🚀 go-pugleaf

John R. Beer writes:
>Jacob Galley writes:
>>Now I think my original question was not specific enough. What bothers
>>me about the theories of meaning that I have seen is that they do not
>>adequately account for how meaningful symbols are linked to WHAT >THEY
>>MEAN. Ie, while we have made a good amount of progress in syntax and
>>taxonomic classifications, which describe how meaningful signs
>>interact with and relate to each other, I have not encountered any
>>theories of how the signifier is linked to the signified.
>
>You should look at:
>Donald Davidson's writing on truth theory and radical interpretation,
>collected in Inquiries into Truth and Interpretation(1983)
>
>Michael Dummett--start with Truth  and Other Enigmas, which will
>prepare you for his classic paper "What is a Theory of Meaning?", in
>Truth and Meaning ,  (McDowell and Evans,eds.)
>
>The Varieties of Reference by Gareth Evans, which sets out an account
>of what abilities one needs in order to make reference to objects in
>the world.
>
>This stuff is hard going, but immensely rewarding.  Although all three
>authors are firmly in the analytic tradition (and you might want to look
>at Frege and Quine before plunging in), they aren't proposing new
>formalisms, but possible ways for grounding these formalisms.

Connectionistic models with unsupervised learning capabilities
seem to be formalisms or even grounding blocks for "theories of
how the signifier is linked to the signified." Professor Teuvo
Kohonen has written: "In attempting to devise Neural Network
models for linguistic representations, the first difficulty
is encountered when trying to find metric distance relations
between symbolic items. [...] it cannot be assumed that encodings
of symbols in general have any relationship with the observable
characteristics of the corresponding items. How could it then
be possible to represent the 'logical similarity' of pairs
of items [...]? The answer lies in the fact that *the symbol,
during the learning process, is presented in context*."
(The Self-Organizing Map, Proceedings of the IEEE, vol. 78, no 9,
September 1990)

The context may consist of words, symbolic features, continuous
features or even pictorial images. The virtue of unsupervised
learning (e.g. the self-organizing map) lies in the fact
that there is no need for any "correct answers" which is the
case with most of the connectionist models.

Timo Honkela (tho@tik.vtt.fi)
(Please, contact if you are interested in an article
"Connectionism and Linguistics" which e.g. discusses
aspects of semantics and pragmatics related to
unsupervised learning.)