This project contains the principal Open Information Extraction (Open IE) system from the University of Washington (UW) and Indian Institute of Technology,Delhi (IIT Delhi). An Open IE system runs over sentences and creates extractions that represent relations in text. For example, consider the following sentence.
The U.S. president Barack Obama gave his speech on Tuesday and Wednesday to thousands of people.There are many binary relations in this sentence that can be expressed as a
triple (A, B, C) where A and B are arguments, and C is the relation
between those arguments. Since Open IE is not aligned with an ontology, the
relation is a phrase of text. Here is a possible list of the binary relations
in the above sentence:
(Barack Obama, is the president of, United States)
(Barack Obama, gave, his speech)
(Barack Obama, gave his speech, on Tuesday)
(Barack Obama, gave his speech, on Wednesday)
(Barack Obama, gave his speech, to thousands of people)The first extraction in the above list is a "noun-mediated extraction", because the extraction has a relation phrase is described by the noun "president". The other extractions are very similar. In fact, they can be represented more informatively as an n-ary extraction. An n-ary extraction can have 0 or more secondary arguments. Here is a possible list of the n-ary relations in the sentence:
(Barack Obama, is the president of, United States)
(Barack Obama, gave, [his speech, on Tuesday, on Wednesday, to thousands of people])Extractions can include more than just the arguments and relation as well. For example, we might be interested in whether the extraction is a negative assertion or a positive assertion, or if it is conditional in some way. Consider the following sentence:
Some people say Barack Obama was born in Kenya.We would not want to extract that (Barack Obama, was born, in Kenya) alone
because this is not true. However, if we have the condition as well, we can
have a correct extraction.
Some people say:(Barack Obama, was born in, Kenya)To see an example of Open IE being used, please visit http://openie.cs.washington.edu/.
Firstly, Open IE 5.1 improves upon extractions from numerical sentences. For example, consider the following sentence.
Barack Obama is 6 feet tall.Open IE 5.1 gives the following extractions:
(Barack Obama, has height of, 6 feet)
(Barack Obama, is, 6 feet tall)Open IE 5.1 can also extract implicit numerical relations from units in sentences. For example, consider the following sentence.
James Valley has 5 sq kms of fruit orchards.The extractions are the following:
(James Valley, has area of fruit orchards, 5 sq kms)
(James Valley, has, 5 sq kms of fruit orchards)Secondly, Open IE 5.1 improves upon conjunctive sentences by breaking conjunctions in arguments to generate multiple extractions. For example, consider the following sentence.
Jack and Jill visited India, Japan and South Korea.Open IE 5.1 gives the following extractions:
(Jack, visited, India)
(Jill, visited, India)
(Jack, visited, Japan)
(Jill, visited, Japan)
(Jack, visited, South Korea)
(Jill, visited, South Korea)Open IE 5.1 is a combination of CALMIE(Extraction from conjunctive sentences), BONIE(Extraction from Numerical Sentences), RelNoun (Noun Relations Extraction) and SRLIE. The relevant papers are:
CALMIE - Swarnadeep Saha, Mausam. "Open Information Extraction from Conjunctive Sentences." International Conference on Computational Linguistics (COLING). Santa Fe, NM, USA. August 2018. [paper]
BONIE - Swarnadeep Saha, Harinder Pal, Mausam. "Bootstrapping for Numerical Open IE". Annual Meeting of the Association for Computational Linguistics (ACL). Vancouver, Canada. August 2017. [paper] [data]
RelNoun - Harinder Pal, Mausam. "Demonyms and Compound Relational Nouns in Nominal Open IE". Workshop on Automated Knowledge Base Construction (AKBC) at NAACL. San Diego, CA, USA. June 2016. [paper]
SRLIE - Janara Christensen, Mausam, Stephen Soderland, Oren Etzioni. "An Analysis of Open Information Extraction based on Semantic Role Labeling". International Conference on Knowledge Capture (KCAP). Banff, Alberta, Canada. June 2011. [paper]
A survey paper summarizing about ten years of progress in Open IE:
Mausam. "Open Information Extraction Systems and Downstream Applications". Invited Paper for Early Career Spotlight Track. International Joint Conference on Artificial Intelligence (IJCAI). New York, NY. July 2016. [paper]
Open IE 5.1 is the successor to Open IE 4.x and Open IE 4.x is the successor to Ollie. Open IE 5.1 improves extractions from noun relations(RelNoun), numerical sentences(BONIE) and conjunctive sentences(CALMIE). Whereas Ollie used bootstrapped dependency parse paths to extract relations (see Open Language Learning for Information Extraction), Open IE 4.x uses similar argument and relation expansion heuristics to create Open IE extractions from SRL frames. Open IE 4.x also extends the defintion of Open IE extractions to include n-ary extractions (extractions with 0 or more arguments 2s).
First, download the standalone jar for BONIE from here and place it inside a lib folder(create the lib folder parallel to the src folder).
Also, download the standalone jar for CALMIE from here and place it inside the lib folder.
CALMIE uses Berkeley Language Model. Download the Language Model file from here and place it inside a data folder(create the data folder parallel to the src folder)
openie uses java-8-openjdk & the sbt build system, so downloading
dependencies and compiling is simple:
Open IE uses scala 2.10.2. In case of a version mismatch problem, try using Scala 2.10.2.
If you are unable to compile the jar locally on your machine, you can directly use the jar from here. Note that you would still need the Language Model file and Wordnet folders in the correct locations.
This jar has been compiled on an ubuntu machine. Thus, it might not work if there's a platform (or version) change, in which case it is recommended to build the jar locally.
You can run openie with sbt or create a stand-alone jar. openie requires
substantial memory. sbt is configured to use these options by default:
-Xmx10G -XX:+UseConcMarkSweepGC
OpenIE's large memory requirements largely accounts to the fact that it currently uses Berkeley Language Model in the background.
For running without jar:
sbt 'runMain edu.knowitall.openie.OpenIECli'First create the stand-alone jar.
sbt clean compile assemblyYou may need to add the above memory options.
sbt -J-Xmx10000M clean compile assemblyThen you can run the resulting jar file as normal.
java -jar openie-assembly.jarYou may need to add the above memory options.
java -Xmx10g -XX:+UseConcMarkSweepGC -jar openie-assembly.jarThe WordNet folder and the data/languageModel file must be placed parallel to the standalone openie jar, while running it.
OpenIE 5.1 can be run as a server. For this, server port is required as an argument.
java -jar openie-assembly.jar --httpPort 8000To run the server with memory options.
java -Xmx10g -XX:+UseConcMarkSweepGC -jar openie-assembly.jar --httpPort 8000To get an extraction from the server use the POST request on '/getExtraction' address. The sentence will go in the body of HTTP request. An example of curl request.
curl -X POST http://localhost:8000/getExtraction -d 'The U.S. president Barack Obama gave his speech on Tuesday to thousands of people.'The response is a JSON list of extractions.
Python Wrapper for OpenIE 5.1, along with instructions and examples can be found here
openie takes one sentence per line unless --split is specified. If
--split is specified, the input text will be split into sentences. You can
either pipe input from Standard Input, specify an input file (an option first
argument), or type sentences interactively. Output will be written to Standard
Output unless a second option argument is specified for an output file.
openie takes a number of command line arguments. To see them all run
java -jar openie-assembly.jar --usage. Of particular interest are
--ignore-errors which continues running even if an exception is encountered, --binary which gives the binary(triples) output and --split which splits the input document text into sentences.
There are two formats--a simple format made for ease of reading and a
columnated format used for machine processing. The format can be specified
with either --format simple or --format column. The simple format is
chosen by default.