Can Computers Detect Sarcasm?

4th Jun 2010
  • +0
Share on
close
  • +0
Share on
close
Share on
close

Sarcsam is most often used as a an expression of humour or emotions, mostly stating the opposite of the intended meaning. More often than ever, sarcasm can be really subtle.

Humans detect sarcasm using many visual cues, like rolling of eyes, context of the discussion, facial expression etc. In today’s world of online communication, sarcasm is usually difficult to dtect. Sarcastic comments are usually tagged by “#sarcasm” to denote that the user is intending on sarcasm. Without the hash tag, sarcasm is difficult to identify.

Three researchers at The Hebrew University in Jerusalem have just developed a unique algorithm that can detect certain sentence patterns and identify them as sarcastic. Named SASI, short for a Semi-Supervised Algorithm for Sarcasm Identification , the program is able to recognize sarcastic comments with 77% precision – making it better at detecting sarcasm than most humans.

The research team scanned over 65 000 product reviews from Amazon.com and each sentence was then tagged independently for sarcasm by three human annotators. Sentences were then put into sub-categories of sarcasm: some mean the opposite of the actual statement and others transmit a sentiment inconsistent with the literal meaning of the statement. The program was able to learn from the annotations certain patterns in sentence structure that make a sentence sarcastic. The algorithm was tested by being instructed to identify sarcasm in sentences that were nothing like the examples; that is where the 77% accuracy rate comes from.

Here are a few examples of product review excerpts used to ‘teach’ the computer to recognize sarcasm.

  • “[I] love the cover”
  • “Be sure to save your purchase receipt”
  • “Trees died for this book?”

Sarcastic assessments of Amazon listings aside, the program genuinely has some very useful applications. It can be used to make personalized web content more accurate or to produce better opinion-mining programs that troll the web to assess public opinion about certain new products or services.

Overall it’s not bad for a computers first shot at interpreting Human Humour.

- Nikhil Mukundakumar

    • +0
    Share on
    close
    • +0
    Share on
    close
    Share on
    close