Read Me

To test SORE this repository contains ~110 papers, which are copied from the [OA-STM corpus](https://github.com/elsevierlabs/OA-STM-Corpus). These papers are sourced from 10 different scientific domains: Agriculture, Astronomy, Biology, Chemistry, Computer Science, Earth Science, Engineering, Materials Science, Mathematics and Medicine.

Train and use SciIE

Clone the SciIE code from [this bitbucket repository](https://bitbucket.org/luanyi/scierc/src/master/), also check this repository for more information on adjusting runtime parameters.

Create a separate SciIE environment:

• conda create –name SCIIE python=2.7

• source activate SCIIE

Install the following dependencies:

• pip install TensorFlow==1.8.0

• pip install tensorflow_hub

• pip install pyhocon

1. Fetch GloVe embeddings and build kernels

   Open the SciIE folder in a terminal/console window and run:

   • ./scripts/fetch_required_data.sh

   • ./scripts/build_custom_kernels.sh

2. Generate ELMo embeddings for FOBIE and train a model

   First copy the SciIE formatted FOBIE files to: [SciIE_folder]/data/processed_data/json/

   In the SORE directory you can find a folder called sciie_scripts/, copy the files to your [SciIE_folder]:

   • Replace [SciIE_folder]/experiments.conf and place embed_and_predict.py and generate_FOBIE_embeddings.py alongside it

   • Create a directory [SciIE_folder]/FOBIE_output/ to store predictions

   Generate the embeddings for the FOBIE dataset:

   • Run python generate_FOBIE_embeddings.py (make sure to be in the correct environment)

   To train a model you can run:

   • python singleton.py fobie_train & python evaluator.py fobie_train

3. Predict using your trained model

   You’re now set to make predictions on unseen data:

   • python embed_and_predict.py fobie_train

   This will loop through the files found in [SciIE_folder]/data/processed_data/json/ and ask you which files to process before starting the embedding (can take a while for big files).

Prepare OpenIE 5

Clone the [github repo for OpenIE 5](https://github.com/dair-iitd/OpenIE-standalone), aslo see this repo for more information about compiling. In the newly created OpenIE-standalone folder, you’ll have to create a new directory called lib/.

You will have to download and place into the lib/ folder:

• [The BONIE standalone jar](https://github.com/dair-iitd/OpenIE-standalone/releases/download/v5.0/BONIE.jar)

• [The CALMIE standalone jar](https://github.com/dair-iitd/OpenIE-standalone/releases/download/v5.0/ListExtractor.jar)

Furthermore, you will have to download and place into the data/ folder:

• [The Berkeley Language Model](https://drive.google.com/file/d/0B-5EkZMOlIt2cFdjYUJZdGxSREU/view?usp=sharing)

OpenIE is compiled using sbt and Java 8:

• Install [SDK](https://sdkman.io/install) by running curl -s “https://get.sdkman.io” | bash

• Open a new terminal window and check that sdk is installed sdk version

• Then install sbt sdk install sbt: * You may need sbt at version 0.13.x , e.g.: sdk install sbt 0.13.18

• Inside the OpenIE-standalone/ directory: * First run bash compile.sh to compile SRLIE and ONRE * Then compile the jar file sbt -J-Xmx10000M clean compile assembly * You may have to [install Java 8](https://www.scala-sbt.org/1.x/docs/Installing-sbt-on-Mac.html) first , e.g.: sdk install java 8.0.252-amzn * And add java to your path, e.g., export PATH=$PATH:~/.sdkman/candidates/java/8.0.252-amzn/bin * You may need to set the Scala version to [version 2.10.2](https://www.scala-lang.org/download/2.10.2.html)

Run SORE

Create an environment and activate, e.g.;

• conda create –name SORE python=3.6

• source activate SORE

Install the requirements found inside [FOBIE_repo]/SORE/:

• pip install -r requirements.txt

• Download required language models * python -m spacy download en_core_web_sm * python -m textblob.download_corpora

Edit the SORE_settings.json doc:

• Set the path_to_OIE_jar, for example: /Users/../../OpenIE-standalone/target/scala-2.10/openie-assembly-5.0-SNAPSHOT.jar

You can run the code for SORE from the FOBIE repository:

• python run_SORE.py

Main settings:

• Prepare_data: Yes/No - convert json data found in [FOBIE_repo]/SORE/data/unprocessed/ to OpenIE and SciIE input files. * Note that to predict with SciIE you’ll have to move these input files to [SciIE_folder]/data/processed_data/json/.

• Parse_narrowIE_predictions: Yes/No - parse SciIE prediction files to a single csv, specified in the settings file under “narrowIE_input_files”. These predictions files should be placed in [FOBIE_repo]/SORE/data/narrowIE/predictions/, note that you’ll have to move predictions made with SciIE here (from the [SciIE_folder]/FOBIE_output/ folder).

• Run_OIE: Yes/No - run OpenIE 5 on each of the files for which SciIE has found relations.

• Filter_OIE: Yes/No - filter the OpenIE 5 extractions using the predicted relations stored in the narrowIE output csv file.

• convert_back_to_BRAT: Yes/No - convert the SORE extractions to a BRAT annotations format, this enables you to visualise the resulting extractions in [BRAT](https://brat.nlplab.org/index.html).

Note that FOBIE focuses on tradeoffs and other types of correlations in the Biology domain. Non-tradeoff relations can be used to filter OpenIE extractions in distinct domains, but whether the arguments used for filtering are useful can vary.

If you have any questions, please let me know!