I find XSLT a bit annoying to work with (for the reasons you describe above). I’ve done some things with it, but nothing like this. For this and some other reasons, what we ended up with was a Python tool that reads data from Word 2007 (and later Excel 2007) documents into internal data structures, and then generates C++ header files containing that data. We had to rebuild to refresh data anyway, and given that this was a DS project the time and space savings were not insignificant. Also, it was simpler: we didn’t need any XML or JSON libraries.

Part of why I chose this approach was that I am more familiar with Python than with XSLT or C#, which naturally is kind of an arbitrary thing. However, I honestly don’t know if some of the more complex logic I later had to write to parse Word files (handling invisible text written by translation tools, or especially handling line breaks removed using Track Changes) could have been implemented using XSLT. (But maybe it would have been easier? I don’t know.)

An additional benefit of the route we took is that we could do a lot of preprocessing and especially analysis and error prevention in the tool, which greatly sped up our workflow.