RelEx Semantic Relation Extractor
Version 1.6.3 circa 2016
RelEx is a dependency parser for the English language. It extracts dependency relations from Link Grammar, and adds some shallow semantic analysis. The primary use of RelEx is as a language input front-end to the OpenCog artificial general intelligence system.
There are multiple inter-related parts to RelEx. The core component extracts the dependency relationships. An experimental module provides some simple anaphora resolution suggestions. Output is provided in various formats, including one format suitable for later batch post-processing, another format suitable for input to OpenCog, and an W3C OWL format. There are also a small assortment of perl scripts for cleaning up web and wiki pages, &c.
The main RelEx website is at
It provides an overview of the project, as well as detailed documentation.
The source code management system is at
Source tarballs may be downloaded from either of two locations:
Build and install of the core package is discussed below.
Installing on Ubuntu/Debian
An installation script for Ubuntu/Debian is provided in the install-scripts directory.
Install and run via Docker
The easiest way to build and run RelEx is with Docker. The Docker system allows sandboxed containers to be easily created and deployed; the typical use of a container is to run some server. See the http://www.docker.io website for more info and tutorials.
To use docker, simply say:
$ docker build -t mine/relex . $ docker run -i -t -p 3333:3333 -w /home/Downloads/relex-master mine/relex /bin/sh plain-text-server.sh
$ docker run -i -t -p 4444:4444 -w /home/Downloads/relex-master mine/relex /bin/sh opencog-server.sh $ docker run -i -t -p 9000:9000 -w /home/Downloads/relex-master mine/relex /bin/sh link-grammar-server.sh
For the first two, simple say:
telnet localhost 3333 This is a test sentence!
The first server just returns a plain-text analysis of the input sentence, while the second returns an opencog-scheme version of the parse.
The raw link-grammar server expects a JSON-formated input, begining
with the 5 letters
text: it returns a JSON-formatted response.
telnet localhost 9000 text:This is a test sentence!
A docker cheat-sheet:
docker ps docker ps -a docker rm docker images docker rmi
Installing on all other systems
For other systems, follow the instructions below. To build and use RelEx, the following packages are required to be installed:
- libgetopt-java (GNU getopt)
- Link Parser
- WordNet 3.0
- JWNL Java wordnet library
- OpenNLP tools (optional, but recommended)
- W3C OWL (optional)
The following packages are required pre-requisites for building RelEx.
Link Grammar Parser. Compile and install the Link Grammar Parser. This parser is described at
and sources are available for download at
Link-grammar version 5.2.1 or later is needed to obtain a variety of required fixes.
The Link Grammar Parser is the underlying engine, providing the core sentence parsing ability.
If the parser is not installed in the default location, be sure to modify
relation-extractor.shand other shell scripts.
GNU getopt. This is a standard command-line option parsing library. For Ubuntu, install the
Wordnet. Wordnet is used by RelEx to provide basic English morphology analysis, such as singular versions of (plural) nouns, base forms (lemmas) of adjectives, adverbs and infinitive forms of verbs.
Download, unpack and install WordNet 3.0. The install directory needs to be specified in
data/wordnet/file_properties.xml, with the
name="dictionary_path"property in this file.
Some typical install locations are:
/opt/WordNet-3.0/datafor RedHat and SuSE
/usr/share/wordnetfor Ubuntu and Debian
C:\Program Files\WordNet\3.0\datafor Windows
relex/Morphy/Morphy.javaclass provides a simple, easy-to-use wrapper around wordnet, providing the needed word morphology info.
didion.jwnl. The didion JWNL is the "Java WordNet Library", and provides the Java programming API to access the wordnet data files. Its home page is at
and can be downloaded from
Verify that the final installed location of
jwnl.jaris correctly specified in the
build.xmlfile. Note that GATE also provides a
jwnl.jar, but the GATE version of
jwnl.jaris not compatible (welcome to java DLL hell).
jwnl.jar: verify the file permisions! Be sure to issue the following command:
chmod 644 jwnl.jar, as otherwise, you'll get strange "java cannot unzip jar" error messages.
Apache Commons Logging. The JWNL package requires that the Apache commons logging jar file be installed. In Debian/Ubuntu, this is supplied by the
libcommons-logging-javapackage. In RedHat/CentOS systems, the package name is
The following packages are optional. If they are found, then additional parts of RelEx will be built, enabling additional function.
OpenNLP. RelEx uses OpenNLP for sentence detection, giving RelEx the ability to find sentence boundaries in free text. If OpenNLP is not found, then the (far) less accurate
java.text.BreakIteratorclass is used. Although Oracle documentation states that "Sentence boundary analysis allows selection with correct interpretation of periods within numbers and abbreviations", this is patently false, as it incorrectly breaks the sentence "Dr. Smith is late." into two sentences. Thus, OpenNLP is recommended.
If you use Maven, this dependency is already managed.
The OpenNLP home page is at
Download and install OpenNLP tools, and verify that the installed files are correctly identified in both
OpenNLP also requires the installation of maxent from
The OpenNLP package is used solely in corpus/DocSplitter.java, which provides a simple, easy-to-use wrapper for splitting a document into sentences. Replace this file if an alternate sentence detector is desired.
Trove. Some users may require the GNU Trove to enable OpenNLP, although this depends on the JDK installed. GNU Trove is an implementation of the java.util class hierarchy, which may or may not be included in the installed JDK. If needed, download trove from:
Since trove is optimized, using it may improve performance and/or decrease memory usage, as compared to the standard Sun JDK implementation of the java.util hierarchy.
IMPORTANT OpenNLP expects Gnu Trove version 1.0, and will not work with version 2.0 !!
After the above are installed, the relex java code can be built.
The build system uses
ant, and the ant build specifications
build.xml. Simply saying
ant at the command line
should be enough to build. Saying
ant run will run a basic
demo of the system. The
ant test command will run several tests
verifying both regular parsing, and the Stanford-parser compatibility
ant install command will install the ssytem.
Build and install is currently not supported with
Maven. You might get lucky -- if you
know how to use maven, then please edit the project model
fix it so that it works, and send us the patches!
It is assumed that RelEx will be used in one of two different ways. These are in a "batch processing" mode, and a "custom Java development" mode.
In the "batch processing mode", RelEx is run once over a large text, and its output is saved to a file. This output can then be post-processed at a later time, to extract desired info. The goal here is to avoid the heavy CPU overhead of re-parsing a large text over and over. Example post-processing scripts are included (described below).
In the "custom Java development" mode, it is assumed that a capable
Java programmer can write new code to interface RelEx to meet their needs.
A good place to start is to review the workings of the output code in
The standard RelEx demo output is NOT SUITABLE for post-processing. It is meant to be a human-readable example of what the system generates; it does not include all required output. For example, if the same word appears in a sentence twice, the demo output will not distinguish between these two words.
This release of RelEx includes an experimental Stanford-parser
compatibility mode. In this mode, RelEx will generate the same
dependency relations as the Stanford parser. This mode is technically
interesting for comparing output; RelEx is more than three time faster
than the lexicalized (factored) Stanford parser, although it is slower
than the PCFG parser. This is described in greater detail in the file
This release of RelEx includes an optional Penn Treebank style part of speech tagger. The tagger is experimental, and has not been evaluated for accuracy. It is probable that the accuracy is low, primarily because it has not been well tested. Because the tagging is based on the syntactic parse, in principle the accuracy could be very high, once fully debugged.
Several example unix shell scripts and MS Windows batch files are
included to show sample usage. These files (
*.sh in unix, or
in Windows) define the required system properties, classpath and JVM
If there are any ClassNotFound exceptions, please verify the paths and values in these files.
The primary usage example is the
Running this will display:
- The link parser output.
- The detected persons, organizations and locations.
- The dependency relations found.
- Anaphora resolutions.
- Parse ranking info.
- (Optionally) Stanford and Penn Treebank output.
Output is controlled by command-line flags that are set in the shell
-h flag will print a list of all of the available
batch-process.sh script is an example batch processing script.
This script outputs the so-called "compact (cff) format" which captures
the full range of Link Grammar and RelEx output in a format that can be
easily post-processed by other systems (typically by using regex's).
The idea behind the batch processing is that it is costly to parse large quantities of text: thus, it is convenient to parse the text once, save the results, and then perform post-processing at leisure, as needed. Thus, the form of post-processing can be changed at will, without requiring texts to be re-processed over and over again.
This perl script provides an example of post-processing: it converts the "cff" batch output format into OpenCog hypergraphs, which can then be processed by OpenCog.
This script starts a relex server that listens for plain-text input (English sentences) on port 4444. It then parses the text, and returns opencog output on the same socket. This server is meant to serve the OpenCog chatbot directly; it is not intended for general, manual use.
doc-splitter.sh file is a simple command-line utility to reformat
a free-form text into sentences, one per line.
Ad-hoc script to scrub Wikipedia xml dumps, outputting only valid English-language sentences. This script removes wiki markup, URL's tables, images, & etc. It currently seems to be pretty darned bullet-proof, although it might handle multi-line refs incorrectly.
If you built RelEx with Maven, these scripts can be used.
They accept additional arguments to be passed to
target/relex/bin/relexd, which runs
java relex.Server ...
target/relex/bin/relexd-relex, which runs
java relex.Server --relex ...
target/relex/bin/relexd-link, which runs
relex.Server --link --relex --verbose ...
Using RelEx in custom code
The primary output of RelEx is the set of semantic relationships of a
sentence. To obtain the list of these relationships, make a copy of
src/java/relex/output/SimpleView.java, and customize it to provide
the relationships that you wish, in the format that you wish.
src/java/relex/RelationExtractor.java should be considered
to be a large example program illustrating all of the various features
of RelEx. For custom applications, this class should be copied and
modified as desired to fit the application.
Speed test results
Performance comparison of RelEx-1.2.0 vs. Stanford-1.6.1, run 11 Oct 2009. Test corpus: first 150 sentences (including preface boilerplate) from Project Gutenberg "Pride and Prejudice". Due to differences in sentence detection, Stanford and RelEx disagree on the sentence count. Due to differences in counting punctuation, the splitting of possessives and contractions, the two disagree on the word count as well.
Since these tests were run, the performance of link-grammar has been improved by a factor of 2x-3x. This update should have a significant effect on relex speeds.
The unix command
wc counts 2609 words in 148 sentences, for
2609/148 = 17.6 words/sent.
Stanford, w/ englishFactored.ser.gz , w/unix
real 10m4.882s user 10m1.974s sys 0m4.208s
Actual: 2609/605= 4.31 words/sec
Stanford, w/ englishPCFG.ser.gz , w/unix
real 2m21.690s user 2m23.165s sys 0m1.056s
Actual: 2609/143 = 18.24 words/sec
Stanford, w/ wsjFactored.ser.gz , w/unix
real 10m5.972s user 10m3.802s sys 0m4.516s
Stanford, w/ wsjPCFG.ser.gz , w/unix
real 2m11.154s user 2m14.312s sys 0m1.144s
Actual: 2609/134 = 19.47 words/sec
real 2m59.739s user 2m36.342s sys 0m22.137s
Actual: 2609/180 = 14.50 words/sec
Ratio: Stanford-englishFactored/RelEx = 605sec/180sc = 3.36x (faster) Ratio: Stanford-wsjFactored/RelEx = 606sec/180sec = 3.36x (faster) Ratio: Stanford-englishPCFG/RelEx = 143sec/180sec = 0.79x (slower) Ratio: Stanford-wsjPCFG/RelEx = 134sec/180sec = 0.74x (slower)
TODO - Comparatives
RelEx is buggy when it comes to handling comparative sentences. This needs fixing.
TODO - Wordnet Install
Windows users consistently have trouble installing Wordnet correctly. In particular, dictionary location appears to be totally random. Try to find some work-around for this.
Sentence splitter: The sentence splitter fails to split the following:
"In such cases, a woman has not often much beauty to think of." "But, my dear, you must indeed go and see Mr. Bingley when he comes into the neighbourhood."
- write paper on wsd by pos-lookup
- write paper on relex overview
- Lexis is the basis of language.
- Language consists of grammaticalized lexis, not lexicalized grammar.
See: Olga Moudraia, "Lexical Approach to Second Language Teaching" http://www.cal.org/resources/digest/0102lexical.html
Alternate names: "gambits", "lexical phrases", "lexical units", "lexicalized stems", "speech formulae".
Definition of Lexical Chunks
Lexis may be single words, and also the word combinations that are a basis of one's mental lexicon. That is, language consists of meaningful chunks that, when combined, produce continuous coherent text; only a minority of spoken sentences are entirely novel creations.
Types of lexical chunks:
- Words (e.g., book, pen)
- Phrasal verbs (e.g. switch off, talk to ... about ...)
- Polywords (e.g., by the way, upside down)
- Collocations, or word partnerships (e.g., community service, absolutely convinced)
- Idioms (e.g. break a leg, back in the day)
- Institutionalized utterances (e.g., I'll get it; We'll see; That'll do; If I were you . . .; Would you like a cup of coffee?)
- Sentence frames and heads (e.g., That is not as . . . as you think; The fact/suggestion/problem/danger was . . .)
- Text frames (e.g., In this paper we explore . . .; Firstly . . .; Secondly . . .; Finally . . .)
(Taken from Lewis, M. (1997b). "Pedagogical implications of the lexical approach." In J. Coady & T. Huckin (Eds.), "Second language vocabulary acquisition: A rationale for pedagogy" (pp. 255-270). Cambridge: Cambridge University Press.)