[developers] Paraphrase grammars and SEMI's

Francis Bond fcbond at gmail.com
Mon May 25 14:59:49 CEST 2009


G'day,

   we would like to solicit opinions (especially from Dan) about the
following question:

If you are making a paraphrase grammar, should it use the language
internal MRS, or the post SEM-I external MRS.

That is:

Internal: TEXT - ERG - MRS - EnEn - MRS - ERG - TEXT
or
External: TEXT -ERG - MRS - MRS' - EnEn - MRS' - MRS - ERG - TEXT


My view is that we should use the external MRS for the following reasons:
(i) the SEM-I is the interface to outside resources, such as WordNet,
which we plan to mine for rules
(ii) the internal MRS is more likely to change without warning
(iii) current MT systems all use the internal rules

However, there are also reasons for choosing the internal MRS
(i) full access to the grammar internals
(ii) the current paraphrasing rules use the internal MRS
(iii) two fewer steps to pass through

Using the external MRS also means that the internal hierarchy is
obscured.  To pick a not at all random example, we want to under
specify B's A and (the) A of B.  This is easy to do in the current ERG
due to two hierarchies:
_the_q_rel := def_q_rel, _def_explicit_q_rel := def_q_rel
_of_q_rel :=of_q_rel, poss_rel := of_q_rel.

By creating an MRS with def_q_rel and of_q_rel, both alternatives (and
more) are generated.
However, this is not obvious from the SEM-I, which only records what
predicates are exposed, and says nothing about their place in the
hierarchy.  It is also not encoded in predicates.erg.tdl, although
perhaps it should be?  Currently (for the erg, rather than terg) it
says:

poss_rel := predsort. (in predicates.tdl)
_of_p_rel := of_p_rel. (in predicates.erg.tdl)
def_explicit_q_rel := quant_rel. (in predicates.erg.tdl)
_the_q_rel := quant_rel. (in predicates.erg.tdl)
def_q_rel := quant_rel. (in predicates.tdl)

So, any thoughts?


-- 
Francis Bond <http://www2.nict.go.jp/x/x161/en/member/bond/>
NICT Language Infrastructure Group




More information about the developers mailing list