An Israeli research team has devel­oped a machine algo­rithm that can rec­og­nize sar­casm. Tested on Ama­zon, SASI (Semi-supervised Algo­rithm for Sar­casm Iden­ti­fi­ca­tion) can detect sar­cas­tic sen­tences in prod­uct reviews with 77 per­cent pre­ci­sion. The results are even bet­ter on Twit­ter where SASI iden­ti­fied sar­casm with 94 per­cent pre­ci­sion. Basi­cally, the researchers had human anno­ta­tors tag­ging sen­tences for sar­casm. The team then iden­ti­fied cer­tain sar­cas­tic pat­terns and cre­ated a clas­si­fi­ca­tion algo­rithm that puts each state­ment into a sar­cas­tic class.

The algo­rithm were then trained and turned loose on an eval­u­a­tion seet. Con­sid­er­ing the over­all pre­ci­sion SASI scored, that’s not a bad shot at inter­pret­ing the human sense of humor. It almost imme­di­ately reminded me of a project called Stu­pid­Fil­ter. If they could incor­po­rate this algo­rithm … but then, they haven’t updated their web­site for nearly two years. Too bad. Any­way, if you really want to know more about how the algo­rithm works, you should check out this PDF (~93 kB).

So what’s the use of rec­og­niz­ing sar­cas­tic state­ments? The researchers believe that SASI could be instru­men­tal in gen­er­at­ing bet­ter per­son­al­ized con­tent and make bet­ter rec­om­men­da­tions to human users. They also think that the algo­rithm could ben­e­fit opinion-mining systems.