This repo hosts the code for

@misc{shen2023wave,
      title={Wave to Syntax: Probing spoken language models for syntax}, 
      author={Gaofei Shen and Afra Alishahi and Arianna Bisazza and Grzegorz Chrupała},
      year={2023},
      eprint={2305.18957},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Clone repo and set up and activate a virtual environment with conda:

conda create --name wav2syn --file requirements.txt
conda activate wav2syn

The exact configuration of the conda environment used to conduct the experiments can be found in spec-file.txt

modify the dataset paths within and run python preprocessing.py to generate a dataset csv file for the textCorpus and Corpus classes.

preprocessing.py also uses the stanza parser to save all the constituency trees for each utterance in the dataset.

This project have implemented the embedding extraction script for LibriSpeech and SpokenCOCO corpus. You can download the two corpora from the links here SpokenCOCO LibriSpeech.

After downloading and extracting the datasets, read the main() function in preprocessing.py and change the root directories of the datasets and splits you want to use.

Running preprocessing.py generates dataset csv files that can be understood by the probing scripts.

The preprocessing script also makes the bag-of-words model at the same time.

This repo uses Huggingface Hub to load models.

If you would like to replicate findings with the FaST-VGS family of models, please check instructions on https://github.com/jasonppy/FaST-VGS-Family.

Feature extraction have been implemented in embedding_generation.py. It might be desirable to modify embedding_generation.py to limit what model you want to investigate on. The script will save the extracted features including the meanpooled layerwise embedding, the treedepth, the annotation, the audio path, audio length and wordcount to a .pt file under embeddings.

Run

python treedepthprobe.py >> treedepth.out

Run

python treekernel_prep.py
python treekernelprobe.py >> treekernel.out