Multiple Teachers Model for Ranking

The repository contains code implementation and dataset for paper Multilingual Serviceability Model for Ranking Help Requests on Social Media during Crisis Events.

1. Download tweets based on ids' presented in the folder data.
2. Create 3 files: train.csv, val.csv, test.csv. Each file should contain two colimns: 'text' and 'label'.

Datasets for behavioral fine-tuning:

• sarcasm and irony detection dataset link
• question type classification dataset link

After downloading datasets, create 3 files train.csv, val.csv, test.csv and save the in the folder inside the data folder. Each file should contain two colimns: 'text' and 'label'.

The code was tested on Python 3.10.

Install necessary dependencies with use of pip:

pip -r requirements.txt

1. Command for task-related fine-tuning:

python finetune.py bert data teacher_1

where

• bert -- name of the model (could be bert or roberta)
• data -- folder with data
• teacher_1 -- name of folder for saving the task-related Teacher model

2. Command for behavioral fine-tuning:

python finetune.py bert data/sarcasm teacher_2

where

• bert -- name of the model (could be bert or roberta)
• data/sarcasm -- folder with data for behavioral fine-tuning
• teacher_2 -- name of folder for saving the behavior-guided Teacher model

Command for model distillation:

python distillation.py model/teacher_1 model/teacher_2 model/teacher_3

where model/teacher_1, model/teacher_2, and model/teacher_3 - paths to finetuned models.