an easy-to-use interface to fine-tuned BERT models for computing semantic similarity. that's it.
This project contains an interface to fine-tuned, BERT-based semantic text similarity models. It modifies pytorch-transformers by abstracting away all the research benchmarking code for ease of real-world applicability.
Model | Dataset | Dev. Correlation |
---|---|---|
Web STS BERT | STS-B | 0.893 |
Clinical STS BERT | MED-STS | 0.854 |
Install with pip:
pip install semantic-text-similarity
or directly:
pip install git+https://github.com/AndriyMulyar/semantic-text-similarity
Maps batches of sentence pairs to real-valued scores in the range [0,5]
from semantic_text_similarity.models import WebBertSimilarity
from semantic_text_similarity.models import ClinicalBertSimilarity
web_model = WebBertSimilarity(device='cpu', batch_size=10) #defaults to GPU prediction
clinical_model = ClinicalBertSimilarity(device='cuda', batch_size=10) #defaults to GPU prediction
web_model.predict([("She won an olympic gold medal","The women is an olympic champion")])
More examples.
- You will need a GPU to apply these models if you would like any hint of speed in your predictions.
- Model downloads are cached in
~/.cache/torch/semantic_text_similarity/
. Try clearing this folder if you have issues.
Clinical models in this project were submitted to the 2019 N2C2 Shared Task Track 1. Implementation and model training in this project was supported by funding from the Mark Dredze Lab at Johns Hopkins University.