Islamophobia and Natural Language Processing update (May 4, 2020)

Over the last few months we’ve been working on an annotation scheme for Islamophobia as it occurs in written text. Progress has been slow given the scope and depth of Islamophobic expression. Beyond that, like the rest of the world this project has been significantly impacted by global pandemic, and will no doubt continue to be. But we did want to let you know that we are continuing with this work and welcome additional perspectives and ideas.

One of the main difficulties is that there is no single or simple definition of Islamophobia that is universally agreed upon, and there are subtle variations even if one settles on a definition.

One of the earliest formal definitions was created by the Runnymede Trust in the UK in 1997. Eight specific characteristics of Islamophobia were identified, and these were found to express themselves in unfounded hostility towards Islam, discrimination against Muslims, and exclusion from political and social life. The Trust revisisted this definition in 2017 and updated it to define Islamophobia as “anti-Muslim racism.”

This may at first seem confusing since Islam is not a race, but this raises several important issues that go beyond Islamophobia. There is no scientific basic for race, rather it is more so a social construction. However, groups of people can be “racialized” in that they are categorized as being somehow distinct or different, and therefore become an “other” that can be used as an identifiable target. In that sense Muslims are seen now as a group that has been racialized, and can therefore be subjected to racism.

Racism causes individuals or groups to be discriminated against and targeted for oppression. In her ICLR-2020 keynote Ruha Benjamin paraphrased Ruth Wilson Gilmore and defined racism as “the fatal coupling of power and difference that creates a vulnerability to premature death.” This definition makes clear that racism (and Islamophobia) is not simply a matter of ignorance or intolerance, it is often a deliberate use of power to harm or control a marginalized population.

There are other definitions of Islamophobia, and all they all share common elements. The Bridge Initiative at Georgetown University defines Islamophobia as an “extreme fear of and hostility towards Islam and Muslims” that can lead to “social and political discrimination.” The Council on American-Islamic Relations (CAIR) also defines Islamophobia as anti-Muslim racism, and articulates impacts that include physical attacks, discrimination, the curtailment of civil rights, and the promotion of anti-Muslim policies.

There is a tendency to think of Islamophobia as exclusively coming from the far right — however, there are also anti-Muslim biases that come from the left. These often represent Muslim women as oppressed victims in need of liberation and saving. In addition, Asma Uddin points out in her July 2019 book “When Islam is not a Religion” that Muslims are not granted the same rights of religious express as Christians in the United States.

Islamophobia encompasses a wide variety of forces and prejudices. It is a highly intersectional problem (Crenshaw, 1991) in that the gender, race, immigration status, class, and country of original all impact upon the biases, discrimination, and threats that a person or group that is perceived to be Muslim may experience.

Given all of these factors a simple annotation scheme or set of rules to decide if a statement is Islamophobic is not likely to exist. Despite these difficulties there has been progress in determining how to annotate text for Islamophobia and problems related to it, particularly hate speech and abusive language.

Vidgen and Yasseri (2018, 2019) arrive at a definition of Islamophobia which considers it a form of hate speech that expresses “indiscriminate negativity” against Muslims and Islams. They classify such statements based on whether they are strong or weak expressions — weak expressions may reflect misconceptions or be in some sense offensive, while stronger expressions are more explicitly hateful or threatening. Vidgen and Yasseri’s annotation scheme is described in a blog post and working paper from December 2018 and a journal article from December 2019.

Beyond strength of negative expression, another important axis is whether such expression is directed against an individual or a group. While not specific to Islamophobia, this distinction is captured in the OffensEval 2019 and 2020 tasks held in conjunction with the International Workshop on Semantic Evaluation (SemEval) where the goal is to identify tweets using offensive language. A corpus of tweets annotated with targeting information is available via Zampieri, et al., 2019.

Prior to both of these works, Waseem, et al., 2017 propose a general typology of abusive language which would classify such statements as being implicit or explicit, and being targeted or not. The consensus this work reflects is that the annotation of hate speech should include both the intensity and direction of such expressions.

Missing thus far is an indication of the producer of a potentially Islamophobic statement, and whether that source is in a position of power. This seems necessary to recognize the potential threat level or danger posed by a statement. Does the speaker command the respect of a sufficient number of followers to influence their actions and attitudes? While any individual may utter hateful and Islamophobic sentiments, there is a crucial difference in potential impact if that individual occupies a position of power. Within social media the power of an individual may be reflected by the number of friends or followers, however this is far from reliable.

While there do not seem to be annotation efforts that directly address the nature of content producers, Awan, 2014 presents a qualitative “typology of offender characteristics” of those who propagate Islamophobia on social media. Vigden, et al., 2019 collect tweets from all the followers of a far-right political party and are thereby able to make some assumptions about the nature of these followers and the content they share.

Beyond questions of who produced content, how intense was its expression, and who was it directed to or against, the more basic question of what should be considered an Islamophobic expression remains unanswered. The 1997 Runnymede Trust criteria present a useful starting point. They identify eight characteristics or beliefs that underlie Islamophobic thought and expression :

  1. Monolithic: There is only one form of Islam that all believers follow.
  2. Separate: Islam is an “other”, it shares nothing with other cultures
  3. Inferior: Islam is backwards, primitive, sexist, and inherently inferior.
  4. Enemy: Islam is a violent threat that seeks to win a “clash of civilizations.”
  5. Manipulative: Islam is not a true religion, instead it is a political ideology used to further worldly ends.
  6. Criticism of the West rejected: Muslim perspectives on the West inherently wrong and misguided.
  7. Discrimination defended: Discriminating against Muslim is necessary to protect and preserve existing culture.
  8. Islamophobia seen as natural: Islamophobia is viewed as a rational reaction.

This represents an important foundation for evaluating Islamophobia, and many subsequent definitions include some of these same factors. It should be noted that characteristics 1–5 are projections upon Islam and Muslims, whereas 6–8 are (for the most part) reactions of the producer.

Another set of characteristics of Islamophobia was identified by Nadel, et al., 2012. This focuses on subtle and overt micro-aggressions directed against Muslim Americans, and was studied via focus group interactions with Muslim Americans. A number of themes emerged.

  1. Endorsing Religious Stereotypes of Muslims as Terrorists: Muslims are evil, violent, and not loyal.
  2. Pathology of the Muslim Religion: There is something wrong with Islam, in particular women wearing hijab or other traditional dress.
  3. Assumption of Religious Homogeneity: The assumption that all Muslims have the same beliefs and practices.
  4. Exoticization: Viewing a religion as exotic or trendy, worthy of imitation because of superficial characteristics such as fashion.
  5. Islamophobic and Mocking Language: Making fun of Islam, and teasing people who believe.
  6. Alien in Own Land: The assumption that Muslims are not loyal, that they have a secret allegiance elsewhere and are working to undermine the current system.

As another example, The Bridge Institute provides a list of common anti-Muslim tropes seen in the news and other media:

  1. Islam and Muslims are inherently violent.
  2. Islam and Muslims are oppressive to women.
  3. Islam and Muslims are intolerant toward other religions.
  4. Islam is a political ideology or totalitarian regime, not a religion.
  5. In the West, Muslims are using non-violent, stealth jihad with the goal of implementing sharia law.
  6. Islam is medieval, foreign, and odds with Western modernity.
  7. Islam is a monolith.
  8. All Muslims are Arab and/or Brown.

There is also the issue of how Muslims are portrayed in popular media, in particular television and film. Many of the Islamophobic expressions mentioned already are widely used. Haqq and Hollywood (2018) and the Riz Test document many such examples.

We see some common themes emerging, particularly that Muslims are different and represent an “other”, and that they can not be fully trusted by the larger society. Islam is seen both as a deviant religion and not a religion at all but rather a political or military force bent on conquest.

Our goal now is to arrive at a single annotation scheme that reflects some (but not all) of what we describe above, and to do so in such a way that reliable reproducible annotation is possible.

We’ll post more about this as it develops. In the meantime, please feel free to read our earliest posts about this project, and also get in touch if you have ideas you’d like to share, questions you’d like to ask, or really for any other reason at all.

For more on the Islamophobia and Natural Language Processing project :

2019 updates:

  • July (project kickoff)
  • August (background reading, Ilhan Omar, Minnesota)
  • December (background reading, Genocide)

Please stay in touch!

Computer Science professor at the University of Minnesota, Duluth. Natural Language Processing and Computational Linguistics. http://www.d.umn.edu/~tpederse

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store