Apppublic

ielabgroup/mesh-suggester

sourceHugging Faceupdated 7mo agoView on Hugging Face

2likes

App README

MeSH Term Suggester

AI-powered MeSH term suggestion for systematic review boolean query construction. Built with Gradio (pure Python — no Node/npm required).

Quick Start (local, pip)

bash

# 1. Create environment
conda create --prefix ./envs python=3.9
conda activate ./envs

# 2. Install dependencies
pip install -r requirement.txt

# 3. Install tevatron (Dense retrieval library)
git clone https://github.com/texttron/tevatron
pip install -e tevatron/

# 4. Download model files  (~433 MB BERT + PubMed-w2v.bin)
python download_models.py

# 5. Launch the app
python app.py
# → open http://localhost:7860

Quick Start (Docker)

bash

# First run — builds image and downloads models into a named volume (~2 GB, once only)
docker compose up --build

# → open http://localhost:7860

After the first run, models are stored in a Docker named volume (mesh_models). Code changes don't require a rebuild — just:

bash

docker compose restart

Dependency changes (requirement.txt) do need a rebuild, but models are still in the named volume so they won't be re-downloaded.

Model files

File	Source	Destination
BERT checkpoint	ielabgroup/mesh_term_suggestion_biobert	`server/Model/checkpoint-80000/`
PubMed-w2v.bin	Google Drive	`server/Model/PubMed-w2v.bin`

download_models.py handles both automatically.

Hugging Face Spaces deployment

HF Spaces has no persistent volumes, so models must be baked into the image.

bash

# Push to your Space remote (HF Spaces uses git)
git remote add space https://huggingface.co/spaces/ielabgroup/mesh-suggester
git push space main

HF Spaces will build the Dockerfile with BAKE_MODELS=true (set this in Space settings under Secrets / Variables → add BAKE_MODELS=true as a build argument, or pass it via the Dockerfile default). Models are downloaded once at build time and baked in.

Alternatively, add this to your Space's README.md build config:

build_args:
  BAKE_MODELS: "true"

Suggestion methods

Method	Description
Semantic-BERT	Groups semantically similar keywords via word2vec, then queries the BERT retriever per group
Fragment-BERT	Treats all keywords as a single query fragment
Atomic-BERT	Runs the BERT retriever independently for each keyword
ATM	Uses the NCBI Entrez API (no local model required)

Use of UMLS and MetaMAP (optional)

To use UMLS or MetaMAP suggestion, you need to deploy the respective services locally. See elastic-umls and the MetaMAP installation guide.

Citing

If you use MeSH Suggester in your research, please cite:

bibtex

@inproceedings{wang2023mesh,
  title={Mesh Suggester: A Library and System for Mesh Term Suggestion for Systematic Review Boolean Query Construction},
  author={Wang, Shuai and Li, Hang and Zuccon, Guido},
  booktitle={Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining},
  pages={1176--1179},
  year={2023}
}

License

[![CC BY-NC-ND 4.0][cc-by-nc-nd-shield]][cc-by-nc-nd]

This work is licensed under a [Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License][cc-by-nc-nd].

[![CC BY-NC-ND 4.0][cc-by-nc-nd-image]][cc-by-nc-nd]

[cc-by-nc-nd]: http://creativecommons.org/licenses/by-nc-nd/4.0/ [cc-by-nc-nd-image]: https://licensebuttons.net/l/by-nc-nd/4.0/88x31.png [cc-by-nc-nd-shield]: https://img.shields.io/badge/License-CC%20BY--NC--ND%204.0-lightgrey.svg