Jigg

Jigg is a natural language processing pipeline framework on JVM languages (mainly for Scala), which is easy to use and extensible. Using Jigg, one can obtain several linguistic annotations on a given input from POS tagging, parsing, and coreference resolution from command-lines. The main features include:

Easy to install: basic components are included in a distributed single jar, so no need to install;
Similar interface to Stanford CoreNLP;
Extensible: easy to add new component is a pipeline;
Parallel processing: sentence-level annotation is automatically parallelized.

Jigg is distributed under the Apache License, Version 2.0.

The core ideas and software designs are described in detail in our paper.

Install

The easist way to start Jigg is to download the latet release package (current version is 0.8.0), which includes the core Jar file, as well as several model files, such as a Stanford CoreNLP model file.

[Download Jigg 0.8.0]

Or you can get it in the command line:

$ wget https://github.com/mynlp/jigg/releases/download/v-0.8.0/jigg-0.8.0.zip
$ unzip jigg-0.8.0.zip

Enter the directory before running the following examples:

$ cd jigg-0.8.0

If you wish to build your own jar from sources, please read here.

If you wish to use docker, please read here.

Usage

The following command launches Jigg in a shell mode, which parses a given input with Berkeley parser after preprocessing (tokenization and sentence splitting).

$ java -cp "*" jigg.pipeline.Pipeline -annotators "corenlp[tokenize,ssplit],berkeleyparser"
[main] INFO edu.stanford.nlp.pipeline.StanfordCoreNLP - Adding annotator tokenize
...
>

Let's write some sentences in a line.

> Hello Jigg! This is the first sentence.
<root>
  <document id="d0">
    <sentences>
      <sentence id="s0" characterOffsetBegin="0" characterOffsetEnd="11">
        Hello Jigg!
        <tokens annotators="corenlp berkeleyparser">
          <token pos="UH" characterOffsetEnd="5" characterOffsetBegin="0" id="t0" form="Hello"/>
          <token pos="PRP$" characterOffsetEnd="10" characterOffsetBegin="6" id="t1" form="Jigg"/>
          <token pos="." characterOffsetEnd="11" characterOffsetBegin="10" id="t2" form="!"/>
        </tokens>
        <parse annotators="berkeleyparser" root="s0_berksp0">
          <span id="s0_berksp0" symbol="INTJ" children="t0 t1 t2"/>
        </parse>
      </sentence>
      <sentence id="s1" characterOffsetBegin="12" characterOffsetEnd="39">
        This is the first sentence.
        <tokens annotators="corenlp berkeleyparser">
          <token pos="DT" characterOffsetEnd="4" characterOffsetBegin="0" id="t3" form="This"/>
          <token pos="VBZ" characterOffsetEnd="7" characterOffsetBegin="5" id="t4" form="is"/>
          <token pos="DT" characterOffsetEnd="11" characterOffsetBegin="8" id="t5" form="the"/>
          <token pos="JJ" characterOffsetEnd="17" characterOffsetBegin="12" id="t6" form="first"/>
          <token pos="NN" characterOffsetEnd="26" characterOffsetBegin="18" id="t7" form="sentence"/>
          <token pos="." characterOffsetEnd="27" characterOffsetBegin="26" id="t8" form="."/>
        </tokens>
        <parse annotators="berkeleyparser" root="s1_berksp0">
          <span id="s1_berksp0" symbol="S" children="s1_berksp1 s1_berksp2 t8"/>
          <span id="s1_berksp1" symbol="NP" children="t3"/>
          <span id="s1_berksp2" symbol="VP" children="t4 s1_berksp3"/>
          <span id="s1_berksp3" symbol="NP" children="t5 t6 t7"/>
        </parse>
      </sentence>
    </sentences>
  </document>
</root>
>

The default output format of Jigg is XML, but it also supports JSON (check -outputFormat option below). One can see that Jigg automatically detects sentence boundaries (there are two sentences), and performs tokenization (e.g, period . is recognized as a single word), on which parse tree (<parse>) is built.

In Jigg, each NLP tool such as corenlp (Stanford CoreNLP) or berkeleyparser (Berkeley parser) is called an annotator. Jigg helps to construct easily a NLP pipeline by combining several annotators. In the example above, the pipeline is constructed by combining Stanford CoreNLP (which performs tokenization and sentence-splitting) and Berkeley parser (which performs parsing on tokenized sentences).

Command-line usage

Basic usage is described in the help message:

$ java -cp "*" jigg.pipeline.Pipeline -help
Usage:
  outputFormat                   < str>: Output format, [xml/json]. Default value is 'xml'. [xml]
  annotators                     < str>: List of annotator names, e.g., corenlp[tokenize,ssplit],berkeleyparser (required) [ssplit,kuromoji,jaccg]
  checkRequirement               < str>: Check requirement, [true/false/warn]. Default value is 'true'. [true]
  file                           < str>: Input file; if omitted, read from stdin []
  props                          < str>: Property file []
  nThreads                       < int>: Number of threads for parallel annotation (use all if <= 0) [-1]
  output                         < str>: Output file; if omitted, `file`.xml is used. Gzipped if suffix is .gz. If JSON mode is selected, suffix is .json []
  customAnnotatorClass           < str>: You can add an abbreviation for a custom annotator class with "-customAnnotatorClass.xxx path.package" []
  help                           < str>: Print this message and descriptions of specified annotators, e.g., -help ssplit,mecab [true]

Currently the annotators listed below are installed. See the detail of each annotator with "-help annotator_name".

  mecab, ssplit, jaccg, cabocha, berkeleyparser, spaceTokenize, kuromoji, syntaxnetpos, dsplit, knp, corenlp, knpDoc, juman, syntaxnetparse, syntaxnet

Some annotators, such as mecab, jaccg, kuromoji, etc. are specific for Japanese processing. As shown here, more specific description for each annotator is described by giving argument to -help option:

$ java -cp "*" jigg.pipeline.Pipeline -help berkeleyparser
...
berkeleyparser:
  requires                             : [Tokenize]
  requirementsSatisfied                : [POS, Parse]

  berkeleyparser.variational     <bool>: Use variational rule score approximation instead of max-rule (Default: false) [false]
  berkeleyparser.grFileName      < str>: Grammar file []
  berkeleyparser.accurate        <bool>: Set thresholds for accuracy. (Default: set thresholds for efficiency) [false]
  berkeleyparser.usePOS          <bool>: Use annotated POS (by another annotator) [false]
  berkeleyparser.viterbi         <bool>: Compute viterbi derivation instead of max-rule tree (Default: max-rule) [false]

  A wrapper for Berkeley parser. The feature is that this wrapper is implemented to be
  thread-safe. To do this, the wrapper keeps many parser instances (the number can be
  specified by customizing -nThreads).

  The path to the model file can be changed by setting -berkeleyparser.grFileName.

  If -berkeleyparser.usePOS is true, the annotator assumes the POS annotation is already
  performed, and the parser builds a tree based on the assigned POS tags.
  Otherwise, the parser performs joint inference of POS tagging and parsing, which
  is the default behavior.

Python wrapper and server

Jigg can be directly used in a Python (other languages would be supported in future) script. See python directory for details of this usage.

Inspired by the wrapper mechanism of Stanford CoreNLP, Jigg's wrapper is based on the server, which can be instantiated by:

$ java -cp "*" jigg.pipeline.PipelineServer

This will launch the server on your local system. Currently the server only supports POST request. See more detail in the help message by:

$ java -cp "*" jigg.pipeline.PipelineServer -help

An example of the call via curl is:

$ curl --data-urlencode 'annotators=corenlp[tokenize,ssplit]' \
       --data-urlencode 'q=Please annotate me!' \
       'http://localhost:8080/annotate?outputFormat=json'

Now using wget would need many special cares, so I recommend to use curl instead.

Requirements

Here, requires and reqruiementsSatisfied describe the role of this annotator (berkeleyparser). Intuitively, the above description says berkeleyparser requires that the input text is already tokenized (Tokenize), and after the annotation, part-of-speech tags (POS) and parse tree (Parse) are annotated on each sentence.

Jigg checks with these kinds of information whether the given pipeline can be performed safely. For example, the following command will be failed:

$ java -cp "*" jigg.pipeline.Pipeline -annotators berkeleyparser
annotator berkeleyparser requires Tokenize
  annotators                     < str>: List of annotator names, e.g., corenlp[tokenize,ssplit],berkeleyparser (required) [berkeleyparser]

The error message says tokenize should be performed before running berkeleyparser.

Parallel processing

In the help message above, we can see that berkeleyparser is implemented to be thread-safe. This means we can run Berkeley parser in parallel, which is not supported in the original software. The most of supported annotators in Jigg are implemented as thread-safe, meaning that annotation can be very efficient in a multi-core environment.

To perform parallel annotation, first prepare an input document (whatever you want to analyze).

$ head input.txt
John Blair & Co. is close to an agreement to sell its TV station advertising representation operation and program production unit to an investor group led by James H. Rosenfield, a former CBS Inc. executive, industry sources said. Industry sources put the value of the proposed acquisition at more than $100 million. John Blair was acquired last year by Reliance Capital Group Inc., which has been divesting itself of John Blair's major assets. ...

Then run Jigg as follows:

$ java -cp "*" jigg.pipeline.Pipeline -annotators "corenlp[tokenize,ssplit],berkeleyparser" -file input.txt

Or you can run Jigg in pipe:

$ cat input.txt | java -cp "*" jigg.pipeline.Pipeline -annotators "corenlp[tokenize,ssplit],berkeleyparser" > output.xml

Parallelization can be prohibited by giving -nThreads 1 option:

$ cat input.txt | java -cp "*" jigg.pipeline.Pipeline -annotators "corenlp[tokenize,ssplit],berkeleyparser" -nThreads 1 > output.xml

By default, Jigg tries to use as many threads as the machine can use. On my laptop (with 4 cores), when annotating about 1000 sentences, annotation with -nThreads 1 takes about 154 seconds, which is reduced to 79 seconds with parallel annotation.

You can also customize the number of threads for each annotator separately. For example, the following restrictes the number of threads of berkeleyparser to 2, while allow the corenlp to use 4 threads.

$ cat input.txt | java -cp "*" jigg.pipeline.Pipeline -annotators "corenlp[tokenize,ssplit],berkeleyparser" -nThreads 4 -berkeleyparser.nThreads 2 > output.xml

Full pipeline

For English, currently the main components in Jigg are Stanford CoreNLP. While many components of Stanford CoreNLP requries a model file, it is included in the directory of the latest Jigg (above link) so no need to download a model by yourself.

The following pipeline is one of full CoreNLP pipeline toward coreference resolution.

$ java -cp "*" jigg.pipeline.Pipeline -annotators "corenlp[tokenize,ssplit,parse,lemma,ner,dcoref]"

This is the usage of Jigg just as a wrapper of Stanford CoreNLP, which may not be interesting. More interesting example is to insert the Berkeley parser into a pipeline of Stanford CoreNLP:

$ java -cp "*" jigg.pipeline.Pipeline -annotators "corenlp[tokenize,ssplit],berkeleyparser,corenlp[lemma,ner,dcoref]"

This command replaces the parser component in a CoreNLP pipeline with Berkeley parser. Jigg alleviates to include a NLP tool into a pipeline. As such, the goal of Jigg is to provide a platform on which a user can freely connect the tools to construct several NLP pipelines.

Programmatic usage

Jigg pipeline can also be incorporated another Java or Scala project. The easiest way to do this is add a dependency to Maven.

In Scala, add the following line in the project build.sbt.

libraryDependencies += "com.github.mynlp" % "jigg" % "0.8.0"

In Java, add the following lines on pom.xml:

<dependencies>
  <dependency>
    <groupId>com.github.mynlp</groupId>
    <artifactId>jigg</artifactId>
    <version>0.8.0</version>
  </dependency>
</dependencies>

Jigg is written in Scala, so Scala is the most preferable choice for a programmatic usage. Jigg provides a very similar interface to Stanford CoreNLP:

import jigg.pipeline.Pipeline
import java.util.Properties
import scala.xml.Node

// The behavior of pipeline can be customized with Properties object, which consists of the same options used in command-line usages.
val props = new Properties

props.setProperty("annotators", "corenlp[tokenize,ssplit],berkeleyparser,corenlp[lemma,ner,dcoref]")

// The path to the model to the Berkeley parser may be necessary.
props.setProperty("berkeleyparser.grFileName", "/path/to/eng_sm6.gr")

// Pipeline is the main class, which eats Properties object.
val pipeline = new Pipeline(props)

// Set the input text to be analyzed here.
val text: String = ...

// Get the annotation result in Scala's XML object (Node).
val annotation: Node = pipeline.annotate(text)

The annotation result is obtained in Scala XML object, on which elements can be searched intuitively with expressions similar to X-path. The followings are an example:

val sentences: Seq[Node] = annotation \\ "sentence" // Get all <sentence> elements.
for (sentence <- sentences) { // for each sentence
  val tokens = sentence \\ "token"  // get all tokens
  val nes = sentence \\ "NE"        // get all named entities
  for (ne <- nes) {
    val tokenIds = ne \@ "tokens"   // get the "tokens" attribute in a NE.

    val neTokens = tokenIds map { id =>
      tokens.find(_ \@ "id" == id).get \@ "form" // get surface form of each token consisting the NE
    }
    println(neTokens mkString " ")  // print the detected NE
  }
}

On the result XML, all annotated elements (e.g., sentence, token, and NE) are assigned unique ids. So element search is basically based on these ids.

Build your own jar (advanced)

$ git clone git@github.com:mynlp/jigg.git && cd jigg
$ ./bin/sbt assembly

The last command may take about 10 or 20 minutes including setup of Scala and sbt. This generates a self-contained jar on target/jigg-assembly-xxx.jar where xxx is the current version number.

Although this assembled jar is rather self-contained, several model files, including the model files for berkeleyparser and jaccg are missing. These can be obtained by:

$ wget https://github.com/mynlp/jigg-models/raw/master/jigg-models.jar

and including it in the class path:

$ java -cp "target/jigg-assembly-xxx.jar:jigg-models.jar" jigg.pipeline.Pipeline -annotators ...

Note that in this usage, the CoreNLP models should also be downloaded from the official homepage manually and included in your class path.

Use docker

To install docker, follow the instruction.

To build and run PipelineServer container:

$ git clone --depth 1 https://github.com/mynlp/jigg.git && cd jigg
$ curl -SL https://github.com/mynlp/jigg-models/raw/master/jigg-models.jar -o jar/jigg-models.jar
$ time docker-compose build
$ docker-compose up -d

An example of the call via curl is:

$ curl --data-urlencode 'annotators=ssplit,kuromoji,jaccg' \
         --data-urlencode 'q=テレビで自転車で走っている少女を見た!' \
         'http://localhost:8080/annotate?outputFormat=xml'

Supported annotators

Supported annotators in the current environment can be listed with the help command. To see more details on each annotator, try -help annotator_name (see also Command-line usage).

Citing in papers

If you use Jigg in research publications, please cite:

Hiroshi Noji and Yusuke Miayo. 2016. Jigg: A Framework for an Easy Natural Language Processing Pipeline. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics: System Demonstrations.

Acknowledgements

Following sample files of SsplitKerasAnnotator/BunsetsuKerasAnnotator is generated by using BCCWJ corpus.

Trained model file
- src/test/resources/data/keras/ssplit_model.h5
- src/test/resources/data/keras/bunsetsu_model.h5
Lookup table file
- src/test/resources/data/keras/jpnLookupCharacter.json
- src/test/resources/data/keras/jpnLookupWords.json

Release note

0.8.0: Many bug fixes; kuromoji is modulated; CoreNLP is upgraded to 3.9.1; support benepar and stanfordtypeddep (combine them to obtain state-of-the-art constituency and dependency parsers!; see -help benear and -help stanfordtypeddep).
0.7.2: Many improvements around CCG parsers, including K-best outputs of depccg and easyccg. Support of udpipe.
0.7.1: Bug fixes; Docker for Jigg server; annotators for CCG parsers (candc, easyccg, and depccg)
0.7.0: Support CoreNLP 3.7.0, server mode, several improvements including support of xml/json inputs.
0.6.1: Bug fixes.
0.6.1: New annotators (syntaxnet, coref in corenlp, etc); JSON output (still incomplete); bug fixes.
0.6.0: The initial official release.

Name		Name	Last commit message	Last commit date
Latest commit History 853 Commits
.checker		.checker
bin		bin
dockers		dockers
jar		jar
project		project
python		python
script		script
src		src
.dockerignore		.dockerignore
.gitignore		.gitignore
.travis.yml		.travis.yml
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
build.sbt		build.sbt
docker-compose.yml		docker-compose.yml

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Jigg

Install

Usage

Command-line usage

Python wrapper and server

Requirements

Parallel processing

Full pipeline

Programmatic usage

Build your own jar (advanced)

Use docker

Supported annotators

Citing in papers

Acknowledgements

Release note

About

Releases 7

Packages

Contributors 8

Languages

License

mynlp/jigg

Folders and files

Latest commit

History

Repository files navigation

Jigg

Install

Usage

Command-line usage

Python wrapper and server

Requirements

Parallel processing

Full pipeline

Programmatic usage

Build your own jar (advanced)

Use docker

Supported annotators

Citing in papers

Acknowledgements

Release note

About

Resources

License

Stars

Watchers

Forks

Releases 7

Packages 0

Contributors 8

Languages

Packages