For some reason, I think that this will only add to our existing problem of having too much data to digest, but no way to reasonably digest it. In other words - we need a way to extract valuable data from millions of articles, not generate millions more.
There are still many cases where there is too much data to digest. What Automated Insights does is take data and turn it into interesting and hopefully insightful bits of digestible content.
I don't know that this helps or hurts anything really, what they're really doing (for this example at least) is translating stat lines to text.
It's pretty easy to say something like:
This week's top quarterback was %qb% from the %team1% who had %qbyards% and %qbtds% while playing the %team2%.
The total amount of information is exactly the same, they just use a little NLP to make it sound 'hand written' but display a box score in a easier to read manner.
The technology is significantly more complicated than simple search and replace a bunch of variables. Trust me, if we did millions of stories by swapping out the same sentence every time it wouldn't work.
Also, the part of the value of what we do is describe "insights" not just spit back raw numbers (which again wouldn't be very valuable).
I didn't mean to demean the work, there is clearly a whole lot going on behind the scenes here. It was more of a rebuttal of the "more data is bad" idea. My point above was unclear, but this idea doesn't necessarily represent more data, just the same data presented in a more human-friendly manner.
I really like what the team has done, this could have a long-lasting impact on anything with heavy use of stats and figures. I honestly have a few enterprise applications that would benefit from a similar treatment. e.g. reports that aggregate certain operational stats and must be hand-written every week.
This is pretty cool. It's very personalized—it would never make sense for Yahoo to hire people to write recap for fantasy games, since those recaps are going to be of interest only to the participants in the league (and maybe only to the two players in the particular game). But it makes sense to have a program write recaps for all those individual games instead of just showing box scores or whatever they had before (I don't play fantasy football myself... it feels too much like work for me, messing around with spreadsheets and numbers and algorithms!).
FYI - Fantasy doesn't require the use/study of algorithms or much work at all, it's more of a way to increase a fan's enjoyment of the game by way of becoming familiar with more teams and players in greater depth. Yeah, you've got an ordered list of players, or tiers of players, at the beginning of the season when you draft your team, but other than that, you'll only spend 15-30 minutes per week checking to see if there's a new player available as an upgrade, and then moving guys in/out of the starting lineup as determined by injuries/bye weeks/rumors/hunches.
I think this company is fascinating in its approach -- most companies are trying to automate the analysis of human-generated content where they are automating the generation of human-analyzable content.
This is awesome because the most fun I have in fantasy leagues is the interaction with the people I am playing against. So this just takes it to a next level and provides a recap like the normal NFL games.
Serious question: how is Google going to deal with the advent of technologies like this being used to create billions of pretty decent pieces of content?
I think this is pretty neat, although it does come with many unintended consequences for spam.
Does anyone have more information on the technology they use to generate these articles? I assume some sort of NLP in reverse (natural language generation, I guess?)
13 comments
[ 2.0 ms ] story [ 34.5 ms ] threadIt's pretty easy to say something like:
This week's top quarterback was %qb% from the %team1% who had %qbyards% and %qbtds% while playing the %team2%.
The total amount of information is exactly the same, they just use a little NLP to make it sound 'hand written' but display a box score in a easier to read manner.
Also, the part of the value of what we do is describe "insights" not just spit back raw numbers (which again wouldn't be very valuable).
I really like what the team has done, this could have a long-lasting impact on anything with heavy use of stats and figures. I honestly have a few enterprise applications that would benefit from a similar treatment. e.g. reports that aggregate certain operational stats and must be hand-written every week.
Good to see a Durham, NC company get some love.
Does anyone have more information on the technology they use to generate these articles? I assume some sort of NLP in reverse (natural language generation, I guess?)