[ACT-R-users] on similarity
Raluca.Budiu at parc.com
Raluca.Budiu at parc.com
Fri Jun 17 13:28:32 EDT 2005
Yes, it's true that the results depend on the number of eigen vectors.
The intuition is that, the more eigen vectors, the more noise (i.e.,
accidental co-occurrences) the system takes into account. With lower
numbers of eigen vectors, more generalization is allowed.
Ayman has run some tests and found out that the optimal number of eigen
vectors is around 200-300 (this is also true for the traditional LSA).
As for the results of the GLSA being called "similarities", that's a
topic very much open to debate indeed. As Peter has said, there is some
theoretical motivation to them being called so (that PMIs can be shown
to roughly be equivalent with the original definition of strengths of
association in ACT-R 4.0). Whether association is the same with
similarity, that's another story. In ACT-R 5.0 I think there was a
tacit understanding (maybe not so tacit, if I remember correctly some
of the discussions at the ACT-R workshops and postgraduate summer
school) that there is some need to experiment with Sji-s and see
exactly what they should reflect (co-occurrence or perhaps some other
measure of semantic similarity).
Raluca
From: Roman Belavkin <R.Belavkin at mdx.ac.uk>
Date: Jun 17, 2005 7:58 AM
Subject: [ACT-R-users] on similarity
To: act-r-users at act-r.psy.cmu.edu
One more observation: The results depend greatly on the number of
eigenvectors used (I understand this controls the number of dimension
PCA reduces the space to). I found the `optimal' value for the word
apple is about 50 dimensions, becuase the results are:
apple carnation 0.6078935
apple fruit 0.5972236
apple blackberry 0.59559596
apple mac 0.5931967
apple orange 0.570567
apple sweet 0.5677523
apple palm 0.5576204
apple radish 0.55353457
apple intel 0.5511009
apple persimmon 0.54269683
They are quite different as you can see.
Cheers!
Roman
-----Original Message-----
From: act-r-users-admin at act-r.psy.cmu.edu on behalf of Roman
Belavkin
Sent: Fri 6/17/2005 15:17
To: Kelley, Troy (Civ,ARL/HRED); act-r-users at act-r.psy.cmu.edu
Cc:
Subject: RE: [SPAM: 4.500] RE: [ACT-R-users] ACT-R output from
GLSA server at PARC
Hi,
I think the word `similar' is not really appropriate here. LSA
just shows co-occurance of the two words really, and higher co-occurance
for word man can be simply explained because the word man means also
human, while woman is more specific and thus less ambiguous.
Cheers,
Roman
-----Original Message-----
From: act-r-users-admin at act-r.psy.cmu.edu on behalf of
Kelley, Troy (Civ,ARL/HRED)
Sent: Thu 6/9/2005 21:52
To: act-r-users at act-r.psy.cmu.edu
Cc:
Subject: [SPAM: 4.500] RE: [ACT-R-users] ACT-R output
from GLSA server at PARC
Here is a sample output from the GLSA server at PARC for
the word Love
love love 0.9999997
love fun 0.16717705
love city 0.103955254
love place 0.2078263
love man 0.36086833
love friend 0.19896048
love neighbor -0.037641484
love woman 0.21863273
love fondness 0.0076576206
Interesting, love is more similar to the word "man" than
"woman", and
more similar to "man" than "fondness" or "friend".
Troy
-----Original Message-----
From: act-r-users-admin at act-r.psy.cmu.edu
[mailto:act-r-users-admin at act-r.psy.cmu.edu] On Behalf
Of
Raluca.Budiu at parc.com
Sent: Thursday, June 09, 2005 3:54 PM
To: act-r-users at act-r.psy.cmu.edu
Subject: [ACT-R-users] ACT-R output from GLSA
server at PARC
Some of you may have heard about PARC's effort to build
an external
GLSA/PMI server; it is now available at:
http://glsa.parc.com/
and it produces ACT-R output (other formats are
supported as well).
GLSA (Generalized Latent Semantic Analysis) is a
LSA-like method of
computing word similarities, but it has the advantage
of an adjustable,
web-based corpus. It takes as input a list of word pairs
and provides
similarities between those words.
Just very recently it started providing ACT-R output;
this is an ACT-R
file that defines a meaning chunk type and sets the
Sij-s between words
to their similarity value as computed by the server.
The server is still in a development phase, but please
feel free to
experiment with it. Comments and suggestions are very
welcome.
Raluca Budiu
_______________________________________________
ACT-R-users mailing list
ACT-R-users at act-r.psy.cmu.edu
http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users
_______________________________________________
ACT-R-users mailing list
ACT-R-users at act-r.psy.cmu.edu
http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users
This message has been checked for viruses but the
contents of an attachment
may still contain software viruses, which could damage
your computer system:
you are advised to perform your own checks. Email
communications with the
University of Nottingham may be monitored as permitted
by UK legislation.
_______________________________________________
ACT-R-users mailing list
ACT-R-users at act-r.psy.cmu.edu
<mailto:ACT-R-users at act-r.psy.cmu.edu>
http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users
_______________________________________________
ACT-R-users mailing list
ACT-R-users at act-r.psy.cmu.edu
http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/act-r-users/attachments/20050617/088fea24/attachment.html>
More information about the ACT-R-users
mailing list