ReVerb is a program that automatically identifies and extracts binary relationships from English sentences. ReVerb is designed for Web-scale information extraction, where the target relations cannot be specified in advance and speed is important.
ReVerb takes raw text as input, and outputs (argument1, relation phrase, argument2) triples. For example, given the sentence "Bananas are an excellent source of potassium," ReVerb will extract the triple (bananas, be source of, potassium).
More information is available at the ReVerb homepage: http://reverb.cs.washington.edu
If you want to run ReVerb on a small amount of text without modifying its source code, we provide an executable jar file that can be run from the command line. Follow these steps to get started:
-
Download the latest ReVerb jar from http://reverb.cs.washington.edu/reverb-latest.jar
-
Run
java -Xmx512m -jar reverb-latest.jar yourfile.txt. -
Run
java -Xmx512m -jar reverb-latest.jar -hfor more options.
Building ReVerb from source requires Apache Maven (http://maven.apache.org). Run this command to download the required dependencies, compile, and create a single executable jar file.
mvn clean compile assembly:single
The compiled class files will be put in the target/classes directory. The
single executable jar file will be written to
target/reverb-core-*-jar-with-dependencies.jar where * is replaced with
the version number.
Once you have built ReVerb, you can run it from the command line.
The command line interface to ReVerb takes plain text or HTML as input, and outputs a tab-separated table of output. Each row in the output represents a single extracted (argument1, relation phrase, argument2) triple, plus metadata. The output has the following columns:
- The filename (or
stdinif the source is standard input) - The sentence number this extraction came from.
- Argument1 words, space separated
- Relation phrase words, space separated
- Argument2 words, space separated
- The start index of argument1 in the sentence. For example, if the value is
i, then the first word of argument1 is thei-1th word in the sentence. - The end index of argument1 in the sentence. For example, if the value is
j, then the last word of argument1 is thejth word in the sentence. - The start index of relation phrase.
- The end index of relation phrase.
- The start index of argument2.
- The end index of argument2.
- The confidence that this extraction is correct. The higher the number, the more trustworthy this extraction is.
- The words of the sentence this extraction came from, space-separated.
- The part-of-speech tags for the sentence words, space-separated.
- The chunk tags for the sentence words, space separated. These represent a shallow parse of the sentence.
- A normalized version of arg1. See the
BinaryExtractionNormalizerjavadoc for details about how the normalization is done. - A normalized version of rel.
- A normalized version of arg2.
For example:
$ echo "Bananas are an excellent source of potassium." |
./reverb -q | tr '\t' '\n' | cat -n
1 stdin
2 1
3 Bananas
4 are an excellent source of
5 potassium
6 0
7 1
8 1
9 6
10 6
11 7
12 0.9999999997341693
13 Bananas are an excellent source of potassium .
14 NNS VBP DT JJ NN IN NN .
15 B-NP B-VP B-NP I-NP I-NP I-NP I-NP O
16 bananas
17 be source of
18 potassium
For a list of options to the command line interface to ReVerb, run reverb -h.
./reverb file1 file2 file3 ...
./reverb < input
The --strip-html flag (short version: -s) removes tags from the input
before running ReVerb.
./reverb --strip-html myfile.html
You may have an entire directory structure that you want to run ReVerb on.
ReVerb takes approximately 10 seconds to initialize, so it is not efficient to
start a new process for each file. To pass ReVerb a list of paths, use the -f
switch:
# Run ReVerb on all files under mydir/
find mydir/ -type f | ./reverb -f
To include ReVerb as a library in your own project, please take a look at the
example class ReVerbExample in the
src/main/java/edu/washington/cs/knowitall/examples directory.
When running code that calls ReVerb, make sure to increase the Java Virtual
Machine heap size by passing the argument -Xmx512m to java. ReVerb loads
multiple models into memory, and will be significantly slower if the heap size
is not large enough.
To modify the ReVerb source code in Eclipse, use Apache Maven to create the appropriate project files:
mvn eclipse:eclipse
Then, start Eclipse and navigate to File > Import. Then, under General, select "Existing Projects into Workspace". Then point Eclipse to the main ReVerb directory.
If you want to start a new project that depends on ReVerb, first create a new skeleton project using Maven. The following command will ask you to fill in the details of your project name, etc.:
mvn archetype:generate
Next, add ReVerb as a dependency. To make sure you are using the latest version
of ReVerb, consult Maven Central. Do this by adding the following XML under the <project> element:
<dependencies>
<dependency>
<groupId>edu.washington.cs.knowitall</groupId>
<artifactId>reverb-core</artifactId>
<version>1.4.1</version>
</dependency>
</dependencies>
Your final pom.xml file should look something like this:
<project xmlns="http://maven.apache.org/POM/4.0.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
<modelVersion>4.0.0</modelVersion>
<groupId>mygroup</groupId>
<artifactId>myartifact</artifactId>
<version>1.0-SNAPSHOT</version>
<packaging>jar</packaging>
<name>myartifact</name>
<url>http://maven.apache.org</url>
<properties>
<project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
</properties>
<dependencies>
<dependency>
<groupId>junit</groupId>
<artifactId>junit</artifactId>
<version>3.8.1</version>
<scope>test</scope>
</dependency>
<dependency>
<groupId>edu.washington.cs.knowitall</groupId>
<artifactId>reverb-core</artifactId>
<version>1.4.1</version>
</dependency>
</dependencies>
</project>
You should be able to include ReVerb in your code now. You can try this out by
including import edu.washington.cs.knowitall.extractor.ReVerbExtractor in
your program.
ReVerb includes a class for training new confidence functions, given a list of
labeled examples, called ReVerbClassifierTrainer. Example code for training a
new confidence function confFunction is shown below - the non-trivial part is
likely to be converting your labeled data to an
Iterable<LabeledBinaryExtraction>.
Example Pseudocode:
// Provide your labeled data here
Iterable<LabeledBinaryExtraction> myLabeledData = ???
ReVerbClassifierTrainer trainer =
new ReVerbClassifierTrainer(myLabeledData);
Logistic classifier = trainer.getClassifier();
ReVerbConfFunction confFunction = new ReVerbConfFunction(classifier);
// confFunction is ready to use here.
double conf = confFunction.getConf(extraction);
If you already have a list of binary labeled ReVerb extractions, it should be
easy to convert them to ChunkedBinaryExtraction objects, and then to
LabeledBinaryExtraction objects (see the constructors for these classes).
Also note that ReVerb includes a LabeledBinaryExtractionReader and Writer
class. You may wish to (re-)serialize your data using
LabeledBinaryExtractionWriter - this will put it in the same format as all
previous data used to train ReVerb confidence functions, and it will be easy to
read in the future with LabeledBinaryExtractionReader.
For more information, please visit the ReVerb homepage at the University of Washington: http://reverb.cs.washington.edu.
-
How fast is ReVerb?
You should really benchmark ReVerb yourself, but on my computer (a new computer in 2011) ReVerb processed 5000 high-quality web sentences in 21 s, or 238 sentences per second, in a single thread. ReVerb is easily parallelizable by processing different sentences concurrently.
- Anthony Fader http://www.cs.washington.edu/homes/afader
- Michael Schmitz http://www.schmitztech.com/
- Robert Bart (rbart at cs.washington.edu)
- Janara Christensen http://www.cs.washington.edu/homes/janara
- Niranjan Balasubramanian http://www.cs.washington.edu/homes/niranjan
- Jonathan Berant http://www.cs.tau.ac.il/~jonatha6
If you use ReVerb in your academic work, please cite ReVerb with the following BibTeX citation:
@inproceedings{ReVerb2011,
author = {Anthony Fader and Stephen Soderland and Oren Etzioni},
title = {Identifying Relations for Open Information Extraction},
booktitle = {Proceedings of the Conference of Empirical Methods
in Natural Language Processing ({EMNLP} '11)},
year = {2011},
month = {July 27-31},
address = {Edinburgh, Scotland, UK}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent