Code to plot activation growth & Mental lexicon/Lexical decision questions

[Peter Pirolli]

I've used one of the large word indexes from the TREC text retrieval 
workshop to compute base strengths and inter-chunk association strengths 
for works.  In principle, this provides a lexicon of 200 million lexical 
items and 55 million association links.  Many of these items, however, have 
little practical use, since they are things like dates, numeric codes, 
etc.  For a specific simulation I construct a network from the subset of 
items that I think are needed for the model.  In practice, this has ranged 
from several hundred to tens of thousands of nodes.

I do not, however, use the ACT-R activation mechanism, so I can't comment 
on specific problems in that regard.

--Pete


At 03:35 AM 7/5/00 -0700, Hedderik van Rijn wrote:

>In the last couple of weeks, I've been investigating how to implement a
>large mental lexicon in ACT-R. (Consisting of all 4 letter words the CELEX
>(http://www.kun.nl/celex/) English orthographic word-forms database.)
>
>** Code to plot activation growth
>
>However, during tests with the model that uses this lexicon, I discovered to
>my shame that I my knowledge of how the activation of chunks is
>calculated/approximated fell short in certain important areas. Because just
>reading formulas doesn't give me "feeling" for fine nuances, I wrote some
>R/S code to plot figures like Figure 4.1 in the Atomic Components of Thought
>book. For those interested, it is available at:
>
>   http://swipc30.swi.psy.uva.nl/~rijn/actr-activations/
>
>(Evaluate fig4.1() and fig4.1.2() to get plots like Fig4.1)
>
>After playing around, a lot of the issues involving activation became a lot
>clearer to me. However, there is one issue still unsolved.
>
>** Approximated activation with d=.9 is too high?
>
>If the decay (d) is set to .5, like in Fig 4.1 of the book, the "real
>base-level activation equation" is closely approximated by the "optimized
>learning base-level activation equation". (As is argued at p.124 of the
>book.) However, if d is set to a very high value, for example .9, the
>approximation equation seems to yield structurally higher base-level
>activations than the real equation. This is illustrated in the two plots
>shown on the same web page as above. (I didn't attach these graphs not
>clutter the mailing list with binary information.) Can someone shed some
>light on this issue? Is the approximation function indeed "better" for
>values of d close to .5? Or is there a bug somewhere in my
>interpretation/code?
>
>** Mental Lexicon/Lexical Decision task questions
>
>The project that was causing this all, is an attempt to model lexical
>decision data. I searched around a bit, but did not find any previous
>ACT-R (or related) modeling for this task. However, if someone can point
>me to relevant information, I would be very pleased.
>
>Alternatively, less specifically, did someone already try to model a large
>mental lexicon (current model has +- 2500 word entries) in ACT-R? I would
>like to discuss some issues involving, for example, representations or
>activations of low frequent words. Same as for the previous question,
>pointers to relevant modeling literature would be very welcome.
>
>  - Hedderik.