SaarlandWordNet

FB: We've already discussed WN, but I'd like to see who's interested and what they're looking for, and discuss specific issues

FB: 2 main reasons for doing this:

improve parse ranking

• Using Goi-Taikei worked fairly well; some but worthwhile improvements.
• Trying to redo this with WN
• But we had few marked with ERG gold standard

Tim: Of the treebanked corpora, the intersection was very small

• WeScience with wordnet first sense, decent baseline

• We're mapping about 94% of predicates

• Standard semantic dependencies, from DMRSs

• "I treat dogs and cats with worms", dogs and cats have worms, not treated with them

  • moving "with worms" to link directly to "and"

• wanted to do grandparenting, but ended up just taking long-distance dependencies (e.g. dogs...worms), turned out it helped

• Adjectives and adverbs not very completely annotated in WN

• parse selection accuracy average ~ 26%

  • using supertypes, up to 28.2
  • oe;fb ... this is a smaller model, some lexicalized features were thrown away

• Zina's work had little gains on SemCor, but small gains on WeScience

...

Oe: You're evaluating extact tree match?

FB: Exact MRS match

• Trained w/ Mallet comparing all good trees to all bad trees, perhaps not the best setup
• Feature engineering and supersenses give some gains
• best model got 40% accuracy

Idioms

• The existing WN taggers do not get MWEs
• If you want MWEs right, you're currently out of luck

Tim: There's the Mark Finlayson that does contiguous MWEs

• Some we get, some we don't know how to deal with
• Liling will try to solve some tokenization problems, then finding if we want to work off of words or MRS predicates.
  • for English and German

Discussion

FB: Anybody else want to do things with WN?

Liling: Does anybody know how to deal with the ERG having several entries for one lexeme, but WN only has two.

FB: It's very uncommon, but it's mostly the frequent English words

Tim: I think the answer is carefully and painstakingly

Oe: It's possible to look at the list, there may be distinctions in the ERG that are not sense distinctions. E.g. depends -> depends_on_..._rel, it's possible there's an intransitive use of depends with a diff predicate. This technique may have lead to an inflation of predictate distinctions.

....

Oe: You should only be looking at predicates, not syntactic elements

FB: Yes, we are only looking at the semantic predicates

Liling: Does anybody else want to map them besides me (and Francis)?

FB: Woodley tried to replicate Zina's early numbers without success..

Woodley: yes...

Sanghoun: Other languages?

FB: Yes, we're working with Japanese and Chinese, ..

Sanghoun: Is there a Chinese wordnet?

FB: Like Poland, China has many wordnets, but they weren't free. Annoyingly, a month before we released ours, (someone...) released theirs. There's some available from Chinese academies, some from Academica Sinica in Taiwan. If you want ours just ask.

Tim: Are you considering cross-lingual syntactic sense disambiguation?

FB: Yes, I'm doing the cross-lingual WSD, someone is working on syntactic disambiguation; the idea is to use the linked WN senses. But we'd like to have more working on it

Tim: The kuchi-wo-hiraku example, you have it sense annotated in English...? You know the lexical entries in Japanese and English, and likely they are not the same, can you use that difference to pre-sense-annotate other alignments

FB: We're finding half of the predicates don't link to anything; they are so far apart we don't want to make a strong claim initially. But yes that's why we're looking at multiple languages, so we can try to find common senses....

Tim: There are some similarities with POS tagging to lower-density langauges by projection, aligning across things and so on... Here you're doing it with synsets, etc.. Perhaps the data can be used to bootstrap such a model

FB: Yes and we're currently doing that.

FB: One issue is that the coverage is far from perfect. We'd like to know how much we are adding for English for each new corpus; are we filling the bucket or pouring water into the ocean...

FB: Good news is we're not the only ones working with wordnets, so things in other areas may expand out.

Tim: Perhaps some triangulation across multiple sense link annotations can help with aligning things...

FB: We'd like to treat things like "hot dog" as not a "dog", unlike say "guard dog". Does anybody have any thoughts about that?

Oe: I'd ask: are there diffs that are grammaticised. "Hot dog" I assume is lexicalized an unambiguous. I don't see how there'd be a distributional difference in syntax. It could be ambiguated at the SEMI level.

FB: I'd like to have "hot dog" as a single word, and "hot dog" as two words in the WN, but we'd almost always select the single word

Oe: We should maintain divisions of labor in these concerns, and decide how to deal with these situations.

Mike: what about variations, like "chili dog" and "tofu dog"

FB: This just means the lexicon gets bigger, unfortunately.. It'd be nice to combine them if possible

Oe: Is "Kick the bucket" in there?

FB: No, surprisingly

Oe: It is, in idioms.mtr

FB: Try parsing

Oe parses; gets idiomatic reading

FB: Oh. I parsed earlier and did not get them. Anyway, it's in WN as a synonym of die.

Oe: These are transfer rules with only something on the input side, and perhaps we need an output side. We could have kick+the+bucket_v_rel

FB: But for "let the cat out the bag" cannot be lexicalized like that. Or "rack your brain", "rack your tired brain", if transfering to "think hard", where do you attach "tired"?

Tim: Interestingly, it's listed as "rack" in wordnet... "to rack"

FB: If we do it as post-processing, research suggests we need access to both the idiomatic and literal meanings, since people can play around with the words.. True for metaphors, as well.

Liling produces example... e.g. getting paint from pigment... etc..

FB: We may find when mapping that the ERG is missing some subcategorization frames

...

FB: So when we map, there's often not a 1-to-1 correspondence. I expect the mapping to show interesting holes---it's one reason to do it. For those have used WN, the noun part is better than the verb part, verbs better than adjectives, etc..

FB: Finish with a plea: if people are interested in tagging more data, I'll put something up on the wiki. We're trying to come up with a small set of things that we'd like to all annotate together, e.g. Cathedral and Bazaar, etc..