Possibly if there was any Perl module to replace pdftohtml -xml , ie to get character clusters by positions.

Then one could try to combine histograms of word positions with further user hints like font or area of position.

Not sure if CAM::PDF can be used to get word positions, the docs only mention "objects", whatever that means...