Multiprocessing PubTator Parsing for Entity Linking

Quick Starts with MedMentions, BC5CDR and NCBI-dataset

$ git clone https://github.com/izuna385/PubTator-Multiprocess-Parser.git
$ cd PubTator-Multiprocess-Parser
$ docker build -t multiprocess_pubtator .
$ docker run -itd multiprocess_pubtator /bin/bash

# In container
$ sh ./scripts/quick_start_Med_full.sh # for MedMentions

• You can run quick_start_NCBI_full.sh, too. If so, before running, make pickled_doc_dir empty.

• Note: If you use Mac, do brew install wget before running above script.

Description

• Preprocessing PubTator-format documents to each mentions.

• If you are japanese, this might be useful for you.

  https://qiita.com/izuna385/items/d673694d25b2cf4efb89

How to run

• Note: The following steps are entirely automated.

  After building container, run sh ./scripts/quick_start_[dataset_name]_full.sh

1. Place PubTator format files to the ./dataset/

• corpus_pubtator.txt, corpus_pubtator_pmids_trng.txt, corpus_pubtator_pmids_dev.txt, and corpus_pubtator_pmids_test.txt must be placed there.

2. run

python3 main.py

3. Check

• Each Pubtator documents is preprocessed and dumped to ./dataset/**pmid**.pkl

  The format is as the below.

  {'title':title,  
   'abst':abst,
   'title_plus_abst': title_plus_abst,
   'pubmed_id': pubmed_id,
   'entities': entities,
   'split_sentence': splitted_sentence,
   'if_txt_length_is_changed_flag':if_txt_lenght_is_changed_flag,
   'lines':lines,
   'lines_lemma':lines_lemma
  }
  

  • The Key component is 'lines', in which all information for entity linking is included.

• Each document takes about 100sec for preprocessing, under en_core_sci_md model.

• Under 24 core cpus and en_core_sci_md model, ~10GB RAM is needed.

LISENCE

MIT