exactly these two features of (a) having a unique identifier and (b) easily finding the term frequencies for the document is what we (i.e. our working group and seemingly others) are missing.
(a) As far as I understand Lucene, there is no such notion as value identity on Document instances. This is a problem if you want to do things like applying set-theory as we did. The workaround is easy: store a unique id in the index and wrap the documents in an object using this field as base for equals/hashCode. But it still has to be done, you need the unique id in the index and it is not really elegant since you have to go through the wrapper all the time.
(b) Lucene allows you to find term frequencies in the index, but not for subsets or single items. Many information retrieval approaches define document similarity using metrics on the term frequencies. The more similar the term frequencies, the more similar the documents are considered. You get different details and levels of complexity (esp. if you try to mix in background knowledge like knowing synonyms and generalizations of the terms), but the basic idea is that documents are similar if they contain the same terms (and maybe even in the same amounts). I failed to find a way to get Lucene to give me this information without hacking this or that. Considering the attention IR techniques like latent semantic analysis (LSA) and others get nowadays (and rightfully so I think), not finding these features in Lucene was a bit of a surprise. I still haven't looked at Mark's code, but I would be surprised if he had to do much. But you still have to do something.
After the more abstract talk a bit of a more concrete answer for your question: one simple way of defining similarity of documents is just treating the term frequencies of some (or all) terms as a vector space and then use a metric in the vector space to define distance. If you have two frequency maps, you can for example go through all keys in them, create all differences of the values attached (assuming null if a term is not in a map) and sum them up (giving you the manhattan metric in R^n), then you divide by the numbers of terms to normalize (the frequency maps are probably of different lengths) and that might give you a reasonable first try. If the result is zero, you consider the documents to be extremely similar, the higher the value, the more different they are suppossed to be.
The approach I described is a bit too naive to be really good -- for example I'd expect some bias towards more similarity on documents with less terms. And there are so many other enhancements you could do. Actually the whole idea is a field of research. And one I am not really expert in, I just sometimes work with people who are. This might help:
http://citeseer.nj.nec.com/cs?q=latent+semantic+analysis&submit=Search+Documents&cs=1
Maybe someone else on the list has better pointers.
Peter
Terry Steichen wrote:
Hi Peter,
I took a look at Mark's thesis and briefly at some of his code. It appears to me that what he's done with the so-called forward indexing is to (a) include a unique id with each document (allowing retrieval by id rather than by a standard query), and to (b) include a frequency map class with each document (allowing easier retrieval of term frequency information).
Now I may be missing something very obvious, but it seems to me that both of these functions can be done rather easily with the standard (unmodified) version of Lucene. Moreover, I don't understand how use of these functions will facilitate retrieval of documents that are "similar" to a selected document, as outlined in my original question on this topic.
Could you (or anyone else, of course) perhaps elaborate just a bit on how using this approach will help achieve that end?
Regards,
Terry
----- Original Message ----- From: "Peter Becker" <[EMAIL PROTECTED]> To: "Lucene Users List" <[EMAIL PROTECTED]> Sent: Thursday, August 21, 2003 1:37 AM Subject: Re: Similar Document Search
Hi all,and
it seems there are quite a few people looking for similar features, i.e. (a) document identity and (b) forward indexing. So far we hacked (a) by using a wrapper implementing equals/hashcode based on a unique field, but of course that assumes maintaining a unique field in the index. (b) is something we haven't tackled yet, but plan to.
The source code for Mark's thesis seems to be part of the Haystack distribution. The comments in the files put it under Apche-license. This seems to make it a good candidate to be included at least in the Lucene sandbox -- although I haven't tried it myself yet. But it sounds like a good candidate for us to use.
Since the haystack source is a bit larger and I actually couldn't get the download at the moment, here is a copy of the relevant bit grabbed from one of my colleague's machines:
http://www.itee.uq.edu.au/~pbecker/luceneHaystack.tar.gz (22kb)
Note that this is just a tarball of src/org/apache/lucene out of some Haystack source. Untested, unmodified.
I'd love to see something like this supported in the Lucene context were people might actually find it :-)
Peter
Gregor Heinrich wrote:
Hello Terry,
Lucene can do forward indexing, as Mark Rosen outlines in his Master's thesis: http://citeseer.nj.nec.com/rosen03email.html.
We use a similar approach for (probabilistic) latent semantic analysis
fixedvector space searches. However, the solution is not really completely
conductyet, therefore no code at this time...
Best regards,
Gregor
-----Original Message----- From: Peter Becker [mailto:[EMAIL PROTECTED] Sent: Tuesday, August 19, 2003 3:06 AM To: Lucene Users List Subject: Re: Similar Document Search
Hi Terry,
we have been thinking about the same problem and in the end we decided that most likely the only good solution to this is to keep a non-inverted index, i.e. a map from the documents to the terms. Then you can query the most terms for the documents and query other documents matching parts of this (where you get the usual question of what is actually interesting: high frequency, low frequency or the mid range).
Indexing would probably be quite expensive since Lucene doesn't seem to support changes in the index, and the index for the terms would change all the time. We haven't implemented it yet, but it shouldn't be hard to code. I just wouldn't expect good performance when indexing large collections.
Peter
Terry Steichen wrote:
Is it possible without extensive additional coding to use Lucene to
bea search based on a document rather than a query? (One use of this would
to refine a search by selecting one of the hits returned from the initial
one.)query and subsequently retrieving other documents "like" the selected
Regards,
Terry
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
