Sentence-USE: Universal Sentence Encoder with Siamese Architecture

This repository contains scripts to finetune the USE for a given semantic textual similarity dataset.

Train and evaluate on STSb dataset

Download STSb dataset

To download and prepare the STSb dataset, run the following script:

python sentence_use/data/stsb.py

It will download the train, dev and test datasets.

Evaluate

To run the evaluation on the STSb test benchmark and calculate the pearson correlation and the spearman`s rank correlation to the groundtruth, one can run the following script:

python evaluate.py --eval-data=stsb_test.csv \
                    --model-name-or-path=https://tfhub.dev/google/universal-sentence-encoder/4

It should output the evaluation metrics:

Pearsons correlation: 0.7873, Spearman`s rank correlation: 0.7709,

Run training

Expects a training and validation dataset. Dataset needs to be provided as csv file with score, sentence1 and sentence2 columns. Scores must be in range [0, 1].

python train.py --train-data=stsb_train.csv \
                --val-data=stsb_dev.csv \
                --model-name-or-path=https://tfhub.dev/google/universal-sentence-encoder/4 \
                --lr=0.0001 \
                --epochs=1 \
                --batch-size=8

The script will save the model under the path provided with the output-path. Per default the model is saved under savedmodel.

Evaluate again

To run the evaluation again after the training, run the evaluate script again and provide the path to the trained model.

python evaluate.py --eval-data=stsb_test.csv \
                   --model-name-or-path=savedmodel

Output:

Pearsons correlation: 0.8042, Spearman`s rank correlation: 0.7915

References

Yang, Y., Cer, D., Ahmad, A., Guo, M., Law, J., Constant, N., ... & Kurzweil, R. (2019). Multilingual universal sentence encoder for semantic retrieval. arXiv preprint arXiv:1907.04307.
Cer, D., Yang, Y., Kong, S. Y., Hua, N., Limtiaco, N., John, R. S., ... & Kurzweil, R. (2018). Universal sentence encoder. arXiv preprint arXiv:1803.11175.
Cer, D., Diab, M., Agirre, E., Lopez-Gazpio, I., & Specia, L. (2017). Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055.