Skip to content

SaarlandMrsWordnet

LilingTan edited this page Jul 29, 2013 · 8 revisions

A page for the discussion of MRS-Wordnet mappings (part of the work on Lexical Semantics).

Motivations:

  • want to get semantic back-off in parse ranking

  • want to get better tokenization for wordnet annotation (three-toed sloth)

  • combine structural and lexical semantics so we can show semantic compatibility across languages

Francis:

  • There is a SemCor mapping to DMRS on SemCor page. There's a long tail of things where the representation in MRS is different from the representation in WN. E.g. "here" in WN is here_a_1, "here" in ERG (e10:loc_nonsp = here_a1; x11:place_n = here_n1).

  • Single words are easy, but Zina says if you have the gold mapping and map to the correct tree in English, then mapping is good. But if the English tree is wrong then the MRS-WN mapping is incorrect.

  • Quantifiers are sometimes in WN and sometimes not, in particular invisible quantifiers. If we use the ERG as tokenizers, then we are not using the words but the predicates.

Ann: Is there anything useful in monolingual WN for closed class word?

Francis: Closed class word are not in WN but things that have synonymy and adj like are in WN.

Ann: Is there anything useful in the linked WN?

Francis: No, the assumption is that closed class words are not in WN.

Francis:

  • Complex semantics with decomposed things
    • WN has here_adv is diff from here_n, but ERG diff is the presence of absence of loc_nonsp.

Woodley: All of the EP are there but the not all of them are marked with the WN tags?

Emily: Why is that ideal to map one of the "here"?

Francis: We want to "e10:loc_nonsp = here_a1" (scribe missed explanation)

Ann: the decomposition only happens in such a way that some of the things in decompose. You don't need to bother to link to WN but we should rather characterize the elements of the decomposed things which are all the things with underscores "_".

Francis: We also don't want to lose info about "hereness", i.e. "x11:place_n = here_n1"

Slayden: How many predicates without the underscore?

Dan: It's >50 < 1000

Francis: WN has very few superlative, things where they are lexically different. ERG splits this into more_good, more_bad. WN has very little to say about such adjectives.

Emily: Is WN producing the same sense for "quick" and "quickly"?

Francis: Maybe but they are linked. If there is an example where the range of senses between adv and adj then we will be losing information.

Francis:

  • For MultiWord Expression (MWE), both ERG and WN has gobble_up

  • Then we have the Noun-Noun compound, which we need to distinguish between "hot dog" as one lexical item and "guard dog" as a composite Noun-Noun compound

Ann: It will be nice to have an inventory where these Noun-Noun compound occurs.

Francis: It's mostly Noun-Noun compounds.

Francis:

  • For idioms, ERG says it's idiomatic but we don't know how to mark the sense of the elements within the idiom.

  • Then we have coordination, "grass and brown snakes" => grass_snakes and brown_snakes

  • And light verbs? Should we just list them and make entries into WN?

  • We're building a token mapper from ERG+WN, where Dan will make the ERG find compatibility when necessary and Christian (from WN) will do the same with WN.

Clone this wiki locally