I'm learning Solr and have become confused trying to figure out ICUCollation, what it does, what it is for and how to use it. From here. I haven't found any good explanation of this online. The doc appear to be saying that I need to use this ICUCollation and implies that it does magical things for me, but does not seem to explain exactly why or exactly what, and how it integrates with anything else.
Say I have a text field in French and I want stopwords removed, accents, punctuation and case ignored and stemming... how does ICUCollation come into this? Do I set solr.ICUCollationField and locale='fr' and it will do everything else automatically? Or do I set solr.ICUCollationField and then tokenizer and filters on this in addition? Or do I not use solr.ICUCollationField at all because that's for something completely different? And if so, then what?
Collation is the organisation of written information into an order - ICUCollactionField (the API documentation also provides a good description) is meant to enable you to provide locale aware sorting, as the sort order is defined by cultural norms and specific language properties. This is useful to allow different sorting based on those rules, such as the difference between Norwegian and Swedish, where a Swede would order Å before Æ/Ä and Ø/Ö, while a Norwegian would order it Æ/Ä, Ø/Ö and then Å.
Since you usually don't want to sort by a tokenized field (exception: KeywordTokenizer) or a multivalued field, these fields are usually not processed any more than allowing for the sorting / collation to be performed.
There is a case to be made for collation filters for searching as well, as search in practice is just comparison. This means that if you're aiming to search for two words that would be identical when compared in the locale provided, it would be a hit. The tokens indexed will not make any sense when inspected, but as long as the values are reduced to the same token both when indexing and searching, it would work. There's an example of this on the wiki under UnicodeCollation.
Collation does not affect stopwords (StopFilterFactory), accents (ICUFoldingFilterFactory), punctuation, case (depending on locale - if the locale for sorting is case aware, then it does not) (LowercaseFilterFactory or ICUNormalizer2FilterFactory) or stemming (SnowballPorterFilterFactory). Have a look at the suggested filters for that. Most filters or tokenizers in Solr does very specific tasks, and try to avoid doing "everything and the kitchen sink" in one single filter.
You normally have two or more fields for one text input if you want to do different things like:
search: text analysis
sort: language sensitive / case insensitive sorting
facet: string
For search use something like:
<fieldType name="textFR" class="solr.TextField" positionIncrementGap="100">
<analyzer>
<tokenizer class="solr.ICUTokenizerFactory"/>
<filter class="solr.ICUFoldingFilterFactory"/>
<filter class="solr.ElisionFilterFactory"/>
<filter class="solr.KeywordRepeatFilterFactory"/>
<filter class="solr.FrenchLightStemFilterFactory"/>
<filter class="solr.RemoveDuplicatesTokenFilterFactory"/>
</analyzer>
</fieldType>
For sorting use:
<fieldType name="textSortFR" class="solr.ICUCollationField"
locale="fr"
strength="primary" />
or simply:
<fieldType name="textSort" class="solr.ICUCollationField"
locale=""
strength="primary" />
(If you have to support many languages. Should work fine enough in most cases.)
Do make use of the Analysis UI in the SOLR Admin: open the analysis view for your index, select the field type (e.g. your sort field), add a representative input value in the left text area and a test value in the right field (in case of sorting, this right side value is not as interesting as the sort field is not used for matching).
The output will show you whether:
accents are removed
elisions are removed
lower casing is applied
etc.
For example, if you see that elisions (l'atelier) are not remove (atelier) but you would like to discard it for sorting you would have to add the elision filter (see example for search field type above).
https://cwiki.apache.org/confluence/display/solr/Language+Analysis
Related
I created my own core on http://localhost:8983/solr and added some documents so I could query. But When I query something like"dog", I want those documents that contains "pooch" will be returned too. So I want to implement SVD algorithm to make some improvement on my results.
Since I am new to the search engine thing. All I know is that I can use Mahout to implement SVD, but it seems a little bit difficult coz I have to install Maven, Hadoop and Mahout.
Any suggestion will be appreciated.
You can use SynonymGraphFilterFactory
This filter maps single- or multi-token synonyms, producing a fully correct graph output. This filter is a replacement for the Synonym Filter, which produces incorrect graphs for multi-token synonyms.
If you use this filter during indexing, you must follow it with a Flatten Graph Filter to squash tokens on top of one another like the Synonym Filter.
Create a file i.e mysynonyms.txt in the directory your_collection/conf/ and put the synonyms with => sign
pooch,pup,fido => dog
huge,ginormous,humungous => large
And Example Schema will be :
<analyzer type="index">
<tokenizer class="solr.StandardTokenizerFactory"/>
<filter class="solr.SynonymGraphFilterFactory" synonyms="mysynonyms.txt"/>
<filter class="solr.FlattenGraphFilterFactory"/> <!-- required on index analyzers after graph filters -->
</analyzer>
<analyzer type="query">
<tokenizer class="solr.StandardTokenizerFactory"/>
<filter class="solr.SynonymGraphFilterFactory" synonyms="mysynonyms.txt"/>
</analyzer>
Source : https://cwiki.apache.org/confluence/display/solr/Filter+Descriptions
The is another way to augment your index with terms not in the content. Synonyms is good as #ashraful says. But there are 2 other problems you will run into:
words used but not in the synonym list
behavioral search: using other user behavior as a hint to what they are looking for
These require you to augment the index with terms learned from 1) other searches, and 2) user behavior. Mahout's Correlated Cross Occurrence algorithm can help with both. You can set it up to find terms that lead to people reading an item and (if you have something like purchase or other preference data) conversion items that correlate with items in the index. In the second case you would add user conversions to the search query to personalize the results.
A blog about the technique here: http://actionml.com/blog/personalized_search
The page on Mahout docs here: http://mahout.apache.org/users/algorithms/intro-cooccurrence-spark.html
You should also look at word2vec, which will (given the right training data) find that "dog" and "pooch" are synonyms regardless of the synonym list because it is learned from the data. I'm not sure how you add word2vec to Solr but it is integrated into Fusion, the closed source product of Lucid.
I have a Solr index filled with documents, with a field named issuer.
There is a document with issuer=first issuer.
I'm trying to implement matching of two consequent words. The first word needs to match completely, the second needs to match partially.
What I am trying to achieve is:
I search for something like: issuer:first\ iss*
I expect it to match "first iss uer"
I tried the following solutions but none is working:
issuer:first\ iss* -> returns nothing
issuer:"first iss"* -> returns everything
issuer:(first iss*) -> also returns "issuer first"
Does anybody have a clue on how to achieve the desired result?
My suggestion is to add a shiringle filter based field type to your schema. Below is a simple definition:
<fieldtype name="shingle">
<analyzer>
<tokenizer class="solr.WhitespaceTokenizerFactory"/>
<filter class="solr.ShingleFilterFactory" minShingleSize="2" maxShingleSize="5"/>
</analyzer>
</fieldtype>
You then add another field with this type as shown below:
<field name="issuer_sh" type="shingle" indexed="true" stored="false"/>
At query time, you can issue the following query:
issuer_sh:"first iss*"
The shingleFilter creates n-gram tokens from your text. For instance, if the issuer field contains "first issue", then Solr will create and index the following tokens:
first
issue
first issue
You can't search with wildcards in phrase queries. Without changing how you are indexing (see #ameertawfik's answer), the standard query parser doesn't provide a good way to do this. You can, however, use the surround query parser to search using spans. This query would then look like:
1N(first, iss*)
Keep in mind, surround query parser does not analyze, so 1N(first, iss*) and 1N(First, iss*) will not find the same results.
You could also construct this query using lucene's SpanQueries directly, of course, like:
SpanQuery[] queries = new SpanQuery[2];
queries[0] = new SpanTermQuery(new Term("issuer","first"));
queries[1] = new SpanMultiTermQueryWrapper(new PrefixQuery(new Term("issuer","iss")));
Query finalQuery = new SpanNearQuery(queries, 0, true);
a pretty basic question but can anyone tell me how to remove special characters from documents while indexing in solr? I went through Solr wiki but couldn't find anything relevant. I saw few tokenizers like WhiteSpaceTokenizerFactory and StandardTokenizerFactory. I am using WhiteSpaceTokenizerFactory in my schema.xml but it doesn't seem to solve the purpose. I am still able to query using "*" and "-" etc.
Consider using the standard tokenizer:
<tokenizer class="solr.StandardTokenizerFactory"/>
It should remove the characters you have mentioned.
Once the words have been tokenized you may apply further processing, like splitting on case change and numerics, using the WordDelimiterFilterFactory for better matching.
Also, very useful almost all the time when dealing with schema configuration, is the solr's analysis page. It gives you a lot of valuable feedback.
I have keyword field in Solr schema.
<fieldType name="text_keyword" class="solr.TextField" positionIncrementGap="100">
<analyzer>
<tokenizer class="solr.SimpleKeywordTokenizerFactory"/>
</analyzer>
</fieldType>
When I try to search this field with default solr query or dismax query
category:(Mouse Pad) it creates query (category:Mouse) AND (category:Pad)
I want to know is there a way not to split terms by whitespaces if it is keyword field or so.
Added:
I need SimpleKeywordTokenizerFactory analyze (which is lowercase without white-space splitting) on query, so raw and term query parser doesn't work for me
You want to enter this query:
category:"Mouse Pad"
The query syntax already provides a way to do this. Quotes are for phrases. Parentheses mean something different. You can write your own query parser if you want, but I don't recommend it.
You could use TermQParserPlugin:
{!term f=category}Mouse Pad
Beware that no analysis is performed, so this will only work if the internal representation of your field is "Mouse Pad" (with title case).
Edit (2012-04-17):
If you still want analysis to be performed, all you need to do is to escape the space by prepending a backslash:
{!lucene}category:Mouse\ Pad
I'm trying to search for partial beginning matches on a big list of lastnames. So Wein* should find Weinberg, Weinkamm etc.
I could do this by creating a special field, and adding
<filter class="solr.EdgeNGramFilterFactory" minGramSize="1" maxGramSize="50" preserveOriginal="1"/>
to its type specification in schema.xml. When I add the line above only to the indexing analyzer and leave it empty for the query analyzer, I can then search by just search special_field:Wein and get the expected results.
Now I see that solr also has a *-syntax. What's the connection between EdgeNGramFilterFactory and the *-syntax?
Am I doing things correctly or is there a better, more regular way?
Thanks!
Or just do a simple wild card match:
name:Pe*
I don't recommend the Wein* query. That is implemented internally as PrefixQuery, which rewrites the original query to include all terms that have prefix equals "Wein". Depending on how large is your index (I mean how many terms), this query rewriting can be a bottleneck.
The EdgeNGramFilter at index time is a better approach. This solution will use more space, but queries will be processed much faster.
Note: I also asked this question in the Lucene forum where I got a good answer:
http://lucene.472066.n3.nabble.com/How-to-do-partial-beginning-matches-td781147.html