GradSeg: Unsupervised Word Segmentation Using Temporal Gradient Pseudo-Labels (ICASSP 2023)

Tzeviya Sylvia Fuchs ([email protected])
Yedid Hoshen ([email protected])

GradSeg is an unsupervised approach for word segmentation using pretrained deep self-supervised features. It uses the temporal gradient magnitude of the embeddings (the distance between the embeddings of subsequent frames) to define psuedo-labels for word centers, and trains a linear classifier using these psuedo-lables. It then uses the classifier score to predict whether a frame is a word or a boundary.

If you find our work useful, please cite:

@inproceedings{fuchs23_icassp,
  author={Tzeviya Sylvia Fuchs and Yedid Hoshen},
  title={{Unsupervised Word Segmentation Using Temporal Gradient Pseudo-Labels}},
  year=2023,
  booktitle={ICASSP 2023},
}

---

Installation instructions

• Python 3.8+

• Pytorch 1.10.0

• torchaudio 0.10.0

• numpy

• boltons

• Download the code:

  git clone https://github.com/MLSpeech/GradSeg.git
  

How to use

In this example, we will demonstrate how to run GradSeg on the Buckeye corpus.

• We use the same experimental setup as in "DSegKNN: Unsupervised Word Segmentation using K Nearest Neighbors (INTERSPEECH 2022)" (Paper, Code, see README file there for data preprocessing).

• Run grad_segmenter.py with the following options:

   python grad_segmenter.py --min_separation 3 
   			 --train_n 100 
   			 --eval_n -1 
   			 --reg 1e7 
   			 --target_perc 20 
   			 --frames_per_word 15
   			 --train_path datasets/buckeye_split/train/
   			 --val_path datasets/buckeye_split/val/
  
  

  Result should be:

   Final result: 45.77675489067894 44.84274602637809 45.30493707647628 -2.040356216886474 53.62285114327586
  

  which are the precision, recall, F-score, OS, and R-value.

• For comparison, the evaluation script eval_segmentation.py used here is by Herman Kamper.