Apache Pig utilities to build training corpora for machine learning / NLP out of public Wikipedia and DBpedia dumps.
This fork attempts to add some functionalities:
- Generating statistics from Wikipedia that can be used in Named Entity Recognition and Named Entity Disambiguation.
This project is alpha / experimental code. Features are implemented when needed.
Some preliminary results are available in this blog post:
Building from source
Install Maven and Java JDK >= version 6, then:
$ mvn package
This should download the dependencies, build a jar in the target/ subfolder and run the tests.
The following introduces some sample scripts to demo the User Defined Functions provided by pignlproc for some practical Wikipedia mining tasks.
Those examples demo how to use pig on your local machine on sample files. In production (with complete dumps) you might want to startup a real Hadoop cluster, upload the dumps into HDFS, adjust the above paths to match your setup and remove the '-x local' command line parameter to tell pig to use your Hadoop cluster.
Extracting links from a raw Wikipedia XML dump
You can take example on the extract-links.pig example script:
$ pig -x local \ -p PIGNLPROC_JAR=target/pignlproc-0.1.0-SNAPSHOT.jar \ -p LANG=fr \ -p INPUT=src/test/resources/frwiki-20101103-pages-articles-sample.xml \ -p OUTPUT=/tmp/output \ examples/extract_links.pig
Building a NER training / evaluation corpus from Wikipedia and DBpedia
To achieve this you can run time following scripts (splitted into somewhat independant parts that store intermediate results to avoid recomputing everything from scratch when you can the source files or some parameters.
The first script parses a wikipedia dump and extract occurrences of sentences with outgoing links along with some ordering and positioning information:
$ pig -x local \ -p PIGNLPROC_JAR=target/pignlproc-0.1.0-SNAPSHOT.jar \ -p LANG=en \ -p INPUT=src/test/resources/enwiki-20090902-pages-articles-sample.xml \ -p OUTPUT=workspace \ examples/ner-corpus/01_extract_sentences_with_links.pig
The parser has been measured to run at a processing of 1MB/s on in local mode on a MacBook Pro of 2009.
The second script parses dbpedia dumps assumed to be in the folder /home/ogrisel/data/dbpedia:
$ pig -x local \ -p PIGNLPROC_JAR=target/pignlproc-0.1.0-SNAPSHOT.jar \ -p LANG=en \ -p INPUT=/home/ogrisel/data/dbpedia \ -p OUTPUT=workspace \ examples/ner-corpus/02_dbpedia_article_types.pig
This step should complete in a couple of minutes in local mode.
The third script merges the partial results of the first two scripts and order back the results by grouping the sentences of the same article together to be able to build annotated sentences suitable for OpenNLP for instance:
$ pig -x local \ -p PIGNLPROC_JAR=target/pignlproc-0.1.0-SNAPSHOT.jar \ -p INPUT=workspace \ -p OUTPUT=workspace \ -p LANG=en \ -p TYPE_URI=http://dbpedia.org/ontology/Person \ -p TYPE_NAME=person \ examples/ner-corpus/03bis_filter_join_by_type_and_convert.pig $ head -3 workspace/opennlp_person/part-r-00000 The Table Talk of <START:person> Martin Luther <END> contains the story of a 12-year-old boy who may have been severely autistic . The New Latin word autismus ( English translation autism ) was coined by the Swiss psychiatrist <START:person> Eugen Bleuler <END> in 1910 as he was defining symptoms of schizophrenia . Noted autistic <START:person> Temple Grandin <END> described her inability to understand the social communication of neurotypicals , or people with normal neural development , as leaving her feeling "like an anthropologist on Mars " .
Building a document classification corpus
TODO: Explain howto extract bag of words or ngrams and document frequency features suitable for document classification using a SGD model from Mahout for instance.
Copyright 2010 Nuxeo and contributors:
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.