Chet,
Aha, that's not how I would do it, but you may be right. I visualise a
bottom-up system in which you first access the specific concept - the
individual lexeme - and then look around for a value for the function
that's in question (e.g. past-tense-of). In a WG network an irregular is
bound to be *closer* than an inherited regular. Imagine the following network:
verb --past--> stem + /ed/
| \
| \
BARK SEE --past--> /saw/
The past of SEE is directly linked to it, while that of BARK is two links
away. So if it's all done by spreading activation, /saw/ should be found
faster than /barked/, because /saw/ will be activated as soon as you
activate SEE, but not so for /barked/. That's why I suspect an experimental
flaw.
Dick
At 11:19 18/11/1999 -0500, you wrote:
>Joe:
>> Steven Pinker, in his latest book (_Words and Rules: The Ingredients
>> of Language_, Weidenfeld & Nicholson, London, 1999) says the following:
>>
>> When you produce an regular form, you not only have to dredge
>> it out of memory but also must rpress the "Add *-ed*" rule so
>> you don't say *breaked* or *broked*. Linguists call this
>> principle *blocking*--the irregular for blocks the rule--and
>> the experiments help us understand how the mind implements it.
>> One possibility is that when we need to utter a past-tense
>> form we first scan our list of irregular verbs to see if it
>> is there, and if it isn't, we turn on the rule. That predicts
>> that the slowest irregular verb (the one at the end of the
>> list) should be faster than the fastest regular veb. The
>> prediction is wrong. Irregular forms usually are slower to
>> produce than regular forms; they are never faster. (p. 130)
>>
>>
>> Pinker bases this claim on unpublished work (Prasada, Pinker & Snyder
>> 1990) and says that these results have been replicated in his lab.
>> Assuming that Pinker's right, how could this situation be represented
>> in WG?
>
>[Preamble: I remember a discussion last year where Dick seemed to have
>a conception which was bottom-up rather than top-down, so perhaps my
>depiction below is _exactly_ wrong. I wasn't convinced last time
>around, but must admit that although it is clear that in some sense
>declarative knowledge is subject to temporal considerations
>somewhere/sometime/somehow, the ways are less obvious than with rules.]
>
>If Pinker isn't careful, he is going to end up a WGian despite his
>claims to the contrary. In WG, the experimental results follow as a
>straightforward example of default inheritance. Presumably the mind
>attempts to apply the higher level propositions first and checks for
>overrides. If one is found then a new proposition is invoked. If I
>remember correctly, this is the way default inheritance is defined in
>WG. Note that the commonly cited experiments with canaries singing and
>canaries flying aren't relevant here. The kind of experiment that
>would be relevant would be one that tests reaction time for "sparrows
>fly" (no override) against "ostriches don't fly" (one override).
>
>I'm not sure if it would be possible to test the WG declarative approach
>against Pinker's rule approach, but one thing to try would be to see
>how both do when confronted with "wugs". The rule approach predicts
>no difference between "wugs" and lexical items already in place whereas
>the WG approach might be taken to require an association of a default
>with the "wug" and thus to take longer with these. Of course care
>would have to be taken to choose lexical items which were relatively
>unfamiliar but real to match against the wugs. In general, familiarity
>effects might swamp any results, and I don't envy someone trying to
>do experimental work of this sort.
>
>I'm sure this posting is dreadfully wrong somewhere, but if I work
>on it longer, it won't be sent.
>
>Chet
>
>
Richard (= Dick) Hudson
Phonetics and Linguistics, University College London,
Gower Street, London WC1E 6BT.
+44(0)171 419 3152; fax +44(0)171 383 4108;
http://www.phon.ucl.ac.uk/home/dick/home.htm
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|