Saturday28 January 2023

When supercomputers learn slang

Reading time is around minutes.

There has been quite a lot written about IBM's Watson. The supercomputer that once won on the famous quiz show Jeopardy!. One of the most covered topics is about the algorithms which provide an excellent understanding of conversational language. This nice bit of programming not only understands what’s spoken, but it also sounds human, with sentences that use informal relaxed syntax and vocabulary of everyday life.

However, it seems that the 35 engineers who regularly maintain the system went a bit too far with the relaxing part because Watson began to swear. Apparently all the blame falls on Urban Dictionary, the online slang dictionary that largely deals with curses and profanity found in everyday speech. As Watson was specially trained for the adoption of this type of communication, he implemented into his everyday speech, or into his responses for communication with the researchers to be precise. Apparently in response to purely technical issues he started using words like shit and bullshit, and that apparently was not acceptable.

As punishment, they erased the entire Urban Dictionary from its memory, and supposedly should be implementing a filter that will generally keep the profanity out of communication. Also under the microscope is Wikipedia, specifically those articles that deal with slang and Urban Dictionary.

[Ed – It was only a matter of time before something like this happened. As we work on pushing more and more information into computer systems they will eventually use this in their communication and the way they interact with us. What is surprising is that no one thought about this beforehand and put filters or something in place to assist the system in determining when this type of language is appropriate. We wonder if there are other AI projects going on where we might see unexpected results…]

Tell us what you think about this on our Forum

Last modified on Tuesday, 15 January 2013 09:24

Leave a comment

Make sure you enter all the required information, indicated by an asterisk (*). HTML code is not allowed.