Bio-AnswerFinder

A question answering system for biomedical literature.

Prerequisites

• Java 1.8+
• Gradle
• An ElasticSearch Endpoint with indexed PubMED abstracts

The datasets generated during and/or analysed during the current study are available in the Zenodo repository.

This includes learned GloVE vectors, vocabulary, named entity databases, nominalization, acronym and morphology databases.

The system is configured by a configuration file. An example configuration file is provided in the project src/main/resources/bio-answerfinder.properties.example.

The code is written in Java (1.8+) and uses Gradle to build the project.

Named entity databases

LookupUtils and LookupUtils2

• Drugs /usr/local/lt/drug_bank.db
• Diseases /usr/local/lt/diseases.db
• Gene names /usr/local/lt/hugo.db
• Scigraph ontology lookup
  • data/scigraph_lt/ontology-classes-with-labels-synonyms-parents.json
  • data/scigraph_lt/scr-classes-with-labels-synonyms-parents.json

NominalizationService

/usr/local/lt/nominalization.db

AcronymService

/usr/local/lt/acronyms.db

Morphology database

org.bio_answerfinder.nlp.morph.Lemmanizer and org.bio_answerfinder.util.SRLUtils

/usr/local/morph/morph.db

Scripts

Python 3 code for the keyword selection classifier and BERT fine-tuning is in the scripts directory of the project. You need Tensorflow (1.12+) and Keras installed into an Python virtual environment. You need to clone Google BERT github project

git clone https://github.com/google-research/bert

After that, follow instructions on BERT github site to download BERTBase model uncased_L-12_H-768_A-12. Similarly, download BioBERT model from https://github.com/naver/biobert-pretrained/releases/tag/v1.1-pubmed.

Copy run_bioasq_classifier.py, train_bioasq.sh, train_bioasq.sh, train_biobert_bioasq.sh, predict_biobert_bioasq.sh scripts to Google BERT clone directory and update the environment variables in the driver Bash shell scripts to match your project and model installation directories.

Annotated Datasets

All the annotated datasets generated during this study is under the project directory data.

Query Keyword Selection Classifier Data

To be used by the Python 3 script scripts/question_keyword_selector_v2.py

• data/bioasq/bioasq_manual_100/qsc/qsc_set_train.txt
• data/bioasq/bioasq_manual_100/qsc/qsc_set_test.txt
• data/bioasq/bioasq_manual_100/qsc/qsc_set_pos_tags_train.txt
• data/bioasq/bioasq_manual_100/qsc/qsc_set_pos_tags_test.txt

BERT reranker training/testing data

Note that there is no development data. Both dev.tsv and test.tsv are same. The file dev.tsv is there for the Google BERT code.

• data/bioasq/bioasq_manual_100/train.tsv
• data/bioasq/bioasq_manual_100/dev.tsv
• data/bioasq/bioasq_manual_100/test.tsv

BioAnswerFinder unsupervised (wRWMD reranking) annotated results

• data/bioasq/bioasq_manual_100/qaengine1/question_results_wmd_defn_focus.txt

BERT Reranker Blind Evaluation

• data/evaluation/annotator_1_bert_rank.csv
• data/evaluation/annotator_1_rwmd_rank.csv
• data/evaluation/annotator_2_bert_rank.csv
• data/evaluation/annotator_2_rwmd_rank.csv
• data/evaluation/annotator_3_bert_rank.csv
• data/evaluation/annotator_3_rwmd_rank.csv
• data/evaluation/rwmd_question_answer_candidates.csv
• data/evaluation/bert_question_answer_candidates.csv

BioBERT and BERT Reranker Comparison

To be used by the Python 3 script scripts/show_perfomance.py

• data/evaluation/bert_biobert_comparison.csv
• data/evaluation/biobert_perf_records.json

Datasets for AKTS workshop paper

Ibrahim Burak Ozyurt, Jeffrey S. Grethe. Iterative Document Retrieval via Deep Learning Approaches for Biomedical Question Answering in 15th International Conference on eScience (2019). (doi: 10.1109/eScience.2019.00072 )

• data/rank_test (training/testing datasets for NN models)
• data/rank_annotations (curator rankings for each method tested)

Exact answers for factoid question evaluation

• data/evaluation/bert_factoid_test_perf.csv
• data/evaluation/rank_test_factoid_perf.csv

Due PubMed license restrictions, Pubmed abstracts ElasticSearch index cannot be provided. It can be generated using our ETL system Foundry available in GitHub https://github.com/biocaddie/Foundry-ES.

Tips for generating ElasticSearch Index from Scratch

• You need about 500GB available space for the processing.
• Install Elasticsearch (https://www.elastic.co/guide/en/elasticsearch/reference/current/install-elasticsearch.html)
• Download PubMed baseline from PubMed FTP site ftp://ftp.ncbi.nlm.nih.gov/pubmed/baseline (all files ending with xml.gz).
• The gzipped XML files contain multiple records for abstracts which needs to be expanded and parsed. The top element of each file is PubmedArticleSet. Each abstract record starts with the XML tag PubmedArticle.
• From each PubmedArticle tag, you need to extract the PMID, title and abstract text from the following nested XML tags,
  • PMID MedlineCitation/PMID
  • title MedlineCitation/Article/ArticleTitle
  • abstract MedlineCitation/Article/Abstract/AbstractText
• To index the PMID, title and abstract text for each PubMed abstract to Elasticsearch, generate a JSON object of the following form

{
"dc": {
   "identifier": "<PMID>",
   "title": "<title>",
   "description": "<description>"
 }
}

• The JSON objects for the PubMed abstracts can be indexed by using Elasticsearch bulk indexing web API (https://www.elastic.co/guide/en/elasticsearch/reference/current/docs-bulk.html)
• Assuming the index is generated at the local ES endpoint http://localhost:9200/pubmed/abstracts, you need to update your bio-answerfinder.properties file to include the line

elasticsearch.url=http://localhost:9200/pubmed/abstracts

Building

gradle clean war

Then deploy to Tomcat or any Java web app container.

For any questions, please contact iozyurt@ucsd.edu.