Modeling Empathetic Alignment in Conversation

This repo contains all code for paper Modeling Empathetic Alignment in Conversation.

Code for the annotation website tool of this project is here.

To use the trained model directly, please check Huggingface:

• Appraisal prediction
• Alignment prediction

General Work-flow

Two models are performed in sequence for the final prediction of whether two appraisal aligned or not:

1. Appraisal prediction
2. Alignment prediction given appraisals

Input Format

The input format is the same for both tasks.

The input file is a json file. Easier to prepare the data by creating a pd.Dataframe with the following labels and convert it to json file:

id	target_id	observer_id	parent_id	subreddit	target_text	observer_text	distress_score	condolence_score	empathy_score	full_text	spans	alignments
int	string	string	string	string	string	string	float	float	float	string	[start, end, label], ... ]	[[(target_start, target_end), (observer_start, observer_end)], ...]

Note: For spans and alignments entry, start and end are set relative to full_text which contains prefixes for target_text and observer_text with the format of:

target:\n\n${target_text}\n\nobserver:\n\n${observer_text}

Appraisal Prediction

Preprocess data

Command is in preprocess/run.sh.

The output file is a pickle file that can be directly used for training appraisal models.

Appraisal models

Baselines

Commands are in baseline/run.sh.

Note: For majority_baseline, Self-other agency is always predicted.

BERT models

Commands are in bert/run.sh. Modify for your own use (data paths, hyper-parameters etc.)

All available arguments are at the top of bert/train.py.

evaluation_only not available.

OpenPrompt models

Commands are in openprompt/run.sh. Modify for your own use (data paths, hyper-parameters etc.)

All available arguments are at the top of openprompt/train.py.

evaluation_only available.

Alignment Prediction

Preprocess data

You can:

1. Directly use the json input format in training script OR;

2. Preprocess the data with snn/run_quick_preprocess.sh and specify the argument --load_processed_dataset in the training script.

Train alignment models

Commands are in snn/run.sh. Modify for your own use (data paths, hyper-parameters etc.)

All available arguments are at the top of snn/train.py.

evaluation_only available.