[Topic-models] "online" inference in LDA

Hal Daume III hal at cs.utah.edu
Tue May 13 14:20:35 EDT 2008


I think it would be pretty easy to construct a particle filter for 
solving this.... I don't know if anyone has done this yet, though.

Sugato Basu wrote:
> Dave,
> 
> Arindam Banerjee and I had a paper where we studied online versions of 
> LDA and a couple of other topic models -- mixture of vMF distributions 
> and the DCM model. That may be useful for you:
> 
> http://www.siam.org/proceedings/datamining/2007/dm07_040Banerjee.pdf
> 
> Note that the online inference we used for LDA in this paper was an 
> efficient heuristic based on MAP estimates, not the true online Bayesian 
> update.
> 
> Best,
> Sugato.
> 
> On Tue, 13 May 2008, Dave Stallard wrote:
> 
>> Dear list,
>>
>> I am interested in applying LDA to a conversational dialog 
>> application, in which the "document" is dialog, composed of a sequence 
>> of utterances.  The application is "online", meaning that we want to 
>> apply inference and find topics in the current utterance, considering 
>> only it and the previous utterances of that dialog.  Crucially, we 
>> can't look at future utterances, because they haven't happened yet!  
>> So we never see the whole "document" until the dialog is over, just 
>> successively longer prefixes of it.
>>
>> An obvious thing to do is to apply inference repeatedly to the 
>> document/dialog as seen so far.  This would redo computation of 
>> course, and once we got far enough into the dialog, the gammas would 
>> probably be pretty stable.  Any thoughts on this technique, or on the 
>> whole enterprise in general? Does it make sense?  Has it been done 
>> before?  I've not been able to find anything in the literature on it.
>>
>> (Thanks for answering my previous query re topic tagging, btw.)
>>
>>  Dave
>>
>> _______________________________________________
>> Topic-models mailing list
>> Topic-models at lists.cs.princeton.edu
>> https://lists.cs.princeton.edu/mailman/listinfo/topic-models
>>
> _______________________________________________
> Topic-models mailing list
> Topic-models at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/topic-models

-- 
  Hal Daume III --- me AT hal3 DOT name  |  http://www DOT hal3 DOT name
  "Arrest this man, he talks in maths."  |  http://nlpers.blogspot.com


More information about the Topic-models mailing list