Taglocity General Discussion Forum

I'm confused by the logic behind the autotags.

I created tag A and marked it to autotag
Now, something like 99% of the messages I receive get tagged with this.  I bcc: myself on each email and EVERY message I send gets autotagged with the tag.  Even if I manually un-tag a message and then reply to it, the reply will get tagged. The message is nearly identical to one that just got manually un-tagged -- why doesn't it learn from this and not apply the tag?  I'm really struggling to find the value in the autotagging logic. 

(In a related note, which is the preferred tag logic that I should be using?)

Hi Ken,

One thing worth pointing out about AutoTags is that they are each 'buckets' of word patterns.

Here's how I think of the algorithm, I hope it helps and I welcome any questions/clarifications:

- When you tag something with TagA then the 'word bucket A' gets filled a little with those patterns. This adding of content to an AutoTag is what we call a 'learn'.

- When you untag something with TagA then the word patterns are taken out of that bucket A - we call this an 'unlearn'.

- When the AutoTag process is run on something new, i.e. when you have something arrive in the inbox and you want to see which AutoTag applies. then that's called a 'classification'.

Classification works by taking the text content of the current message and seeing which 'bucket of word patterns' is the best pattern match. It looks at all of the AutoTags that are currently configured and have something in their 'buckets'.

- Once it has this statistically best pattern match (i.e. which bucket) then it sees if the results are 'sure enough' based on how high the 'pass threshold' is for that particular AutoTag (this is the % threshold you can allocate to a tag.

So in your case Ken a single AutoTag will probably always pass, in that it's the 'winner' of the best match contest every time, plus the default pass threshold is pretty low too (1% I think).

Things to fix this would be to set up more AutoTags and/or raise the 'pass threshold' much higher for the few you have. The thinking around making the threshold default quite low was that the 'unlearn' process really helps the accuracy, so people would see that it was tagged wrong and unlearn it until it corrected itself.

Does that make sense?

There is also another description here: http://www.taglocity.com/help/How_AutoTags_Work.htm which you can read online or from the Taglocity / Help subjects.

Please let us know if this helps, as we are keen to improve this help contents.

I believe it does, thanks.  I was missing the fact that, with some autotags enabled, a classification is nearly always going to occur when new mail arrives.  The match doesn't have to be even remotely "good", it just has to be better than all other matches for autotags.  I was hoping to start with just a few autotags and then add more, if I found them useful.  It sounds, however, if I might actually have more luck if I created a larger number of tags initially and worked to train them all simultaneously.

I'll try to read the article you pointed me to as well.

Thanks for the clarification.

Cool - thanks.

Don't forget, even if you only have two or three AutoTags then you can reduce the 'false passes' you get by increasing the 'Pass Threshold' that they have to score (Configuration / AutoTags (tab))

If you try setting it very high, i.e. say 90%, then reduce it over time, it might help even with a couple of AutoTags.