[ACT-R-users] Need for Multi-Level Activation Spread in ACT-R

Mon Sep 14 13:45:57 EDT 2009

We are in the process of mapping the linguistics representations that
are generated by our language comprehension model into a situation model
based semantic representation. We are trying to do this in a
representationally reasonable way within the ACT-R architecture. The
problem we face is the many-to-many mapping between words and concepts.
Individual words may map to multiple concepts, and individual concepts
may may to multiple words. Given this many-to-many mapping, we would
like to use mapping chunks to map from words to concepts. The mapping
chunks would encode a single mapping relationship (e.g. a separate
mapping chunk to map from the word "bank" to the financial institution
concept; from the word "bank" to the river bank concept; from the
concept dog to the word "dog"; from the concept dog to the word
"canine"). When processing a word, the goal is to retrieve the
contextually relevant concept. We would like to accomplish this in a
single retrieval, however, we do not know how to do this given the
single-level activation spreading mechanism in ACT-R. Since there is no
direct link between a word and a concept if mapping chunks are used
(i.e. there is no slot in the concept that contains the word), the word
will not spread activation to the concept. Instead, given the use of
mapping chunks, it appears that two retrievals are needed: 1) given the
word, retrieve a mapping chunk, and 2) given a mapping chunk, retrieve a
concept. Since our model of language comprehension is already slower
than humans at processing language, any extra retrievals are
problematic. In fact, we have already eliminated an extra retrieval in
determining the part-of-speech of a word. Previously, two retrievals
were needed: 1) retrieve the word corresponding to the perceptual input,
and 2) given the word (and context) retrieve the part-of-speech of the
word. While we were successful in eliminating a retrieval, the resulting
word-pos chunks contain a mixture of word form information (e.g. the
letters and trigrams in the word) and pos information. Even so, they do
not yet contain any representation of phonetic, phonemic, syllabic or
morphemic information. With just letter and trigram information, long
words contain many slots. Ideally, we would like to represent letter and
trigram information independently of each other and POS information
(allowing them to interact in retrieving a word), but given the
single-level activation spreading mechanism in ACT-R doing so would
necessitate multiple independent retrievals, which would fail to capture
the interaction of letter and trigram information that leads to
successful retrievals of words in the face of variability in the
perceptual form (e.g. "arispeed" should retrieve "airspeed").

      The fall back for mapping words to concepts is to embed all the
possible concepts as slot values in a word and vice versa. While we
consider this a representationally problematic solution -- word and
concept chunks will wind up needing many extra slots, we do not know how
else to work around the single-level activation spread in ACT-R.

      The primary empirical argument against the need for multi-level
activation spread in ACT-R is based on studies which show no activation
from words like "bull" to words like "milk", even though "bull"
activates "cow" and "cow" activates "milk". Even if it is true that
there are no instances of "indirect" activation from "bull" to "milk",
this does not rule out the need for multi-level activation spread. There
is a hidden assumption that "cow" and "bull" are directly associated,
and that "cow" and "milk" are also directly associated. Such direct
associations may seem reasonable in small-scale models addressing
specific spreading activation phenomena, but they are questionable in a
larger-scale model. Do we really want to include all the direct
associates of "cow" as slot values in the "cow" chunk, and do the same
for all other chunks? 

      We understand that the inclusion of a multi-level activation
spreading mechanism in ACT-R would be computationally explosive.
However, we would like to have the capability to explore use of such a
mechanism and to look for ways to keep it computationally tractable. We
have already dealt with the problem of computational explosion in our
word retrieval mechanism. Originally, we attempted to use a "soft
constraint" retrieval mechanism for words. All words in DM were
candidates for retrieval--the most highly activated word being
retrieved. With just 2500 words in DM, the activation calculations
slowed the model down considerably. To manage retrievals in a tractable
manner we implemented a disjunctive retrieval capability combined with a
new perceptual span mechanism -- the model first tries a hard-constraint
retrieval on the entire perceptual span (which is larger than a word)
using the "get-chunk" function (and chop-string under the covers). If
get-chunk succeeds (indicating that there is a chunk in DM corresponding
to the entire perceptual span) a retrieval is constructed using the
entire perceptual span as a hard constraint to retrieve the
corresponding multi-word unit in DM, if this fails, the model backs-off
and uses the first space delimited word (using chop-string) in the
perceptual span to check for a corresponding word in DM -- if a match is
found with get-chunk, a retrieval is constructed to retrieve the word.
If all else fails, we construct a retrieval that imposes a hard
constraint on the first letter (this is less than ideal, but a
reasonable compromise). The overall effect is a (nearly) soft-constraint
retrieval implemented in a computationally tractable way.

      A similar capability to effect multi-level activation spread in a
computationally tractable manner would be highly desirable.

Jerry

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/act-r-users/attachments/20090914/699bac8f/attachment.html>