Lexical Semantics (Was: Translation Theory)

   Date: Mon, 2 May 94 20:07:00 PDT
   From: "KENNETH D. LITWAK" <klitwak@vnet.ibm.com>

					    The other point that I want
   to raise is the issue of words having one meaning.  Obviously,
   few words convey one and only one meaning.

This depends on your notion of what kind of data structure a
``meaning'' is.  For example, take a ``figure-ground'' nominal like
``door'' which can refer either to the opening (as in ``John walked
through the door'') or to the door frame and other material that make
up the door (``John painted the door'').  Are these two different
meanings (or, as I would prefer to say, two different senses) or two
aspects of a single sense?  A good friend and sometimes collaborator
has been developing a theory of lexical semantics over the past
several years in which the lexical semantic structure even for simple
``entity'' nominals such as ``door'' is already a complex object, so
that one sense may subsume multiple, predictable ``sub-senses''.  (Cf.
container nominals, which can either refer to the container itself, as
in ``I opened the bottle'', or to the contents, as in ``I poured out a
bottle of the lotion'', for another example of the same sort of
generative phenomenon.  The example could be multiplied, both of
members of the classes of nominals already mentioned, as well as of
new classes of nominals.)

					       On the other hand,
   I don't think it is accurate to say that claiming a word has
   a limited set of meanings that a translator can choose from
   is unreasonable or "positivistic".  Obviously, sarx may have a certain
   semantic range, i.e., flesh, body, and some metaphorical meanings
   as seen in Paul's letters, but that does not mean that sarx can
   be rendered "cheese pizza".  So I guess I'm a little confused
   over the issue of a word having a univocal meaning.  Granted
   you cannot look at sarx in isolation from its rhetorical, social,
   cultural, religious context and in isolation from the
   words around it in determining what the best equivalent in the
   target language is, but surely that does not mean that words don't
   have a limited set of possible meanings and thus it is not
   "positivistic" to assert otherwise.

You are confusing two different issues: (1) do lexical items have a
fixed number of meanings? and (2) does a given lexical item, in the
context of an utterance, which itself appears within the context of a
given text or discourse, set within a given domain, and so on,
typically have a single, or at least relatively small, set of
interpretations?  One reasonable approach to natural language
semantics is to view interpreting a natural language utterance as
being akin to solving a set of simultaneous equations, in that we must
come up with the most likely interpretation(s) of each lexical item
that makes up the utterance, consistant with the interpretations of
all the other lexical items in that utterance, so that we tend to
assign mutually reinforcing interpretations and to discard mutually
contradictory interpretations.  Given this view, the answer to
question (2) can be ``Yes'' while the answer to (1) is ``No'', in the
sense that the most descriptively and explanatorily adequate theory of
lexical semantics may be one which does not assign to each lexical
item an enumerated set of meanings, but rather one which assigns each
lexical item a generative meaning representaton that can produce an
appropriate interpretation for that nominal when it is composed with
other lexical items.  For example, personal proper names can refer to
the person himself/herself, as in ``Aretha Franklin is the daughter of
a preacher'', or to that person's work, as in ``Put on some Aretha
Franklin''.  But it can have even more extended senses, as in:
``Aretha Franklin is booked into the Orpheum from May 20 through May
22'', in which ``Aretha Franklin'' means a performance by Aretha
Franklin.  Presumably, we don't want to say that every proper name
includes an explicit performance sense, but, obviously, we must allow
such an interpretation to be available when the context demands it.

Given the appropriate generative mechanisms, we do not need to imagine
that all possible interpretations of a lexical item are enumerated in
the (mental) lexicon.  Rather we can allow them to be generated on the
fly.  This results in a theory of lexical semantics which allows, at
least in potentia, for many more interpretations of a lexical item
than a theory of lexical semantics that views senses as being
enumerated in the lexicon.  However, this theory also allows far fewer
interpretations for a given lexical item in a particular context than
a radically indeterminate approach like deconstruction, given that the
interpretation chosen must be globally consistent with the rest of the
utterance, discourse, and overall pragmatic context.  This approach
also has the nice consequence that, while it does not view
interpretation as radically underdetermined, it does view the
intepretation procedure as non-deterministic; i.e. it allows for
misinterpretation.  In the language of stochastic classifiers, such as
Boltzmann (sp?) machines, simulated annealing, etc., under such a
theory, it is possible to get caught in a local optimum state, which
is sub-optimal with respect to the utterance as a whole.  (Compare
this with the familiar syntactic notion of the ``garden path
sentence'' as in ``The horse raced past the barn fell'', in which we
first interpret ``The horse raced past the barn'' as a complete
utterance, with ``raced'' as an intransitive past tense verb (i.e. a
local optimum state).  On encountering ``fell'', we must backtrack and
re-interpret ``raced'' as a passive participle.)

Bob Ingria