Levenshtein transducers accept a query term and return all terms in a dictionary that are within n spelling errors away from it. They constitute a highly-efficient (space and time) class of spelling correctors that work very well when you do not require context while making suggestions. Forget about performing a linear scan over your dictionary to find all terms that are sufficiently-close to the user's query, using a quadratic implementation of the Levenshtein distance or Damerau-Levenshtein distance, these babies find all the terms from your dictionary in linear time on the length of the query term (not on the size of the dictionary, on the length of the query term).
If you need context, then take the candidates generated by the transducer as a starting place, and plug them into whatever model you're using for context (such as by selecting the sequence of terms that have the greatest probability of appearing together).
For a quick demonstration, please visit the Github Page, here. There's also a command-line interface, liblevenshtein-java-cli. Please see its README.md for acquisition and usage information.
The library is currently written in Java, CoffeeScript, and JavaScript, but I will be porting it to other languages, soon. If you have a specific language you would like to see it in, or package-management system you would like it deployed to, let me know.
| Branch | Description |
|---|---|
| master | Latest, development source |
| release | Latest, release source |
| release-3.x | Latest, release source for version 3.x |
| release-2.x | Latest, release source for version 2.x |
Issues are managed on waffle.io. Below you will find a graph on the rate at which I've been closing them.
Please visit Bountysource to pledge your support for ongoing issues.
When it comes to documentation, you have several options:
liblevenshtein has been developed against Java ≥ 1.8. It will not work with prior versions.
Add a Maven dependency on Artifactory. For example, in a
Gradle project, you would modify your repositories as follows:
repositories {
maven {
url 'https://oss.jfrog.org/artifactory/oss-release-local'
}
}Add a Maven dependency on one of the following:
<dependency>
<groupId>com.github.universal-automata</groupId>
<artifactId>liblevenshtein</artifactId>
<version>3.0.0</version>
</dependency>'com.github.universal-automata:liblevenshtein:jar:3.0.0'<dependency org="com.github.universal-automata" name="liblevenshtein" rev="3.0.0" />@Grapes(
@Grab(group='com.github.universal-automata', module='liblevenshtein', version='3.0.0')
)compile 'com.github.universal-automata:liblevenshtein:3.0.0'libraryDependencies += "com.github.universal-automata" % "liblevenshtein" % "3.0.0"[com.github.universal-automata/liblevenshtein "3.0.0"]% git clone --progress [email protected]:universal-automata/liblevenshtein-java.git
Cloning into 'liblevenshtein-java'...
remote: Counting objects: 8117, done.
remote: Compressing objects: 100% (472/472), done.
remote: Total 8117 (delta 352), reused 0 (delta 0), pack-reused 7619
Receiving objects: 100% (8117/8117), 5.52 MiB | 289.00 KiB/s, done.
Resolving deltas: 100% (5366/5366), done.
Checking connectivity... done.
% cd liblevenshtein-java
% git pull --progress
Already up-to-date.
% git fetch --progress --tags
% git checkout --progress 3.0.0
Note: checking out '3.0.0'.
You are in 'detached HEAD' state. You can look around, make experimental
changes and commit them, and you can discard any commits you make in this
state without impacting any branches by performing another checkout.
If you want to create a new branch to retain commits you create, you may
do so (now or later) by using -b with the checkout command again. Example:
git checkout -b <new-branch-name>
HEAD is now at 4f0f172... pushd and popd silently
% git submodule init
% git submodule update
Let's say you have the following content in a plain text file called, top-20-most-common-english-words.txt (note that the file has one term per line):
the
be
to
of
and
a
in
that
have
I
it
for
not
on
with
he
as
you
do
at
The following provides you a way to query its content:
import java.io.InputStream;
import java.io.OutputStream;
import java.nio.file.Files;
import java.nio.file.Path;
import java.nio.file.Paths;
import com.github.liblevenshtein.collection.dictionary.SortedDawg;
import com.github.liblevenshtein.serialization.PlainTextSerializer;
import com.github.liblevenshtein.serialization.ProtobufSerializer;
import com.github.liblevenshtein.serialization.Serializer;
import com.github.liblevenshtein.transducer.Algorithm;
import com.github.liblevenshtein.transducer.Candidate;
import com.github.liblevenshtein.transducer.ITransducer;
import com.github.liblevenshtein.transducer.factory.TransducerBuilder;
// ...
final SortedDawg dictionary;
final Path dictionaryPath =
Paths.get("/path/to/top-20-most-common-english-words.txt");
try (final InputStream stream = Files.newInputStream(dictionaryPath)) {
// The PlainTextSerializer constructor accepts an optional boolean specifying
// whether the dictionary is already sorted lexicographically, in ascending
// order. If it is sorted, then passing true will optimize the construction
// of the dictionary; you may pass false whether the dictionary is sorted or
// not (this is the default and safest behavior if you don't know whether the
// dictionary is sorted).
final Serializer serializer = new PlainTextSerializer(false);
dictionary = serializer.deserialize(SortedDawg.class, stream);
}
final ITransducer<Candidate> transducer = new TransducerBuilder()
.dictionary(dictionary)
.algorithm(Algorithm.TRANSPOSITION)
.defaultMaxDistance(2)
.includeDistance(true)
.build();
for (final String queryTerm : new String[] {"foo", "bar"}) {
System.out.println(
"+-------------------------------------------------------------------------------");
System.out.printf("| Spelling Candidates for Query Term: \"%s\"%n", queryTerm);
System.out.println(
"+-------------------------------------------------------------------------------");
for (final Candidate candidate : transducer.transduce(queryTerm)) {
System.out.printf("| d(\"%s\", \"%s\") = [%d]%n",
queryTerm,
candidate.term(),
candidate.distance());
}
}
// +-------------------------------------------------------------------------------
// | Spelling Candidates for Query Term: "foo"
// +-------------------------------------------------------------------------------
// | d("foo", "do") = [2]
// | d("foo", "of") = [2]
// | d("foo", "on") = [2]
// | d("foo", "to") = [2]
// | d("foo", "for") = [1]
// | d("foo", "not") = [2]
// | d("foo", "you") = [2]
// +-------------------------------------------------------------------------------
// | Spelling Candidates for Query Term: "bar"
// +-------------------------------------------------------------------------------
// | d("bar", "a") = [2]
// | d("bar", "as") = [2]
// | d("bar", "at") = [2]
// | d("bar", "be") = [2]
// | d("bar", "for") = [2]
// ...If you want to serialize your dictionary to a format that's easy to read later, do the following:
final Path serializedDictionaryPath =
Paths.get("/path/to/top-20-most-common-english-words.protobuf.bytes");
try (final OutputStream stream = Files.newOutputStream(serializedDictionaryPath)) {
final Serializer serializer = new ProtobufSerializer();
serializer.serialize(dictionary, stream);
}Then, you can read the dictionary later, in much the same way you read the plain text version:
final SortedDawg deserializedDictionary;
try (final InputStream stream = Files.newInputStream(serializedDictionaryPath)) {
final Serializer serializer = new ProtobufSerializer();
deserializedDictionary = serializer.deserialize(SortedDawg.class, stream);
}Serialization is not restricted to dictionaries, you may also (de)serialize transducers.
Please see the wiki for more details.
This library is based largely on the work of Stoyan Mihov, Klaus Schulz, and Petar Nikolaev Mitankin: Fast String Correction with Levenshtein-Automata. For more information, please see the wiki.
liblevenshtein-java is maintained by@dylon ([email protected])