Machine learning countering hate speech

We all have heard of hate speech. It’s prevalent in our current western society, having gained prominence after the rise of various nationalist parties in Europe and an increasingly polarized political field in general.

Hate speech is described in The Council of Europe’s Committee of Ministers Recommendation as follows:

“Hate speech covers all forms of expression which spread, incite, promote or justify racial hatred, xenophobia, antisemitism or other forms of hatred based on intolerance.“

Hence, hate speech is an umbrella term for many forms of offensive language, but in Finland has no legal definition. Instead, in legislation, hate speech falls under defamation or incitement to hatred against a group.

In recent years, social media has proven a most prolific environment for hate speech. It is practised not only by individual users, but by various actors from political parties to flat out racist groups. While most of us agree that hate speech in itself is bad and thus a problem, I will not go into the multitude of ethical and actual problems that follow, but instead focus on the methods available to counter it, namely machine learning.

To put it simply, machine learning is a method where a computer is given a large data set of existing material, messages in this case, of which each message is rated by a human as containing, in this case, either hate speech or not hate speech. With a large enough data set the machine eventually “learns” to recognize messages containing hate speech and is thus automagically able to pick them out. After that we, the humans, are able to do what we deem necessary.

A simplified human comparison of the process could be learning to recognize a certain bird, say crow. Anyone can probably recognize the common crow when seen close up in broad daylight, but what about a dark evening, when you only see it very briefly? You probably couldn’t tell, or would be very unsure at the very least. However, after seeing the bird time and time again in such circumstances and being told that it is indeed a crow, you would eventually “learn” to recognize it yourself. There would be times when you are wrong though, and that is the case with a computer as well.

Utilizing such methods is not only useful, but rather the only option. On Twitter alone, over 700 million tweets are posted daily, essentially meaning that the numbers are way beyond human moderation. Trying to filter out hate speech from such a huge pool in, at least close to, real time is something only a machine is capable of achieving. And the good news is that this is working. In 2018 tech companies including Facebook, Google and Twitter managed to remove 72 percent of hate speech on their platforms, the EU found. This is significantly up from the past years, where Facebook, for example, managed to remove only 26 percent of such content. [source]

What this all means, is that we have a meaningful way of tackling hate speech online. While it still isn’t perfect, the algorithms are getting better, literally, by the minute. Naturally, there will be countermeasures from the people deliberately spreading hate speech, but that is something we are going to have to deal with then.