People on the web should admit: online trolls are practically everywhere. From social media networks with the most influence to the very least popular blogs, they can be anywhere. And what makes them difficult to comprehend is because those trolls and abusers often use naturally-spoken words, decreasing the chances for them to be caught by filters.
And speaking of filters and computers to do manage the job, even us humans are pretty terrible at identifying those people.
While we have to fight an endless horde of those pesky people with our current arsenal, Yahoo! came up with something better. The company came up with an algorithm that is able to correctly identify abusive comments with a 90 percent success rate in test cases.
This level of accuracy is unmatched by humans, and marks another milestone in machine-learning artificial intelligence (AI).
What the company did, is using a combination of machine-learning and crowdsourced suggestions to build the AI's knowledge. The algorithm was set free to crawl sections at Yahoo! News and Finance, learning and finding new things, catching up the many things we shouldn't do (say) online.
After a million of Yahoo! article comments were scanned, the AI was able to give Yahoo! a curated database of online hate speech.
The algorithm isn't just like some ordinary filtering mechanism that searches for specific words or phrases that will trigger when a corresponding word is detected. Most other AIs have been bad in identifying posts,
This is because the keyword-based filtering mechanism is similar to spam filters found on emails. They aren't actually that good in catching hate speech. Online trolls, especially those that trolls out regularly, use obscure abusive words to sneak past filters. Their hate speech might be full of "hatred", but can come without even using a particular abusive word.
For this reason, the pattern can be somehow unpredictable, and this has made existing algorithms, and humans, to fail in labeling posts. By failing to recognize the expression of hate can lead to a real sarcasm from those being flagged or abused.
Computers have failed to spot trolls because they often can't understand the meaning of messages.
As a workaround, Yahoo! approached the trolls in a more sophisticated manner. Yahoo!' algorithm starts similarly, but does the rest differently. The AI collects more information about a given post by seeking combination of words, as well as overall post length, punctuation and other metrics. This has made it essentially better than any other filtering mechanisms.
To make it better in tracking those trolls down, the AI gets help from humans in its learning process. Trained humans from Yahoo! are helping the AI by rating the comments. Here the AI can spot subtle nuances that it missed, making it better on the next search. And as for the crowdsourced additional ratings, Yahoo! uses Amazon's Mechanical Turk from humans that weren't professional comment moderators.
In short, Yahoo!'s AI got everything it needs to learn and act.
At the moment of its introduction, Yahoo!'s anti-abuse AI hasn't been used outside the company's assets and dataset. But the company is confident that the ability of this AI is a "major step forward" in the field of natural language processing.
Described as pests, internet trolls lurk in every corner of the internet, and they delight in ruining anyone's day. Yahoo!'s AI has past its tests in getting to know the pattern, and now the company aims to give it a real test: setting it loose outside its original place, out to the wilderness of the troll-infected internet.