next up previous
Next: Applying Constraints Up: The Semantic Analyzer Previous: The Semantic Analyzer

Generating Constraints

Step 0 in the semantic analysis process is acquiring the syntactic analysis of the input sentence. To avoid duplication of effort, uK uses the output of the Pangloss MT syntactic analysis module (or Panglyzer), also developed at NMSU. gif Since, at the present time, the Panglzer makes all attachment decisions, uK is limited to deciding between word sense meanings. gif

   
Figure 6: Possible Word Senses for Example Sentence.

The first real step for uK is to gather up all of the possible lexicon entries for each of the words. Figure 6 lists all the word-sense mapping possibilities for the example sentence. For ``adquirir,'' the two lexicon entries shown in Figure 5 are retrieved, with mappings into ACQUIRE and LEARN word senses. For each word sense, the SYN-STRUC zone must be examined to see if it fits the current sentence. If it does, then the VARs must be bound to their corresponding word instances in the current input sentence. For ``adquirir,'' both word senses have identical SYN-STRUC zones, so the variable binding process displayed in Figure 7 applies to both.

   
Figure 7: Variable binding for ``adquirir''.

After variable binding, the semantic analyzer examines the SEM zone of each word sense in order to construct a list of constraints that must be satisfied for that word sense. Constraints can arise from five sources:

  1. The ontological definition of the current word-sense restricts the semantics of its slot fillers. The definitions for ACQUIRE and LEARN are shown in Figure 4. ACQUIRE and LEARN both require a HUMAN AGENT. ACQUIRE requires a non-HUMAN OBJECT for its THEME, while LEARN requires an INFORMATION THEME.
  2. The ontological definition of the word-sense that will fill the slot restricts the kind of slots it may be the filler of. Type 1 constraints ask ``What kind of fillers do I allow?'' Type 2 constraints ask about the fillers, ``What kind of concepts can this filler modify with the given slot?'' For instance, HAMMER, when used as the filler for an INSTRUMENT slot usually modifies some sort of BUILD event. In the example, ORGANIZATION (from Grupo-Roche-1) as an AGENT filler currently gif does not select for any specific type of event, nor do ORGANIZATION (Dr-Andrew-1) or HUMAN (Dr-Andrew-2) as THEMEs select for a specific event.

  3. The ontological definition of the slot (the property name that is being added) restricts what its DOMAIN and RANGE can be. Sometimes, in the absence of more specific constraints from 1 and 2 above, uK can find default values by looking up the slot itself in the ontology. An AGENT slot requires its DOMAIN (adquirir, in this case) to be an EVENT and its RANGE (Grupo-Roche) to be HUMAN. gif A THEME slot REQUIRES an EVENT for the DOMAIN (adquirir) and any OBJECT or EVENT for its RANGE (Dr-Andrew). These constraints are always very general, but still can help eliminate wrong attachments and word meanings.
  4. The lexicon entry explicitly includes constraints that override or add to the above ontological constraints. Section 4.2 gave an example of when this is necessary. In our example, however, the two word-senses for ``adquirir'' have no explicit constraints in their lexicon entries.
  5. Other structures in the sentence that are not explicitly specified by the lexicon entry can nonetheless modify the word in question. For instance, adjectives and PPs typically add slots to the TMR corresponding to the word they modify, even though they rarely are included in its lexicon entry explicitly. In this case, ``adquirir'' is modified by ``a-traves-de,'' which, depending on the meaning used, will either add a LOCATION slot or an INSTRUMENT slot to the TMR resulting from adquirir's analysis. In both cases, the slot will be filled by the TMR that results from ``compania,'' gif which maps into either a CORPORATION or a SOCIAL-EVENT (as in ``companionship''). The only interesting constraints that arise out of these combinations is that for the LOCATION meaning of ``a-traves-de,'' the DOMAIN (adquirir) must be a PHYSICAL-OBJECT (which it is not), whereas the INSTRUMENT meaning requires an EVENT. Although the LOCATION meaning of ``a-traves-de'' can be eliminated using these constraints, it does not help to further disambiguate ``adquirir.''


next up previous
Next: Applying Constraints Up: The Semantic Analyzer Previous: The Semantic Analyzer



Kavi Mahesh
Sun Nov 12 15:24:36 MST 1995