Tracking down hate speech

We are still a long way off from developing artificial life forms that can understand and communicate with humankind on all subjects, according to Michael Wiegand, computer linguist at the Digital Age Research Center (D!ARC) at the University of Klagenfurt. He specialises in the detection of hate speech and tells us what algorithms need to learn in order to reliably recognise insults.

On the one side we have the machine, which works with formal language, and on the other we have natural language. What complicates the communication between the two?
Natural language is notoriously ambiguous, whereas formal language, as found in programming languages, for example, invariably works with unequivocal commands. Let’s take the term “bank” as an example. If I casually mention it, you won’t know whether I’m talking about a financial institution or a seating arrangement. But if I frame it in a certain context, things become clearer. This sentence illustrates what I mean: “I just went to the bank to withdraw some cash.” The real difficulty lies in describing the meaning of the term within its given context in such a way that a computer is able to interpret it. That is our fundamental challenge.

Can computational linguistics offer an approach to teaching a machine the correct interpretation of ambiguous terms?
We often come across a major misunderstanding with respect to computational linguistics. In our field, we do not aim to build a programme that universally comprehends all facets of language. That would be far too complicated. Instead, we concentrate on more narrowly defined areas, where we can make valuable improvements that will benefit many fields of application.

In your opinion, is the notion that we will one day be able to discuss world affairs and politics with a robot still somewhat utopian?
Yes, we are still a long way off that point. If we look at chat bots like Alexa or Siri, we can see that they are closely tailored to specific applications. The programme can tell me what the weather will be like tomorrow or it can read aloud from my schedule. But it can’t hold a conversation on an arbitrary topic. To render those kinds of components possible, we need more than linguistics, we need to imbue the machine with world knowledge. There are a few debating programmes, research prototypes, which can deliver pro and con arguments on certain issues, based on representations in corresponding Internet forums. In other words, they are solely capable of reproducing what a person has already voiced or written. The encoding of world knowledge, which is essential for any realistic conversation conducted on equal terms, is extremely challenging. I have yet to see a promising approach to this problem.

One of your core areas of research revolves around the automatic detection of hate speech, i.e. offensive statements, in the digital sphere. Where do you see the opportunities and the limits of machine recognition in this context?
Basically, we have to acknowledge this: An algorithm is not capable of thinking any better than a human being. The advantage of automation is that a machine is often able to work more quickly and more comprehensively in terms of quantity. Where two people disagree over whether a statement is offensive or not, no computer programme will ever render a final decision that satisfies both sides. Given that there is no universal definition of what constitutes an insult, we first have to invest a lot of effort in manually tagging statements as “offensive” and “non-offensive”. Ultimately, this tagging relates not only to the subjective experience, but also to the prevailing social and cultural background. And this is the point where it really starts to get complicated.

Social networks like Facebook or Twitter in particular face a huge problem when it comes to recognising hate speech. Yet, in many countries, the legislator holds companies responsible for deleting such remarks or for pursuing criminal prosecution. What can be done to support companies in this regard?
Due to the sheer volume of posts, the platform operators are completely overwhelmed by the task of deploying human screeners to manually scan the websites for insults. The idea of hate speech detection is one of pre-filtering, whereby the human screener is presented with a pre-selection of potentially offensive remarks for her or his final evaluation. Of course, in many comments the insult is plain to see because of the expletives used; but often things are more complicated and for this we need a human being.

How complicated is it to search a platform for specific words?
Unfortunately, it is more complicated than most people think. First of all, we need to gather offensive terms. Even that is far from easy, since the collection ideally needs to be as exhaustive as possible – and also needs to cover different languages. In addition, language itself is in a constant state of flux and new terms such as “covidiot” are continually being added. 18 months ago, this term did not even exist. Today, it is associated with a very specific meaning. Some 70 years ago, certain expressions that carry racist connotations today were regarded as part of normal everyday communication. Clearly, natural language is highly dynamic.

It’s also possible to cause offence without using defamatory language, isn’t it?
Yes, it is possible to be insulting without using offensive words. In this context, we are dealing with a very heterogeneous group of statements. This is another complex task for computational linguistics. If you say, “You’re not very intelligent, are you?”, most systems will fail to recognise this as an insult. So, in addition to word recognition, we also need to make an effort to recognise the linguistic patterns that underpin remarks like these, which will help us to identify them as offensive. At present, we do not have the data sets necessary to teach this skill to a machine and subsequently validate the procedures.

What is the best way to generate such data sets?
Again, besides requiring data sets containing statements, we also need human beings to label these utterances as “offensive” or “non-offensive”. The issue of tagging is not unique to hate speech detection, it also exists in other problem areas addressed by computational linguistic. Once we have the connoted data, we use machine learning techniques that allow the machine – in a relatively autonomous way – to use the observations to discern signals, interactions and linguistic structures that make up an insult. So, once the machine recognises that the term “dimwit” occurs in remarks that are judged to be offensive, it also learns from this that the word is likely to be insulting. It is important, however, that the learning techniques do not only refer to individual words, but also consider more complex patterns.

Does this guarantee accuracy?
At the moment that depends very much on the data set. We clearly have too few representative data sets to date which comprise adequate data for both “offensive” and “non-offensive”. Very often, learning methods are still finding random indicators. If, for instance, the term “allotment” occurs in ten offensive remarks and coincidentally does not appear in any of the non-offensive remarks included in the data set, then the method will conclude that “allotment” is an offensive word. The data sets involved are so vast that it is often difficult to find the source of the error and to eliminate random correlations. What we are looking to do, is to develop procedures that can reliably detect hate speech in general – and can do so independently of a specific data set.

Can you tell us more about the data tagging? Who performs this task?
Some years ago, along with a few colleagues, I set up a data base for German-language insults, which has since become well established in the community. To move beyond that to diversified data sets, we need to annotate on a much greater scale. One way to do this is through crowdsourcing services. There, people may, for a fee of course, assign certain tags to linguistic expressions. The advantage here is that we end up with more than one evaluation per instance, perhaps 5 evaluations, and thus we no longer have to rely so heavily on what one individual finds offensive.

It would seem that a lot of basic research is still needed to move forward here. In your estimation, how are big companies like Facebook or Twitter dealing with the problem of hate speech?
Linguistic research questions are of secondary importance for these companies, as they are primarily concerned with finding quick solutions to the problem of hate-filled comments on their platforms. It is their aim to automatically spot anything that might constitute a criminal offence. They find it easier to ignore elegant insults or ironic comments, both of which require a great deal of background world knowledge in order to be recognised as such, as they tend to be less grievous than blatant defamation. In essence, the role of computational linguistics within the academic sphere is to generate fundamental insights in order to make advancements in the detection of hate speech. In this regard, we are dealing with spheres that have different priorities, different speeds and different levels of urgency.


About Michael Wiegand

Michael Wiegand joined the Digital Age Research Center (D!ARC) as Professor of Computational Linguistics with a focus on Digital Humanities in September 2020. Alongside Elisabeth Oswald and Katharina Kinder-Kurlanda, he is one of three professors at D!ARC.

Born in 1982, Michael Wiegand studied computational linguistics at Saarland University and the University of Edinburgh. From 2007 to 2010, he was the recipient of a doctoral scholarship at the international post-graduate college at Saarland University. He completed his doctorate in 2011 with a thesis on “Hybrid Approaches for Sentiment Analysis”. Michael Wiegand served as research fellow at the Chair of Speech and Signal Processing at Saarland University from 2010 to 2018. Prior to his appointment to the University of Klagenfurt, he led a research group at the Leibniz Science Campus “Empirical Linguistics and Computational Language Modelling” in Mannheim in 2019. His research focuses on sentiment analysis, hate speech detection, lexical semantics and relation extraction.

for ad astra: Romy Müller, translation: Karen Meehan