The problem is, the more roles you have, the more you, and your annotators, will be confused about what the right one is, and annotator agreement will go down the drain (FrameNet is a good example of this, the roles for similar words are often incommensurable). Point is, there are only a few real high-level thematic roles (7 seems in the ball park), and anything more will
I agree that looking at John Sowa’s work is probably a good starting point, an due has a good set of basic ones (http://www.jfsowa.com/ontology/thematic.htm is probably a good simple version).
To be honest, we tried to do something at Powerset, and eventually gave (more or less) up and went with whatever the XLE put out with some minor massaging.
So: vote for Sowa
Cheers Martin
On Feb 2, 2014, at 1:45 PM, Adam Przepiorkowski adamp@ipipan.waw.pl wrote:
Dear All,
[Apologies for cross-posting.]
In the context of enriching the Polish LFG grammar with semantic representation, we are looking for a set of semantic roles (Agent, Patient, Beneficiary, etc.) that could be used to mark arguments (and possibly adjuncts) of verbs and other predicates. This set should be exhaustive in the sense that it should be possible to assign – more or less deterministically – a semantic role to any argument (and possibly adjunct) of any predicate. For this reason the standard – in LFG textbooks – sets of some 7 semantic roles do not seem sufficient. Instead, we are looking at larger repertoires proposed in VerbNet, in FrameNet and in John W. Sowa's work on Knowledge Representation.
We don't have any strong views about any particular set of semantic roles, as long as it is exhaustive and applicable to real texts (as opposed to being merely theoretically interesting). Has anybody in the LFG community faced a similar task? If so, what set of semantic roles would you recommend? At the moment, we are wavering between VerbNet and Sowa's system, both being more manageable than numerous roles offered in FrameNet, but we are open to other solutions.
Many thanks, best regards,
Adam P.
-- Adam Przepiórkowski ˈadam ˌpʃɛpjurˈkɔfskʲi http://clip.ipipan.waw.pl/ ____ Computational Linguistics in Poland http://jlm.ipipan.waw.pl/ ___________ Journal of Language Modelling http://zil.ipipan.waw.pl/ ____________ Linguistic Engineering Group http://nkjp.pl/ _________________________ National Corpus of Polish _______________________________________________ ParGram mailing list ParGram@mailman.uni-konstanz.de https://mailman.uni-konstanz.de/mailman/listinfo/pargram