The Whys and Hows of Ontological Semantics
OS work has
been reported to this Colloquium since it started back at IU in 2004. It has
dealt primarily with the currently funded
projects, most of them in applying NLP
to IAS. This work, along with other
developments and applications, is
continuing and being submitted for
presentation at MCLS-
The NLP Lab at Purdue University in cooperation with CS/Colgate
(1982-87), CMT/CMU (1987-1994), and
CRL/NMSU (1994-2000) has established and tested in a number of applications a knowledge- and meaning-based approach to NLP called ontological semantics (OS). Since 1999, NLP Lab
has cooperated with CERIAS in applying the
approach to information assurance and security
(IAS) tasks. The term 'ontological semantics'
is also used by Sergei Nirenburg's ILIT/UMBC for a somewhat different approach,
sharing some theoretical premises reported in Nirenburg and Raskin (2004) but differing in more attention to surface syntax, somewhat shallower
semantics of closed-class items, and aversion
to incremental implementations. OS is a
mature phase of computational semantics approach initiated by Raskin at the Computational Linguistics Lab at
The difference between the two schools is critical. Infected by
"fear of semantics," SSA tries to use the standard surface parsing and elaborate statistical
methods to devine the meaning of text from the observable syntactic and
statistical behavior of its units without any attempt to represent
meaning directly, and it prides itself on not developing knowledge resources, using
none or grabbing opportunistically whatever is easily available on the Web, such
as WordNet, even if the resource is not particularly suitable or designed for
NLP.
OS, committed to the cause of weak AI, has created large knowledge
resources modeling the ones used by humans processing language and information:
the constructed ontology representing the human knowledge of the world; the
large lexicons representing the lexical entries in terms of pertinent ontological concepts; the
ontological parser which represents the
meaning of text, clause by clause and
sentence by sentence, to closely
approximate the way humans understand
them. Supporting the resources is the elaborate acquisition platform ensuring smooth homogeneous acquisition of lexical entries and supporting concepts by various acquirers with varying degrees of training, seamless integration of fully automated
methods with a constrained and tightly limited human intelligence, and the incremental progress towards full
automation in acquisition.
The most publicized use of the term 'semantic' is probably The Semantic
Web (SW), whose visionary founder was recently knighted by Queen Elizabeth II,
a major NLP expert. But even this initiative remains largely non-semantic.
An enormous effort by many SW visionaries and talented scholars has
been spent on developing and perfecting the formalism for letting the semantic
content of various Web sites to interact with each other. So why don't they already? For the same reason they never will— unless the SW
developers realize that they have done
nothing to ensure that there is a methodology
for translating the Web content into
their formalisms.
For various sociological and academic reasons, SW—just like
early NLP—has been developed by computer scientists, logicians, and
mathematicians. As the first half-century of NLP development shows, these
groups tend to possess
an unlimited amount of naivete about natural
language, regularly confusing their competence
in their native tongue with the linguistic
savvy about it. As a result, they lack descriptive techniques, are
commonly unable to determine the meaning of
the sentence, and dismiss meaning anyway because they feel comfortable only with the "objective"
methods of counting observable words
and word combinations and—not so
well—analyzing the surface syntax. Their inability to assess meaning results in
the unavailability of accurate working systems for processing natural language,
and the people who need such systems
dismiss them for poor quality: the US
Government, the major funder of NLP
research, has a pathetic record of deployment of the delivered systems. The few
linguists who were allowed to join these efforts were selected for their formalistic inclinations and tried hard to
support the others' efforts in avoiding
semantics; many of them had no semantic
training.
Even since the late 1990s, when the US Government started funding
primarily semantic projects in NLP, often using the very language of
ontological semantics in the calls for proposals, the anti-semantic forces have
revived their doomed attempts to approximate meaning without investing in the
ontologies and lexicons but rather by recycling syntactic and statistical
methodologies in different packaging. The SW developers are akin to these forces,
and if they ever worry about the migration of Web content at all—and there is
preciously little evidence of any awareness of this problem in print—they
probably assume that it will be done somehow by somebody like the
statisticians/syntacticians in NLP.
SW has generated an enormous white elephant called Web
Ontology Language, for which the acronym is, strangely, OWL. The large number of a
priori and explicitly undefined rules have been developed for expressing
content in the recommended formalism. Rarely, are these illustrated with a few convenient
examples. OWL comes without a methodology for training in acquisition or for acquisition itself. It is tacitly assumed that the Web content owners
will somehow learn OWL by themselves and will voluntarily spend a considerable effort in translating their content into OWL and they will do it uniformly, bravely solving the problems of homonymy and ambiguity in unexplored ways. The closest precedent to this wishful thinking was Chairman Mao's idea of increasing steel production in PRC by obligating regular citizens to manufacture
steel in their backyard vats. The Chinese
did. The Web content owners have and will not—but if they attempt it, the
result will be the same as with the
Chinese—unacceptably low quality of the product and its consequent non-usability.
Contrary to this approach, OS offers a large developed ontology, with multiple properties interrelating concepts to each other in accordance
with human intuition and with easy and
tried extensibility to new domains, a battery of lexicons, one for each language, with each sense of a lexical entry clearly and systematically defined in ontological terms, and the
ontological parser, producing text meaning representations
and increasingly approximating human
understanding. Incredibly importantly, OS
also comes with an explicit acquisition toolbox, combining human and
computerized limited training with an
increasingly automated hybrid human-computer system for uniform acquisition
of ontological concepts and lexical entries.
The lack of such system has rendered CYC,
a lovely idea, unusable. Failure to incorporate OS will also leave SW on the drawing tables.
There are several reasons for the rejection of computational/ontological
semantics.
First, the sociological reason, mentioned above in connection with
the Semantic Web: the prevalence of non-linguists in the NLP effort. This is
compounded by the educational reason: neither the non-linguists nor the linguists they hire are
liberated of the fear of semantics by their education: most linguists think that semantics is "awfully
hard," intuitive and unformalizable—and
that after a decade of the moribund formal semantics in the 1990s, which
formalized everything that could be formalized and ignore most of natural language semantics.
Occasionally, one hears grumbles about the subjectivity of the engineered ontology and ontological lexicons. There are no grumbles about the subjectivity and idiosyncracy of natural language, which actually only exists as
idiolects. All of these fears were discussed and dismissed as early as 1995
(see Nirenburg et al. 1995).
The full paper will deal with these in more
updated detail.
There is yet another group of people who are moving into NLP and filling the positions that would more successfully be manned by linguists:
psycholinguists, like psychologists a group
whose approach is that of statisticians with underspecified theories
that often operates under the
moniker 'cognitive science.'
As statisticians, they
bolster the statistical emphasis of
SSA and, simultaneously, as bad theoreticians, they either confirm the
fears of computer scientists about useless
theories muddling their clean formalistic approaches or enforce the anti-theoretical stance by letting them fall by the wayside as well. The use of a new jingoism: vector-space models with singular value decomposition, inappropriately termed 'latent
semantic analysis' (LSA), to replace semantics, is a prominent example
for this development.
7. References
Nirenburg, S., and V. Raskin 2004. Ontological Semantics.
Nirenburg, S., V.
Raskin, and B. Onyshkevich 1995. Apologiae ontologiae. In: J. Klavans,
B. Boguraev, L. Levin, and J. Pustejovsky (eds.), Symposium: Representation and
Acquisition of Lexical Knowledge: Polysemy,
Ambiguity, and Generativity. Working Notes. AAAI Spring Symposium
Series,