Hate speech [detection] as a service

These are notes from a talk about hate speech and the technological and linguistic problems that arise in trying to define and detect it. This talk given by @TimothyQuinn, technology lead at Hatebase, at the Beyond Tech conference, CodeNode, London. Discussing hate speech is difficult without looking at hate speech so triggering examples may be in these notes.

The general consensus is that hate speech is on the rise. While it should be noted that this is happening at the same time as online speech in general is on the rise, there is sufficient evidence that hate speech is increasing, particularly on social media and other online communities. A study from the University of Warwick shows that there are strong statistical correlations between hate speech and violence (Source: 'Fanning the Flames of Hate: Social Media and Hate Crime' by Karsten Müller and Carlo Schwarz). This analysis was performed using Hatebase data.

No one would disagree that there is value in tracking hate speech on a national level but it is also valuable to do so in online communities. Different companies are finding commercial value in ending hate speech. Firstly, it drives away users but also there are increasing legal liabilities. The UK has discussed fining companies 4% of global revenue for keeping hate speech on a platform; Australia has discussed 10%. It is also demotivating for developers. Despite this, most moderation is done manually. Around half of Facebook's content is being moderated by human beings as of September 2018.

Despite significant gains in natural language processing and artificial intelligence, automatically detecting hate speech is extremely difficult. It requires an understanding of context. It cannot just be a library of banned words. For example, the word 'aboka' in Nigeria varies by context. when a Hausa talks to another Hausa it means 'friend', when another ethnicity talks about a Hausa, 'aboka' means uneducated. Additionally, language contexts are more complicated than just meaning. Satire, homonyms, misspellings and ambiguity make it difficult, though not impossible to identify hate speech. As such, for companies who want to end hate speech in their platforms, they either need to build tools to identify hate speech and keep them updated as language evolves or hire a service which does this for them. Hatebase was created to cater to these companies. Hatebase's goal is to make removing hate speech from platforms as simple as installing antivirus software.

Hatebase is currently being used by some of the world's largest social media platforms, by law enforcement, governments, publishers, academia and nonprofits. They currently ingest around 10k unique datapoints every 24 hours. And now they have almost 100 languages tracking in over 200 countries. This data comes from different online communities. Most users interact with the Hatebase API and researchers use Hatebase to understand how hate speech propagates. Online communities use Hatebase as a first line of defence. Creating a technical solution to a hate speech problem is preferable to having humans work on reducing hate speech as it can be a stressful and emotionally challenging task to read and identify hate speech. It is currently a job which is offshored to countries with employees with low protections and qualities of life.

How it works: it is built around a NLP engine which recognises hate speech terms, even if obfuscated, eliminates homonyms using rudimentary language detection, recognises clinical, non-hateful, contexts and can return the probability that a piece of text contains hate speech looking at indicators (e.g. blacklisted emoji, xenophobic references, negative adjectives, etc.) which are likely to come with hate speech.

So what is hate speech? Insults? jokes? Criticism of a specific country? What about if a group tries to re-appropriate a term to become more positive? Does something have to be a threat? What about general remarks about specific targeted groups? Justice Stewart once said of hate speech "I know it when I see it." It's a fair remark but hard to explain programmatically.

Hatebase's definition of hate speech:

Any expression, regardless of offensiveness, which broadly characterizes a specific group of people based on malignant, qualitative and/or subjective attributes -- particularly if those attributes pertain to: ethnicity, nationality, religion, sexuality, disability or class.

Finally, there are people who accuse Hatebase of ending free speech. Hatebase does not support censorship or the criminalization of speech but they do have a few caveats. They believe that online communities and platforms have the right to moderate user activity and individual rights to speech do not carry over to using private companies' platforms to proliferate that speech. They believe that everyone has the right to express their own opinion, unless it carries over into a threat. Lastly they believe that governments have a responsibility to monitor hate speech that can potentially lead to a threat.

Hatebase is available here: Hatebase.org

You can watch this talk for yourself on the skillsmatter website. You will have to register to their community, it is free to do so and if you're London based, I can strongly recommend some of their events.