Re: [UAI] Encode semantics of words into a big baysian network

From: Francisco J. Diez (fjdiez@dia.uned.es)
Date: Mon Apr 02 2001 - 10:22:21 PDT

  • Next message: Richard Dybowski: "[UAI] Machine Learning Journal Special Issue on Fusion of Knowledge CFP"

    ruan tong wrote:
    >
    > When I work on information filtering, I feel if I could encode
    > semantics of words, such as synonym, antonomy, something like WordNet
    > and above it(For example, term clustering, context sensitive) into a
    > big bayesian network, information filtering and other similar tasks,
    > will be made easier and effiecient. I found there are someone who are
    > working on symonym relationships. I do not know if I could do more, as
    > I metioned above, is it will be significant? Although I feel it is, I
    > can not decide. It will cost long time work and many difficulties, may
    > be very dull. Is there some suggestions? Thank you very much if
    > somebody could talk with me about that.
    >
    > Tong.

    I do not think Bayesians network are the appropriate method for
    enconding that kind of knowledge. The first reason I find is the
    inherent asymmetry of Bayesian networks--please note that d-separation
    means "directional separation". This shortcoming might be avoided by
    using Markov networks instead of Bayesian networks.

    However, semantics has little to do with probability, and even if you
    were able to establish a relationship between semantics and probability,
    which led to a meaningful definition of conditional probability among
    words, it would be difficult to find such probabilistic dependencies and
    independencies in practice.

    Certainly, there are some people using Bayesian networks for information
    retrieval, but their models do not adhere to the axioms of Bayesian
    networks (at least when those networks are built by experts rather than
    learnt from databases), and the only justification for using them is
    that the experimental results are good in some cases, in the same way
    that MYCIN's advice was quite good in spite of the inconsistencies of
    its mathematical model.

    Sincerely,
      Javier Díez

    - --------------------------------------------------------------------
    F. J. Diez Phone: +34-91-3987161
    Dpto. Inteligencia Artificial Fax: +34-91-3986697
    UNED. Senda del Rey, 9 E-mail: fjdiez@dia.uned.es
    28040 Madrid. Spain WWW: http://www.ia.uned.es/~fjdiez



    This archive was generated by hypermail 2b29 : Mon Apr 02 2001 - 10:24:01 PDT