There may be some overlap with the Procedure Committee here, but presumably no contradiction.

• Back to April 2-3 Meeting page
• Back to Content Committee Page

Here is a version of Rebecca's list of interesting paraphrase possibilities and where we normalize them. Here, "normalize" means give the two variants exactly the same representation.

We give no examples here, see Rebecca's list of interesting paraphrase possibilities for examples (perhaps open that file in another window).

Phenomenon	Where Normalized
Lexical synonymy	IL1
Different sentence types	not normalized
Different argument realizations	IL2 (requires ontology work)
Overlapping meanings	IL2 (coercion and adding new nodes )
Syntactic variation	IL0
Clause subordination vs. anaphorically linked sentences	IL2 (with coref annotation)
Noun-noun phrases	IL2
Comparatives vs. superlatives	not normallized
Inference	not normalized
Inverse relationship	not normalized
Head switching	IL2
Viewpoint variation	not normalized
Morphological derivation	IL2

Foreign Language IL0

The only issue for the content committee is that all IL0s should be as similar as possible. In general, this means that there should be no function words, words should be morphologically normalized, syntactic argument structure alternations such as passive should be normalized, and so on. In particular cases, if English has made a choice which also makes sense for another language, follow that choice, even if other choices are possible. (Always remember: there are often many ways of representing the syntax of a language, and it is often hard to choose among them.)

IL1

Near-Synonymy

Near synonyms should be mapped to the same meaning at IL1. We need ot make precise what we mean by near synonyms.

Extending IL1

See the procedure committee report for a definition of IL1+ (Section 4), which includes multi-word expressions as one node, categorization of named entities, and a speciallight verb treatment. The latter is repeated here.

Light Verbs

Light verbs are verbs which do not contribute any meaning in a sentence, because an argument (usually the direct object) carries the entire meaning of the verb-argument combination. Typically, this verb is leixcally idiosyncratic to the noun and it is not semantically predictable; as a result, it is frequently different in different languages. Examples: take an oath (German: `discard an oath') or ask a question (German: `put a question'). Most light verbs come from a list of about ten verbs, but there are also more. (Note: a verb is not intrinsically light, only in combination with an argument.)

We annotate light verbs at IL1 as follows. The verb gets a special "Light Verb" meaning, and the dependent noun gets the full meaning from Omega. The arguments and adjuncts attached to the verb are understood as referring to the menaing of the argument. The verb-argument combination becomes one node at IL2.

IL2

Annotation by Hand or Automatic?

Much of the gap between IL1 and IL2 can be done automatically, using various heuristics. We should explore this as much as possible, to make the final human annotation step as simple as possible (ideally, just a check that the automatic IL2 is correct).

How Large are the IL2 Graphs? Are they Trees?

There are several options. There seems to be no desire to force IL2 to be a tree, so we should assume a dag. The other question is whether we get one dag per clause, one per sentence, or one per paragraph. Since the sentence is the most artificial unit from a semantic perspective, this would seem to be the least desirable. There was not consensus for teh paragraph-seze dag, so this should be discussed again.

Requirements on Ontology

The following requirements on the ontology arise from plans for IL2. If the ontology does not supply this information, we may need to ask IL2 annotators to supply it.

The ontology must reveal which verbs refer to events, states, or qualities.

Coreference

Coreference is crucial for IL2. Since it is fairly well understood (in the sense that other projects have annotated coreference), we need not worry about it before Aug 2004, as we will not be in production mode before then. However, we should keep our eyes and ears open for any annotation manuals and tools.

Scope

The reprepsentation of scope is deferred to Year 3, when Christof will solve the problem elegantly. In the meantime, some logically equivalent sentences will have different IL2s. We all collect interesting examples of this type as we encounter them.

Noun-Noun Compounds

The meaning of the compounding will be made explicit at IL2. For example the velvet dress example in Rebecca's list is the made-of relation, or in monastery visit, the relation is whatever the relation between visit and monastery would be in they will visit the monastery. It may be possible to get at this relation from other IL1s.

Overlapping Meanings

"Overlapping Meanings" is the term from Rebecca's list for some of the most brutal of Bonnie's divergences. There are several options at IL2:

• Coerce IL2 to be equivalent to one or the other of the two possible representations. For example, swim across the lake would look like traverse the lake swimmingly (or vice versa).

• Add extra nodes. This is the LCS solution: the meaning is decomposed, either until all paraphrases are represented the same at IL2, or until some pain threshold is reached.

Relative Clauses

Many ways of representing them -- the choice depends also on the size of the dags.