rbroc / sbert-align

Compute alignment using SentenceBERT

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