BioLAMA
BioLAMA is biomedical factual knowledge triples for probing biomedical LMs. The triples are collected and pre-processed from three sources: CTD, UMLS, and Wikidata. Please see our paper Can Language Models be Biomedical Knowledge Bases? (Sung et al., 2021) for more details.
* The dataset for the BioLAMA probe is available at data.tar.gz
Getting Started
After the installation, you can easily try BioLAMA with manual prompts. When a subject is "flu" and you want to probe its symptoms from an LM, the input should be like "Flu has symptom such as [Y]."
# Set MODEL to bert-base-cased for BERT or dmis-lab/biobert-base-cased-v1.2 for BioBERT
MODEL=./RoBERTa-base-PM-Voc/RoBERTa-base-PM-Voc-hf
python ./BioLAMA/cli_demo.py \
--model_name_or_path ${MODEL}
Result:
Please enter input (e.g., Flu has symptoms such as [Y].):
hepatocellular carcinoma has symptoms such as [Y].
-------------------------
Rank Prob Pred
-------------------------
1 0.648 jaundice
2 0.223 abdominal pain
3 0.127 jaundice and ascites
4 0.11 ascites
5 0.086 hepatomegaly
6 0.074 obstructive jaundice
7 0.06 abdominal pain and jaundice
8 0.059 ascites and jaundice
9 0.043 anorexia and jaundice
10 0.042 fever and jaundice
-------------------------
Top1 prediction sentence:
"hepatocellular carcinoma has symptoms such as jaundice."
Quick Link
Installation
# Install torch with conda (please check your CUDA version)
conda create -n BioLAMA python=3.7
conda activate BioLAMA
conda install pytorch=1.8.0 cudatoolkit=10.2 -c pytorch
# Install BioLAMA
git clone https://github.com/dmis-lab/BioLAMA.git
cd BioLAMA
pip install -r requirements.txt
Resources
Models
For BERT and BioBERT, we use checkpoints provided in the Huggingface Hub:
- best-base-cased (for BERT)
- dmis-lab/biobert-base-cased-v1.2 (for BioBERT)
Bio-LM is not provided in the Huggingface Hub. Therefore, we use the Bio-LM checkpoint released in link. Among the various versions of Bio-LMs, we use `RoBERTa-base-PM-Voc-hf'.
wget https://dl.fbaipublicfiles.com/biolm/RoBERTa-base-PM-Voc-hf.tar.gz
tar -xzvf RoBERTa-base-PM-Voc-hf.tar.gz
rm -rf RoBERTa-base-PM-Voc-hf.tar.gz
Datasets
The dataset will take about 78 MB of space. Download data.tar.gz and uncompress it.
tar -xzvf data.tar.gz
rm -rf data.tar.gz
The directory tree of the data is like:
data
├── ctd
│ ├── entities
│ ├── meta
│ ├── prompts
│ └── triples_processed
│ └── CD1
│ ├── dev.jsonl
│ ├── test.jsonl
│ └── train.jsonl
├── wikidata
│ ├── entities
│ ├── meta
│ ├── prompts
│ └── triples_processed
│ └── P2175
│ ├── dev.jsonl
│ ├── test.jsonl
│ └── train.jsonl
└── umls
├── meta
└── prompts
Important: Triples of UMLS is not provided due to the license. For those who want to probe LMs using triples of UMLS, we provide the pre-processing scripts for UMLS. Please follow this instruction.
Experiments
We provide two ways of probing PLMs with BioLAMA:
Manual Prompt
Manual Prompt probes PLMs using pre-defined manual prompts. The predictions and scores will be logged in '/output'.
# Set TASK to 'ctd' for CTD or 'umls' for UMLS
# Set MODEL to 'bert-base-cased' for BERT or 'dmis-lab/biobert-base-cased-v1.2' for BioBERT
TASK=wikidata
MODEL=./RoBERTa-base-PM-Voc/RoBERTa-base-PM-Voc-hf
PROMPT_PATH=./data/${TASK}/prompts/manual.jsonl
TEST_PATH=./data/${TASK}/triples_processed/*/test.jsonl
python ./BioLAMA/run_manual.py \
--model_name_or_path ${MODEL} \
--prompt_path ${PROMPT_PATH} \
--test_path "${TEST_PATH}" \
--init_method confidence \
--iter_method none \
--num_mask 10 \
--max_iter 10 \
--beam_size 5 \
--batch_size 16 \
--output_dir ./output/${TASK}_manual
Result:
PID [email protected] [email protected]
-------------------------
P2175 9.40 21.11
P2176 22.46 39.75
P2293 2.24 11.43
P4044 9.47 19.47
P780 16.30 37.85
-------------------------
MACRO 11.97 25.92
OptiPrompt
OptiPrompt probes PLMs using embedding-based prompts starting from embeddings of manual prompts. The predictions and scores will be logged in '/output'.
# Set TASK to 'ctd' for CTD or 'umls' for UMLS
# Set MODEL to 'bert-base-cased' for BERT or 'dmis-lab/biobert-base-cased-v1.2' for BioBERT
TASK=wikidata
MODEL=./RoBERTa-base-PM-Voc/RoBERTa-base-PM-Voc-hf
PROMPT_PATH=./data/${TASK}/prompts/manual.jsonl
TRAIN_PATH=./data/${TASK}/triples_processed/*/train.jsonl
DEV_PATH=./data/${TASK}/triples_processed/*/dev.jsonl
TEST_PATH=./data/${TASK}/triples_processed/*/test.jsonl
PROMPT_PATH=./data/${TASK}/prompts/manual.jsonl
python ./BioLAMA/run_optiprompt.py \
--model_name_or_path ${MODEL} \
--train_path "${TRAIN_PATH}" \
--dev_path "${DEV_PATH}" \
--test_path "${TEST_PATH}" \
--prompt_path ${PROMPT_PATH} \
--num_mask 10 \
--init_method confidence \
--iter_method none \
--max_iter 10 \
--beam_size 5 \
--batch_size 16 \
--lr 3e-3 \
--epochs 10 \
--seed 0 \
--prompt_token_len 5 \
--init_manual_template \
--output_dir ./output/${TASK}_optiprompt
Result:
PID [email protected] [email protected]
-------------------------
P2175 9.47 24.94
P2176 20.14 39.57
P2293 2.90 9.21
P4044 7.53 18.58
P780 12.98 33.43
-------------------------
MACRO 7.28 18.51
Acknowledgement
Parts of the code are modified from genewikiworld, X-FACTR, and OptiPrompt. We appreciate the authors for making their projects open-sourced.
Citations
@inproceedings{sung2021can,
title={Can Language Models be Biomedical Knowledge Bases},
author={Sung, Mujeen and Lee, Jinhyuk and Yi, Sean and Jeon, Minji and Kim, Sungdong and Kang, Jaewoo},
booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
year={2021},
}
3 Oct 14, 2022
163 Dec 31, 2022
116 Dec 27, 2022
2 May 26, 2022
147 Jan 06, 2023
205 Dec 16, 2022
5.3k Jan 01, 2023
20 Apr 30, 2022
153 Jun 22, 2022
3 May 23, 2022
552 Nov 28, 2022
9 Jun 27, 2022
3 Dec 18, 2021
743 Jan 08, 2023
46 Dec 28, 2022
41 Dec 15, 2022
112 Nov 06, 2022
32 Dec 29, 2022
4.8k Dec 30, 2022
91 Nov 29, 2022