Comment by burtonator | Hacker News Reader

Comment by burtonator | Hacker News Reader

Parent Full thread

I've been working on a tag-centric document manager for the last 90 days:

https://getpolarized.io/

It allows you to tag and manage everything you're reading and the primary management metaphor is the 'tag'.

It's a central repository for all your reading. PDFs, web content, and it supports annotation (highlights, comments, etc).

There are a bunch of clear wins here.

Related tags are awesome. You can build a bunch of really nice derived metadata just by managing your tags on the documents.

For example, I've been reading a lot about growth hacking as I'm trying to get Polar to 100k users. (We're officially 10% of the way there today!)

Since I have a bunch of documents named 'growth' Polar recommends startups, venturecapital, and marketing as secondary tags.

Also, since Anki supports tags (and Polar supports Anki sync) these systems are compatible.

Polar tags are also Twitter tags. While we haven't enabled this feature yet your annotations could be exported to Twitter and the systems are compatible.

We're still getting users that want hierarchy though. I'm not sure what I want to do here.

I'm considering building out algorithmic hierarchy.

For example, if you have two documents with the first tagged 'tech, microsoft' and the second tagged 'tech, linux' we can infer that tech is the parent of 'microsoft' and 'linux'

... and it would still support tags.

the problem I have now is that folder hierarchy is still being requested by our users but I'm resistant to implement it yet.

Tags do not fit spatial memory. Tags are utterly worthless in any spatial context.

You're at the mercy of a search engine's implementation and index foibles for finding things through search queries.

A hierarchy maps onto our spatial memory, which we use to figure out where we've been and go back to it.

There's a lot to be said for maintaining spatial metaphors in information space.

Another little detail here is that spatial organisation (in meatspace) is a shared environment. A single 'shared' ontology is useful even if it's nonsense, because it's still a shared frame of reference in which navigational directions can be shared.

If you use modern NLP you can embed tags into vectors and then just use dot product between query and targets to rank them. There's even a fast library for that in Python called 'annoy'. It is capable of relating any query to any tag.

Did you read the last part of my comment?

I clarified that there's a workaround to give hierarchy.

> For example, if you have two documents with the first tagged 'tech, microsoft' and the second tagged 'tech, linux' we can infer that tech is the parent of 'microsoft' and 'linux'

How do you account for the availability effect [0]?

To use your example, let's say that I tagged a document 'tech, microsoft' and the second is tagged 'linux, tech', because of the availability effect -- it's been a long time since I tagged documents causing me to add the 'tech' tag as an afterthought rather than as the first tag?

Better to keep things simple.

0: https://en.wikipedia.org/wiki/Availability_heuristic

> We're still getting users that want hierarchy though. I'm not sure what I want to do here.

Have you considered hierarchical tags?

As an example:

https://dmitryfrank.com/projects/geekmarks/article

I use it for my bookmarks, and I find it is a suitable balance between having some kind of hierarchy, while allowing for resources to be classified under several categories.

Reply on news.ycombinator.com