Comedy Slam: How can machines tell if something is funny?
Ever wondered if Google can tell if content is funny or not? Sanketh Shetty writes about YouTube’s attempt to gauge comedy levels of video content by observing its content and comments. In summary, it’s a combination of crowd-sourced algorithm feeding (75,000 users and 700,000 votes), material recognition and comment analysis.
Some types of audio-visual clues are easy to detect, for example audible laughter or shaky camera motion. Video content which lacks those properties may be a bit trickier to understand by AI, especially as humour is considered to be one of the most subjective preferences much like music choice.
For starters YouTube selects videos which are already uploaded in the “Comedy” category. Next set of things being observed are elements such as titles, video file name, tags and description. These are author-specific much like the selected category so additional clues need to be used to verify.
Apparently YouTube (Google) take video comments seriously, unless they are flagged as spam. What’s interesting is that they look for instances of onomatopoeic laughter and go through a number of applicable variants.
Here are a few examples including standard laughter, country and language specific and internet acronyms:
Other elements observed in comments are instances and variations of the word “funny”, emoticons :) and then attempt to classify them into categories which may be pranks, stand-up comedy, commercials, bloopers, animals… etc.
Algorithm is employed to gauge the comedy potential and that goes beyond raw video views which are largely influenced by video age. What YouTube team figured out was that extra emphasis can be used to highlight extra funny content.
Examples of that are:
- loooool / lolololol / lolllll!!!!
The challenge here was in the evaluation of the level of amusement based on transformation of the above acronym variants. Algorithm was developed to work with various variants of “amusement” keywords and rank in a passive-aggressive manner using ‘human-annotated pair wise ground truth’ in conjunction with text and audio-visuals. So called ‘Comedy Slam‘ operates similarly to already popular ‘Music Slam‘.
- Opinion mining and sentiment analysis
- ICWSM – A Great Catchy Name: Semi-Supervised Recognition ofSarcastic Sentences in Online Product Reviews
- That’s What She Said: Double Entendre Identification