GGPONC 1.0 - Pre-Processing and Basic Analysis

This repository contains the code to reproduce the results in: Florian Borchert, Christina Lohr, Luise Modersohn, Thomas Langer, Markus Follmann, Jan Philipp Sachs, Udo Hahn, Matthieu-P. Schapranow. GGPONC: A Corpus of German Medical Text with Rich Metadata Based on Clinical Practice Guidelines. ArXiv:2007.06400 [Cs]. [arXiv] [Code on GitHub] [data-request@DKG] accepted at [LOUHI@EMNLP'20) https://arxiv.org/abs/2007.06400

Prerequisites

Requesting text data

• GGPONC source files:
  • Follow the instructions of the GGPONC website (Access & Download)
  • Copy cpg-corpus-cms.xml into src/main/resources
• PubMed Abstracts from German Case Reports and Case Descriptions
  • Install Entres API from NCBI or EDirect, the commandline tools requesting the PubMed infrastructure
  • Open a terminal and type esearch -db pubmed -query "Case Reports[Publication Type] AND GER[LA]" | efetch -format xml > allGermanPubMedCaseAbstracts.xml (This step could take an hour.)
  • export the extracted file allGermanPubMedCaseAbstracts.xml into src/main/resources
• JSYNCC v1.1: follow the instructions of https://github.com/JULIELab/jsyncc or contact Christina Lohr
• 3000PA: no public access
• KRAUTS Corpus (Strötgen et al):
• WikiWarsDe Corpus (Strötgen et al)

UMLS Terminology data

• You need files from the UMLS.

• You need a registration at UTS, you can download the UMLS files from the U.S. National Library of Medicine (NIH).

• For our current work, we used the UMLS release 2019AB and you need the following files:

  • 2019AB MRCONSO.RRF
  • 2019AB MRSTY.RRF (only accessible from the full release zip file.)
  • unzip the files.

• More information of UMLS releases

• More information on the UMLS can be found in the UMLS® Reference Manual.

Software requirements

• Java 11 - We prefer Open JDK
• Apache Maven (mvn)
• Python 3 => We prefer to use Eclipse IDE or IntelliJ IDEA

Configuration after downloading of this repository

• Configure the project as a Maven project
  • In Eclipse: right click on project => Configure => Convert to Maven Project
  • Command line: mvn compile

Processing the data

Conversion of GGPOnc corpus XML file to plain text and preprocessing

• Run mvn compile before executing mvn exec:java -Dexec.mainClass="de.hpi.guidelines.reader.GGPOncXMLReader" -Dexec.args="<Path to cpg-corpus-cms.xml>" or run GGPOncXMLReader.java (in package de.hpi.guidelines.reader) in Eclipse (Run As => Java Application)
• Wait a minute
• Look into the directory /output

Create PubMed abstract text files

• We download PubMed data at February 21 2020, if you download PubMed data by esearch commands, you will receive a larger text corpus than our export. The file src/main/resources/usedPubMedIds_20200221.txt contains a list with the used PubMed identifiers from February 21 2020.
• If you want to create the described data set from PubMed, import your extracted XML file and run the src/main/extractPubMedCaseAbstracts.java. This code is able to filter our used PubMed text data from your new created download.

Processing dictionaries

Filtered dictionaries from UMLS by JuFiT

• We worked with JuFit v1.1 - you can find the right jar file in this repository.
• If you want to work with the real JuFit, follow the steps below:
  • Download JuFit from https://github.com/JULIELab/jufit
  • create the jar file by Apache Maven and run mvn clean package
  • run java -jar JuFiT.jar MRCONSO.RRF MRSTY.RRF GER --grounded > UMLS_dict.txt
• Run the Java Code RequestJuFiT.java (package de.julielab.dictionaryhandling) or the Python script extended_script_dictionaries/request-jufit.sh

Gene Dictionary

• We used a list of gene names compiled from Entrez Gene and UniProt with the approach originating from Wermter et al.
• Code of JULIELab/gene-name-mapping
• The integration of this code in the GGPOnc Repository is coming soon.

Connect Dictionaries

• For the usage of JCoRe Pipelines you will need one large file global_dictionary.txt
• Run the script extended_script_dictionaries/createDics.py to create on large dictionary (before run: adapt path names in the script file)
• Or run the Java Code CreateLargeDictionary.java (package de.julielab.dictionaryhandling) (before run: adapt path names in the script file)

JCoRe Pipeline

• Unpack the *.zip files in jcore-pipelines, there are 2 pipelines:
  • dectectUMLSentries
  • detectStopwords
• Create the folder data/files in the pipeline directories and put the data to be analyzed in the directory data/files (subdirectories are not read, be carefully with *.tar files)
• Put the global dictionary file into jcore-pipelines/detectUMLSentries/resources
• Adapt filename of the dictionary and the stopword dictionary in the following files:
  • desc/GazetteerAnnotator Template Descriptor with Configurable External Resource.xml
  • descAll/GazetteerAnnotator Template Descriptor with Configurable External Resource.xml
• Open a terminal and root into one of the pipeline directories
• Start the pipeline with java -jar ../jcore-pipeline-runner-base-0.4.1-SNAPSHOT-cli-assembly.jar run.xml
• Results
  • offsets.tsv
  • data/outData/output-xmi
• This JCoRe pipeline is derived of the JULIE Lab own jcore-pipeline-modules (see also https://zenodo.org/record/4066619#.X3sPVS8Rp-U)

Evaluation of Annotations

• To calculate the inter-annotator-agreement between human annotators follow the instructions of bratiaa
• To calculate precision and recall between automatically created annotations and the human annotated data run:
  • pip install bratutils
  • python src/main/python/umls_evaluation.py <path to gold annotations> <path to automatic annotations>