lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Tommaso Teofili (JIRA)" <>
Subject [jira] [Commented] (LUCENE-4345) Create a Classification module
Date Thu, 13 Sep 2012 10:55:08 GMT


Tommaso Teofili commented on LUCENE-4345:

bq. I agree with robert you should build a small inverted index instead of retokenizing. I'd
use a BytesRefHash with a parallel array as we use during indexing, if you have trouble with
this I am happy to update your patch and give you an example.

+1, not trouble but actually not much time in the latest days to work on it, however if you
have time that'd be surely nice (I've committed this on trunk at r1384219 so it should be
easier to update it).

bq. I suggest to move the into the while() part like while((next =
!= null) for consistency (in assignClass)


bq. Can you use BytesRef for fieldNames to safe the conversion everytime.

actually this depends on "who"'s calling the Classifier, if, for example, it's Solr then it
makes sense to keep Strings but since this lives in Lucene it may make sense to use ByteRefs,
however from an API point of view I'm not sure what could be better.

bq. Instead of specifying the document as a String you should rather use IndexableField and
in turn pull the tokenstream from IndexableField#tokenStream(Analyzer)

I think this is not always ok as often the document to be classified is not supposed to be
in the index immediately but _may_ get indexed right after the classification, however we
could provide that with as an additional method. 

bq. I didn't see a reason why you use Double instead of double (primitive) as return values,
I think the boxing is unnecessary

yes, I agree.

bq. in assignClass can't you reuse the BytesRef returned from the termsEnum for further processing
instead of converting it to a string?

actually, after reading your comment above I realized converting to a String is not a good
idea, so I'll change the methods (#calculateLikelihood and #calculatePrior) to use ByteRef
rather than String.

bq. in getWordFreqForClass you might want to use TotalHitCountCollector since you are only
interested in the number of hits. That collector will not score or collect any documents at
all and is way less complex that the default TopDocsCollector

very good point, thanks :)

bq. I have trouble to understand why the interface expects an atomic reader here. From my
perspective you should handle per-segment aspect internally and instead just use IndexReader
in the interface.

as a first implementation I thought it made sense to keep the complexity of explicitly doing
distributed probabilities calculations out, also AtomicReaders expose more internals that
can be leveraged in a classification algorithm.

bq. The interface you defined has some problems with respect to Multi-Threading IMO. The interface
itself suggests that this class is stateful and you have to call methods in a certain order
and at the same you need to make sure that it is not published for read access before training
is done. 

it'd raise an exception if #assignClass() is called before #train()

bq. I think it would be wise to pass in all needed objects as constructor arguments and make
the references final so it can be shared across threads and add an interface that represents
the trained model computed offline? In this case it doesn't really matter but in the future
it might make sense. We can also skip the model interface entirely and remove the training
method until we have some impls that really need to be trained.

I'm +1 for making the references final while I put the #train() method so that a Classifier
could be trained multiple times. In this implementation that doesn't make much difference
but it may not be the case for other implementations.
Therefore we could (maybe should) mark this API _@experimental_ and just evolve it form the
different implementations we have so finally moving parameters to the constructor may be a
nice idea here.
On the contrary removing the #train() method from the API would remove any reference to Lucene
APIs in the Classifier interface leading to question if that's too much generic.
> Create a Classification module
> ------------------------------
>                 Key: LUCENE-4345
>                 URL:
>             Project: Lucene - Core
>          Issue Type: New Feature
>            Reporter: Tommaso Teofili
>            Assignee: Tommaso Teofili
>            Priority: Minor
>         Attachments: LUCENE-4345_2.patch, LUCENE-4345.patch, SOLR-3700_2.patch, SOLR-3700.patch
> Lucene/Solr can host huge sets of documents containing lots of information in fields
so that these can be used as training examples (w/ features) in order to very quickly create
classifiers algorithms to use on new documents and / or to provide an additional service.
> So the idea is to create a contrib module (called 'classification') to host a ClassificationComponent
that will use already seen data (the indexed documents / fields) to classify new documents
/ text fragments.
> The first version will contain a (simplistic) Lucene based Naive Bayes classifier but
more implementations should be added in the future.

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see:

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message