InterSent

Code for our paper Bridging Continuous and Discrete Spaces: Interpretable Sentence Representation Learning via Compositional Operations at EMNLP 2023

Requirements

• transformers == 4.18.0
• pytorch-lightning == 1.6.1

Data

• ParaNMT
• Discofuse Wikipedia balanced portion
• Wikisplit
• Google Sentence Compression The data folder should have a similar structure as the following:

└── data 
    └── paranmt
        └── para-nmt-5m-processed.txt
    └── discofuse
        ├── discofuse-train-balanced.txt
        └── discofuse-valid-balanced.txt
        └── discofuse-test-balanced.txt
    └── wikisplit
        ├── wikisplit-train.txt
        └── wikisplit-valid.txt
        └── wikisplit-test.txt
    └── google
        ├── sent-comp-train.txt
        └── sent-comp-test.txt

Training

To train InterSent from scratch, run the following:

bash train.sh

Evaluation

To evaluate InterSent on interpretability, run the following with your checkpoint path:

bash test.sh

To evaluate InterSent on STS, run the following with your checkpoint path:

bash stseval.sh