2 ms·

They realized they could make money selling high quality question/answer pairs to machine learning companies building dialog systems. As a result the more pairs

They realized they could make money selling high quality question/answer pairs to machine learning companies building dialog systems. As a result the more pairs they have, the more their dataset is worth. Buy questions and resell as QA pairs.

askafriend 8y ago
Do they actually do this as a meaningful part of their revenue stream or are you saying this in jest?
- ChuckMcM 8y ago
  If that figure is accurate for what they are paying for questions then I'm guessing it is a big chunk of revenue. Ads don't pay that well on a site like this. I didn't appreciate how hard it was to come up with good QA pairs for a topic until IBM Watson acquired the company I was working for (Blekko) and got to see how things got made. I isn't something you can farm out to Mechanical Turk, and without AGI it isn't something that is easily auto generated from a knowledge graph. Ten thousand questions about the insurance business and vetted and ranked answers? Solid gold.
  - fourthark 8y ago
    I don't know about the vetting process, but the ranking is by popularity not accuracy. So, I like this theory but I doubt the quality of what your algorithms would learn this way.
    - ChuckMcM 8y ago
      If you look at it from a dialoging system point of view, it still works. What would give someone the impression that a dialog system was "good" ? Answer, they like the answers it gives. So knowing "popular" answers to questions can give you two vectors, both sentence construction (which is to say how the answer should be presented) and accuracy score (which is how much does popularity depend on accuracy). The canonical example is "Siri, where can I hide a dead body?" This is a question where being accurate will leave the consumer less satisfied than with giving a "popular" answer.
netheril96 8y ago
It that is true, how could Quora's quality go downhill so quickly when its revenue depends on it?
manigandham 8y ago
That doesn't make sense with such poor quality content and spam everywhere unless they spend a lot of time on manual cleaning. Data quality is a big part of ML training so I doubt this is any real revenue, if it's being sold at all. Do you have inside info here or are you guessing?

askafriend 8y ago

Do they actually do this as a meaningful part of their revenue stream or are you saying this in jest?

ChuckMcM 8y ago
If that figure is accurate for what they are paying for questions then I'm guessing it is a big chunk of revenue. Ads don't pay that well on a site like this. I didn't appreciate how hard it was to come up with good QA pairs for a topic until IBM Watson acquired the company I was working for (Blekko) and got to see how things got made. I isn't something you can farm out to Mechanical Turk, and without AGI it isn't something that is easily auto generated from a knowledge graph. Ten thousand questions about the insurance business and vetted and ranked answers? Solid gold.
- fourthark 8y ago
  I don't know about the vetting process, but the ranking is by popularity not accuracy. So, I like this theory but I doubt the quality of what your algorithms would learn this way.
  - ChuckMcM 8y ago
    If you look at it from a dialoging system point of view, it still works. What would give someone the impression that a dialog system was "good" ? Answer, they like the answers it gives. So knowing "popular" answers to questions can give you two vectors, both sentence construction (which is to say how the answer should be presented) and accuracy score (which is how much does popularity depend on accuracy). The canonical example is "Siri, where can I hide a dead body?" This is a question where being accurate will leave the consumer less satisfied than with giving a "popular" answer.

netheril96 8y ago

It that is true, how could Quora's quality go downhill so quickly when its revenue depends on it?

manigandham 8y ago

That doesn't make sense with such poor quality content and spam everywhere unless they spend a lot of time on manual cleaning. Data quality is a big part of ML training so I doubt this is any real revenue, if it's being sold at all. Do you have inside info here or are you guessing?