Lexical semantics discussion notes
Francis: Who sees the potential use for lex sem in a system?
Rebecca: For QA (fuzzy predicate matching) Yi: In Chinese grammars not enough syntactic info for disambiguation Francis: Likewise in Japanese -- currently have some controversial rules that the syntax can't disambiguate Woodley: QA Glenn: thai-language.com MT Emily: Add to ANC to increase appeal to broader IJCAI/AAAI community Francis: MASC is being annotated by Princeton for WordNet Mike: Generation and paraphrasing; Prescott would likely also be interested for summarization. Justin: Have a strong feeling that lexical information impacts syntax. Stephan: Back in the MT days it seemed like it would have been useful to have predicates in a hierarchy, but that became less urgent with work on learning transfer correspondences. But also generally think we should know something in this space.
Francis: Shared synsets across languages -> can train a disambiguation model on Spanish tagged data and use it to do WSD on Japanese. The fact that dogs bark is not necessarily just a fact about English.
Stephan: Aware of some unhappiness with the definition of the task, if conceived of as resolving to English WN senses. OntoNotes comes with a small upper model.
Francis: Philosophy of OntoNotes was to generalize up the hierarchy until they could get to 90% IAA. Generally, the level of granularity in any lex resource is inconsistence and for WN has been viewed as too fine-grained. Hierarchy allows for some generalization, but also looking to something more coarse grained could be good. SUMO (de facto IEEE standard, top parts of WN and some other things).
Francis: Would be good to have a flexible framework allowing us to use multiple ontologies, but in the short term will I will be working with WN. Level of granularity is an issue analogous to comparing deep v. shallow parsing. We estimated we needed 2000 for verb disambiguation, 20-30,000 for noun disambiguation.
Glenn: Advantage of WN is active development work on it. But: complaint has been too many overlapping senses at any given level; just going up the hierarchy doesn't necessarily help.
Stephan: How much hierarchy is there in FrameNet?
Francis: WN doesn't have everything, doesn't have selectional restrictions on verbs. I think we'll want to learn that from a corpus and then store it somewhere separately. Other things as well like VerbNet. We'll often get a gain (for disambiguation and generation) by knowing semantic class of arguments preferred by verbs. Mappings between FrameNet, VerbNet, WordNet can help.
Stephan/Yi: Sergio Roa has worked on mapping to ERG predicates, a couple of years ago.
Glenn: What about Cyc?
Francis: Yes, Cyc is a contender. One level further on from lex semantics, more like frames. Slightly bigger step, maybe not ready to take it yet, but could be an interesting task. E.g.,: "Water will spill if you turn a glass upside down." But also "A dog is an animal". Experiments on predicting countability of English nouns did as well as (slightly better than?) WN.
Glenn: License is Apache.
Francis: Also lot of work on bio-specific hierarchies.
Francis: Summary so far. We have a notion of why this could be useful. For QA and paraphrasing ISA hierarchy is what you want. One reason WSD has been found unsatisfying is tendency to look at only single word entries, but 50% of the entries are MWEs, and that part has a lot of discriminating power. Advantage for us because of our model of MWEs.
Stephan: That's an argument of operating at the level of MRSs, where we have already reprsented the MWEs as single things.
Rebecca: Not using MRS for parse selection at this point.
Emily: Feasible to project MRS info back onto trees?
Rebecca: Would prefer reranking which uses MRS information as first pass approach.
Yi: Master's student compared to syntactic parse selection and didn't see big gain from using MRS, except maybe some domain-portability improvement. But that was without lex sem.
Stephan: Woodley recently tried some MRS derived features and found some gains, but it's not been a silver bullet.
Francis: Baldridge et al also found some gain, but better to build separate syn and sem models and combine them, rather than building on model. We found some gains, but only if we were using the type hierarchy. Only looked at Japanese for that one, where the grammar is less syntactically constrained than the ERG. (Based on Oracle WSD results.) Tim and Egirre have also found gains, using automatically disambiguated senses.
Emily: Backing up to the controversy, just because we want to take advantage of MWEs doesn't mean we have to work at the level of MRS. Can ambiguate at the level of the MWE entries. (Except maybe true idioms, but would speculate that those only have one idiomatic sense.)
Francis: Not ready to accept that speculation without checking.
Dan: PET doesn't yet handle the idiom machinery.
Stephan: Could see how many of the multiword WN entries could be accommodated early on in the lexicon.
Dan: Ann's position on noun-noun compounds is that they are deceptive in looking like a completely productive process in English when in fact they are are not. There are lots of unattested noun-noun compounds that are surprising. She assumes a long list of noun-noun subregularities --- semiregular patterns constrained by semantic sorts etc. A good lexicon would provide a characterization of those subregularities that are not fully productive. ERG contains none of that insight allowing almost any noun-noun combinations.
Stephan: We don't currently have the machinery (additional technical devices) which would support this.
Stephan: Not only NN, also A-N combinations (strong tea, strong drink, tall building, tall person, high building, *high person [different sense]).
Dan: Collocation constraints are very clear to speakers, but not clear how we would or should encode them.
Francis: Science would advance if looked at how many of the noun compounds can be handled with simple constraints.
Dan: Example of constrained NN compounds: "doctor's office", "dentist's chair". The generalization seems to be that if the left member of the compound (human?) than you can do it. Also: *doctor office *men room.
Emily: Other classes of NN compounds don't have a morphological mark, but do have characteristic semantic relationship between the parts.
Francis: Another example is in matching numeral classifiers to nouns. But need synecdote: "three person-CL of Microsoft" DFKIが三人来ました。
Francis: Was completely convinced that we can't do it as restrictions on sorts. Something like "machine translation" where there is a single node for that in WN would like to be able to map to that as a node. "Gaurd dog" is a "dog", "dog" is a "dog". Putting in things like compounds doesn't necessarily pack as easily.
Emily: "machine translation" as a single node shouldn't block the compositional reading (with a different sense of translation --- translating the machines). So maybe that one we'd need to ambiguate from the beginning.
Stephan: What is the external distributional difference?
Dan: "machine translation of books" but not "machine translation of that ENIAC"
Francis: If we accept paraphrase as a test, "translation by machine" and "translation of machines".
Stephan/Emily: That's not a syntactic test.
Francis: Still useful as a test to determine what should be in the semantics.
Stephan: Back to question of whether to ambiguate early. It's entrenched in our technology not to use the semantic features while constructing the forest. Further postpone construction of the actual MRS until we're done unpacking, because final feature structure is needed for that. So there should be strong arguments for taking the ambiguate early position.
Francis: If we don't need it along the way, then doing it at the end makes more sense.
Emily: One place it could be used along the way is in chart pruning/best first parsing.
Francis: And incremental parsers might also want it.
Emily: [Disclaimer: Not necessarily an advocate for this position.] Another possible objection is the linguistic notion that the lexicon should be a single holistic thing. Two separate lists of lexical entries would need to maintain their alignment with each other. Also, some things would be in one but not the other (e.g., "machine translation").
Stephan: Proposal is that the lexicon distinguishes between syntactic and semantic information. Same as the one that parsing with ambiguity packing makes. Would be possible at least that some entries have an empty syntactic part (e.g., "machine translation" the collocation). Parallel to derivational morphology.
Francis: Would like to believe that the bulk of the mapping can be done largely automatically. Some might need to be done by hand, e.g., because the internal structure of MWEs is not represented in WordNet. Will require some effort to build, but then will be stored, and updated (somehow) as one or both changes.
Emily: Some things in one but not the other begins to sound like motivation for two separate lists anyway.
Francis/Stephan: Link is from dog lexical entry to one node in WN, could be stored in lex entry if we wanted.
Woodley/Francis: But actually the word node in WN will collapse distinctions that the ERG makes and can't say that ERG _dog_n_rel has all of those meanings linked to the dog word node. Even in the noun space, we might have subcat information that will partition the WN sense space, and it is an interesting research question whether we can get at that from the semantic hierarchy (e.g., relational nouns do one thing).
Francis: This is one of the reasons we didn't do this 10 years ago --- not just an engineering problem.
Laurie: But these issues may be informed by the practicality of trying to make this generalizeable and not chained to WordNet.
Woodley: I agree. It's going to be a complicated mapping, and WordNet might not stay so attractive.
Francis: OTOH, can't really do the experimentation without a mapping. WordNet appears like it will at least be maintained in five years' time. There will be a cost for adding any new lex resource, the more generic the mapping, the less we can make that cost, but...
Dan: Doesn't have to be a change in the original grammar source file. Could well have core.smi and core.wn.semi, where core.smi is as currently, core.wn.smi is a mapping from ERG to WN, could have separate mappings for other resources like core.cyc.smi... Those should be partially automatically constructed to the extent possible, while also allowing for hand-work.
Woodley/Dan: May turn into a more general MRS to MRS mapping task (transfer rules) that takes advantage of context.
Emily: Still need lexical resource of some sort (even if only a set of transfer rules) and there could be multiple versions of these for different external lex semantic resources.
Stephan: Compatible with the view that the grammar provides all the information that is constrained by the syntax.
Dan: predicates in the ERG are publicly defined hooks that can be used to go get more information out of lex sem resources.
Francis: I have a feeling of a general shared consensus that we look for the post-semi mapping to WN using the transfer machinery or some other new machinery. Possibly there will be some fix up on both sides as I start working on this mapping.