Having found this out last summer, extracting citations from papers reliably is hard. If you're getting the references direct (no pre-cleaning), then after you solve the semi-colon problem you're likely to find about 1-5% of your citations fail on something else because of how an author has formatted the reference in an intelligible, but machine-confounding, fashion (or mistyped or just missed a bit). You end up with regex after regex, just to catch all the perversities. The ADS project has likely had to solve this problem. You may be able to find some of their utilities on github.

Out of almost 3000 papers, I had to find references for about 50 by hand using the search interface and some intuition. The most innocent unfindable reference was in this paper where A. Einstein, Sit. Preus. Akad. (1919). would have been identifiable if Albert hadn't gone and published 2 papers in that journal that year.

Update: I found it! Six months and no looming deadline has really improved my German.