In some pdfs, paragraph detection is too fine or too coarse. Why don't you implement your own paragraph detection logic?