📅 September 28, 2011 • 🕐 08:44 • 🏷 Blog

*Update*
I added a graph which shows the ratio of articles containing the word ‘Cancer’ to total articles per year. It sadly still suffers from the incomplete data of earlier years:

*Original post*

This is my first attempt to get some data to get some data out of the BioMedCentral dataset, the freely available, Open Access archive of over 40 years of Biomedical research articles. I’ll use this set as a training corpus for my thesis, to extract domain-specific features to use when comparing the similarity between two documents. The dataset consists out of 103.782 articles from 1969 to today.

My text-mining experiment was a very simple one: count the occurrence of the word ‘cancer’ in every article of the journal. My expectation was that the term would occur more frequently as time progresses: as a science journalist I frequently came across (obscure) biomedical research which concluded its findings by in some way linking to (promising a potential way to discover a potential cure for:) cancer. I always figured it had to do with funding. But I’m no expert.

Anyway, to test this I threw together a simple Python script to parse each (xml-formatted) article and extract its date and the frequency of the word cancer, and output this data to a csv-file. I averaged the amount of counts per year per article. Resulting in the following graph:

I hoped to be able to provide an overview of the frequency of the word in ~40 years of BMC. I wasn’t. The first couple of years seem very incomplete: there aren’t many articles (in the hundreds instead of in the thousands as in later years), and lots of “(To access the full article, please see PDF)”-references (yay to Open Access). Anyway, I figured the last 10 years WERE okay, so I graphed the average occurrence of the word cancer of those last couple of years.

Some initial thoughts:

  • The average (word count per article) might be the wrong metric here. Articles dedicated to cancer-related topics skew the average too much. I am actually looking for the papers which do not contain the word frequently.
  • A better metric could be the ratio of articles that DO contain the word (at least once). I’ll give that a shot later and update this post.
  • There does seem to be some increase in occurrence, however I wouldn’t say it’s enough to support my observation.