This is a project completed as part of the CENG596 - Information Retrieval course at METU. In this project, we aim to implement a search engine to efficiently search through a set of COVID-19-related articles from the medical field, while also providing a simple user interface to make this tool accessible to as many healthcare professionals and the general public alike as possible.
-
Can Ufuk Ertenli, ufuk.ertenli@metu.edu.tr
-
Onat Ozdemir, onat.ozdemir@metu.edu.tr
An example search from the UI can be seen below:
The overall data flow is presented below:
Supported models:
- TF-IDF
- BM25
- BERT
- SLEDGE (https://arxiv.org/abs/2010.05987)
-
Create a conda virtual environment and activate it.
conda create -n coveng python=3.8 -y conda activate coveng
-
Install PyTorch and torchvision following the official instructions, e.g.,
conda install pytorch torchvision -c pytorch
-
Install the requirements.
pip install -r requirements.txt
-
Install OpenNIR.
pip install -q --upgrade git+https://github.com/Georgetown-IR-Lab/OpenNIR
-
Run main.py for experimentation and coveng_search.py to trigger the search engine system launching a UI.
Note that JAVA_PATH environment variable must point to the local Java installation.
The necessary settings can be controlled from the given config file. The best results can be obtained using SLEDGE with a full re-ranking approach with stemming, stop word removal, and query expansion.
| Method | map | P(R=2)@10 | nDCG@10 | nDCG@100 |
|---|---|---|---|---|
| TF_IDF | 0.2579 | 0.6020 | 0.6878 | 0.5408 |
| BM25 | 0.2426 | 0.5540 | 0.6390 | 0.5084 |
| BERT(c=10) | 0.0135 | 0.5500 | 0.6307 | 0.1385 |
| BERT(c=100) | 0.0752 | 0.4019 | 0.4695 | 0.4775 |
| BERT | 0.1294 | 0.0840 | 0.0894 | 0.1697 |
| SLEDGE(c=10) | 0.0142 | 0.5500 | 0.6508 | 0.1429 |
| SLEDGE(c=100) | 0.0943 | 0.6700 | 0.7474 | 0.5296 |
| SLEDGE | 0.2568 | 0.6359 | 0.7228 | 0.5629 |