6
$\begingroup$

There is a feature of showing suggested tags after the body of a question is included. These suggestions are based on the body of the question. More details can be found in these posts:

I will quote part of one of the answers, which gives a reasonable summary.

  1. It uses similar-to-Bayesian reasoning on the text from the title and body of a post
  2. It has been trained on older questions
  3. A minimum of around 20k questions are required for accurate results
  4. There is no substitute for material to learn from, so it can't be rolled out to all sites

If I understand this correctly, this feature is used on Stack Overflow. (In one of the post you can see also a screenshot.) But it can be added to any site which has sufficiently many old questions for training the algorithm. (Ideally, most of the questions should be correctly tagged.)

Still, it seems that some non-trivial work is involved when such a feature is added to a site.

Question. Do you think feature like would be useful on MathOverflow? Should we request Stack Exchange staff to add the feature also to this site?

I have brought this up because a somewhat similar feature request was posted here recently: Why isn't there an automated supertagging? Since these two features are IMO related, it probably makes sense to discuss them at the same time. The difference is basically that in the other feature request the added tags are chosen manually - here by an algorithm; the other feature request suggests to add tags - this one only makes suggestions, the poster might choose some of the suggested tags by clicking on them.


Another post on Meta Stack Exchange explicitly mentions that:

For each individual stack, you'd have to pick at least 20,000 (old) questions to train the classifier. If you do not have enough questions, it would not be accurate enough.

If this feature is considered on MO, perhaps the training set could be chosen in such way that those questions are less likely to be incorrectly tag. (E.g., only pick questions with a top-level tag, do not include questions with a deprecated tag.)

$\endgroup$
  • 1
    $\begingroup$ I probably should mention that I have posted basically the same feature request on Mathematics Meta: What about suggest tags feature? Of course, the two sites are different, it is quite possible that the outcome of the discussion on this site and on that site will be different in the end. $\endgroup$ – Martin Sleziak Mar 15 '18 at 9:16
  • $\begingroup$ minimum of 20k questions... no tag achieves this so far. Do I misunderstand? $\endgroup$ – YCor Jan 12 at 18:42
  • $\begingroup$ @YCor The way I understood it, it's 20k questions on the site. (When I looked into chat, balpha's wording was: "The training corpus should be at least around 20k questions.".) $\endgroup$ – Martin Sleziak Jan 12 at 19:30
  • $\begingroup$ "On the site" as opposed to what? $\endgroup$ – YCor Jan 12 at 19:33
  • $\begingroup$ @YCor The way I understood it, you need 20k question on the site in total. Your comment suggests that the algorithm needs 20k questions in a specific tag. $\endgroup$ – Martin Sleziak Jan 12 at 19:35
  • $\begingroup$ Thanks for the clarification! Of course I would like to see what such an algorithm proposes in a representative sample of recent existing questions before implementation. It might also encourage some existing unnecessary tags (e.g., adding matrices to linear algebra questions when this is plainly unnecessary, adding polynomials to questions about polynomial rings unrelated to individual polynomials, etc.) $\endgroup$ – YCor Jan 12 at 19:43
  • $\begingroup$ Of course, we should keep in mind that many questions on MO are tagged incorrectly (perhaps 15-20%). The quality of algorithm very likely depends also on the quality of the set of questions on which it is trained. $\endgroup$ – Martin Sleziak Jan 12 at 20:20
  • 1
    $\begingroup$ @MartinSleziak If we elected you to be a moderator, could you handle this stuff yourself? You do seem like the ideal candidate fwiw. $\endgroup$ – Harry Gindi Jan 14 at 15:23
4
$\begingroup$

Yes, I think this could be a good thing! Tagging is often a thankless or even forgotten task. I think machine assistance can reduce the cognitive load needed for the task, hence encourage it.

$\endgroup$
  • 1
    $\begingroup$ As long as the machine suggestions don't make the same mistakes in confusing mathematical meanings of words with non-technical meanings, as some new users do. $\endgroup$ – David Roberts Jan 11 at 23:10
  • 1
    $\begingroup$ These are suggestions, no? we can hope most users to ignore phoney suggestions. $\endgroup$ – YCor Jan 12 at 18:40
  • $\begingroup$ Well, when searching by tag, getting irrelevant results is better if it is rare. $\endgroup$ – David Roberts Jan 13 at 8:05

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .