This post is written for advanced users. If you do not know what SVN (Subversion) is or if you’re not ready to get your hands dirty, there might be something more interesting to read on Wikipedia. As usual. This is an introduction to how to get a Lucene development environment running, a Solr environment and lastly, to create your own Snowball stemmer. Read on if that seems interesting. The receipe for regenerating the Snowball stemmer (I’ll get back to that…) assumes that you’re running Linux. Please leave a comment if you’ve generated the stemmer class under another operating system.
When indexing data in Lucene (a fulltext document search library) and Solr (which uses Lucene), you may provide a stemmer (a piece of code responsible for “normalizing” words to their common form (horses => horse, indexing => index, etc)) to give your users better and more relevant results when they search. The default stemmer in Lucene and Solr uses a library named Snowball which was created to do just this kind of thing. Snowball uses a small definition language of its own to generate parsers that other applications can embed to provide proper stemming.
By using Snowball Lucene is able to provide a nice collection of default stemmers for several languages, and these work as they should for most selections. I did however have an issue with the Norwegian stemmer, as it ignores a complete category of words where the base form end in the same letters as plural versions of other words. An example:
The base form is “elektriker”, while “elektrikere” and “elektrikerene” are plural versions of the same word (the word means “electrician”, btw).
Lets compare this to another word, such as “Bus”:
Here the base form is “buss”, while the two other are plural. Lets apply the same rules to all six words:
buss => buss
busser => buss [strips “er”]
bussene => buss [strips “ene”]
elektrikerene => “elektriker” [strips “ene”]
elektrikere => “elektriker” [strips “e”]
So far everything has gone as planned. We’re able to search for ‘elektrikerene’ and get hits that say ‘elektrikere’, just as planned. All is not perfect, though. We’ve forgotten one word, and evil forces will say that I forgot it on purpose:
elektriker => ?
The problem is that “elektriker” (which is the single form of the word) ends in -er. The rule defined for a word in the class of “buss” says that -er should be stripped (and this is correct for the majority of words). The result then becomes:
elektriker => “elektrik” [strips “er”]
elektrikere => “elektriker” [strips “e”]
elektrikerene => “elektriker” [strips “ene”]
As you can see, there’s a mismatch between the form that the plurals gets chopped down to and the singular word.
My solution, while not perfect in any way, simply adds a few more terms so that we’re able to strip all these words down to the same form:
elektriker => “elektrik” [strips “er”]
elektrikere => “elektrik” [strips “ere”]
elektrikerene => “elektrik” [strips “erene”]
I decided to go this route as it’s a lot easier than building a large selection of words where no stemming should be performed. It might give us a few false positives, but the most important part is that it provides the same results for the singular and plural versions of the same word. When the search results differ for such basic items, the user gets a real “WTF” moment, especially when the two plural versions of the word is considered identical.
To solve this problem we’re going to change the Snowball parser and build a new version of the stemmer that we can use in Lucene and Solr.
To generate the Java class that Lucene uses when attempting to stem a phrase (such as the NorwegianStemmer, EnglishStemmer, etc), you’ll need the Snowball distribution. This distribution also includes example stemming algorithms (which have been used to generate the current stemmers in Lucene).
After extracting the file you’ll have a directory named
snowball_code, which contains among other files the
snowball binary and a directory named
algorithms. The algorithms-directory keeps all the different default stemmers, and this is where you’ll find a good starting point for the changes you’re about to do.
But first, we’ll make sure we have the development version of Lucene installed and ready to go.
You can check out the current SVN trunk of Lucene by doing:
svn checkout http://svn.apache.org/repos/asf/lucene/java/trunk lucene/java/trunk
This will give you the bleeding edge version of Lucene available for a bit of toying around. If you decide to build Solr 1.4 from SVN (as we’ll do further down), you do not have to build Lucene 2.9 from SVN – as it already is included pre-built.
If you need to build the complete version of Lucene (and all contribs), you can do that by moving into the Lucene trunk:
cd lucene/java/trunk/ ant dist (this will also create .zip and .tgz distributions)
If you already have Lucene 2.9 (.. or whatever version you’re on when you’re reading this), you can get by with just compiling the snowball contrib to Lucene, from lucene/java/trunk/:
cd contrib/snowball/ ant jar
This will create (if everything works as it should) a file named
lucene-snowball-2.9-dev.jar (.. or another version number, depending on your version of Lucene). The file will be located in a sub directory of the build directory on the root of the lucene checkout (.. and the path will be shown after you’ve run ant jar):
If you got the lucene-snowball-2.9-dev.jar file compiled, things are looking good! Let’s move on getting the bleeding edge version of Solr up and running (if you have an existing Solr version that you’re using and do not want to upgrade, skip the following steps .. but be sure to know what you’re doing .. which coincidentally you also should be knowing if you’re building stuff from SVN as we are. Oh the joy!).
Getting and building Solr from SVN is very straight forward. First, check it out from Subversion:
svn co http://svn.apache.org/repos/asf/lucene/solr/trunk/ solr/trunk/
And then simply build the war file for your favourite container:
cd solr/trunk/ ant dist
Voilá – you should now have a apache-solr-1.4-dev.war (or something similiar) in the build/ directory. You can test that this works by replacing your regular solr installation (.. make a backup first..) and restarting your application server.
Editing the stemmer definition
After extracting the snowball distribution, you’re left with a
snowball_code directory, which contains
algorithms and then
norwegian (in addition to several other stemmer languages). My example here expands the definition used in the norwegian stemmer, but the examples will work with all the included stemmers.
Open up one of the files (I chose the iso-8859-1 version, but I might have to adjust this to work for UTF-8/16 later. I’ll try to post an update in regards to that) and take a look around. The snowball language is interesting, and you can find more information about it at
the Snowball site.
I’ll not include a complete dump of the stemming definition here, but the interesting part (for what we’re attempting to do) is the main_suffix function:
define main_suffix as ( setlimit tomark p1 for ([substring]) among( 'a' 'e' 'ede' 'ande' 'ende' 'ane' 'ene' 'hetene' 'en' 'heten' 'ar' 'er' 'heter' 'as' 'es' 'edes' 'endes' 'enes' 'hetenes' 'ens' 'hetens' 'ers' 'ets' 'et' 'het' 'ast' (delete) 's' (s_ending or ('k' non-v) delete) 'erte' 'ert' (<-'er') ) )
This simply means that for any word ending in any of the suffixes in the three first lines will be deleted (given by the (delete) command behind the definitions). The problem provided our example above is that neither of the lines will capture an "ere" ending or "erene" - which we'll need to actually solve the problem.
We simply add them to the list of defined endings:
among( ... 'hetene' 'en' 'heten' 'ar' 'ere' 'erene' 'eren' ... ... (delete)
I made sure to add the definitions before the shorter versions (such as 'er'), but I'm not sure (.. I don't think) if it actually is required.
Save the file under a new file name so you still have the old stemmers available.
Compiling a New Version of the Snowball Stemmer
After editing and saving your stemmer, it's now time to generate the Java class that Lucene will use to generate it base forms of the words. After extracting the snowball archive, you should have a binary file named
snowball in the
snowball_code directory. If you simply run this file with
snowball_code as your current working directory:
You'll get a list of options that Snowball can accept when generating the stemmer class. We're only going to use three of them:
-j[ava] Tell Snowball that we want to generate a Java class -n[ame] Tell Snowball the name of the class we want generated -o <filename> The filename of the output file. No extension.
So to compile our NorwegianExStemmer from our modified file, we run:
./snowball algorithms/norwegian/stem2_ISO_8859_1.sbl -j -n NorwegianExStemmer -o NorwegianExStemmer
(pardon the excellent file name stem2...). This will give you one new file in the current working directory:
NorwegianExStemmer.java! We've actually built a stemming class! Woohoo! (You may do a few dance moves here. I'll wait.)
We're now going to insert the new class into the Lucene contrib .jar-file.
Rebuild the Lucene JAR Library
Copy the new class file into the version of Lucene you checked out from SVN:
Then we simply have to rebuild the .jar file containing all the stemmers:
/contrib/snowball/ ant jar
This will create
<lucenetrunk>/build/contrib/. You now have a library containing your stemmer (and all the other default stemmers from Lucene)!
The last part is simply getting the updated stemmer library into Solr, and this will be a simple copy and rebuild:
Inserting the new Lucene Library Into Solr
From the build/contrib directory in Lucene, copy the jar file into the lib/ directory of Solr:
Be sure to overwrite any existing files (.. and if you have another version of Lucene in Solr, do a complete rebuild and replace all the Lucene related files in Solr). Rebuild Solr:
Copy the new
apache-solr-1.4-dev.war (check the correct name in the directory yourself) from the
build/ directory in Solr to your application servers home as solr.war (.. if you use another name, use that). This is webapps/ if you're using Tomcat. Remember to back up the old .war file, just to be sure you can restore everything if you've borked something.
Add Your New Stemmer In schema.xml
After compiling and packaging the stemmer, it's time to tell Solr that it should use the newly created stemmer. Remember that a stemmer works both when indexing and querying, so we're going to need to reindex our collection after implementing a new stemmer.
The usual place to add the stemmer is the definition of your text fields under the <analyzer>-sections for index and query (remember to change it BOTH places!!):
Change NorwegianEx into the name of your class (without the Stemmer-part, Lucene adds that for you automagically. After changing both locations (or more if you have custom datatypes and indexing or query steps).
Restart Application Server and Reindex!
If you're using Tomcat as your application server this might simply be (depending on your setup and distribution):
cd /path/to/tomcat/bin ./shutdown.sh ./startup.sh
Please consult the documentation for your application server for information about how to do a proper restart.
After you've restarted the application server, you're going to need to reindex your collection before everything works as planned. You can however check that your stemmer works as you've planned already at this stage. Log into the Solr admin interface, select the extended / advanced query view, enter your query (which should now be stemmed in another way than before), check the "debug" box and submit your search. The resulting XML document will show you the resulting of your query in the parsedquery element.
Download the Generated Stemmer
If you're just looking for an improved stemmer for norwegian words (with the very, very simple changes outlined above, and which might give problems when concerned with UTF-8 (.. please leave a comment if that's the case)), you can simply download NorwegianExStemmer.java. Follow the guide above for adding it to your Lucene / Solr installation.
Please leave a comment if something is confusing or if you want free help. Send me an email if you're looking for a consultant.