Discussion: MRS Post-Processing
Moderator: Berthold Crysmann; Scribe: Stephan Oepen
There is currently some de facto recognition of the need for further harmonisation of MRS outputs beyond what is standardly delivered by DELPH-IN grammars:
- The LOGON MT architecture caters for two accommodation phases: one on the source language side (post-parsing), one on the target language side (pre-generation). MRS accommodation includes full MRS term rewriting in addition to variable property mapping.
- Plain output of current DELPH-IN grammars relies on MRS post-processing in order to reach full linguistic adequacy: in the ERG, treatment of idioms is deferred to a post-parsing MRS sanitisation step. Other grammars (e.g. GG) maintain some idiosyncrasies in their output, inter alia for reasons of efficiency in generation (_.*_x_sel_rel etc.). Languages with total reduplication (e.g. Indonesian, Hausa etc.) may need to eliminate the superfluous PRED of the reduplicant..
Consumers of MRS are not limited to MT transfer components. Still, such consumers (e.g., multi-lingual IE applications) will certainly benefit from more harmonic MRS output.
- MRS harmonisation should be a logical part of the grammar. In real life, however, standardisation of output representation is only one of several design criteria in grammar development. MRS Matrix++ compliance is probably best achieved with an MRS processing step.
- Integration of full MRS harmonisation with the grammar/parser benefits other content-oriented multilingual applications, besides MT
- Decoupling of MRS harmonisation from the parsing phase permits grammars to be written more flexibly.
- Processing systems (LKB/PET) already provide (some of) the machinery needed for post-parsing/ pre-generation accomodation. E.g., PET's chart mapping fromalism is in part modelled after the MRS transfer machinery. It may therefore be straightforward to reuse the code for output manipulation.