[ACT-R-users] similarity

[Duncan Brumby]

Hi Ayman, 

Hope all is going well. 

So your point that: 

>The reason why we get better results with entropy has more 
>to do with the corpus than with technique.

Still presents the challenge for, say modeling a user searching a web
page, because how can one make an a priori assumption about that users
knowledge (or in this case, which training corpus should be selected).
Further, if a model using a particular corpus does not fit the
empirical data, it is then unclear whether this is because the model
is incorrect or that the corpus was incorrect ... a  similar case can
of course be made when attributing the success of a model to fitting
the data.



--Duncan 

~~~~~~~~~~~~~~~~~~~~~~
Duncan Brumby
Intern Microsoft Research 
One Microsoft Way, Building 113
Redmond, WA 98052
phone: +1 (425) 706 8259 x68259
email: BrumbyDP at cardiff.ac.uk
web: http://www.cardiff.ac.uk/psych


On 6/17/05, Ayman Farahat <farahat at parc.com> wrote:
> Hello
> The apple example is interesting because it illustrates how the corpus (or
> domain knowledge) can influence similarity.  My feeling is that if we asked
> Steve Jobs (or for that matter a random person in the bay area) about the
> most similar term to apple, MAC would come up very high.
> An interesting example to try is "entropy". You will see that the similarity
> server captures the two senses of the term; the information theory and
> thermodynamics.
> The reason why we get better results with entropy has more to do with the
> corpus than with technique.
> Ayman
> 
> > From: Peter Pirolli <pirolli at parc.com>
> > Date: Fri, 17 Jun 2005 09:50:05 PDT
> > To: Roy Wilson <rwilson+ at pitt.edu>, act-r-users at act-r.psy.cmu.edu
> > Cc: Ayman Farahat <farahat at parc.com>, Raluca.Budiu at parc.com, royer at parc.com
> > Subject: Re: [ACT-R-users] similarity
> >
> > As I point out
> > in the Pirolli (2005) Cognitive Science paper, PMI is approximately the
> > same thing as association strength in ACT-R. PMI (like LSA) is very good at
> > generating scores that correlate well with synonym judgements (e.g., as
> > tested by the TOEFL test), which might be one way to define
> 
> 
> _______________________________________________
> 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
>