Neural search engine for ML papers.
Usage is simple: input an abstract, get the matching papers. The following demo also showcases the finetuning functionality (notice how the paper marked as "irrelevant" is assigned a lower score after finetuning).
We used a stripped-down version of the Kaggle arXiv Dataset in which only the following categories are retained: cs.AI, cs.CL, cs.CV, cs.LG, cs.MA, cs.NE
Clone the repository
git clone https://github.com/fissoreg/papers-search/
cd papers-search
For both the folders frontend
and backend
, run the following commands
cd folder_to_go_into/ # `folder_to_go_into` is either `frontend` or `backend`
python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
The app works by suggesting papers whose abstract is similar to the one you provided. The suggestions come from a database of published papers: you need to index all the suggestions for the system to be able to function. This is a lenghty operation, but it needs to be performed only once:
cd backend
python src/app.py --index
For testing, you can index a small number of papers providing the --n
argument:
python src/app.py --index --n 10
This can be run after indexing (section above).
Run the backend
cd backend
python3 src/app.py
In a new terminal, run the frontend
cd frontend
streamlit run app.py
Connect to http://localhost:8501/
(with your favourite browser).
Refer to the Makefile
for the specific commands
To format code following the black
standard
$ make format
Code linting with flake8
$ make lint
Testing
$ make testdeps
$ make test
Testing with coverage analysis
$ make coverage
Format, test and coverage
$ make build
This project is in its starting phase. If you are interested in contributing, don't hesitate to get in touch! (Or go straight to the Issues ;)).
Made possible by:
- Jina AI
- Sentence-Transformers
- arXiv: Thank you to arXiv for use of its open access interoperability.
- Kaggle