https://github.com/rbroc/sbert-align

Compute alignment using SentenceBERT
https://github.com/rbroc/sbert-align

Last synced: 28 days ago
JSON representation

Compute alignment using SentenceBERT

• Host: GitHub
• URL: https://github.com/rbroc/sbert-align
• Owner: rbroc
• Created: 2022-12-09T08:31:13.000Z (about 2 years ago)
• Default Branch: main
• Last Pushed: 2023-12-18T08:55:07.000Z (about 1 year ago)
• Last Synced: 2024-10-24T11:52:07.866Z (3 months ago)
• Language: Jupyter Notebook
• Size: 11.1 MB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# sbert-align

Compute parent-child alignment using SentenceBERT

### Usage

1. Create a virtual environment (not necessary).

You can do so by typing:

``` 

python3 -m venv PATH_TO_ENV

source PATH_TO_ENV/bin/activate

```

Replace `PATH_TO_ENV` with path for virtual environment

2. Install requirements

```pip install -r requirements.txt```

3. Run the `align.py` script.

`python3 align.py --lag 1 --model all-mpnet-base-v2`.

Arguments are customizable.

Note that the script will be looking for a `transcripts.txt` or `surrogates.txt` file in the `data` folder, and outputs will be saved in an `outputs` folder.

4. Deactivate once you're done, by running ```deactivate```.

### Output columns

- Turn metadata: (`ChildID|ID`, `Visit`, `Turn`)

- `Lag`: 1 if alignment is computed with previous turn, 2 if two turns back. Note that even numbers compute alignment with previous turns from same speaker;

- `ModelId`: Which SentenceBERT checkpoint we are using, see https://www.sbert.net/docs/pretrained_models.html for available models; 

- `SemanticAlignment`: cosine similarity between sequence encodings;

- `AlignmentType`: 'child2caregiver' or 'caregiver2child'

### Notes on study 2

- We keep the second iteration of a conversation, when the Turn ID is repeated

- Turns where previous index is missing are not coded for current-to-1back alignment, and for 1back-to-2back alignment

- Turns where the preceding turn does not follow the previous one are not coded for  1back-to-2back alignment

### Potential expansion:

- Make synthetic raw data for better reproducibility