If you’ve immersed yourself in the culture surrounding anime more than a little, you’re probably aware that Japanese video sharing sites operate their comment sections a little differently from the way Crunchyroll does. For the uninitiated, while sites like Youtube stick their comments underneath the video for viewers to scroll down and see if they want to, sites like Nico Nico Douga have user-submitted comments scrolling across the same screen playing the video. I personally prefer this style of commentary for a variety of reasons,* and I sometimes miss the feature when transitioning to official releases. Just recently, I realized fully functional logs of these comment tracks are actually available for download at himawari douga, a fairly large** and in-no-way-legal video streaming site.
These himado comment tracks have a particularly neat (if totally necessary) feature – each comment keeps its 1/100th of a second timestamp, which usually corresponds fairly precisely with the thing in the video being commented on.*** Thus, these tracks, properly analyzed, have the potential to provide an interesting window into which moments in an episode most engage the audience. I built a very basic code**** to analyze the data, and I came up with some fun plots.
The first plot is for Carnival Phantasm’s Kenny McCormick tribute episode, Final Dead Lancer (an episode where a character tries to avoid various comedic deaths). The episode has been taken down from here, but it amassed 1962 total comments, enough to do basic analysis:
The largest spikes in comments in the 13-minute episode occur at 90-100 seconds (the holy grail popping up out of nowhere while Lancer’s sitting on a bench), and 590-600 seconds (the punchline of the episode), with various smaller-scale gags getting more minor spikes (especially between 4 and 6 minutes, when they’re rapid-fire montaging through various death scenes). This seems pretty well in line with what I’d expect the pacing of the episode to produce.
Let’s try it again with a more recent show. Satou Keiichi’s Shingeki no Bahamut has a fairly engaging first episode, one which introduces its love of sweeping set pieces and delightfully petty duo of male leads. When I pulled the comments last night from here, it had a total of 8694 comments (unlike FDL, the video is still available). Here’s how they shook down:
Note: The comment counts are nonzero after 1500 seconds=25 minutes because people can timestamp their comments for after the video stops running.
The intervals of time that produced a comment count over 100/10sec were the following:
180-200 seconds – Beginning the moment in the Favaro-Kaisar chase scene where they jump their horses off of an aqueduct onto a rooftop.
270-280 seconds – Kaisar and Favaro ride a waterwheel through the town streets.
470-480 seconds – Amira falls from the sky, wearing nothing.
640-650 seconds – The end of Favaro’s fight with the show’s first bounty
860-870 seconds – This is probably the most complicated one. Kaisar, having cornered Favaro in a pub, gets distracted by a bunch of women and instead of shaking them off, instinctively plays the flirt. He then immediately regrets this and makes what I would describe as a fun face:
990-1010 seconds – The show’s second bounty summons a massive demon, but the demon takes time to get out of the summoning circle and gets its face licked by a dog in the meanwhile.
1310-1330 seconds – A character realizes he’s just grown a tail, but has no clue why. This transitions into…
1330-1350 seconds – The ending credits start to roll, people share their impressions of the show.
It actually seems here like comments may peak in two distinct areas – scenes that impress the audience (either through action or humor), and breath-catching moments at the end of bigger fight scenes.
The above is mainly just some preliminary eye-test stuff to make sure this method is something I can use. Hopefully, I’ll be able to do some more interesting things in this vein in the very near future.
*If you’ve never seen episode 9 of Blood-C with multicolor scrolling comments, you’ve never really lived.
**One report for the Anime Manga Guardians Project pegged it as the 22nd-largest streaming site relevant to anime piracy, with over 15 million visitors, and a lot of the larger ones (youtube, animefreak.tv) don’t have scrolling comments.
***Also helpful is the fact that these data stay available after the episode itself gets the eventual copyright takedown, so even older series can be analyzed provided enough people commented while it was up.
****For people who care, I wrote a python script that found the variable “vpos” in the comment stream, ripped the string of digits after it, converted that string of digits to a floating point variable, then binned that data into a histogram with 10-second-interval bins.