Brian MacWhinney macw at CMU.EDU
Fri Jul 30 22:37:46 UTC 1999

Dear FunkNet,
  It was a busy day, or I would have chimed in earlier regarding this useful
discussion of emergence.  I found Ă–sten Dahl's note useful.  In particular, I
very much agreed with the idea that one tends to evaluate emergence
differently according to whether one emphasizes the novelty of the emergent
system or the degree to which its structure can be partially predicted from
its predecessors.
  Wolfgang Schulze's claim that "subjecthood is the construction of emergent
activities exerted by information flow (word order), NP semantics, case
marking (if present), agreement (if present) and much more" is certainly a
correct application of the notion of emergence to language, as Liz Bates and
I argued in a series of papers in the 1980s.  In those papers, we
distinguished three levels of emergence: diachronic, acquisitional, and
synchronic (sentence processing).  More recently, I have found it useful to
think in terms of five levels of emergence:
1.  Evolutionary
2.  Embryological
3.  Developmental
4.  On-line
5.  Diachronic
It seems to me that it is crucial to view emergence in terms of these
varieties and to keep in mind the fact that the time scales for these
varieties of emergence differ vastly.
   Simply listing these varieties of emergence does not answer Tom Givon's
question.  He wants to know which type of emergence is operative in a given
behavior. To do that, we have to dig into the guts of specific models.  For
example, in their chapter in "The Emergence of Language" (Lawrence Erlbaum,
1999, B. MacWhinney (Ed.)), Prahlad Gupta and Gary Dell show that "a good
lexicon" has words that share neighbors in the rime, but which differ in the
onset. (CAT, BAT, SAT, RAT, MAT, FAT ..)  They then proceed to show how this
aspect of lexical structure arises directly from the Dell Speech Error model
-- a data-driven connectionist model of speech production.
  What Tom wants to know, if I read him correctly, is whether the automated
phonotactic processes that Gupta and Dell have studied are supported by
simple general purpose learning mechanism or instead by some more dedicated
neural hardware.  The likely answer to this is that general purpose learning
mechanisms are themselves based on specific neural adaptations.  For example,
Kohonen self-organizing feature maps demonstrate behaviors that emerge from
the neural preference for lateral inhibition and short connections.  These
models can help explain the neurological grounding of the emergence of the
preference for word onsets during development.
  Once models are formulated in this way, we start to realize that behaviors
arise from an interaction of sources of emergence on varying time scales.
Although automaticity is typically a fact of developmental emergence, it
builds on embryologically-emergent structures and can be modified by on-line
   The moral:  the importance of emergence becomes clear once we begin to
think in terms of models of specific behaviors.  If we don't think in terms
of concrete models, it seems like semi-TT gobbledy-gook.

--Brian MacWhinney

More information about the Funknet mailing list