Skip to content

Latest commit

 

History

History
263 lines (190 loc) · 10.1 KB

README.md

File metadata and controls

263 lines (190 loc) · 10.1 KB
Raw

Monnet Translate

Monnet translation system for translation of ontologies based on statistical machine translation and ontological semantics. The system is composed of the following modules

  • core: Interfaces used between components
  • controller: Central controller for translations
  • chunker: Input analysis and tokenization
  • decomposer: Compound noun analysis for German and Dutch
  • sqlpt: Phrase table extraction from SQL databases
  • sources: Additional candidate extraction from lexico-semantic resources
  • feat: Featurizers for semantic analyzers
  • opm: Observed phrase model for partial translation
  • phrasal: Decoder forked from Stanford Phrasal
  • phrasal.test: Integration tests of Phrasal
  • langmodel: Language models
  • fidel: Alternative decoder
  • evalmetrics: Evaluation metrics
  • quality: Quality estimation
  • jmert: Ranking SVM parameter tuning
  • evaluation: Command-line evaluation tools
  • controller.web: REST web service for translation

Configuring the system

Standard configurations can be found here

com.mysql.jdbc.cfg

This is the file that specifies the details of the MySQL database the format is as follows:

username=monnet
password=???
database=acquis
server=localhost

eu.monnetproject.translation.langmodel.cfg

This file contains the location of the language models to be used (in ARPA format). The files are of the form lang=file, e.g.,

es=/home/shared/es/lm
de=/home/shared/de/lm
en=/home/shared/en/lm
nl=/home/shared/nl/lm

eu.monnetproject.translation.wts.$src-$trg.cfg

This file contains the weights on the log-linear model. The file is of the form featureName=value, note : must be back-slashed in feature names! There must be a weights file for each translation direction and features not in the weights file will not be used

LinearDistortion=0.12499999999999997
TM\:phi(t|f)=0.12499999999999997
TM\:lex(t|f)=0.12499999999999997
TM\:phi(f|t)=0.12499999999999997
TM\:lex(f|t)=0.12499999999999997
TM\:phrasePenalty=0.12499999999999997
LM=0.12499999999999997
WordPenalty=0.12499999999999997
UnknownWord=0.0
SentenceBoundary=0.0

eu.monnetproject.seals.cfg (Optional)

To enable results to be stored with SEALS specify the name of the SEALS repository as follows:

resourceLocator=http://pinto.dia.fi.upm.es:8080/rrs-web/

Running the translation system

The standard evaluation script is e.m.t.evaluation.Evaluate. It can be invoked using the following command (see Maven FAQ), for the eu.monnetproject.translation.evaluation project folder.

mvn exec:java -Dexec.mainClass="eu.monnetproject.translation.evaluation.Evaluate" -Dexec.args="folder srcLang trgLang"

Where

  • folder: Is the folder containing all ontologies to be translated
  • srcLang: Is the source language (as ISO code, e.g., "en")
  • trgLang: Is the target language (as ISO code, e.g., "es")

If run like this all results will be printed to STDOUT

To save the results the parameter -runName must be specified, e.g.,

mvn exec:java -Dexec.mainClass="eu.monnetproject.translation.evaluation.Evaluate" -Dexec.args="-runName example MultilingualOntologies/ en es"

The result will be written as

  • results/runName_srcLang_trgLang_YYYY-MM-DD_hh:mm.xml

Where YYYY-MM-DD_hh:mm is the time of execution e.g.,

  • results/example_en_es_2012-01-17_14:01.xml

Example output is

<results time="2012-01-16T16:31:27" sourceLanguage="en" targetLanguage="es" executionTime="1049934">
  <ontology uri="http://trajano.us.es/~isabel/EHR/Demographic_RM.owl" labelCount="22">
    <result metric="BLEU" score="0.0"/>
    <result metric="BLEU-2" score="0.0"/>
    <result metric="METEOR" score="0.0"/>
    <result metric="NIST" score="0.0"/>
    <result metric="PER" score="-1.0"/>
    <result metric="TER" score="-1.0"/>
    <result metric="WER" score="-1.0"/>
  </ontology>
  <ontology uri="http://xbrl.iasb.org/taxonomy/2009-04-01/ifrs" labelCount="37">
    <result metric="BLEU" score="0.0"/>
    <result metric="BLEU-2" score="0.05431183395717798"/>
    <result metric="METEOR" score="0.051048875835972445"/>
    <result metric="NIST" score="0.7366279797295097"/>
    <result metric="PER" score="-0.854251012145749"/>
    <result metric="TER" score="-0.9291497975708503"/>
    <result metric="WER" score="-0.9311740890688259"/>
  </ontology>
  ...
  <summary labelCount="311">
    <result metric="BLEU" score="0.0"/>
    <result metric="BLEU-2" score="0.01721014745117195"/>
    <result metric="METEOR" score="0.03802367416584292"/>
    <result metric="NIST" score="0.5087312368265963"/>
    <result metric="PER" score="-0.9775997190472975"/>
    <result metric="TER" score="-0.9961771043049569"/>
    <result metric="WER" score="-1.0054506062045343"/>
  </summary>
  ...
</results>

Tuning

Weights may be learnt as follows:

mvn exec:java -Dexec.mainClass="eu.monnetproject.translation.controller.evaluate.Tune" \
   -Dexec.args="MultilingualOntologies/ en es nist 5 load/eu.monnetproject.translation.wts.en-es.cfg" \
   -Djmert.method=SVM

This optimizes NIST using 5 iterations:

Where jmert.method is one of

  • DUMMY: Does nothing
  • MERT: Och-Ney linear search
  • OLS: Linear regression
  • SVM: Ranking SVM (default)

Cross-folded evaluation

You may also run cross-folded evaluation, which divides the ontologies into n even sets, and then uses n-1 of these sets to train and tests on the held out test. The command is as follows

mvn exec:java -Dexec.mainClass="eu.monnetproject.translation.evaluation.EvaluateNFolds" \
     -Dexec.args="-runName example MultilingualOntologies/ en es 10 5 nist"

This runs using 10 folds and tuning using 5 iterations on the NIST metric

Single-ontology translation

For debugging the translation can read a single ontology as follows

mvn exec:java -Dexec.mainClass="eu.monnetproject.translation.controller.Translate" \
     -Dexec.args="ontology.rdf en translation.rdf"

Interactive mode

For debugging, the decoder may be used directly by reading from STDIN

mvn exec:java -Dexec.mainClass="eu.monnetproject.translation.controller.RTPL" \
  -Dexec.args="en es 1"

1 is number of candidates to print.

Web interface

The web interface is compiled at eu.monnetproject.translation.controller.web, and can be deployed by uploading the WAR file to Tomcat or Jetty.

A REST API is provided with support for the following formats

  • SKOS
  • SKOS-XL
  • RDFS
  • Lemon

All can be provided in both RDF-XML as well as turtle format.

Input format is determined by the HTTP Content-Type header value.

The translation service accepts the following parameters which must be offered as a POST multipart request:

  • ontology (Required)
    • The source file containing the source ontology and the lexica
  • target-language (Required)
    • IETF code, eg. "nl"
  • scope
    • Comma separated list of URI's (by default all URI's in the ontology will be translated)
  • name-prefix
    • The name prefix for URI's of new entities created by SKOS-XL or Lemon (defaults to blank nodes)
  • custom-label
    • Comma separated list of URI's for custom label properties
  • n-best
    • The maximum amount of candidates to return for a single label (defaults to 1)
  • accept-vocabularies
  • include-source
    • Include the source ontology and lexicon in the result. If true all existing candidates that were not created by this translation services will be untouched. (defaults to true)
  • esimate-confidence
    • Provide estimates of confidence of translation (Lemon Metadata only) (defaults to false)
  • speed
    • Optimize for speed not quality of translations, one of {fast,normal} (defaults to fast)

Implementation specific features:

  • Automatically infer the language of the ontology to be translated

Error conditions:

  • The ontology contains labels without language tags
  • The ontology contains no labels

The translation service will provide a maximum of n candidates for each label in the source language. The translation service can use translation candidates that are already present in the source lexicon but only those that are marked as accepted. Candidates with a draft status should be ignored.

Metadata included per translation candidate

  • confidence
  • logprob
    • Logaritmic probability of the translation
  • translation-service-identifier
    • Identifier of the translation service, e.g. Monnet
  • source
    • Source of the translation (used by the translation service)
  • status
    • Either accepted, for-review, rejected. If no status is available it's presumed to be accepted. (Not sure if we really return this?)

Testing with CURL

curl -F "[email protected]" -F "target-language=de" http://monnet01.sindice.net:8080/translate/

Using from Java

HttpClient httpClient = new HttpClient();
PostMethod method = new PostMethod(config.getProperty("http://monnet01.sindice.net:8888/translate"));
HttpClientParams params = new HttpClientParams();
int timeout = 60000;
String ontologyAsString = "<your ontology goes in here>";
params.setSoTimeout(timeout);
  Part[] parts = {
    new StringPart("ontology", ontologyAsString),
    new StringPart("n-best", "5"),
    new StringPart("target-language", "nl"),
    new StringPart("estimate-confidence","false")
};
method.setRequestEntity(new MultipartRequestEntity(parts, method.getParams()));
httpClient.setParams(params);
httpClient.executeMethod(method);
String result = method.getResponseBodyAsString();