Re: Similar Document Search
> As a user of Lucene I missed some features. Part of the OSS culture is > for me to tell others about this and maybe to try to find solutions. > Mark's code seems to be one, so I proposed to consider adding it into > some spot with better exposure for testing. And I don't seem to be the > only person with the need for these features. I think Lucene would be > better if these features were easily available. If the Lucene team > doesn't think so -- fair enough, it is their project. But asking me to > stop requesting features in a (hopefully) sensible way is pretty much > against the spirit of OSS and hacker culture as far as I understand it. > > Does that answer your questions? > Yep. I guess we have different philosophies but are more or less heading toward the same goal :) I've been on other projects where the response to questions has been "look at the source". So maybe I have gotten into a bad habit of hacking first and asking questions second. thanks for the excellent reply, btw. brian > Peter > - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
RE: Similar Document Search
Hi Terry, the suggestion of Haystack's Lucene was a hint to give you an additional alternative to reach your goal. Depending on the definition of your notion "similar document", this solution does or does not make sense. My definition of similar document (and term) is maybe more general than yours: It supports rather generic similarity metrics and needs to cover cosine similarity according to vector-space model (VSM; can be achieved using unmodified Lucene code), semantic similarity according to a generative model like latent semantic indexing or Bayesian approaches etc. and even semantic similarity according to a taxonomy. If you want such a flexibility (like I do for my research), you should consider this approach because you can relatively easily work on the forward document vectors. If all you need is vanilla VSM cosine similarity, you are probably best off with the suggestion that was sent in this list, to submit the document content in the query and throw it through the same Analyzer that was used to create the index, thus finding best matches using Lucene's standard matching scheme. Good luck, Gregor -Original Message- From: Terry Steichen [mailto:[EMAIL PROTECTED] Sent: Thursday, August 21, 2003 2:54 PM To: Lucene Users List Subject: Re: Similar Document Search 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, > > 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 and > >vector space searches. However, the solution is not really completely fixed > >yet, 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 frequen
Re: Similar Document Search
Brian Mila wrote: amounts). I failed to find a way to get Lucene to give me this information without hacking this or that. Considering the attention IR Excuse me if this is off-topic, but isn't hacking the code what open source software is all about? Not always, but quite often :-) I mean, its always better to try to do it with existing methods but if it can't, why not hack the source? Because you might need to put quite some effort into getting it right? Because you might do something someone else already did better -- which is not really against the spirit of hackerism, but I have so many other things to hack where I think I can do better than most people. Inverted file indexes is not my particular domain. If it works and people use it then it should probably be incorporated into the main source tree. If poeple don't use it (or the hack is terribly ugly, which may be what you were referring to) then it doesn't make the cut. That needs exposure. If some Lucene code is hidden in the Haystack project, it won't get enough exposure IMO. In either case, I'm just wondering why I see many questions or answers include this almost standard reply. I hack the source regularly to acheive a needed goal. Sure its not forward-compatible, but if I waited for the feature to be added on its own, our project would never get off the ground. One of the important things about OSS for me is resuse and collaboration. If you hack things again and again without trying to turn it into something reusable, I'd say you constantly create small proprietary forks based on open source code but you are not part of any OSS effort. That's of course my point of view on OSS, but then you asked for it :-) As a user of Lucene I missed some features. Part of the OSS culture is for me to tell others about this and maybe to try to find solutions. Mark's code seems to be one, so I proposed to consider adding it into some spot with better exposure for testing. And I don't seem to be the only person with the need for these features. I think Lucene would be better if these features were easily available. If the Lucene team doesn't think so -- fair enough, it is their project. But asking me to stop requesting features in a (hopefully) sensible way is pretty much against the spirit of OSS and hacker culture as far as I understand it. Does that answer your questions? Peter - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: Similar Document Search
> amounts). I failed to find a way to get Lucene to give me this > information without hacking this or that. Considering the attention IR Excuse me if this is off-topic, but isn't hacking the code what open source software is all about? I mean, its always better to try to do it with existing methods but if it can't, why not hack the source? If it works and people use it then it should probably be incorporated into the main source tree. If poeple don't use it (or the hack is terribly ugly, which may be what you were referring to) then it doesn't make the cut. In either case, I'm just wondering why I see many questions or answers include this almost standard reply. I hack the source regularly to acheive a needed goal. Sure its not forward-compatible, but if I waited for the feature to be added on its own, our project would never get off the ground. Brian - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
RE: Similar Document Search
Apologies for asking the obvious, but could someone explain why Documents.Document is a sealed class? Seems like many of us would love to implement UniqueDocument to support this oft-requested uniqueness field. Would still have the task of implementing an IndexWriterEx.AddDocument(UniqueDocument), but all the "uniqueness" tests could be encapsulated in UniqueDocument... Along the same lines, Document.Field is also sealed... Have others wished to derive from that, too? In our case, it would be to have a notion of field "schemas" so the isstored, isindexed, istokenized values, etc. could be defined once for the field. More exotic things like field aliases, etc. could be implemented there, too. Just curious... Wrapping the classes works fine for now. -Eric - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: Similar Document Search
Hi Terry, 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, 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
Re: Similar Document Search
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, > > 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 and > >vector space searches. However, the solution is not really completely fixed > >yet, 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 conduct > >> > >> > >a search based on a document rather than a query? (One use of this would be > >to refine a search by selecting one of the hits returned from the initial > >query and subsequently retrieving other documents "like" the selected one.) > > > > > >>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]
Re: Similar Document Search
Hi all, 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 and vector space searches. However, the solution is not really completely fixed yet, 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 conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) 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]
RE: Similar Document Search
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 and vector space searches. However, the solution is not really completely fixed yet, 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 conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) > >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]
Re: Similar Document Search
Hi Peter, I guess you are right. I've implemented this for a index with ten millions of really small documents that all are stored in the index. The documents are never more than a thousand words so re-indexing is quick enough. However it is probably not advisable to do this with bigger documents or documents that need additional parsing. /magnus Peter Becker wrote: Hi Magnus, thanks for the offer, but unfortunately I can't/don't want to make the assumption that I can easily access the documents to re-index them. And I don't think this approach would be feasible unless you can keep the documents in memory somehow. Storing the other/non-inverted/normal/whatever index would be expensive for indexing, but querying should be a lot faster than having to re-index documents. That is in our situation preferable. Peter Magnus Johansson wrote: Hi Peter If the original document is available. You could extract keywords from the document at query time. That is when someone asks for documents similar to document a. You re-analyze document a and in combination with statistics from the Lucene index you extract keywords from document a that can then be used as a query for findining similar documents. I've got some sample code if anyone is interested. /magnus - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: Similar Document Search
Hi Magnus, thanks for the offer, but unfortunately I can't/don't want to make the assumption that I can easily access the documents to re-index them. And I don't think this approach would be feasible unless you can keep the documents in memory somehow. Storing the other/non-inverted/normal/whatever index would be expensive for indexing, but querying should be a lot faster than having to re-index documents. That is in our situation preferable. Peter Magnus Johansson wrote: Hi Peter If the original document is available. You could extract keywords from the document at query time. That is when someone asks for documents similar to document a. You re-analyze document a and in combination with statistics from the Lucene index you extract keywords from document a that can then be used as a query for findining similar documents. I've got some sample code if anyone is interested. /magnus Peter Becker wrote: 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 conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) 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]
Re: Similar Document Search
Ok, here it is. It's part of a JSP that prints out all keywords in a document. /magnus <%@ page import="org.apache.lucene.index.IndexReader, org.apache.lucene.document.Document, com.technohuman.search.language.SwedishAnalyzer, java.io.StringReader, org.apache.lucene.analysis.TokenStream, org.apache.lucene.analysis.Token, org.apache.lucene.index.Term, org.apache.lucene.index.TermEnum, java.util.*"%> <%! class Entry implements Comparable { public double score; public String termText; public Entry(double score, String termText) { this.score = score; this.termText = termText; } public int compareTo(Object o) { Entry e = (Entry) o; if (e.score < score) return -1; else return 1; } } %> <% IndexReader reader = IndexReader.open(application.getRealPath("/WEB-INF/index")); Document d = reader.document(Integer.parseInt(request.getParameter("docId"))); Map m = new HashMap(); // Count all terms in the description field of the given document String description = d.getField("Parser.DESCRIPTION").stringValue(); final java.io.Reader r = new StringReader(description); final TokenStream in = new SwedishAnalyzer().tokenStream(r); for (; ;) { final Token token = in.next(); if (token == null) { break; } if (m.containsKey(token.termText())) { int a = ((Integer)m.get(token.termText())).intValue(); m.put(token.termText(), new Integer(a + 1)); } else { m.put(token.termText(), new Integer(1)); } } ArrayList tm = new ArrayList(); // Calculate inverse document frequency * term frequency Iterator it = m.keySet().iterator(); while (it.hasNext()) { String termText = (String) it.next(); TermEnum te = reader.terms(new Term("Parser.DESCRIPTION", termText)); double idf = Math.log(reader.numDocs() / (te.docFreq() + 1)) + 1; double tf = Math.sqrt(((Integer)m.get(termText)).intValue()); tm.add(new Entry(idf * tf, termText)); } Collections.sort(tm); // Print the keywords and the score for each keyword Iterator it2 = tm.iterator(); while (it2.hasNext()) { Entry e = (Entry) it2.next(); out.println(e.score + " " + e.termText + ""); } reader.close(); %> Rociel Buico wrote: hello magnus, can i ask your sample script? --buics Hi Peter If the original document is available. You could extract keywords from the document at query time. That is when someone asks for documents similar to document a. You re-analyze document a and in combination with statistics from the Lucene index you extract keywords from document a that can then be used as a query for findining similar documents. I've got some sample code if anyone is interested. /magnus Peter Becker wrote: 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 conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) 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] - Do you Yahoo!? The New Yahoo! Search - Faster. Easier. Bingo. - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: Similar Document Search
hello magnus, can i ask your sample script? --buics Hi Peter If the original document is available. You could extract keywords from the document at query time. That is when someone asks for documents similar to document a. You re-analyze document a and in combination with statistics from the Lucene index you extract keywords from document a that can then be used as a query for findining similar documents. I've got some sample code if anyone is interested. /magnus Peter Becker wrote: > 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 >> conduct a search based on a document rather than a query? (One use >> of this would be to refine a search by selecting one of the hits >> returned from the initial query and subsequently retrieving other >> documents "like" the selected one.) >> >> 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] - Do you Yahoo!? The New Yahoo! Search - Faster. Easier. Bingo.
Re: Similar Document Search
Hi Peter If the original document is available. You could extract keywords from the document at query time. That is when someone asks for documents similar to document a. You re-analyze document a and in combination with statistics from the Lucene index you extract keywords from document a that can then be used as a query for findining similar documents. I've got some sample code if anyone is interested. /magnus Peter Becker wrote: 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 conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) 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]
Re: Similar Document Search
Hi Peter, What got me thinking about this was the way that Lucene computes similarity (or scoring). After the boolean keyword matches have been found, Lucene then computes relevance. What Lucene does, I think, is to process the query into some intermediate internal representation and computes the similarity between the query (now a kind of a pseudo-document) and each of the matching hits. I was wondering if there might not be a way to internally process a selected document (rather than the query per se) and then, in effect, compute the similarity between that document and all the other documents (which have already been pre-processed in the indexing process). So, what you'd be doing is not a boolean keyword match, but a ranking of all the documents in the repository on the basis of relevance or similarity to the target document. (If that's not too far off in terms of reality, maybe Doug could comment?) Regards, Terry - Original Message - From: "Peter Becker" <[EMAIL PROTECTED]> To: "Lucene Users List" <[EMAIL PROTECTED]> Sent: Monday, August 18, 2003 9:05 PM 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 conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) > > > >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]
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 conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) Regards, Terry - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: Similar Document Search
Using the QueryFilter would help with the refining a search based on hits from a previous search, but it wouldn't help with the "like" part your asked about. I'm interested in what you turn up with this though. Erik On Monday, August 18, 2003, at 01:11 PM, Terry Steichen wrote: Is it possible without extensive additional coding to use Lucene to conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) Regards, Terry - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Similar Document Search
Is it possible without extensive additional coding to use Lucene to conduct a search based on a document rather than a query? (One use of this would be to refine a search by selecting one of the hits returned from the initial query and subsequently retrieving other documents "like" the selected one.) Regards, Terry