Using machine learning to detect bad behavior on the internet

One of the main problems for networks like Twitter over the years, as I have commented on frequently, is preventing harassment and insults.

Since its beginnings, Twitter has sold itself as a defender of freedom of expression, but has ended up creating an environment where that supposed freedom of expression has been severely limited by the activities of trolls and the like.

Over time, this harmful environment has caused serious difficulties for Twitter, from slower-than-expected growth to the decision by growing numbers of people not to take part in the conversation and simply lurk. Its future is now in doubt, given that many potential buyers have been put off by the poisonous dynamics on the platform.

After many attempts to correct these dynamics, most of which have ignored the real problem Twitter is trying something new: a collaboration With IBM to make its machine learning system, Watson, detect, through the study of conversational patterns, harassment and abusive behavior before they are reported. Can artificial intelligence really help to detect harassment or verbal abuse? This is a complex challenge: insults can be detected using a dictionary, but what about irony, double meanings, or innuendo, or more subtle use of language. Harassing someone can be done in many ways.

To make its job easier, Twitter probably has one of the best online files of abusive behavior. Throughout its eleven years of history, the company has been involved in all kinds of high-profile scandal and in a wide variety of infinitely lesser- known situations that have affected all types of users. The company can draw on its immense files of harassment, insults, bullying, sexism, incitement to hatred, etc. It could even label profiles based on their behavior. That type of data is precisely what a machine learning algorithm needs to be trained correctly, considering that semantics and human language analysis are already carried out algorithmically perfectly well. Obviously, there will be some situations, such as the use of images, that may be prove more difficult to process, but this is no longer outside the capabilities of artificial intelligence, and in the final analysis, human evaluators could help in establishing what is going on. 

Could Watson be the judge that decides whether somebody is being offensive? As somebody who has suffered from offensive behavior on Twitter and who has seen how the company has done nothing to combat it (or even made the problem worse), I think machine learning can provide a means to at least identify bad behavior, classify it and help manage it, as well as helping detect multi-identity management by people whose accounts have been closed for bad behavior.

Is there any conflict here with freedom of expression? It all depends on how we want to define freedom of expression. If we think the social networks are places where anything goes, then yes. But we live in society, which is regulated by certain rules. By now we should all understand that the adjective “social” applied to the noun “network” should mean more than it currently does. At least, in the case of Twitter …

(En espa?ol, aquí)