The continuing battle of parsing Word documents for reading list material
At this year’s Meeting The Reading List Challenge (MTLRC) workshop, Gary Brewerton (my boss) showed the delegates one of our LORLS features: the ability to suck citation data out of Word .docx documents. We’ve had this for a few years and it is intended to allow academics to take existing reading lists that they have produced in Word and import them relatively easily into our electronic reading lists system. The nice front end was written by my colleague Jason Cooper, but I was responsible for the underlying guts that the APIs call to try parsing the Word document and turn it into structured data that LORLS can understand and use. We wrote it originally based on suggestions from a few academics who already had reading lists with Harvard style references in them, and they used it to quickly populate LORLS with their data.
Shortly after the MTRLC workshop, Gary met with some other academics who also needed to import existing reading lists into LORLS. He showed them our existing importer and, whilst it worked, it left quite alot entries as “notes”, meaning it couldn’t parse them into structured data. Gary then asked me to take another look at the backend code and see if I could improve its recognition rate.
I had a set of “test” lists donated by the academics of varying lengths, all from the same department. With the existing code, in some cases less than 50% of the items in these documents were recognised and classified correctly. Of those, some were misclassified (eg book chapters appearing as books).
The existing LORLS .docx import code used Perl regular expression pattern matching alone to try to work out what sort of work a citation referred to this. This worked OK with Word documents where the citations were well formed. A brief glance through the new lists showed that lots of the citations were not well formed. Indeed the citation style and layout seemed to vary from item to item, probably because they had been collected over a period of years by a variety of academics. Whilst I could modify some of the pattern matches to help recognise some of the more obvious cases, it was clear that the code was going to need something extra.
That extra turned out to be Z39.50 look ups. We realised that the initial pattern matches could be quite generalised to see if we could extract out authors, titles, publishers and dates, and then use those to do a Z39.50 look up. Lots of the citations also had classmarks attached, so we’d got a good clue that many of the works did exist in the library catalogue. This initial pattern match was still done using regular expressions and needed quite a lot of tweaking to get recognition accuracy up. For example spotting publishers separated from titles can be “interesting”, especially if the title isn’t delimited properly. We can spot some common cases, such as publishers located in London, New York or in US states with two letter abbreviations. It isn’t fool proof, but its better than nothing.
However using this left us with only around 5% of the entries in the documents classified as unstructured notes when visual checking indicated that they were probably citations. These remaining items are currently left as notes, as there are a number of reasons why the code can’t parse them.
The number one reason is that they don’t have enough punctuation and/or formatting in the citation to allow regular expression to determine which parts are which. In some cases the layout doesn’t even really bear much relation to any formal Harvard style – the order of authors and titles can some time switch round and in some cases it isn’t clear where the title finishes and the publisher ends. In a way they’re a good example to students of the sort of thing they shouldn’t have in their own referencing!
The same problems encountered using the regular expressions would happen with a formal parser as these entries are effectively “syntax errors” if treating Harvard citations as a sort of grammar. This is probably about the best we’ll be able to do for a while, at least until we’ve got some deep AI that can actually read the text and understand it, rather that just scan for patterns.
And if we reach that point LORLS will probably become self aware… now there’s a scary thought!
Comments are closed.