In order to demonstrate some stuff I wanted to have a set of tagged data to play with. Delicious, flickr, that sort of thing. After some digging around on places like theinfo.org I found out that CiteULike (like delicious but targetted at academic papers) makes a dump of their data available. Unfortunately, it’s a bit messy. Not the data format itself, which is a simple pipe separated text file, but the quality of the tags themself. This is more or less to be expected for user reported tagging. It would be nice to have something a bit cleaner though.

I thought it might be illustrative to clean it up in public. It’s a completely hacky process, and not a particularly smart one, but it might be interesting or helpful to someone.

So, first things first. Get the data: http://static.citeulike.org/data/current.bz2. It’s zipped, so not too large, but will be about 300M unzipped.

It looks roughly like this:

42|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:25:05.373798+00|ecoli
42|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:25:05.373798+00|metabolism
42|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:25:05.373798+00|barabasi
42|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:25:05.373798+00|networks
43|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:25:51.839281+00|control
43|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:25:51.839281+00|engineering
43|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:25:51.839281+00|robustness
44|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:26:33.156319+00|networks
44|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:26:33.156319+00|strogatz
44|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:26:33.156319+00|survey
44|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:26:33.156319+00|review
45|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:27:38.983179+00|pleiotropy
45|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:27:38.983179+00|barabasi
45|61baaeba8de136d9c1aa9c18ec3860e8|2004-11-04 02:27:38.983179+00|notsmall

A couple things to note:

As mentioned, it’s pipe separated. We have a document id, an anonymized user id, a date and a tag. There can be multiple tags for the same (document, user) pair.

Another thing to note is that sometimes it contains concatenatedwords. e.g. “notsmall” (a tag which seems to appear only once, for “Wrestling with pleiotropy: genomic and topological analysis of the yeast gene expression network” for some reason. I think it’s a mistag and was meant to go on “The metabolic world of Escherichia coli is not small”).

Let’s get a sense of what the tags for this look like. We’ll plot a distribution graph like so:

 cat Desktop/current | ruby -ne 'puts $_.split("|")[-1]' | sort | uniq -c | sort -g -r > citeulike_tags

What’s this doing? Not much. We’re catting the data to standard out, feeding it through ruby to split off the last column and sorting the results, giving us a big list of tags with repetitions. We then pipe this through uniq -c, which collates consecutive unique lines with a count prepended (that’s what the -c does). We then sort again in generalised numerical order, reversed, and save the output to a file.

Unix is fun.

Here’s what the results look like:

[email protected]:~$ head citeulike_tags
 212611 bibtex-import
 156901 no-tag
  27926 elegans
  27886 celegans
  27825 c_elegans
  27795 nematode
  27738 wormbase
  27736 caenorhabditis_elegans
  18897 review
  15280 all-articles
[email protected]:~$ tail citeulike_tags
      1 00301512
      1 0025
      1 00208
      1 001287275ep1114608epa00128727epb00128727
      1 0010521342
      1 0009811908
      1 0009390946
      1 000
      1 -------------
      1 ___

So there’s clearly a bunch of random noise there. The top two are unimportant – they’re some sort of autogenerated thing – and the bottom lot are garbage. So, we’ll throw away the top two and everything with only 1 occurrence.

[email protected]:~$ vim citeulike_tags

Let’s look at the data again.

27926 elegans
27886 celegans
27825 c_elegans
27795 nematode
27738 wormbase
27736 caenorhabditis_elegans
18897 review
15280 all-articles
14597 evolution
13694 meeting_abstract
[email protected]:~$ tail citeulike_tags
2 035
2 0325
2 0315
2 021
2 01012
2 004692
2 003
2 0022
2 —-
2 __
The top is looking better. I’m a bit skeptical of “all-articles”. Looking at a few examples it seems to be something generated along with bibtex-import. But we’ll leave that for now.

We’ve got a bunch of duplication there. “celegans”, “c_elegans”, “caenorhabditis_elegans”. Citeulike seems to have a nematode obsession. Not much we can do about that right now though.

The bottom half is the more serious issue. It consists entirely of numbers, which is rubbish. So let’s filter out anything that doesn’t contain some text:

[email protected]:~$ mv citeulike_tags citeulike_tags_old
[email protected]:~$ cat citeulike_tags_old | ruby -ne ‘puts $_ if !($_ =~ /^[^a-z]+$/)’ > citeulike_tags
[email protected]:~$ tail citeulike_tags
2 0graphs
2 0ds-flicker
2 0-doek
2 0compas
2 0cath
2 05-wise-00-01
2 05matheco30_theoretic
2 05matheco25_theoretic
2 04-sose-00
2 041500040u1

Hm. Those are still pretty shitty. At this point I’m tempted to filter out everything which appears only twice. But after a quick trawl through with less I shall resist the temptation, as one finds things like this:

2 accessibility-technology-for-the-deaf
2 accessibilitystandards
2 accesory
2 acceptorphysiology
2 acceptormetabolismphysiology
2 acceptorgeneticsmetabolism
2 acceptability-judgments
2 accents
2 accent_learning
2 accentedness
2 accelerator-physics
2 acceleration-measurement
2 accelerationadverse
2 accelerating-admixture
2 accelerated-combinatorial-synthesis

Which it would be a shame to lose out on if we can avoid it.

At this point I noticed something annoying: There’s absolutely no consistency in how people space things in these tags. Even ignoring the people who concatenatewords, some people use _, some use -, some even have things like “academic-libraries—-collection-development”, which is just aggravating.

Let’s try to get some consistency out of this.

At this point I break out of the console and move to irb for some more interactive hacking. Unfortunately trying to record this turned out to be a pain due to IRB’s tendency to print the whole giant hash. So here’s it is as a ruby script. This looks for all tags which differ only in terms of the presence and type of _s and -s and conflates them all. In each case it chooses the most commonly occuring one, takes that as canonical and gives it the sum of all occurrences of equivalent tags. It then normalises the separator to an underscore.

I ran this as

[email protected]:~$ cat citeulike_tags | ruby group_duplicates.rb > normalized_tags

Now, we’ve still got quite a few tags left:

[email protected]:~$ wc -l normalized_tags
118785 normalized_tags

Let’s see if we can figure out some good ways of reducing this (or at least cutting out noise).

I dug around in it for a bit and noticed that there were a lot of tags of the form “file-import-something”. Not that many (291) but it’s a start. We’ll probably continue blacklisting things as we fine them.

Here’s an example of where we’ve got redundant tags: We’ve got genomic and genomics. statistic and statistics. population and populations. i.e. plurals. Let’s fix that.

For this step we’ll use a stemmer. Snowball has a decent binding for ruby, so we’ll use that. We’ll consider two tags to be equivalent if their parts stem to the same words.

We’re going to repurpose the spacing script above. Software reuse by cut and paste, yay. :-) I’ll need to do this all properly later, so I’ll clean it up for then, but now we’re just experimenting with data. So here’s a rewrite to identify things by stemming.

At this point we’re down to 106229 tags, from 118494 prior to stemming and 272919 originally. Not doing badly, given that most of what we’ve thrown away is junk or redundant information. If we threw out the tags which only appear twice we’d lose another 30352 tags (cat normalized_tags_2 | grep ”    2 ” | wc -l). I was resisting doing that because some of them are quite good quality, but really I don’t think we have enough information to clean the remainder up.

We’re nearly at the point where we’ve run out of what we can do with frequency and word based information – there’s still plenty more we could do in principle, but we’ll hit the point of diminishing returns pretty rapidly from here on out. One thing I have noticed though is that there are a bunch of tags like “of” and “and”. At this point I shall put out the favourite hacky linguistic hammer: The stopword list.

The one I tend to use was compiled for the SMART system by Chris Buckley and Gerard Salton at Cornell University. You can download a copy here. All we’ll use this for is to filter out any tags which happen to be stopwords. Here’s an obvious script to use it to remove stopwords.

This loses us 46 tags. Doesn’t sound like many, but they were mostly fairly high frequency ones, so it’s a nice win.

At this point I declare this part of the work done. We’ve compiled a good list of tags and, although I’ve not actually written the code to do so, because each tag was arrived at by grouping a bunch of existing tags, it’s easy to see how you could figure out which documents are tagged with which of the new tags (if you can’t see, don’t worry, I’ll be tidying this up and using it to generate a tag list next time). In order to go further, we’ll need to actually look at sets of tagged documents. And, to be honest, I don’t know how much further it will take us. It may be that there’s not much more to add. But I’m hopeful.

If you want to have a play with the data, here’s the end result.