billion monthly active users ( March 31 2016) and 323 million in Europe 4. In an analysis of traffic data by Alexa & SimilarWeb 15 among the 137 countries analyzed in 127 of them Facebook is the prevailing social network, while in China, Russia, Japan the leader is QZone, V Kontacte and Twitter respectively (Appendix C). Among the second ranked social networks is Instagram Odnoklassniki, Twitter, Linkedin and Reddit (Appendix C). Undoubtedly, social media preference differs in a worldwide view, however we can see that in Europe, except from the Russian territories, the prevalent social networking sites are sort of homogeneous.
It has to be considered that social networking sites according to the group of users and the context they support. Twitter, is usually characterized as a microblogging site, has every second 9.100 5 posts uploaded, known as Tweets while LinkedIn is a business-oriented social networking service with a rate of more than two new members per second .
2.4.2 Mining
Provided the information above it can be observed that the user-generated content of social media can provide a valuable information source. Literature has shown that Social Media are widely used as an information source, researchers extract information for various fields such as politics [27], marketing [29] health [31].
Violence has also been a topic of research but neither in Europe nor on a multi-country level. Most of the related work to identify hate considers online communities[9][30]. Only one US research papers try to identify online as well as incidents of bullying for social analysis[23]. Interesting, it is also that most research about violence was focused on preserving the in transit safety of the users. For instance, in Brazil there was research with social media mining in order to assist users when they navigate in the city [10]. However, none of the research is European nor focuses on a particular group, even if there has been shown interest in mapping of violence [10][7] and analyzing violent behaviors[23][7]. From various research papers there are considerations regarding the amount of data that is geotagged. In a 2 year study which was intended to identify the Who, What, Why, Where and When of bullying from the 32.477.558 Tweets that were labeled as bullying traces, only the 9.764.583 were actually bullying traces, from which only the 2% were geotagged in the US [23]. Throughout that research, human coders and machine learning were utilized continuously in order to identify the verbal structure of the bullying Tweets[23]. Also Facebook location tags namely “check-ins” consist the 39% of the Facebook mobile app users16
In 14 https://en.wikipedia.org/wiki/List_of_social_networking_websit es 15 http://vincos.it/world-map-of-social-networks 16 http://marketingland.com/survey-check-ins-dominated-by-facebook-30-percent-geotagging-posts-58792
another research 22.32% of Tweets are geo-tagged in North America[26]. Thus, given this significant variance, it is hard to identify the location of the users [23][26], as we can see that the geotagging might depend on the content of the extracted Tweets. As for Instagram only the 5% of the posts are geo-tagged 17This can be regarded as an issue for content and location related analysis, when researchers are initially making assumptions about the geotagged results.
An important subject of research is how representative and valid the mined data is compared to the total amount of data. More specifically, one of the most open social media data sources, the Twitter Streaming API allows for extraction of a specific number of Tweets 46 namely 150 per hour18. Also due to the
scarcity of extraction rules [26] the Social Media API’s mining is restrictive. This fact usually poses questions on how much this data, for example, can represent the complete set of Twitter data. Researchers refer to the problem of restricted amount of data from social media, as sample bias [23] [26] [28]. Having access to full Social Media data is often expensive, time-consuming and is used for validation of the sample analysis 19. Significant variance can be observed between the Twitter Streaming and Firehose (Table 2). However, even Streaming API is only the 1% of the Firehose, most researchers do projections to the results which are can be considered as valid [26] . Also, Facebook Graph gives access to a low amount of public posts that works when friends authorize it. Even for crawling public users data Facebook is restrictive20 due
to constraints such as rate limits21.
Table 2. Comparison of the daily number of Tweets between the sample data (Streaming API) and the full data(Twitter Firehose). Source: [26] 17 https://www.quora.com/What-percentage-of-Instagram-users-geo-tag-their-locations 18 https://dev.twitter.com/rest/public/rate-limiting 19 https://twittercommunity.com/t/how-do-i-get-firehose-access/7490 20 https://yro.slashdot.org/story/10/03/31/1430256/facebook-kills-dataset-of-crawled-public-profiles 21 https://developers.facebook.com/bugs/1725507231023944/
5
Table 3. Applications on violence that use mining techniques
2.5 Existing Applications
In this section, the functionality of other applications regarding the violence (Table 3) or minority topic is described.
Most of the existing mobile applications like Crowdsafe let the user report and explore crime information. The data is provided by the users, who engage in the crowd-sourcing of crime incidents. This application calculates the most safe route and provides crime hotspots, crime clusters and statistics in the city map by using users location data. [4]
Treads mobile application also provides route suggestions in order to accommodate the transportation of visitors in the Washington DC area. The data is leveraged from Twitter and Yelp , and through text summarization techniques provides a link in the map with relevant information [8]. This application requests users points of interest and accordingly mines summarized violence incident data from Twitter [8].
CityWatch mobile application intends to prevent crime at a community level. It uses data from reliable sources as well as from users. Machine learning algorithms make predictions and advice the users according to their personal profiles characteristics [5]. This application leverages data from trustworthy sources like the national census and the insurance industry and also from its users [5]. In this app the users create a profile and request information such as their current address, date of birth, gender, type of housing . In order to add an event a when (exact time and date), what (incident type and optional picture), where (location on the map), how (short textual description) is inquired also. Transafe is a mobile application that emphasizes in making the users feel safer. Specifically users input their feelings on a particular place; an emometer. The input data for this application are from users and law enforcement agencies[6]. This application, requests from users not only the date, time and location but also the emotions of users through an “emometer”, such as sad/happy, bored/excited, scared/safe, angry/peaceful feelings of specific
places[6] with this way the crowd’s mood for a specific block can be measured.
Furthermore, the South African police aims to combat crime through community policing forums (CPF's) in Facebook. In these forums users can report crime and also get information from the police [9].
Fearsquare intends to inform the users about the type and level of crime in particular places by using the location-based service of the Foursquare location reviewing app. It is an added service in Foursquare and it indicates the number violent incidents such as anti-social, theft and violent crime, that have happened in a specific location. The data used for this application is available in the UK police API. The UK monthly crime data discloses the location (latitude/longitude), the crime type (anti-social behaviour, robbery, burglary, possession of weapons, theft, violence and sexual offences etc) and date of the reported crime 3. Fearsquare also requests the public offices to release the data answering where, what, when the incident did happen, as practiced the UK police crime maps and the NYC crime maps22. There are plenty minority applications which intend to bring together the minority members. One of them is Minority Times , a magazine presenting visionaries, success stories and other minority issues, as well as other e-books that indicate the rights of particular minorities like "The Hindu Minority Act 1956" and "Rights of Minorities in Islam"23.
The trend of mobile dating apps has also focused in specific minorities. For example popular apps in the gay community are Gay Romeo24 , GrindR 25 and Jack'd 26. These are examples of minority specialized applications with more than 1.000.000 22 https://maps.nyc.gov/crime/ 23 http://minority-times-magazine.soft112.com 24 www.planetromeo.com 25 http://www.grindr.com 26 www.jackd.mobi
Application Explanation Dataprovided by Data Country Type Active Fearsquare a map with violent
incidents UK government Location, description, time only UK Foursquare, no mobile application yes
Crowdsafe reporting, searching users - every Mobile yes
CityWatch predicts unsafe regions, advices national census, insurance industry, users Users’ personal information, time date, location, type of violence, optional picture, description
depends on the amount of accessible data
Mobile no
Treads safe route recommendation
Twitter, YELP Points of interest (POI)
US only Washington DC Mobile no
Transafe make the users feel safer users, law enforcement date, time, location, emotions
Australia only Melbourne Mobile no
Data Analysis how CPFs use information
Facebook CPFs Police
6 downloads. Also, Muzmatch 27 , Minder 28 and Salaam Swipe 29
are muslim marriage applications which focus on hiding their identity and matching minority members according to their beliefs.
Current applications either aim to inform the users about the violent incidents in cities request from users [4], [5], [6], [7] or mine 3 [5], [8], [9] almost the same categories of data. Most of the
data correspond to the where, what and when the violent incident happened. Another researcher, Fokaeos [13] is an exception to the other applications as she focuses on minorities, the data that she requests are: the friend’s name (if he/she reports the incident), location, date, time, violence type, if the police interfered, the number of victims, gender, against a type of minority, the number of offenders, the description of the offenders, incident description and scale [7].
Regarding the photos of violent incidents that can be uploaded in some applications such as City Watch [5], Garbett et al. 2014 indicate that the unsightly aesthetics of them [20] and Fokaeos’ [13] survey indicates that the 80% of the potential users do not wish to view photos. However, US & Canada governmental authorities use photos to depict the victimizer [21] this could leave potentially a stigma and violates fundamental human rights (UDHR Article 7) [20].
As can be clearly seen, current violence applications focus on crime against anybody and not on specific minorities. Furthermore, regarding the minority applications, it can be identified that minority members largely use apps created for these minorities. This could indicate the existing gap for developing a violence against minorities application. Overall, the majority of the applications about violence have as a preliminary data requirement, the exact location of the incident as well as when and what type of violent incident happened.
3. PROBLEM STATEMENT & RESEARCH
QUESTION
As seem in the previous section, there is interest in applications in violence and in minority field. Most applications mine personal information and interests, and the violence data are requested from other sources. Social media are an information source that can surpass barriers that countries and statistical organizations pose.
Many countries' data collection mechanisms and statistical reports provide insufficient information, and therefore they cannot make hate crime visible in large city centers throughout Europe. Even the Netherlands have better data collection mechanisms than other countries (Table 1), the annual reports on criminal descrimination (Criminaliteitsbeeld discriminatie) are not published because the Police’s National Expertise Centre on Diversity (Landelijk Expertisecentrum Diversiteit van de Politie) has closed on January 201530. Also, after a 6 week period through contacting the CBS the data sheet “Ondervonden delicten;
27 http://muzmatch.com/ 28 www.minderme.co/ 29 www.salaamswipe.com/ 30 www.politieacademie.nl/lecd
persoonskenmerken”31 was found, however the statistics are
generalized and cannot serve a location-based application. Moreover, as far as the different data collection mechanisms cannot cooperate, it is beyond refutation that a social media mining application for tracking the violence against specific minority groups would be beneficial.
Based on the difficulty to accumulate the required statistics for an informative violence against minorities application, we were led to investigate which data types can be actually retrieved by social media. A preliminary and significant question could be if an efficient data set could be extracted to support such an application (feasibility) and if so, which parameters could be extracted automatically. Also, in order to extract parameters the most efficient searching terms or statements should be identified. Another issue that needs to be investigated is about the frequency on which the results will updated and if that could be supported in near real-time.
The main research question intended to be answered is:
"Can Social Media support the visualization of violence against minorities on a local basis?"
This question can be fragmented in the following sub-questions: •Sq1: How can social media data support such an interface? •Sq2: What kind of data could be automatically supported by such an interface?
4. APPROACH
In this section is explained the approach that intends to answer the research question.
As discussed in the literature section, there is a problem on acquiring the violent-incident data that would support a violence against minorities application. In this paper it is aimed to use social media in order to surpass those problems and identify which data can be extracted automatically.
Initially, the violence and minority concepts we have our investigation on will be considered as a starting point. We assume that we have an application that intends to inform minority members about violent incidents against them. In this research we follow Fokaeos [7] application User Interface (Table 4 ) and its respective data requirements. The application has some initial requirements, the input has to contain where the incident happened and against which minority (Who), and also what and when it happened. Regarding the "what happened" question the type of violence is required, regarding the "when" the exact time the incident occurred, regarding where the exact point in the map and for which minority the type of minority is required. From this moment this set of data will be mentioned as: Where-What-Who-When (4W’s). 31 http://statline.cbs.nl/Statweb/publication/?VW=T&DM=SLNL& PA=83095ned&D1=5-31&D2=a&D3=0&D4=a&HD=160413-2147&HDR=G1,G2,G3&STB=T
Table 4. Application User Interface Fokaeos [7]
In this research, as an information source, we consider the social media user who posts about a violent action against a specific minority, the post contains specific data that construct the social media post description of the violent incident. This data is valuable for an application that intends to inform other minority members about violent incidents. Answers to the 4W’s application requirements will be given if they exist in a social media post. Given that social media users post on Social Media a description about a violent incident against a specific minority, the first step is to identify which social media are more likely to contain these posts. Another import issue is to examine if those data are accessible to developers through a specific social media API. In this research, first we define what we look for, this is an essential part and we call it as pre-analysis. The definitions and the categories of this part will be followed throughout this research. This data will be intended to be extracted from social media.
Before proceeding to the extraction, we conduct a survey with a primary goal to identify the keywords that social media users utilize in order to post about a violent incident. The survey also asks other useful information in order to investigate the users approach in this issue such as social media use and their eagerness to provide information.
As we pointed out each Social Media site has its own API32, we
look at all of them and we focus on one. Having chosen a social media API and identified the violence and minority keywords, we can extract relevant posts. After the extraction we initiate a basic filtering, this is also characterized as the pre-phase of the filtering as we get rid off of the non posts such as the replies and other. In this phase the non-geo-tagged posts are not deleted even if the location is a prerequisite for the application. In the second filtering phase, it is intended to extract the posts which have an incident based structure. For this phase Natural Language processing basic structures are used. Thereupon, the 4 W’s are attempted to be extracted. From the posts that survived, an investigation is examined on if those data can support the application.
Table 5 shows the process that we described in this part. The social media user reports an incident on a specific social media site. Before we extract this information we identify the terms that users utilize. Then we extract those information from the social media API and filter the data according to their context (phase 1) and relevance to the incident basis (phase 2).
32
www.programmableweb.com/category/social/apis?category=200 87
8
Table 5. Representation of the final system complemented with the detailed components.
Various approaches is tried to find if they give the same image, this is called as triangulation. All in all, this is the followed approach and in the remaining part of the thesis we look at each step of our investigation in more detail.
5. PRE-ANALYSIS
In this section, first the data structure is finalized (Section 5.1). Based on those data requirements the types of visualization were concluded (Section 5.2)
5.1 Definitions
As for the violence, the definition by WRVH6 is followed. Regarding the violent types the “Government of Canada” 7
approach is followed however within the context of minorities the physical, sexual, emotional, psychological, religious, cultural and verbal is selected because they are equivalent with Krug (2002) and Fokaeos (2013) but more detailed as it contains the religious, emotional and cultural type. Most researchers [7] [11] refer to the psychological violence as equal to emotional violence. In this research, minorities are defined by the 5 features of Feagin [13]. The types of minorities chosen is the Ethnicity, Race, Sexual, Religion, Disability [7] and illness10.
5.2 Define Data Structure
Through years social scientists change their view and the categorization of violence and minorities, as it is a societal evolving subject [10]. Thus, before starting defining, it should be considered that the categories in an application about violence against minorities are not stable and should be redefined and change through time.
Regarding the information that should be included, Fokaeos [7] conducted a survey among minority members in which they were asked to prioritize Krug’s nature of violence [11], only 7% of respondents would rank as important the deprivation incidents [7], while Physical, Sexual and Psychological noted 33%, 27% and 27% respectively [7]. This means that including these types of violence is highly recommended. Another categorization of the types of abuse is physical, sexual, emotional, psychological, spiritual/religious, cultural, verbal, financial and neglect 33. The “Government of Canada” approach is followed however within the context of minorities the physical, sexual, emotional, psychological, religious, cultural and verbal is selected because they are equivalent with Krug [11] and Fokaeos [7] but more detailed as it contains the religious, emotional and cultural type. Most researchers [7] [11] refer to the psychological violence as equal to emotional violence. In this research it is intended to make a distinction between them as psychological violence focuses on provoking fear with threats in order to gain control, whereas the emotional violence only affects the victim’s feelings and self-esteem. Regarding the choice to consider the cultural violence, it has to be stated that many cultures punish because of their culture’s principles and it “occurs when a person is harmed as a result of practices that are part of her or his culture, religion or tradition”33
.
As far as we know which data is used in the current applications and literature (Section 4.1), it is intended to investigate which data has to be requested in this research. Afterwards the data requirements will be defined in order to investigate if these data requirements remain almost the same throughout the research. The parameters defined from this step will be utilized in the next one which is the data mining. It is intended to examine on if all these data can be extracted automatically from social media.
By considering the definitions (Section 3.1, Section 3.2), an initial data structure set of requirements follows. The main data requirements, such as where, what and when an incident happened remain the same in almost every application, however some additions and exclusions of data exist. It is considered that the external appearance of the victimizer is of interest of the public rather than of the victim. It would be ideal to mine data about the incident in order to inform the users. By considering these, the data structure will be defined as:
Violent Types 1. Physical 2. Sexual 3. Emotional 4. Psychological 5. Spiritual/Religious 33 http://www.gov.nl.ca/VPI/types/index.html
9 6. Cultural 7. Verbal Abuse Minority Types 1. Race 2. Religion 3. Ethnicity 4. Sexuality 5. Disability 6. Illness Location (Where)
Time and Date (When)
Description of the incident (What)
Number of victims
Number of offenders
Description of victimizer (sex, age, external appearance)
Description of the victim (sex, age)
In the next parts this data structure is followed and those data are tried to automatically harvest from the social media API.
5.3 Define the type of visualization
In a survey Fokaeos [7] asked the users, members of minorities for the more helpful visual representation of the incidents. In this survey the visual representation was preferred by 87% of the users, while only 7% preferred sound or haptic. Another important finding is that the 67% of the users state that it should be located in a neighborhood level and be colored according to the minority classification. Also 73% of the users wanted to pin the incidents. All participants stated that they would prefer to have access to combination of data [7]. This means that the introduction of a visual filter can be important.
First of all crime and, secondly, minority applications are examined regarding their current visual representations. This approach is followed to show how data must be visualized regarding the required data structures that were defined in the previous step of the methodology. So, the main aim is to utilize a methodological approach to visualize the data in order to enhance its dissemination, accuracy and ease of comprehension [20]. Almost every mobile or web application that aims to inform the users about a violent incident in a city or a country level use a map 3 Error! Bookmark not defined. Error! Bookmark not
defined. [4][8][5][6]. Crime mapping through Graphic Information Systems (GIS) is widely used as to map, model, query and analyze data 34. Tuft’s (1990) points provided significant influence in the transition from data to visual representations. In this visualization, it is aimed to “escape from the flatland” [22] by increasing the dimensions of the incident map from latitude and longitude to add also time, and areas grouped by the emotions and the minority areas, this will result in a desirable augmented amount of information per unit area. Also, the “Micro/Macro Readings” [22] attribute will be implemented through the description tile, raised when clicking in a pin of the map to clarify and add detail. The tile description will appear only if the user
34
www.colorado.edu/geography/gcraft/notes/intro/intro.html
chooses to gain insights on the incident description, the type of violence, the number of victims and offenders. The ” Layering and Separation” [22] attribute was regarded in order to clarify the relationships among information layers and not lead to confusions, thus it needs to be clarified that only the violent incidents against specific minorities are examined. It is important to use the “Small Multiples” [22] point to underline the change by making comparisons between the violent types through time. Undoubtedly, color will provide in an easy way information about the different emotion and minority kinds and their respective extend-area projected in the map.
The information should be filtered and made visible through a map with pins that represent a specific incident, according to the users’ preference. The filter would consider the type of violence , the minority type and the year in order to let the users customize and personalize the data for their own use.
Violent Types
The violent type will be displayed as a part of the incident description of the pins of the map. There will be also a different frame in which the data will be clustered per violent type. The different violent types will be also compared in the timeline
Minority Types
The minority type will be displayed as a part of the incident description of the pins of the map. Zones of minorities will be displayed in order to make the crime against them visible.
Location (Where)
The location will be visualized through a map. Time and Date (When)
The time will be visualized through a timeline and will be shown in the incident description tile. There will be also a time option in the filter in order to eliminate the results according to time.
Description of the incident (What)
The description of the incident will be visible when the user clicks to a specific pin.
Number of victims In the description tile. Number of offenders
In the description tile.
Description of victimizer (sex, age, external appearance)
In the description tile. It is considered that the external appearance of the victimizer is of interest of the public rather than of the victim.
Description of the victim (sex, age) In the description tile.
Given the parameters defined here, in the next section we will utilize them in order to come closer to their use in social media. With these grounds, through a survey we aim to gain insights on if these data can be automatically extracted from social media extraction, in order to support an application that intends to inform users about violent incidents against minorities.
10
6. SURVEY
In this part it is intended to identify what people think in case they had available an application that informs them towards violent incidents against minorities. We aim to come closer to user behavior and identify which of them could be automatically extracted as well as which terms is more likely to be used when users are talking about those incident concepts on social media.
6.1 Purpose
It is essential to reveal social media users’ behavior in the context of violence against specific minority groups. To capture an incident of violence against a specific minority, the concepts and the specific words must be identified through a survey. Through questions it is intended to gain a better understanding of the concepts and capture the words or phrases that social media users are more likely to include in a post about experienced violence. This aims to identify what type of mining terms the application should be provided with, in order to supply its requirements and establish an appropriate visualization of location-based violence. Furthermore, a keyword survey facilitates the process of prioritizing and thus determining the mining terms that will, potentially, bring us closer to how people express and thus to more complete and valid, to our initial requirement (4W’s), data sets. Social media user-centered keyword choices will assist to find enough violence against minorities incident traces without exceeding the rate limit 35 and thus avoiding delay.
6.2 Structure of the survey
To serve the preceding purposes the survey followed a general to specific structure in its questions and was divided in four parts. In the first part, a few demographic questions were asked in order to classify the respondents in specific groups and identify potential trends in specific minority groups. In the second part, the use of different social media and social media utilities was questioned. Therefore, the users were introduced to consider that they have available an application that aims at informing users about violent incidents against minorities in the current location and asked on being informed and on providing information according to seven types of violence (Physical, Verbal, Cultural, Religious, Psychological, Emotional and Sexual) and to whom the violent incident occurred (Themselves, Friend and Stranger). The definitions of the types of violence were mentioned in the questions in order to clarify respondents perception of questioned terms. Finally, they were asked to think that they are posting about a violent incident on social media and to select or write the words that are the most relevant to the terms or classes of terms in the context of reporting on different forms of violence against their minority.
6.3 Survey Distribution
The survey was distributed online and was spread through emails and social media from the 27th of April until the 9th of
June, 2016. No other limitation except from belonging in a minority was taken. Emails were sent to the COC Netherland36, Jewish community in Netherlands37, Muslim community of Germany, UK, US, Canada38 and many more. Among the minority 35 https://support.twitter.com/articles/160385 36 http://www.coc.nl/ 37 http://www.nik.nl/english/ 38 https://www.alislam.org/contactus/
groups that were approached on Social Media were Jews Against Circumcision39, European Forum Of Muslim Women40 Greeks in Amsterdam41, Gay Guide42, Erasmus UVA 2016-1743, A.M.E.A (Greek Disabled people)44, Blacks fighting back against racism45 and others.
6.4 Survey Results
The questionnaire was answered completely by 56 minority members. The results of the research are presented in the aforementioned categories and are completely illustrated in Appendix X.
The biological gender of the respondents is equally distributed with 52% of them being a male and 48% of them being a female. With respect to the minority that they can classify themselves 46% of the respondents selected ethnicity, either of itself (27%) or with combination with other minorities. The second most popular option was religion (29%) from the religious minorities th 18% of the respondents belong also in ethnic minorities. Another 21% chose the “Sexuality” option while 12% the “Disability” and other 9% the “Race” option. In the “Other” option respondents indicated sexism and political.
Regarding use of social media of minority members, the vast majority of respondents has a Facebook account (96%), second ranks the business oriented network service LinkedIn (73%). Approximately half of the minority respondents have a social media account in Google+ (45%), in Twitter (43%) and on Instagram (43%), whereas all the Pinterest (21%) and Tumblr (13%). The frequency of social media use is another important factor, the respondents use more often Facebook (83%), as a second choice is LinkedIn (30%), as a third Twitter (30%) then follow Google+, Instagram, Pinterest, Tumblr in a descending sequence of preference. It can be observed that Twitter is ahead from Google+ as 17 respondents opted Twitter as their third choice instead of 9 who selected Google+. Another question was whether minority members use hashtags with 41% of them to use considering that only 44% of them “hashtag” with related to the content words. Moreover, 25% of the respondents have a public account, from those 25% have Facebook, 21% Twitter and Instagram, it should be regarded that 11% opted “do not know” in this question. Concerning the location utility, 32% of the respondents have their location utility on and 9% opted “Do not know”. In the case of posting on social media about a violent incident that occurred to a stranger, to a friend, and to themselves, the percentages decreased and noted 57%, 36% and 32% respectively.
If they had available an application that aims at informing users about violent incidents against minorities in the current location, 39 https://www.facebook.com/JewsAgainstCircumcision/?fref=ts 40 https://www.facebook.com/EuropeanForumOfMuslimWomenEf omw/?fref=ts 41 https://www.facebook.com/groups/greeks.in.amsterdam/?ref=gro up_browse_new 42 https://www.facebook.com/gayguide.gr 43 https://www.facebook.com/groups/391656567688570/ 44 https://www.facebook.com/groups/360660249398/?ref=group_b rowse_new 45 https://www.facebook.com/groups/1457450201144571/
73% of the potential users would like to use it in order to get informed. While 32% would not be eager to provide information
about the violent incident.
Trends Physical Verbal Psychological Cultural Sexual Emotional Religion Occurred to you 66% 62% 55% 54% 52% 50% 45% Occurred to a friend 71% +5% 66% -4% 53% -2% 59% +5% 55% +3% 55% +5% 59% +14% Occurred to a stranger 77% +6% 71% +5% 60% +7% 66% +7% 68% +13% 60% +5% 66% +7% Average 71% 66% 56% 59% 58% 55% 57% Ranking 1st 2nd 3rd 4th 5th 6th 7th Occurred to you Physical 66% Verbal 62% Psychological 55% Cultural 54% Sexual 52% Emotional 50% Religion 45% Occurred
to a friend Physical 71% Verbal 66% Cultural 59% Religion 59% Sexual 55% Emotional 55% Psychological 53% Occurred to a stranger Physical 77% Verbal 71% Sexual 68% Cultural 66% Religion 66% Psychological 60% Emotional 60% Average Psychological 71% Verbal 66% Cultural 59% Sexual 58% Religion 57% Emotional 57% Psychological 56%
Table 5. Eagerness to provide information about a violent incident regarding to the relation with the victim and the type of violence.
We also analyze the eagerness of the users to provide information of specific types of violence (Table 5). From the survey results, it was clear that users prefer in general to provide information (68%). However, the eagerness to report a violent incident differs according to violent type (Physical, Verbal, Cultural, Sexual, Religion, Psychological, Emotional) and according to the victim (themselves, friend, stranger). The prevalent violent types that respondents are more eager to report is the Physical and the Verbal Violence in every victim category and with an average of 71% and 66% respectively. In the subsequent positions of the ranking the results differ, per victim type, with an average of 59% cited Cultural, 58% Sexual and 57% Religious Violence. Regarding a violent incident that occurred to themselves, the third highest rating goes to Psychological violence which is quoted by 55% of the poll sample while 54% cited Cultural, 52% Sexual, 50% Emotional and 45% Religious Violence. In the case of a violent incident occurred to a stranger, the order of preference changes, Sexual violence scored 68% while Cultural, Religion, Psychological and Emotional violence achieved 66%, 66%, 60% and 60% respectively.
As can be seen clearly, the most striking feature of this question is that there is a trend to be more eager to provide information regarding a specific violent incident that occurred to a stranger than to a friend and even less about themselves. More specifically, 66% of respondents would provide information about a violent incident that occurred to themselves and its type is Physical
Violence, the percentage increased by 71% for a friend and 77% for a stranger. Similar rising rates can be identified in Verbal Violence, Cultural Violence, Religious Violence, Emotional Violence and Sexual Violence. A minor exception is the Psychological violence which was quoted by 55% of the respondents and slightly declined by 2% and then increased to 60%.
In the third part of the survey, the words that are the most relevant to the terms or classes of terms in the context of reporting on different forms of violence against their minority were questioned. Regarding the related terms to violence ,”Attack” and ”Abuse” have the highest rating as each is being mentioned by the 75% of respondents. The term “Threat” is chosen by the 61% of the respondents, while “Assault”,” Crime”, ” Rough Treatment”,” Offence” are stated here in descending order and were mentioned by above the 30% of minority members. Similar, regarding the Physical violence incidents, the leading terms were “Hitting”, “Injury”,” Beating”, “Weapon” quoted by 65% of respondents, 63%, 63%, 63% respectively, while “Murder”,” Stabbing”, “Punching”, “Strangling, ”Pain”, “Kicking”, “Burning”, “Slapping”, “Choking”, “Throwing” scored above 30% of the survey sample. In the cultural violence field, “Sexual slavery” and “Female circumcision” was rated by 76% and 62%of respondents respectively.
Thereafter, the classes of terms were surveyed. Regarding the sexual violence, the amount of responses skyrocketed, as the 83%
12 of the participants mentioned “Forced sexual intercourse” and mentioned relevant words such as “Rape”, high rates were also denoted in the classes of terms “Forcing a person to perform sexual acts that may be degrading or painful”, “Forced prostitution” by 74% and 70% of participants respectively. Of the members of minorities surveyed, 78% chose in the emotional violence field the “Humiliating or making fun of the person” while 76% selected “Intimidating the person, causing fear to gain control”. With regard to the psychological violence, the leading class is “Threatening to harm the person or her or his family” given by 85% of the respondents. In relation to the verbal violence “Threatening” and “Insulting” were the dominant classes of terms being rated by 72% and 70% respectively of the minority members. In the religious violence field “Not allowing the person to follow her or his spiritual or religious tradition” and “Forcing a spiritual or religious path or practice on another person”” was opted by the 79% and 71% of the people surveyed.
Taking into consideration all the aforementioned survey results, estimations about the social media mining can be projected, however the most important finding from this survey was the mining terms that will be utilized in the next step of this research which is the Twitter data mining.
6.5 Decide & Set Keywords
The results of the survey let us prioritize the key words according to popularity and determine which classes of terms were preferred instead of others. The keyword terms are used directly for the mining.
Regarding the classes of terms, conclusions on keywords were made upon the filled fields, however some classes of terms were reviewed as irrelevant. For example some respondents in the Religious violence topic, were influenced of the choices so much that were led in filling out feeling terms such as “Unfair” and “Cult” instead of concept descriptive. However those terms described their feelings instead of the terms. Moreover, some times the respondents filled out general terms that could not describe distinctively that kind of violence such as the “Degrading” term in the Psychological violence field.
Also, in the case of there was not substantial input, individual terms research was conducted. Finally, the terms are prioritized because of Twitter Streaming API restrictions 46 . The prioritization made, was according to the users preference regarding the terms and the concepts. On the concept fields the terms that were filled out by the respondents had the respective fconcept weight. As a result, the ”*” means that all the suffixes are covered, for instance for the term violen*, violen-ce and violen-t is covered. Table 6 presents the most popular words of the survey of each category in a descending order, as key set in the upcoming input of our research.
Overall, the terms’ list that will be utilized for the social media mining is in Table 6. In the first column is depicted the category violence and in the second the prioritized set of terms. From this moment a word that belongs in a violence category is reliable for tagging this category. These key terms will be used to capture the posts, the process of the data mining is described in the next part.
46
https://twitter.com/Politie
Categories Prioritized terms
Violence Violen*,Abus*, Attack*, Threat*, Assault, Crim*, Rough Treat*, Offen*, Forc*, Insult, Harm, Bull*, Accuse, Harassment, Harm
Violent types:
Prioritized terms
Physical injury, hitting, beating, weapon, murder, stabbing, punching, strangling, pain, kicking, burning, slapping, choking, throwing, pushing, biting, arm-twisting, pinching, hair- pulling, shoving Sexual rape, coercion, trafficking, intercourse, forced
prostitution, stalker, humiliating, without consent, exhibitionism, leering, obliging, touching inappropriate, abuse, assault, tease, sexist. Emotional humiliate, intimidate, blame, disrespect, bully,
dominating, threatening, fear, asshole, prohibit, destroying possession.
Psychological blackmail, psychological disorder, bully, isolate, abuse, harassement, control, maltreatment. Religious disrespect, proselytize, abuse, belittling, overpower
Cultural offence, discrimination, oppression, female circumcision, slavery, rape marriage, murder, lynching, stoning, witchcraft.
Verbal threaten, blackmail, insult, yelling, swearing, cheap, worthless, blame, accuse, minimize, damn, blocking, trivializing, degrading, abusive .
Table 6. Prioritized key terms
7. SOCIAL MEDIA MINING
The proliferation of social media and the large amounts of data uploaded daily by all kinds of users is a powerful information source. This has orientated us to examine on if the data can be mined by the most popular and relevant social media in the context of minorities. The selection of them will be made based on multiple factors that will be analyzed thoroughly. Until this moment, the anticipated social media are Facebook, Twitter and Instagram , on them a tool analysis will be executed. Then, it is aimed at finding which parameters can be mined from them. This step requires multiple information retrieval experiments. Once mined, a distinction will be made on the data that are considered important and unimportant. Otherwise, if the result will be complete datasets of the required information, then it will be concluded that the investigated API is potentially providing the information required. After the analysis of the APIs it is intended to identify if the amount of the data that were mined is sufficient, and their quality too in order to start building an analysis demo for data mining. Then there will be a realization of the actual percentage of the data mined, compared to the Fokaeos [7] required data.
The data is extracted from the Twitter Streaming Application Programming Interface (API). Twitter is chosen because of its broad user base and public nature. To address these questions various aspects of the data present within the 140 characters in discrete posts were coded and analyzed
13
Parameters Is it expected to find data on Social Media? Violent Type Yes, with the restriction that only keyword related
material will be mined. Minority
Type
Yes, with the restriction that only keyword related material will be mined.
Location (Where)
Little possibility, in previous research only 2% of the bullying traces was geotagged [39]. May lead to sample bias [41]
Time and Date (When)
Yes, Tweets. Questions about the actual date the incident happened.
Description of the incident
(What)
Yes, if and only of the incident is identified.
Number of victims Potentially Number of offenders Potentially
Table 7. Preliminary expectations for Twitter Streaming API
Of course, whatever was asked in the survey was related to the incidents, as our intention is to identify the violent incidents against the minority groups.
7.1 The structure of a violence against
minorities incident
In order to lead to a successful contextually data mining, an analysis of the structure of the violent incident against a specific minority is regarded of paramount importance. An event is something that happened 47, in this context a violent episode against a minority member. An incident is a situation that is accompanied by its unique characteristics, those components will be analyzed in the next part.
7.1.1 Who is the victim and on which minority
belongs.
One of the most significant attributes of an incident is who the victim is and more specifically, regarding the minority context, to which particular type of minority he or she belongs. Primarily there is a social media user who reports through a Tweet a violent incident against a specific minority, there are options to be the victim itself, a friend or a stranger [7]. It cannot be excluded though that offenders may describe and boast about their achievements in social media, because the offenders can be information contributors through their hate speech [3][23]. Notwithstanding, victimizers who are members of the same minority are not regarded [24], as violence directed to a minority itself cannot be considered against the minority. This does not restrict that an inter-minority violent behavior has to be extracted because this is in line with the purpose of our research which is to inform the minority members about violence against them. On the grounds that we have previewed the potential involved individuals, the language structure is determined. Having defined the language structure, the minority types and their respective keywords, those keywords are used to deduct to the victim’s type
47
https://en.wiktionary.org/wiki/event
of minority. For instance, the word transsexual resides from the LGBT minority type, hence this Tweet can be classified in the LGBT category.
7.1.2 What form of violence against minorities is
mentioned on Twitter.
Another important theme needed to investigate is what form of violence is mentioned in Twitter. As previously defined, violence against minorities includes many forms of violence such as Physical, Sexual, Emotional, Psychological, Spiritual/Religious, Cultural, Verbal. Except from the Physical violence which can only be practiced offline, every other violent type can be also mentioned online. Again in this part a deductive way of classification was used by matching the defined keywords in the respective types of violence. In this way, through the mining it is investigated whether these genres of violence can be detectable in Tweets.
7.1.3 Where did the violent incident happen.
In order to support a location based application which intends to inform the minority members in a neighborhood level, the location of the violent incident must be identified. An answer must be given to where was the victim attacked followed by whether the incident happened online or offline. In case of an online harassment , answer must be provided on where the individuals who post are. This provides information of a potential trend of specific areas where people seem against specific minorities. Thus, through geo-tagged Tweets, the exact location of the person who posts is identified. In case of an offline violent incident standard natural language processing techniques are utilized. The location identification in Tweets can result in a real time information source and may reveal continuously updated hotspots of violent activities against specific minorities. This will facilitate our understanding of the spread of violence against minorities across the world and in our research in Europe.
7.1.4 When did the violent incident happen.
The date and time dimension of a violent incident against a minority must be specified, in order to let the application users create a holistic or a more specified view throughout time. The when values are also used to identify the trends of violence . Insights will be also given on when are people posting about them. The time mining process is the same with the location as previously discussed.
7.1.5 Why people may report about violent incidents
against minorities.
Deep insights must be taken on whether Twitter constitutes an information source that contains the violent incidents against specific minorities and why individuals are posting on Twitter. As discussed, individuals use Twitter to talk about their daily activities and to seek or share information [25] and socialize with other users through it. Moreover, members of minorities may seek social support through reporting a violent incident against them, and victimizers may aggress against others [23]. Knowing why people Tweet is an asset that can bring us closer to the user behavior and thus gain insights on the extraction results.
