How often are these titles changing?

More titles are added each day, but I re-analyze the data once a month.

How important is it to get "this article belongs to this cluster" right?

It is important, every failure in the process is a bug in the system.

How "tight" are the clusters - these articles all talk about social networking sites vs these articles all have to do with lawsuits against Facebook?

All are articles describing the analysis of metagenomic samples (samples that contains many bacterial species), i.e. where the samples were collected (geographical place), how the sample was treated, what results did they obtained...

A couple of title examples:

1 "Comparative analysis of bacterial communities passing through the g
+uts of three earthworm species with rRNA-based techniques"
2 "Prokaryotic Diversity of An Inland Salt Habitat Investigated by usi
+ng Two Different Molecular Approaches"
3 "High diversity of fungi recovered from the roots of mature tanoak (
+Lithocarpus densiflorus) in northern California"
[download]

Thanks for your interest and help

citromatik,
You really didn't answer my last question which means I probably didn't ask it well. Is your goal only to identify duplicate articles (typos and what not) or is it to group related articles as well? If it is to group related articles, what criterion are you using. I gave an example relating to social network sites. A "loose" group might be anything having to do with a social networking site where a "tighter" group might only have to do with lawsuits against Facebook.

What do you intend to do if an article can belong to more than one cluster? You mentioned that it would be considered a bug if an article ended up in the wrong cluster - how would you know? Is there a subjective element to this task or can you define the criterion for clustering in black and white terms. In other words, if randomly selected 1000 articles and were to cluster them by hand - could you write code that would produce the same results (where computational time is not a factor)?

Cheers - L~R

You really didn't answer my last question which means I probably didn't ask it well.

Ok, sorry for the misunderstanding

Is your goal only to identify duplicate articles (typos and what not) or is it to group related articles as well?

No, it is just to group together articles with the same title (allowing misspellings).

What do you intend to do if an article can belong to more than one cluster?

It should be very improbable a typo that "converts" a title into another one (i.e. different article titles should have an edit distance greater than 5). Nevertheless I can imagine cases where this is not true (for example: Blah, blah, blah typeI VS Same Blah, blah, blah typeII. It will be very difficult to differentiate such cases (not solved yet).

Is there a subjective element to this task or can you define the criterion for clustering in black and white terms. In other words, if randomly selected 1000 articles and were to cluster them by hand - could you write code that would produce the same results (where computational time is not a factor)?

Probably there will be errors in the computational approach that can be ironed out by manual inspection. However, a reasonable small amount of errors will not degrade the final result.

I hope I answered your doubts.

Thanks in advance

citromatik