Ollie is a program that automatically identifies and extracts binary relationships from English sentences. Ollie is designed for Web-scale information extraction, where target relations are not specified in advance.
Ollie is our second-generation information extraction system . Whereas ReVerb operates on flat sequences of tokens, Ollie works with the tree-like (graph with only small cycles) representation using Stanford's compression of the dependencies. This allows Ollie to capture expression that ReVerb misses, such as long-range relations.
Ollie also captures context that modifies a binary relation. Presently Ollie handles attribution (He said/she believes) and enabling conditions (if X then).
You can now run Ollie with a single Docker command.
docker run -it schmmd/ollie:latest
To configure Ollie, you can drop into a bash shell with docker run -it schmmd/ollie:latest /bin/bash
and run Ollie from the command line.
If you want to run Ollie on a small amount of text without modifying the source code, you can use an executable file that can be run from the command line. Please note that Ollie was built using Scala 2.9 and so it requires Java 7. Follow these steps to get started:
-
Download the latest Ollie binary from http://knowitall.cs.washington.edu/ollie/ollie-app-latest.jar.
-
Download the linear English MaltParser model (engmalt.linear-1.7.mco) from http://www.maltparser.org/mco/english_parser/engmalt.html and place it in the same directory as Ollie.
-
Run
java -Xmx512m -jar ollie-app-latest.jar yourfile.txt
. The input file should contain one sentence per line unless--split
is specified. Omit the input file for an interactive console.
An enabling condition is a condition that needs to be met for the extraction to be true. Certain words demark an enabling condition, such as "if" and "when". Ollie captures enabling conditions if they are present.
sentence: If I slept past noon, I'd be late for work.
extraction: (I; 'd be late for; work)[enabler=If I slept past noon]
An attribution clause specifies an entity that asserted an extraction and a verb that specifies the expression. Ollie captures attributions if they are present.
sentence: Some people say Barack Obama was not born in the United States.
extraction: (Barack Obama; was not born in; the United States)[attrib=Some people say]
sentence: Early astronomers believe that the earth is the center of the universe.
extraction: (the earth; is the center of; the universe)[attrib=Early astronomers believe]
Some relations are expressed without verbs. Ollie can capture these as well as verb-mediated relations.
sentence: Microsoft co-founder Bill Gates spoke at a conference on Monday.
extraction: (Bill Gates; be co-founder of; Microsoft)
Often times similar relations will specify different aspects of the same event. Since Ollie captures long-range relations it can capture N-ary extractions by collapsing extractions where the relation phrase only differs by the preposition.
sentence: I learned that the 2012 Sasquatch music festival is scheduled for May 25th until May 28th.
extraction: (the 2012 Sasquatch music festival; is scheduled for; May 25th)
extraction: (the 2012 Sasquatch music festival; is scheduled until; May 28th)
nary: (the 2012 Sasquatch music festival; is scheduled; [for May 25th; to May 28th])
Building Ollie from source requires Apache Maven (http://maven.apache.org). First, clone or download the Ollie source from GitHub. Run this command in the top-level source folder to download the required dependencies, compile, and create a single jar file.
mvn clean package
The compiled class files will be put in the base directory. The single
executable jar file will be written to ollie-app-VERSION.jar
where VERSION
is the version number.
Once you have built Ollie, you can run it from the command line.
java -Xmx512m -jar ollie-app-VERSION.jar yourfile.txt
Omit the input file for an interactive console.
Ollie takes sentences, one-per-line as input or splits text into sentences if
--split
is specified. Run Ollie with --usage
to see full usage.
The Ollie command line tool has a few output formats. The output format is
specified by --output-format
and a valid format:
- The
interactive
format that is meant to be easily human readable. - The
tabbed
format is mean to be easily parsable. A header will be output as the first row to label the columns. tabbedsingle
is similar totabbed
but the extraction is output as (arg1; relation; arg2) in a single column.- The
serialized
is meant to be fully deserialized into anOllieExtractionInstance
class.
Ollie works ontop of a subcomponent called OpenParse. The distinction is largely technical; OpenParse does not handle attribution and enabling condition and uses a coarser confidence metric. You can use a GUI application to visualize the OpenParse extractions in a parse tree. To use it, you will need to have graphviz installed. You can run the GUI with:
java -Xms512M -Xmx1g -cp ollie-app-VERSION.jar edu.knowitall.openparse.OpenParseGui
By default, this application will look for graphviz's dot
program at
/usr/bin/dot
. You can specify a location with the --graphviz
parameter.
You can try out your own models with Options->Load Model...
. To see an
example model, look at openparse.model
in src/main/resources
. Your model
may have one or more patterns in it. If you want to see pattern matches
(without node expansion) instead of triple extractions, you can choose to show
the raw match with Options->Raw Matches
. This will allow you to use patterns
that do not capture an arg1, rel, and arg2.
Ollie is packaged to use Malt Parser, one of the fastest dependency parsers
available. You will need the model file (engmalt.linear-1.7.mco
) in the
directory the application is run from or you will need to specify its location
with the --malt-model
parameter. Malt Parser models are available online.
http://www.maltparser.org/mco/english_parser/engmalt.html
Ollie works with any other parser in the nlptools
project. For example, it
is easy to swap out Malt for Stanford's parser. Stanford's parser is not a
part of the Ollie distribution by default because of licensing conflicts, but
the Stanford parser was used as the execution parser for the results in the
paper. Malt Parser was used to bootstrap the patterns. We are interested
in Clear parser as an alternative, but it's not a trivial change because Clear
uses a slightly different dependency representation.
To modify the Ollie source code in Eclipse, use the M2Eclipse plugin along with ScalaIDE. You can then import the project using the following.
File > Import > Existing Maven Projects
Add the following as a Maven dependency.
<groupId>edu.washington.cs.knowitall.ollie</groupId>
<artifactId>ollie-core_2.9.2</artifactId>
<version>[1.0.0, )</version>
The best way to find the latest version is to browse Maven Central.
ollie-core
does not include a way to parse sentences. You will need to use a
parser supplied by the nlptools
project. The source for for ollie-app
is an excellent example of a project
using ollie-core
as a dependency. ollie-app
supplies a parser from
nlptools.
There is an example project that uses Ollie in the example
folder of the
source distribution.
While Ollie comes with a trained confidence function, it is possible to retrain the confidence function. First, you need to run Ollie over a set of sentences and store the output in the serialized format.
echo "Michael rolled down the hill." | java -jar ollie-app-1.0.0-SNAPSHOT.jar --serialized --output toannotate.tsv
Next you need to annotate the extractions. Modify the output file and
change the first column to a binary annotation--1
for correct and 0
for
wrong. Your final file will look similar to ollie/data/training.tsv
. Now
run the logistic regression trainer.
java -cp ollie-app-1.0.0-SNAPSHOT.jar edu.washington.cs.knowitall.ollie.confidence.train.TrainOllieConfidence toannotate.tsv
When operating at web scale, parallelism is essential. While the base Ollie extractor is immutable and thread safe, the parser may not be thread safe. I do not know whether Malt parser is thread safe.
-
How fast is Ollie?
You should really benchmark Ollie yourself, but on my computer (a new computer in 2011), Ollie processed 5000 high-quality web sentences in 56 seconds, or 89 sentences per second, in a single thread. Ollie is easily parallelizable and the Ollie extractor itself is threadsafe (see Concurrency section).
To contact the UW about Ollie, email [email protected].
If you use Ollie in your academic work, please cite Ollie with the following BibTeX citation:
@inproceedings{ollie-emnlp12,
author = {Mausam and Michael Schmitz and Robert Bart and Stephen Soderland and Oren Etzioni},
title = {Open Language Learning for Information Extraction},
booktitle = {Proceedings of Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CONLL)},
year = {2012}
}