7 ms·

Crowdsourcing Moderation Without Sacrificing Quality

36 points by paulchristiano 10y ago

lightedman 10y ago
So... basically, trying to automate what Slashdot already does. The machine-learning program would need to be an expert in every subject to make this even remotely workable.
- paulchristiano 10y ago
  The hope is to use features like "person X upvoted this content," rather than to try to use domain-specific features or to do natural language processing or so on. Over the long run, algorithms may play a larger role in that process, but at first it would just be people.
  - lightedman 10y ago
    "The hope is to use features like "person X upvoted this content,"" Bad idea. They should focus on WHY such a post got upvoted, not that it was upvoted.
    - paulchristiano 10y ago
      You seem to be imagining very impressive machine learning. I agree that's the right thing in the very long run. I don't understand why you say the simple version is "unworkable" though. Currently many forums just use a heuristic based on date and "total number of upvotes" to decide what to display. Would you describe that as unworkable?
      - ohyeshedid 10y ago
        > Currently many forums just use a heuristic based on date and "total number of upvotes" to decide what to display. That's why these systems often fall victim to brigading by organized trolls.
        samplatt 10y ago
        I'd change "often" to "always", given enough viewers/exposure/popularity.
        paulchristiano 10y ago
        The proposed mechanism makes brigading harder (e.g. a voter is only weighted highly if they are adding useful information, such that increasing their weight improves predictive accuracy), but doesn't make it impossible. In any case, it seems extreme to call existing systems "unworkable."
        lightedman 10y ago
        "a voter is only weighted highly if they are adding useful information" Just how do you propose one determines if the information is useful? Again, back to my original point, you'd need a system trained in EVERY FIELD. In other words, such systems are pretty much unworkable unless you're giving it full access to the entire ACCURATE knowledge of mankind. You happen to have a nice accurate database of everything man's ever accomplished and discovered for such a machine to use? Even Watson isn't even close to that level.
        paulchristiano 10y ago
        By "useful" I mean "improves our ability to predict what the moderator would say." The mechanism for determination is just the usual approach in machine learning, set the weights to minimize prediction error. (Plus some regularization, which may incorporate a prior against new users offering useful feedback.)
      - lightedman 10y ago
        "You seem to be imagining very impressive machine learning" It doesn't even need machine learning. I've got an image aggregator that can recommend images to me based upon what I like, why I like it, and IT LEARNS. Only a couple of times of telling it I don't like this subject matter in the picture, and it won't show me pictures with that subject matter in it. It's nothing more than a game of association and having a reliably fast database with proper tagging. "Currently many forums just use a heuristic based on date and "total number of upvotes" to decide what to display." That's not a heuristic. That's a simple if post count upvotes > x number, display here, if < this number, don't show. That's simple child's play TI-BASIC level programming.
MarkPNeyer 10y ago
Why Must there be one "correct" view of the comments on a document? This insistence seems to be what hobbles us. People will disagree on what is or isn't worth reading. That's fine. We should see the problem as sorting comments per viewer, not for all viewers. If 10 of my close friends think a comment is worth reading, then I personally would like to read it. It doesn't matter what the moderator or a quoroum of random internet people think at that point. We should be using the social graph to filter out nonsense, and to surface constant that we are likely to enjoy. If anyone is interested, I've written code to do this here: https://github.com/neyer/respect https://github.com/neyer/respect
- paulchristiano 10y ago
  I think that the correct behavior in the long run is to give personalized recommendations, analogous to this proposal: https://sideways-view.com/2016/12/01/optimizing-the-news-feed/ https://sideways-view.com/2016/12/01/optimizing-the-news-fee... The main motivation for having a unified view is to decide what to show people who visit but don't rate things / log in / etc. I'm currently finishing my dissertation on collaborative learning algorithms, and I agree that it is not generally possible to have a single global answer. This paper does something similar to your respect matrix: https://arxiv.org/abs/1411.1127 https://arxiv.org/abs/1411.1127 (The same technique can be modified to incorporate information from an external social network.)
- Yen 10y ago
  I'd argue that there is exactly one "correct" view of comments on a document, per-community. The correct view is the one that best represents the range of opinions of the community, and bests represents that community's cultural norms. To be specific, if I'm reading the comments section on Hacker News, I don't want to see comments only from people who agree with me or think how I do, and I don't only want to see comments from people I disagree with. I want to see what the community of HN readers thinks, in the particular way they express it. Additionally, I want to know that what I'm seeing is also what they're seeing. In a sense, this is still social filtering, except averaged over all members of a community, and with the same results for all members. The community responds to posts with approval or disapproval, through votes or comment responses, and this gets averaged (messily) by the moderators. I think, really, that there's an almost objective set of metrics for what makes a good comment thread. * The thread covers most of the held opinions & responses, with few to no 'duplicate' comments * The posts are well-written and polite * Assertions about the world are based on agreed-on facts, and justifiable (even if not agreed-on) conclusions stemming from those. * The posts respect any other cultural norms particular to that community. These sound subjective, but I think that neutral bystanders would tend to score a post similarly on these different metrics - i.e., there's actually a measurable quality of a post being 'polite', or 'reasonably argued', and it's not entirely relative to an individual reader (or even a single community!) On a polarizing political post, I don't want to read poorly written insults from the opposing position, but I also want to know that the other side won't be reading poorly written insults that misrepresent my side. I'd rather the best, most rational, and most empathetic posts from both sides be displayed as representative of the discussion.

So... basically, trying to automate what Slashdot already does. The machine-learning program would need to be an expert in every subject to make this even remotely workable.

paulchristiano 10y ago
The hope is to use features like "person X upvoted this content," rather than to try to use domain-specific features or to do natural language processing or so on. Over the long run, algorithms may play a larger role in that process, but at first it would just be people.
- lightedman 10y ago
  "The hope is to use features like "person X upvoted this content,"" Bad idea. They should focus on WHY such a post got upvoted, not that it was upvoted.
  - paulchristiano 10y ago
    You seem to be imagining very impressive machine learning. I agree that's the right thing in the very long run. I don't understand why you say the simple version is "unworkable" though. Currently many forums just use a heuristic based on date and "total number of upvotes" to decide what to display. Would you describe that as unworkable?
    - ohyeshedid 10y ago
      > Currently many forums just use a heuristic based on date and "total number of upvotes" to decide what to display. That's why these systems often fall victim to brigading by organized trolls.
      - samplatt 10y ago
        I'd change "often" to "always", given enough viewers/exposure/popularity.
      - paulchristiano 10y ago
        The proposed mechanism makes brigading harder (e.g. a voter is only weighted highly if they are adding useful information, such that increasing their weight improves predictive accuracy), but doesn't make it impossible. In any case, it seems extreme to call existing systems "unworkable."
        lightedman 10y ago
        "a voter is only weighted highly if they are adding useful information" Just how do you propose one determines if the information is useful? Again, back to my original point, you'd need a system trained in EVERY FIELD. In other words, such systems are pretty much unworkable unless you're giving it full access to the entire ACCURATE knowledge of mankind. You happen to have a nice accurate database of everything man's ever accomplished and discovered for such a machine to use? Even Watson isn't even close to that level.
        paulchristiano 10y ago
        By "useful" I mean "improves our ability to predict what the moderator would say." The mechanism for determination is just the usual approach in machine learning, set the weights to minimize prediction error. (Plus some regularization, which may incorporate a prior against new users offering useful feedback.)
    - lightedman 10y ago
      "You seem to be imagining very impressive machine learning" It doesn't even need machine learning. I've got an image aggregator that can recommend images to me based upon what I like, why I like it, and IT LEARNS. Only a couple of times of telling it I don't like this subject matter in the picture, and it won't show me pictures with that subject matter in it. It's nothing more than a game of association and having a reliably fast database with proper tagging. "Currently many forums just use a heuristic based on date and "total number of upvotes" to decide what to display." That's not a heuristic. That's a simple if post count upvotes > x number, display here, if < this number, don't show. That's simple child's play TI-BASIC level programming.

MarkPNeyer 10y ago

Why Must there be one "correct" view of the comments on a document? This insistence seems to be what hobbles us. People will disagree on what is or isn't worth reading. That's fine. We should see the problem as sorting comments per viewer, not for all viewers. If 10 of my close friends think a comment is worth reading, then I personally would like to read it. It doesn't matter what the moderator or a quoroum of random internet people think at that point. We should be using the social graph to filter out nonsense, and to surface constant that we are likely to enjoy. If anyone is interested, I've written code to do this here: https://github.com/neyer/respect https://github.com/neyer/respect

paulchristiano 10y ago
I think that the correct behavior in the long run is to give personalized recommendations, analogous to this proposal: https://sideways-view.com/2016/12/01/optimizing-the-news-feed/ https://sideways-view.com/2016/12/01/optimizing-the-news-fee... The main motivation for having a unified view is to decide what to show people who visit but don't rate things / log in / etc. I'm currently finishing my dissertation on collaborative learning algorithms, and I agree that it is not generally possible to have a single global answer. This paper does something similar to your respect matrix: https://arxiv.org/abs/1411.1127 https://arxiv.org/abs/1411.1127 (The same technique can be modified to incorporate information from an external social network.)
Yen 10y ago
I'd argue that there is exactly one "correct" view of comments on a document, per-community. The correct view is the one that best represents the range of opinions of the community, and bests represents that community's cultural norms. To be specific, if I'm reading the comments section on Hacker News, I don't want to see comments only from people who agree with me or think how I do, and I don't only want to see comments from people I disagree with. I want to see what the community of HN readers thinks, in the particular way they express it. Additionally, I want to know that what I'm seeing is also what they're seeing. In a sense, this is still social filtering, except averaged over all members of a community, and with the same results for all members. The community responds to posts with approval or disapproval, through votes or comment responses, and this gets averaged (messily) by the moderators. I think, really, that there's an almost objective set of metrics for what makes a good comment thread. * The thread covers most of the held opinions & responses, with few to no 'duplicate' comments * The posts are well-written and polite * Assertions about the world are based on agreed-on facts, and justifiable (even if not agreed-on) conclusions stemming from those. * The posts respect any other cultural norms particular to that community. These sound subjective, but I think that neutral bystanders would tend to score a post similarly on these different metrics - i.e., there's actually a measurable quality of a post being 'polite', or 'reasonably argued', and it's not entirely relative to an individual reader (or even a single community!) On a polarizing political post, I don't want to read poorly written insults from the opposing position, but I also want to know that the other side won't be reading poorly written insults that misrepresent my side. I'd rather the best, most rational, and most empathetic posts from both sides be displayed as representative of the discussion.