pratt3000 / transcript-interpretor

When AI agents interface with a customer, it extracts the implied date that the customer will make a payment. These conversations are unstructured and dates can be communicated in a nebulous fashion.

How to run

Installing requirements

pip install -r requirements.txt

Training the model

train.ipynb
Note: This will train the model on data from chatgpt_gen_date.json

Predict

predict.py --conversation <str> --date <str> --label_type <str> --ensemble <bool>
Sample:
predict.py --conversation "Agent: Hi, I'm Taylor, calling from Westlake Financial on a recorded line. Unfortunately, we did not receive your monthly payment! Would you be able to make a payment today?\nCustomer: yeah i did i did it through text but that that bring the house for this one and then press two or whatever\n" --date "2020-01-01" --label_type "label" --ensemble

Evaluation

eval.py
Note: This will evaluate the model on data from test_data.json

Make data to proper format:

format_data.py

Other details

1. Trained model weights are stored on hugging face.
2. Final model used - Finetuned FlanT5-large
3. Extra model weights (for model 1, 2, 3 (described in Approach_explanation.pdf)) uploaded to huggingface just in case. Although a few other changes would need to be made to the code to run using these weights.

Approach Explanation:

The details are in approach_explanation.pdf

Compute used

Kaggle - GPU: P100 (16 GB)