Hi Alex,
Thank you for this explanation. This really helped me to understand how it
works, and now I managed to get results I was expecting just after setting
max_query_terms value to be 0 or some very high value. With these results
in my tests I was able to identify duplicates. I noticed couple of things
though.
- I got much better results with web pages when I indexed attachment as
html source and use text extracted by Jsoup in query, then when I indexed
text extracted from web page as attachment and used text in query. I
suppose that difference is related to the fact that Jsoup did not extract
text in the same way as Tika parser used by ES did.
- There was significant improvement in the results in the second test when
I have indexed 50 web pages, then in first test when I indexed 10 web
pages. I deleted index before each test. I suppose that this is related to
the tf*idf.
If so, does it make sense to provide some training set for elasticsearch
that will be used to populate index before system is started to be used?
Could you please define "relevant" in your setting? In a corpus of very
similar documents, is your goal to find the ones which are oddly different?
Have you looked into ES significant terms?
I have the service that recommends documents to the students based on their
current learning context. It creates tokenized string from titles,
descriptions and keywords of the course lessons student is working at the
moment. I'm using this string as input to the mlt_like_text to find some
interesting resources that could help them.
I want to avoid having duplicates (or very similar documents) among top
documents that are recommended.
My idea was that during the documents uploading (before I index it with
elasticsearch) I find if there already exists it's duplicate, and store
this information as ES document field. Later, in query I can specify that
duplicates are not recommended.
Here you should probably strip the html tags, and solely index the text in
its own field.
As I already mentioned this didn't give me good results for some reason.
Do you think this approach would work fine with large textual documents,
e.g. pdf documents having couple of hundred of pages? My main concern is
related to performances of these queries using like_text, so that's why I
was trying to avoid this approach and use mlt with document id as input.
Thanks,
Zoran
On Wednesday, 7 May 2014 06:14:56 UTC-7, Alex Ksikes wrote:
>
> Hi Zoran,
>
> In a nutshell 'more like this' creates a large boolean disjunctive query
> of 'max_query_terms' number of interesting terms from a text specified in
> 'like_text'. The interesting terms are picked up with respect to the their
> tf-idf scores in the whole corpus. These later parameters could be tuned
> with 'min_term_freq', 'min_doc_freq', and 'min_doc_freq' parameters. The
> number of boolean clauses that must match is controlled by
> 'percent_terms_to_match'. In the case of specifying only one field in
> 'fields', the analyzer used to pick up the terms in 'like_text' is the one
> associated with the field, unless specified specified by 'analyzer'. So as
> an example, the default is to create a boolean query of 25 interesting
> terms where only 30% of the should clauses must match.
>
> On Wednesday, May 7, 2014 5:14:11 AM UTC+2, Zoran Jeremic wrote:
>>
>> Hi Alex,
>>
>>
>> If you are looking for exact duplicates then hashing the file content,
>> and doing a search for that hash would do the job.
>> This trick won't work for me as these are not exact duplicates. For
>> example, I have 10 students working on the same 100 pages long word
>> document. Each of these students could change only one sentence and upload
>> a document. The hash will be different, but it's 99,99 % same documents.
>> I have the other service that uses mlt_like_text to recommend some
>> relevant documents, and my problem is if this document has best score, then
>> all duplicates will be among top hits and instead recommending users with
>> several most relevant documents I will recommend 10 instances of same
>> document.
>>
>
> Could you please define "relevant" in your setting? In a corpus of very
> similar documents, is your goal to find the ones which are oddly different?
> Have you looked into ES significant terms?
>
>
>> If you are looking for near duplicates, then I would recommend extracting
>> whatever text you have in your html, pdf, doc, indexing that and running
>> more like this with like_text set to that content.
>> I tried that as well, and results are very disappointing, though I'm not
>> sure if that would be good idea having in mind that long textual documents
>> could be used. For testing purpose, I made a simple test with 10 web pages.
>> Maybe I'm making some mistake there. What I did is to index 10 web pages
>> and store it in document as attachment. Content is stored as byte[]. Then
>> I'm using the same 10 pages, extract content using Jsoup, and try to find
>> similar web pages. Here is the code that I used to find similar web pages
>> to the provided one:
>> System.out.println("Duplicates for link:"+link);
>> System.out.println(
>> "************************************************");
>> String indexName=ESIndexNames.INDEX_DOCUMENTS;
>> String indexType=ESIndexTypes.DOCUMENT;
>> String mapping = copyToStringFromClasspath(
>> "/org/prosolo/services/indexing/document-mapping.json");
>> client.admin().indices().putMapping(putMappingRequest(
>> indexName).type(indexType).source(mapping)).actionGet();
>> URL url = new URL(link);
>> org.jsoup.nodes.Document doc=Jsoup.connect(link).get();
>> String html=doc.html(); //doc.text();
>> QueryBuilder qb = null;
>> // create the query
>> qb = QueryBuilders.moreLikeThisQuery("file")
>> .likeText(html).minTermFreq(0).minDocFreq(0);
>> SearchResponse sr = client.prepareSearch(ESIndexNames.
>> INDEX_DOCUMENTS)
>> .setQuery(qb).addFields("url", "title",
>> "contentType")
>> .setFrom(0).setSize(5).execute().actionGet();
>> if (sr != null) {
>> SearchHits searchHits = sr.getHits();
>> Iterator<SearchHit> hitsIter = searchHits.iterator();
>> while (hitsIter.hasNext()) {
>> SearchHit searchHit = hitsIter.next();
>> System.out.println("Duplicate:" + searchHit.getId()
>> + " title:"+searchHit.getFields().get("url"
>> ).getValue()+" score:" + searchHit.getScore());
>> }
>> }
>>
>> And results of the execution of this for each of 10 urls is:
>>
>> Duplicates for link:http://en.wikipedia.org/wiki/Mathematical_logic
>> ************************************************
>> Duplicate:Crwk_36bTUCEso1ambs0bA URL:http://
>> en.wikipedia.org/wiki/Mathematical_logic score:0.3335998
>> Duplicate:--3l-WRuQL2osXg71ixw7A URL:http://
>> en.wikipedia.org/wiki/Chemistry score:0.16319205
>> Duplicate:8dDa6HsBS12HrI0XgFVLvA URL:http://
>> en.wikipedia.org/wiki/Formal_science score:0.13035104
>> Duplicate:1APeDW0KQnWRv_8mihrz4A
>> URL:http://en.wikipedia.org/wiki/Starscore:0.12292466
>> Duplicate:2NElV2ULQxqcbFhd2pVy0w URL:http://
>> en.wikipedia.org/wiki/Crystallography score:0.117023855
>>
>> Duplicates for link:http://en.wikipedia.org/wiki/Mathematical_statistics
>> ************************************************
>> Duplicate:Crwk_36bTUCEso1ambs0bA URL:http://
>> en.wikipedia.org/wiki/Mathematical_logic score:0.1570246
>> Duplicate:pPJdo7TAQhWzTdMAHyPWkA URL:http://
>> en.wikipedia.org/wiki/Mathematical_statistics score:0.1498403
>> Duplicate:--3l-WRuQL2osXg71ixw7A URL:http://
>> en.wikipedia.org/wiki/Chemistry score:0.09323166
>> Duplicate:1APeDW0KQnWRv_8mihrz4A
>> URL:http://en.wikipedia.org/wiki/Starscore:0.09279101
>> Duplicate:8dDa6HsBS12HrI0XgFVLvA URL:http://
>> en.wikipedia.org/wiki/Formal_science score:0.08606046
>>
>> Duplicates for link:http://en.wikipedia.org/wiki/Formal_science
>> ************************************************
>> Duplicate:8dDa6HsBS12HrI0XgFVLvA URL:http://
>> en.wikipedia.org/wiki/Formal_science score:0.12439237
>> Duplicate:--3l-WRuQL2osXg71ixw7A URL:http://
>> en.wikipedia.org/wiki/Chemistry score:0.11299215
>> Duplicate:Crwk_36bTUCEso1ambs0bA URL:http://
>> en.wikipedia.org/wiki/Mathematical_logic score:0.107585154
>> Duplicate:2NElV2ULQxqcbFhd2pVy0w URL:http://
>> en.wikipedia.org/wiki/Crystallography score:0.07795183
>> Duplicate:pPJdo7TAQhWzTdMAHyPWkA URL:http://
>> en.wikipedia.org/wiki/Mathematical_statistics score:0.076521285
>>
>> Duplicates for link:http://en.wikipedia.org/wiki/Star
>> ************************************************
>> Duplicate:1APeDW0KQnWRv_8mihrz4A
>> URL:http://en.wikipedia.org/wiki/Starscore:0.21684575
>> Duplicate:2NElV2ULQxqcbFhd2pVy0w URL:http://
>> en.wikipedia.org/wiki/Crystallography score:0.15316588
>> Duplicate:vFf9IdJyQ-yfPnqzYRm9Ig URL:http://
>> en.wikipedia.org/wiki/Cosmology score:0.123572096
>> Duplicate:--3l-WRuQL2osXg71ixw7A URL:http://
>> en.wikipedia.org/wiki/Chemistry score:0.1177105
>> Duplicate:Crwk_36bTUCEso1ambs0bA URL:http://
>> en.wikipedia.org/wiki/Mathematical_logic score:0.11373919
>>
>> Duplicates for link:http://en.wikipedia.org/wiki/Chemistry
>> ************************************************
>> Duplicate:--3l-WRuQL2osXg71ixw7A URL:http://
>> en.wikipedia.org/wiki/Chemistry score:0.13033955
>> Duplicate:2NElV2ULQxqcbFhd2pVy0w URL:http://
>> en.wikipedia.org/wiki/Crystallography score:0.121021904
>> Duplicate:8dDa6HsBS12HrI0XgFVLvA URL:<span style="colo
>>
>
> Here you should probably strip the html tags, and solely index the text in
> its own field.
>
--
You received this message because you are subscribed to the Google Groups
"elasticsearch" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to elasticsearch+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/elasticsearch/a92beaad-05bf-431b-9a37-f51512f50aa8%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
