I ask because I've only just noticed a few questions to which the tag has been newly applied, and I don't really see why the old "books" tag wasn't enough.

I think the consensus around here is that clean-up activities like retagging should be minimally intrusive. Since retagging kicks questions to the top of the active queue, it would be best if you refrained from retagging outside the following situations:

1. There is a strong need. In this case it might be best to contact the moderators first, or open a thread here on meta.
2. The question is already near the top for some other reason (e.g., someone recently submitted an answer).

I think there is a case for having either the "books" tag or the "textbook-recommendation" tag, but not both. I also don't believe in trying to fine-tune the tags, except for cases such as those David has mentioned (homological-dimension vs topological-dimension; I don't think one should have a "hausdorff-dimension" tag because there are other notions of topological/metric dimension and I don't want separate tags for all of them). For that reason I don't really want a divide between sheaf-cohomology and sheaf-theory; I would personally use both the tags "cohomology" and "sheaf-theory".

By the way, I view the "reference-request" as being exactly that, and not a "book-recommendation" tag.

Some random thoughts: 1) I find it far easier to use Google to search MO than to use MO to search MO, myself. 2) since I use MO almost exclusively through the "active" page, bumping of old stuff tends to jar me.

Using the site more or less in the way Yemon Choi described, I never paid too much attention to the tags. Sparked by this discussion I looked around a bit more. Here are my thoughts. (Some of what I write could be read as critizising others behavior; this is not my intention. I merely wish to describe my perception of the situation. I can well imagine and understand why the situation is at it is, and I also it is in no way problematic for me.)

My general impression is that there are way too many tags (about a thousand).
A key question to me is what is a good minimal size for a tag to make it useful as a tag. I see two and two half use-cases for a tag:

1. Automatic filtering by tags as supported by the software.
2. Easy access to all question with a certain tag.
3. Tags as clearly visible keywords of a question.
4. Searching.

The last two are the half-one, as I do not consider them at so important. For the first, a well-chosen title / the first lines (that are visible without opening the question) can by and large fulfil this purpose. And for the latter, I do not really know, but the general sentiment seems to be that serching the site with Google is anyway the way to go.

For 3. I actually do not really see any non-trivial lower bound for the usefullness, and it seems to me that various tags were (unfortunately) created in this spirit. However, as said I do not consider 3. as so important, and for 1. and 2. and also 4. I think there clearly is a lower-bound on the size, below which the tag is not really useful (and its existence is mildly harmful, on the one hand as it is more or less noise, and on the other hand and more importantly as the existence of, say, five too small, and thus in my opninion. not useful tags prevents the existence of one larger and thus useful tag).

Let me elaborate on the first line of the preceeding pargaraph. Why is there, for example, a thanks-tag ? (To answer my own question, it seems somebody found it a nice idea to create it, and then it was never deleted. Both, in particular the latter, understandable, and locally no problem. But the accumulation of such things creates in the long run masses of tags.) To give a somewhat less extreme examples, there is dedekind-domain, prufer-domain, euclidean-domain, ufds, pre-schreier used in total (not individually!) ten times; and some are dually used so the number of question with such a tag should be a bit smaller still. In general, about a third of the tags was used less than 5 times, many even only once.

For 1. and 2. I think everything that is not used at least somewhat frequently is not of much use; perhaps 20 rather more as lower bound (so roughly and adjusting a bit to later cretaion, used once every two weeks, and if I look for question with the tag I get a page full of questions). Even with 20, only a third of the existing tags would make the cut.

To be cont.

Afterthought: Perhaps I should add, as this might get lost or not be really visible from what I wrote, that I do appreciate (from one user to another on a to some extent community moderated site) the effort of David White. And, in theory, I could see some merits of a sophisticated tagging-system. Yet, in practise, I am more than a bit worried that anything too complex will simply not be used properly/consistenly and then loose its usefulness and become (mildly) harmful instead.

David, I'm afraid we do not have the power to change the underlying software. However, I do not know about the particulars of the tag autocomplete box.

The current tagging system, which involves an ArXiv designation for general area plus some keywords, has worked reasonably well for more than a year. As far as I can tell, we haven't run into serious scaling problems, and it doesn't seem to be a danger in the near future. The solutions you propose would be potentially convenient if they were possible, though.

Dear Emerton,

I would agree that many of the low-use tags individually are helpful, yet I also believe that in there totally they tend to cause confusion and/or inconsitencies.

To illustrate my point, let me return to the rings/domains example: in itself I see nothing wrong with dedekind-domain, prufer-domain, euclidean-domain, ufds, pre-schreier, and also cohen-macauly-ring is fine, and perhaps over time we will get (or perhaps already have for some I did not check) the tags for, say, Bezout, Mori, Krull, Laskerian, gcd, Cohen-Kaplansky, Prüfer v multiplication, Prüfer star multiplication domains or rings. However, what I do find slightly unfortunate is that I believe none of the questions tagged with the above mentioned tags is tagged (in addition) with commutative-rings. [Reading the thread you linked to now completely, before I knew only the end-part, I learned that at somme point even this commutative-rings tags was at some point considered as to narrow and to be subsumed in ac.commutative-algebra, though some disagreed.] Faced with the choice of having them all tagged commutative-rings and not a more specific one and having them all tagged with a more specific one and not commutative-rings I prefer the former. (If they had both tags, well, I would not mind so much.)

Now, I could tag them in addition with the commutative-rings tag, however as said above by Scott Carnaham, one should not do so. Thus, I might consider suggesting to merge them.

In the very end, as said above, for me personally the tags are of little relevance. However, what I wanted to point out is that, in my opinion, tag-proliferation is not completely for free. And, I would thus appreciate if at least tags like the I believe recently created, by an experienced user by the way, 'misapplied' tag (I already have doubts regarding the usefullness of applied, but misapplied how will this ever be useful as a tag?) would not be created.

There's a lot to digest and respond to from various people, which I sadly can't do properly right now. Hence this post of mine is going to be even less coherent than usual, I'm afraid.

Yemon seems to want fewer tags, even if this means they are less specific in filtering questions.

That's probably fair. I just don't see why the situation ``as people post more and more specific research questions it's clear our existing tags are going to be overloaded'' is a problem. If one were to start parsing the fine detail of the functional analysis tags we could easily create something like Borges's system with, IMHO, nothing gained. Since I don't search MO to find answers to questions (except to check that no one has already answered it) and I find Googling sufficient for my search requirements, I personally wouldn't gain from more precise taxonomy.

I guess I still don't understand why one wants finer tags as a general principle, rather than something to do on a trial ad hoc basis. I don't think this is going to significantly affect the likelihood or speed with which someone's question is answered. (See my recent questions about finite groups: I am happy to have general tags and a hopefully specific question.) If someone tags a question "Banach algebras" and writes a clear, well-defined question, that's good enough for me, I don't really see how matters would be improved by having separate tags for "Gelfand theory", "Fourier algebras", "uniform algebras", "function algebras", "nonself-adjoint operator algebras", "radical Banach algebras", "strawberry flavoured amenability", "exponential spectrum", "Beurling algebras", "topological centres", "semigroup algebras", "group algebras", "pseudofunction algebras", "spectral synthesis", "operator synthesis", "Q-algebras", "dual Banach algebras", "measure algebras", "approximate homomorphisms", ...

If someone has a question in algebraic topology related to spectra, do we need a tag for "spectra"? (Do we already have one?) If so then I agree one should differentiate between that and the spectrum of sp.spectral-theory, or the prime ideal spectrum; but I am not sure such specificity is needed.