Designing a Semantic-Enhanced Hybrid Search Engine for Biomedical Events

New site: https://metadb.riken.jp/BEHF

The Biomedical Event Hybrid Finder (BEHF) is a semantic search engine tailored for the discovery of biomedical events (BE) in PubMed documents. It combines Elasticsearch's exact match search capabilities with expanded keywords provided by SPLADE, and leverages Faiss ANN for semantic similarity search. BEHF utilizes a unique dataset of network graphs, derived from the PubMed Baseline Database (BD) repository, and offers a flexible scoring mechanism to balance between exact match and semantic similarity searches.

Creating the pfiles with pubmed titles and abstracts

python create_pubmed_dataset.py --dict_file /home/julio/repos/pubmed_processed_data/dicts/pmid2info_full.pickle 
--outdir /home/julio/repos/event_finder/data/pubmed/

Extracting events with deep event mine

Move create directies to Deepeventmine data. And start extracting events on each gpu:

sh pubmed.sh e2e rawtext pubmed_1 cg 0
sh pubmed.sh e2e rawtext pubmed_2 cg 1
sh pubmed.sh e2e rawtext pubmed_3 cg 2
sh pubmed.sh e2e rawtext pubmed_4 cg 3

Creating embeddings from events

python create_embeddings_sbert.py --indir=/home/julio/repos/ddegk/results/pubmed_small_4/ \
    --outdir=/home/julio/repos/event_finder/data/pubmed_small_graph/  \
    --outname=sbert_embeddings_all-MiniLM-L12-v2.json  \
    --sbertmodel=all-MiniLM-L12-v2 

Creating Faiss index

python create_faiss_index.py --indir=../data/pubmed_small_graph/ \
                            --embs_file=sbert_embeddings_all-MiniLM-L12-v2.json \
                            --indexname=faiss_sbert_all-MiniLM-L12-v2

New

Create graph

python -m bef.data.text2events2graphs --yaml data/pubmed_2000s/predict-pubmed_epi.yaml --gpu 0 --start_year 2000 --end_year 2020

python -m bef.data.text2events2graphs --yaml data/pubmed_2000s/predict-pubmed_cg.yaml --gpu 0 --start_year 2020 --end_year 2021

Deduplicate the graphs

python -m bef.data.graph_deduplicator data/pubmed_70s/cg data/pubmed_70s/cg

Create faiss index:

python -m bef.data.graphs2faiss  --graphfile events_graph.json --base_dir data/pubmed_70s/cg

Creating the ElasticSearch index:

## Don't add '/' at the end 
python -m bef.index.create_elasticsearch_index /home/julio/repos/event_finder/data/pubmed_70s/id/events_graph.json pubmed_70s_id

Creating all the faiss and elastic search indices from dir:

 ./create_indices.sh data/pubmed_70s

Batch

Extracting all graphs years from pubmed.

nohup ./text2graphs_new.sh &

Pre-installation

$pip install -e .
$git clone [email protected]:jcrangel/dkouqe.git
$cd dkouqe 
$ pip install -e .


pip install git+https://github.com/naver/splade.git

Download data

TODO

Installing pubmed database: pmid2text

python -m bef.create_pmid2text data/pubmed/json_pmid2sents_year
./createdb_pubmed.sh

Running DjangoFront end

Run hybrid API:

python web_app/server_hybrid.py --data_path data/pubmed_70s

or nohup python web_app/server_hybrid.py --data_path data/pubmed_70s &

Then run the Django server

cd django_frontend
python manage.py runserver 0.0.0.0:8000

Moving to other server

git clone [email protected]:jcrangel/event_finder.git
cd event_finder
git clone [email protected]:jcrangel/DeepEventMine_fork.git
mv DeepEventMine_fork/ DeepEventMine