currently the ML parser removes most punctuation and
binary stuff from text as it returns it. This is good
for the indexer, but removes too much information for
things like summarisation and URI parsing. There
should be a way to bypass this junk removal. Ideally
it would not affect the efficiency of the current
parsing, and would parse text into the same entities as
currently, just with more information. There are some
implications of this request that haven't been covered
here, such as whether to split text into words
(probably), how to handle whitespace (probably return a
whitespace entity) and so on.