From text to facts: Relation Extraction on German Websites

This repository holds a collection of three Open Information Extraction approaches for the German language:

ReVerb This system is based on the ReVerb system available under https://github.com/knowitall/reverb, which is published under the ReVerb Software License Agreement. We adapted the system to the German language.

German ReVerb German ReVerb is an extension of the previously mentioned system also based on the ReVerb system available under https://github.com/knowitall/reverb. This system is specialized to the German language by adding rules dealing with specific aspects of the German language to the system.

Dep ConIE This system is a new approach based on dependency parsing. It yields the best results.

Requirements

To successfully run one of the above systems, you need

• Java JDK 1.8 or later
• Maven 3.0.0 or later

Set Up

To set up the system locally, you have to

1. Download and install TreeTagger (http://www.cis.uni-muenchen.de/~schmid/tools/TreeTagger/). Set the environment variable 'TREETAGGER_HOME' to the installation directory.
2. Put the following models to the resource folder (src/main/resources):

• de-sent.bin, de-pos-maxent.bin from http://opennlp.sourceforge.net/models-1.5/
• lemma-ger-3.6.model from the archive ger-tagger+lemmatizer+morphology+graph-based-3.6+.tgz, which can be downloaded from https://code.google.com/archive/p/mate-tools/downloads

3. Download Vollformenlexikon 2011-07-22 from http://www.danielnaber.de/morphologie/ and put the extracted morphy-export-20110722.xml-file to the resource folder (src/main/resources)
4. Download and install ParZu (https://github.com/rsennrich/ParZu). Set the environment variable 'PARZU_HOME' to the installation directory.
5. Download the compact version of Zmorge from http://kitt.ifi.uzh.ch/kitt/zmorge/ and put it to the resource folder (src/main/resources).

Running the system

To run the system from the command line execute the following command:

./relation_extraction -e <system> -s <sentence> [<args>]`

 -a,--argCand <arg>       Consider also arguments, which are not directly connected to the relation phrase. 
                          (value: false|true, default: false, extractor: Dep ConIE)
 -c,--constraints <arg>   Use the lexical and syntatic constraint. 
                          (value: false|true, default: true, extractor: ReVerb)
 -e,--extractor <arg>     The extractor to use: 'ReVerb', 'German ReVerb', or 'Dep ConIE'.
 -f,--minFreq <arg>       Each relation must have at least this number of distinct arguments in a large corpus. 
                          (value: integer, default: 20/20/0, extractor: all)
 -g,--greedyArg <arg>     Extract relations from sentences with more than two arguments.
                          (value: false|true, default: false, extractor: Dep ConIE)
 -m,--morph <arg>         Use a morphological lexicon to determine the case of the first argument .
                          (value: false|true, default: true, extractor: German ReVerb)
 -p,--pronouns <arg>      Extract relations, which have just a pronoun as argument.
                          (value: false|true, default: false, extractor: Dep ConIE)
 -r,--refVerbs <arg>      Add reflexive verbs to the relation phrase.
                          (value: false|true, default: true, extractor: German ReVerb)
 -s,--sent <arg>          The sentence to extract the relations from. (value: string)
 -u,--subSent <arg>       Extract relations only from parts of the sentences, which do not contain any comma or conjunction.
                          (value: false|true, default: true, extractor: German ReVerb)
 -v,--combVerbs <arg>     Combine the individual parts of a verb.
                          (value: false|true, default: true, extractor: German ReVerb)

This command extracts the relations from the given sentence using the named extractor with the given arguments. For example, you can execute ./relation_extraction -e "Dep ConIE" -s "Dep ConIE ist ein Relationsextraktionssystem." to obtain the relation (Dep ConIE # ist # ein Relationsextraktionssystem).

To include our system as a library in your own project, please take a look at the example class RelationExtractionExample in the src/main/java/de/hpi/examples directory.

When running code that calls our system, make sure to increase the Java Virtual Machine heap size by passing the argument -Xmx2g to java. Our system loads multiple models into memory. Therefore, not increasing the heap size will slow down the system.