Authorization

Researchers have developed an extremely effective sarcasm detector

Researchers have developed an extremely effective sarcasm detectorSarcasm might be all over the internet, but its still hard to recognize. Researchers want to change that.

A new research paper from two professorsDavid Bamman from UC Berkeley and Noah A. Smith from the University of Washingtonsays algorithms can figure out if someone is being sarcastic or not, using social media sites like Twitter.

But what makes this project different than others is its focus on contextual information. When people use sarcasm, theres quite a bit of background knowledge thats shared between the two. According to the research paper, which Bamman and Smith worked on during their doctoral studies at Carnegie Mellon University, by understanding contextual informationlike who wrote the text and if its being shared with friends or the general publicitll be easier to recognize sarcasm.

The algorithm trains itself to recognize sarcastic posts by scanning tweets with a #sarcasm hashtag. Using machine learning, the algorithm gets better at pointing out sarcasm over time. Then it was put to the test.

The experiment yielded stellar resultsthe algorithm recognized sarcastic texts 85% percent of the time. Who wrote the tweet was the most important factoraccuracy rose the most when the authors information was incorporated into the data. But all of the contextual information included in the algorithmresponses, the audience, etcmade the results more accurate than merely analyzing the text on its own.

The data says some people are more likely to use sarcasm than others, like unverified, male, Twitter users from the US (which isnt shocking, if youre an avid Twitter user). But the researchers have more commercial and mainstream uses for this technology, which is still a work in progress, including national security and sorting through reviews, says The Horizons Tracker.

Source: Quartz
.
See also:
Leave a comment
TOP Video
News
  • Latest
  • Read
  • Commented
Calendar Content
«     2016    »
 1234
567891011
12131415161718
19202122232425
262728293031