Why Ontological Semantics Is neither Ontological nor Semantics
This criticism focuses on Sergei
Nirenburg and Victor Raskin’s
Ontological Semantics. An approach in
automated speech processing, summarily presented in their corresponding book Ontological Semantics (2004).
In the philosophy of language and linguistics
the idea of ‘bringing ontology back to semantics’ is linked to Situation Semantics, the approach
introduced by Jon Barwise, John Perry, Keith Devlin
and their co-authors in the 1980s. ‘ontology’ in this approach meant on the one
hand a theory of the kinds of entities that reality is composed of, and on the
other hand modelling semantics and information flow using these entities, the
models also containing a meta-theory how they relate and refer to parts of
reality. ‘ontology’ here is used in the traditional
(philosophical) sense.
In the artificial intelligence community,
however, ‘ontology’ is unfortunately often used in quite another sense. (This
leads – as one may witness occasionally at conferences – to talking past each
other.) ‘Ontology’ here, and so in Nirenburg and Raskin’s Ontological
Semantics, is introduced as a canonical description
of the world. This alludes to the first sense of ‘ontology’ mentioned above,
but involves no claims about linking models framed in the model language to
reality or even withholds commitment as to the real existence of the kinds of
entities used in the model language. The ‘world’ that is modelled is a more or
the less comprehensive representation (true or not), and is not reality.
Ontology in the AI sense, therefore, can freely make use of fictional entities
and kinds of entities supposedly not belonging to the ultimate furniture of the
universe. Ontology in Ontological
Semantics, in fact, turns out to be a conceptual structure, which is
outlined in a meta-language. Ontological modelling means to use a conceptual
scheme of an explicitly defined format: ‘it belongs in epistemology’ (p.135).
Ontological
Semantics is employed to derive a canonical
interpretation given some input text. As it should be used by automata this
means that such an artificial system comprises sub-systems responsible for
(syntactic) parsing and analysis of text structure and other sub-systems
deriving an interpretation by making use of a lexicon and a stored knowledge
repository. The input text is ultimately translated into a canonical output
text, which expresses the proper interpretation of the input text.
In ontological semantics, …,
sentential meaning is defined as an expression, text-meaning representation
(TMR), obtained through the application of the set of rules for syntactic
analysis of the source text, for linking syntactic dependencies into
ontological dependencies, and for establishing the meaning of source-text
lexical units.
(p.125)
The aim of Ontological
Semantics is to enable this treatment of text not just for toy text-book
examples, but for ordinary – say newspaper – texts. Nirenburg
and Raskin see the scope of envisioned application as
one of the outstanding features of Ontological
Semantics.
Nirenburg and Raskin introduce the
reader both to the methodology and the development of Ontological Semantics, a research programme of about twenty years.
The first part of the book provides an
overview and comment on current linguistics and its methodology and places Ontological Semantics within the field. Nirenburg and Raskin see
themselves as contributing to the methodology of linguistics and the proper
methodology of semantics. They see the explicitness of its methodology as the
other outstanding feature of Ontological
Semantics.
The second part of the book provides basic
insights into the workings of Ontological
Semantics. Nirenburg and Rasking
outline their conception of ontology and their theories of the lexicon and the use of the fact repository. The structure of
the ontology resembles an inheritance hierarchy as one may find in object
oriented programming. Examples of processing are presented,
more are advertised as web-resources of the research group. Although somewhat
repetitive this overview allows some fascinating insights into Nirenburg and Raskin’s approach
to canonical text interpretation by artificial systems.
Whether Ontological Semantics is ‘semantics’ is quite another matter. Nirenburg and Raskin claim to deliver a lexicalist
account of meaning, so that the
system uses a lexicon, which provides for each lexical item an entry with that
entries definition and syntactic as well as semantic features.
A theory of meaning, however, demands far more
than a system deriving some interpretation of a text. Ontological Semantics ultimately is a version of inferential role semantics: a lexical
item is claimed to be defined by its interrelations to other lexical items. If
that was so, one should not use as labels of lexical entries words of ordinary
language (like ‘pay’), which may trick the casual reader into importing their
linguistic knowledge, but new labels (like ‘pon’). As
Nirenburg and Raskin note,
the technical system does not make a distinction between labels we know and new
labels, but once we – as readers – make use of new meaningless labels we
recognize that the examples given by Nirenburg and Raskin are far from a comprehensive model of inferential
roles that are distinct enough for lexical items of natural languages.
Their lexical entries, further on, have a
specific slot for ‘definition’, which again is said for the human user of the
system only. Now, on the one hand these definitions use other words, the
entries of which use other words – a circle to be explained; on the other hand
these definitions show again the insufficieny of
definitions in semantics – as sometimes complained in the linguistics
literature. For example, ‘say’ is taken as ‘inform’ (p.169), ‘pay’ is defined
as ‘to compensate somebody for goods or services rendered’ (p.199). This does
neither account for uses like ‘pay with one’s life’ nor does it distinguish
‘pay’ from ‘remunerate’, ‘reimburse’, ‘pay off’… A theory of interpretation,
which derives a canonical reading of a text in a context, does not need to do
these two things, a theory of meaning should. The definitional slots in the
lexical entries play no role in the processing, thus one may downplay this
problem. An even more serious problem, however, is the status of the
meta-representations. The elements of the ontology (vis.
the concepts like AGENT or ANIMATE) are connected by relations and structured
in some format. These relations and structures (‘metaontological
predicates’, p.224) are part of a meta-language in which the ‘ontology’ is
expressed. If the ‘ontology’ is the conceptual system, where do we have to
place this meta-language? For a technical system employing such a further layer
of analysis or description may be allowed, for a theory which aspires to
account for natural language processing in humans as well, which claims to be
part of our account of cognition, this will not do.
There cannot be, on pains of vicious regress,
a conceptual layer beneath/beyond the
conceptual system.
So, Ontological
Semantics should not be seen as either ontology or semantics in their
respective strict (traditional) sense, but as an approach of Canonical Interpretation making use of a
handcrafted conceptual scheme. That does not make it nor its aspirations, nor Nirenburg and Raskin’s book less
interesting.
Manuel Bremer, 2008.