Ambiguity Resolution (3)

Selectional Restrictions

• Proper assignment of meaning to lexemes can depend on argument relations

  1. Pick one of several meanings:

     Playing Quake ate up 5 hours. 
     
     Bob ate up all my Doritos.

  2. Rule out nonsense meanings [Knight & Rich, p. 386]:

     The satellite orbited Mars.
     
     *Mars orbited the satellite.

  SOLUTION Encode argument relations, restrictions on fillers

  (CHRONO_CONSUME ACTIVITY-OBJECT TIME-QUANTITY) 
                  theme           duration
  
  (PHYSICALLY_INGEST ANIMATE-OBJECT FOOD-OBJECT)
                     agent          theme
  
  (PHYSICALLY_ORBIT UNMANNED-SPACECRAFT PLANET-OBJECT)
                    orbiter             central-body
                    (mass_greater_than central-body orbiter)
  

• O n! complexity for some syntactic constructions

  heads:     1       2             3             4     
       John saw the boys with the boys with the boys with the boys
   mods:                  2             3             4
  
   n = 3 (number of prepositional phrases after the direct object)
   
   (n + 1)! = the number of possible structures (4 x 3 x 2)
  

  IMPLICATION Practical applications may not be able to rely on "parse now, prune later" approaches to syntax + semantics.

• Proper choice of syntactic structure can depend on meaning

  The man saw the boy with the dog.
  
  The man saw the boy with the telescope.

  IMPLICATION Attachment preference can be used to select one structure over another. Even though both attachments may work, one may be much more likely in the given domain.

Conclusion: Deploying semantic constraints during syntactic parsing can eliminate undesirable parses, prefer more likely parses, reduce complexity, and make ambiguity more manageable overall.

Example from KANT/CATALYST

• The KANT system uses semantic domain knowledge to prefer some PP attachments over others. The syntactic parser produces all the attachments, but for each attachment a score assigned; after syntactic parsing, the f-structure with the lowest aggregate score is selected (see example trace). The system can derive a single structure for the example sentence, even though it contains 3 PPs:

  "The oil flows from the relief valve through the tube to the orifice."

• The CATALYST domain model contains triples, which encode binary relations between head concepts (verbal, nominal) and modifiers (nominal prepositional object):

  (*A-FLOW (SOURCE_FROM *O-RELIEF-VALVE))
  (*A-PATH_THROUGH-VERB (PATH_THROUGH &ALL))
  (*A-FLOW (GOAL_TO *O-ORIFICE))
  

• The head and modifier each receive a subscore, depending on how closely they match triple data in the DM; the attachment receives a score which is the sum of the two subscores, which are calculated as follows:
  1. If the head/modifier matches a triple exactly, then the score is 0
  2. If the head/modifier matches a generalized concept, then the score is 1
  3. If the head/modifier matches a default concept (e.g., "&ALL"), then the score is 2

• Check out the f-structure for the example sentence. There are three PP attachments, and the score for each is stored in the TSCORE ("triple score") slot:

  "from" (TSCORE 0) 
  (SHEAD *A-FLOW) (SEMSLOT SOURCE_FROM) (SMOD *O-RELIEF-VALVE)
  
  "through" (TSCORE 3)
  (SHEAD *A-FLOW) (SEMSLOT PATH_THROUGH) (SMOD *O-TUBE)
  
  "to" (TSCORE 0)
  (SHEAD *A-FLOW) (SEMSLOT GOAL_TO) (SMOD *O-ORIFICE)  
  

  Because the first and third PPs match the triples shown above, exactly, they receive a score of "0". Since there is no exact triple for the PATH_THROUGH semantic role, it matches on the generalized head *A-PATH_THROUGH-VERB (partial score = 1) and the default filler &ALL (partial score = 2), for a TSCORE of 3.

• The example f-structure is chosen because it has the lowest aggregate TSCORE, which means that it most closely matches the specific knowledge in the domain model. The original f-structures (before the domain model is used for filtering) show that the other attachment possibilities receive lower aggregate scores.

Heuristic Methods

It's possible to create heuristic filters that prefer one type of syntactic structure over another. For example, in the Caterpillar domain there are many multi-word noun phrases in the lexicon which describe various technical objects, actions and conditions in the heavy equipment domain. It makes sense to prefer f-structures which use these multi-words rather than f-structures which have a compositional treatment of the same words. Nevertheless, overlapping constructions in the domain can create problems:

Oil [flows] from the [bearing seal].
     V-intrans        NP

[Oil flows] from the bearing [seal].
 NP                           V-intrans

Interactive Approaches

Interactive approaches to disamiguation typically involve asking a human operator to select among ambiguous readings, which are presented in human-readable form. The operator's input can be used on-the-spot to filter the set of interpretations for translation; or it can be saved as a form of annotation along with the original source sentence.

• During translation, after source analysis [Brown, 1991]
  Create all possible interlinguas, gloss them for the user, ask the user to select; filter set of interlinguas passed to generation.
• Before translation, during source analysis [Mitamura & Nyberg, 1995]
  Identify the causes of ambiguity, gloss them for the user, ask the user to select; annotate original source sentence.
  1. Example: Lexical Disambiguation (meaning)

  2. Example: Lexical Disambiguation (part-of-speech)

  3. Example: Structural Disambiguation (PP attachment)

• Lexical vs. structural disambiguation [Mitamura & Nyberg, 1995]
  Since lexical ambiguity involves potential differences in the part-of-speech of certain words, it can have a dramatic effect on the syntactic structure. In some systems, therefore, it is best to resolve lexical ambiguity before asking about structural ambiguity.

• Remembering operator choices for later processing [Mitamura & Nyberg, 1995]
  Since some systems decouple the document authoring process from the document translation process (e.g., Caterpillar's AMT system), it is important to store the operator's disambiguation choices during authoring so they are available later when the text is translated. In CATALYST, the author's choices are stored as "invisible" SGML processing instructions (PIs) in the original source text:

  'Release 
   the pressure until 
   the oil flows from the orifice.'  

• Confirming the author's usage
  In some cases, a term may not be ambiguous but authors have been known to use the term to imply a meaning which is not represented by the system. In such cases, it's possible to prompt the author with a confirmation dialog.

Lingering Problems

• True domain ambiguity

  Oil flows from the orifice
  Current flows from the switch
  Service the truck in the garage
  

• Unintended ambiguity (overgeneration)

  Notify the dealer if the engine is damaged. 
  Observe the mechanic while the engine is serviced. 
  Open the valve until the tank is filled.
  Do not start the engine while the tank is filled.