Skip to content


Repository files navigation

Table of Contents generated with DocToc


This repository contains code for converting QA-SRL annotations to Open-IE extractions and comparing Open-IE parsers against a converted benchmark corpus. This is an implementation of the algorithms described in our EMNLP2016 paper.


If you use this software, please cite:

  author    = {Gabriel Stanovsky and Ido Dagan},
  title     = {Creating a Large Benchmark for Open Information Extraction},
  booktitle = {Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  month     = {November},
  year      = {2016},
  address   = {Austin, Texas},
  publisher = {Association for Computational Linguistics},
  pages     = {(to appear)},


Leave us a note at gabriel (dot) satanovsky (at) gmail (dot) com


  • Python 3
  • See required python packages here.

Additional help can be found in the FAQ section.


Since the publication of this resource, we made several changes, outlined below. The original version of the corpus, with 10,359 extractions, as reported in the paper, is available here.

Filtering Pronoun Arguments

We removed extractions that had only pronouns as one of the arguments, and that the same extraction appeared using the entity the pronoun was referring to.
Consider, For example: John went home, he was hungry The original corpus would have had both extractions:

  1. ("John", "was", "hungry")

  2. ("he", "was", "hungry")

    In the current version of the corpus, extraction (2) is omitted, following the observation that we penalize some systems for not extracting it, where they may be doing so as a design choice.

Changing the Matching Function

Somewhat similar to the first point, we changed the evaluation scripts slightly to be more lenient. Overall, while this changes the absolute performance numbers of the different systems, it does not change the relative performance of any of the tested systems.

Converting QA-SRL to Open IE

To run the code, you should first obtain the full QA-SRL corpus and place it under QASRL-full. After obtaining the QA-SRL corpus, run:


If everything runs fine, this should create an Open IE corpus (split between wiki and newswire domain) under oie_corpus. A snapshot of the corpus is also available.

Expected Folder Structure

The script currently expects the following folder structure:

newswire/  wiki/



Please make sure that your folders adhere to this structure and naming conventions.

Otherwise, you can invoke the conversion separately for each QA-SRL file by running --in=INPUT_FILE --out=OUTPUT_FILE. Where INPUT_FILE is the QA-SRL file, and the OUTPUT_FILE is where the Open IE file will be created. You can see that the script above just makes separate calls to this file, and then concatenates the outputs.

Evaluating an Open IE Extractor

After converting QA-SRL to Open IE, you can now automatically evaluate your Open-IE system against this corpus. Currently, we support the following Open IE formats:

To compare your extractor:

  1. Run your extractor over the raw sentences and store the output into "your_output.txt"

  2. Depending on your output format, you can get a precision-recall curve by running

   benchmark --gold=GOLD_OIE --out=OUTPUT_FILE (--stanford=STANFORD_OIE | --ollie=OLLIE_OIE |--reverb=REVERB_OIE | --clausie=CLAUSIE_OIE | --openiefour=OPENIEFOUR_OIE | --props=PROPS_OIE)

  --gold=GOLD_OIE              The gold reference Open IE file (by default, it should be under ./oie_corpus/all.oie).
  --out=OUTPUT_FILE            The output file, into which the precision recall curve will be written.
  --clausie=CLAUSIE_OIE        Read ClausIE format from file CLAUSIE_OIE.
  --ollie=OLLIE_OIE            Read OLLIE format from file OLLIE_OIE.
  --openiefour=OPENIEFOUR_OIE  Read Open IE 4 format from file OPENIEFOUR_OIE.
  --props=PROPS_OIE            Read PropS format from file PROPS_OIE
  --reverb=REVERB_OIE          Read ReVerb format from file REVERB_OIE
  --stanford=STANFORD_OIE      Read Stanford format from file STANFORD_OIE

Evaluating Existing Systems

In the course of this work we tested the above mentioned Open IE parsers against our benchmark. We provide the output files (i.e., Open IE extractions) of each of these systems in systems_output. You can give each of these files to, to get the corresponding precision recall curve.

For example, to evaluate Stanford Open IE output, run:

python --gold=./oie_corpus/all.oie --out=./StanfordPR.dat --stanford=./systems_output/stanford_output.txt


You can plot together multiple outputs of, by using

   pr_plot --in=DIR_NAME --out=OUTPUT_FILENAME 

  --in=DIR_NAME            Folder in which to search for *.dat files, all of which should be in a P/R column format (outputs from
  --out=OUTPUT_FILENAME    Output filename, filetype will determine the format. Possible formats: pdf, pgf, png

Finally, try running:


This will create the Precision Recall figure using the output of OIE parsers in systems_output.