Differences between revisions 2 and 3
Revision 2 as of 2006-06-17 08:15:53
Size: 3615
Editor: ValiaKordoni
Comment: FeforMRSinTranslation
Revision 3 as of 2006-06-17 08:49:57
Size: 687
Editor: ValiaKordoni
Comment:
Deletions are marked like this. Additions are marked like this.
Line 1: Line 1:
MRSs in Translation: interlingua vs. semantic structure

-- Semantic objects and their relation to languages: are they language specific or shared among languages. Rels are shared among languages.

a. Granularity in semantic representation

b. Structure of the semantic space: none, hierarchy, network structure: antonymy

c. Architecture: Pipelined or semantic components

-- DELPH-IN approach:

a. Advantage: grammars, parsers, etc., can be combined via MRSs (see LOGON)

b. MRSs are langage-specific

c. MRSs are shared among languages

d. Grammar does all the work

e. Stochastic Ranking

f. Transfer may be used when a paraphrase-set is monolingual

g. Paraphrase sets for ambiguous sentences

-- Pros and cons

1. Flexibility and practical application: Transfer rule model translational relations are much more flexible than MRSs, but

2. Transfer rules do not model semantic relations

-- Transfer and semantics

a. Words vs. constructions: the problem is the same for text understanding, or question answering with a single language

b. The relation of translation to semantics

c. Grammar harmonisation

d. Translatability and the nature of languages: all meanings are not shared

e. TL may require grammatical distinctions that are not part of SL

Structure of a semantic space

a. Matching on proximity

-- Issues for discussions

a. Semantic types and features

b. Granularity of semantic representation

c. Structure of the semantic space

d. Architecture

-- Contributions from the floor:

EB: Is MRS interlingua? Of what level of granularity

FB: To do the interlingua, you need to have more info about senses than we currently have. There will be very few concepts that are markable from all languages.
Some of the problems require us to go one step more abstract. Perhaps MRS can't be an interlingua.

JT: Some experience from LOGON: transparent rules vs. non-transparent rules. It's possible to go from LFG to ERG. In Norwegian MRSs are simpler than the MRSs
in English.It's interesting to write transfer rules when there are differences among the languages. But if there are not, then what might be needed more is a better harmonisation.
Semantics closer to languages or abstract away from the argument structure.

AF: Why would you need an interlingua? Would you want to write a machine translation system by writing manual rules?

LA: What are the relevant semantic aspects that we should encode?

FB: 1. In the short term, going to a high level interlingua might not be a good idea. But it might be the final goal.
      2. How often don't you know what a correct MRS in 2 languages is? Question to DF.

DF:There are cases where there are MRSs in the 2 languages (Norwegian and English), that are all correct, and one has to make an expert decision of which one is the best one according to the
preferred analysis.




RandomQuote()

RecentChanges

last, but not least, a randomly chosen page: RandomPage(1)

Icon(diffrc)

marks older pages that have at least one backup version stored (click for an author diff)

Icon(updated)

marks pages edited since you set your bookmark (click for a bookmark diff)

Icon(new)

marks pages created since you set your bookmark, and were not edited after creation


This page contains a list of recent changes in this wiki of PageCount pages (more system information on SystemInfo). For changes on other wikis, see WikiSites/Aggregation.

RecentChanges (last edited 2011-10-08 21:12:17 by localhost)

(The DELPH-IN infrastructure is hosted at the University of Oslo)