Documents
Resources
Learning Center
Upload
Plans & pricing Sign in
Sign Out

Formalization of World Knowledge By Precisiated Natural Language

VIEWS: 8 PAGES: 19

									               World Knowledge Enhancement
                        Using Tools of
                Precisiated Natural Language


                       Yu Sun (PAMIR)
                         Supervisors
                        Prof. F. Karray
                        Prof. O. Basir




October 2003                              Yu Sun (PAMIR)
    Previous Works on Term Related
    Unified Semantic Tree (TRUST)

      Previous work on TRUST is based on
       pure statistical approach
      Disadvantages:
              Huge dimensions lead to complexity
              Estimated parameters are based on
               insufficient data and therefore imprecise



October 2003                                   Yu Sun (PAMIR)
    Previous Works (cont)

      Questions left:
              How to impart human Knowledge into the
               statistical approach
              How to use such human Knowledge to
               tune the TRUST




October 2003                                Yu Sun (PAMIR)
    How to Tackle the Issues

  The novel approach is implemented
    through:
   Epistemic Lexicon (EL)



      Precisiated Natural Language (PNL)



October 2003                       Yu Sun (PAMIR)
    Epistemic Lexicon (EL)
                         rij
                                lexinej

               lexinei

    A lexine uses granule fuzzy technique to define
     a term and its attributes, its relations to other
     lexine terms.



October 2003                              Yu Sun (PAMIR)
    Examples of EL

      EL is a static representation of knowledge
        Buyer:
                   institute (company, bank)   / very possible,
                   management-staff            / possible,
                   government                  / not very possible

              Seller:
                   institute (company, bank)    / very possible,
                   management-staff            / possible,
                   government                  / not very possible,



October 2003                                      Yu Sun (PAMIR)
    Precisiated Natural Language
    (PNL)

   EL is a static dictionary. There need generalize
    constraints to compensate the insufficiency.
   PNL is a sub-language of precisiable
    propositions in NL which is equipped with a
    dictionary (EL) defined by domain experts and
    Generalized Constraint (rules of deduction)
   GC dynamically applies EL on NLU



October 2003                            Yu Sun (PAMIR)
      Examples of GC

      Generalized Constraint in PNL is used to enhance
       the static EL through rules
              Any lexine in EL (word) is bi-sense-directed: belongs to one
               of two directions: in/out or up/down dependent on its context;
              One lexine might belong to multi concepts (classes) [Self];
              If Positive(+, such as agree, pro) meets Negative(-, such as
               against, con), the result is Negative(-) [combination];
              Lexine’s relationship can be inherited by its attributes. For
               instance, Institute has the following attributes: management-
               staff, performance, legal, scandal (which are defined by the
               original EL). If action Against(Institute), then
               Against(Institute.management-people, Institute.performance,
               Institute.legal, Institute.scandal). And vice versa [Inheritance].

October 2003                                                  Yu Sun (PAMIR)
    Example of GC Enhancing NLU

This example shows how GC enhances the
  system to correctly understand NL:
     “Beech-Nut Corp. damaged its image over the sale of apple juice
     that turned out to be water”
    EL: Performance-up(sale) [default]; Against(damage);
     Company(image, sale)
    GC: Against(damage)Against(company.image)
     Against(company)Against(company.sale)
     Against(Performance-up(sale)) Performance-
     down(sale) [inferred]

October 2003                                         Yu Sun (PAMIR)
    Experiment

      Pre-processing (manual annotation)
              5 documents from WSJ with unique-topic:
               “company-takeover”
              Epistemic Lexicon: including 35
               concepts(classes), such as action-in,
               performance-up, buyer, seller, etc.
              Generalized Constraints: rules governing
               combinational operations of concepts.

October 2003                                 Yu Sun (PAMIR)
    Results of Applying PNL
       Experimental results showed the tuned TRUST much
        closer to human common sense
       Enlarged and Adjusted EL after tuning TRUST
              Buyer:
                   action-in          / very possible,
                   agree              /very possible,
                   management-staff   /not very possible
                   legal              / not very possible
              Seller:
                   action-out          / very possible,
                   legal              / possible
                   management-staff   / possible
                   sale               / hard to tell
October 2003                                                 Yu Sun (PAMIR)
     Possible Applications

   To identify the semantic meaning of a
    given term (word sense disambiguation):
    “The juice scandal forced Beech-Nut to pay a $ 2.2
    million fine and $ 7.5 million to settle a lawsuit”
        Problem: Beech-Nut may be buyer or
        seller
        Analysis:
               Against(scandal)  closer to Seller;
               Against(fine)     closer to Seller;
               Legal(lawsuit)    closer to Seller;
        Solution:
                  Beech-Nut is a Seller
October 2003                                           Yu Sun (PAMIR)
    Future Possible Applications

      Work as an assistant to a semantic parser:
       For instance:           anaphor resolution

       First Pennsylvania had agreed to be acquired by Marine Midland in
       several month ago…. Midland decides to step out the acquisition and it
       starts a lawsuit. (lawsuit is closer to seller than to buyer)

      Work as an assistant to a syntactic parser:
       For instance:                  pp attachment
       Ralston Co. agreed to buy Beech-Nut Nutrition Corp. with a favorite
       offer.



October 2003                                               Yu Sun (PAMIR)
    Conclusions

      Present a novel approach imparting human
       knowledge into statistical approach;
      The preliminary experiment shows PNL and
       EL improve the statistical approach;
      The experiment uses unique-topic documents
       and in future, research work will expand to
       more complex documents.


October 2003                           Yu Sun (PAMIR)
    Publications

      Submitted to Journal of Fuzzy Sets and
       Systems




October 2003                        Yu Sun (PAMIR)
    Overview of Architecture




October 2003                   Yu Sun (PAMIR)
    Comparing Procedure of Untuned
    TRUST and Tuned TRUST




October 2003              Yu Sun (PAMIR)
October 2003   Yu Sun (PAMIR)
October 2003   Yu Sun (PAMIR)

								
To top