PetKaz at SemEval-2024 Task 3: Advancing Emotion Classification with an LLM for Emotion-Cause Pair Extraction in Conversations

This repository provides the code necessary to replicate the paper PetKaz at SemEval-2024 Task 3: Advancing Emotion Classification with an LLM for Emotion-Cause Pair Extraction in Conversations from SemEval-2024.

Our pipeline has two parts: 1) (non-)emotional utterances classification with fine-tuned gpt-3.5-turbo-1106; 2) cause extraction with a simple CauseExtractor neural network.

First, please install dependencies:

pip install -r requirements.txt

Note: itilially we divided an original training data (data/Subtask_1_train.json) into data/Subtask_1_train_real.json and data/Subtask_1_dev.json for our needs.

Emotion classification

If you want to fine-tune an emotion classifier based on gpt-3.5-turbo-1106, you need to go through the following steps.

1. Prepare data, run:

   python src/data4finetuning.py
   

   This will create files data/train_main_2.jsonl and data/test_main_2.jsonl obtained from data/Subtask_1_train_real.json and data/Subtask_1_dev.json respectively. They contain prompts and responses. The prompt template is in data/prompt_main.txt. We use the target utterance and the previous one as a context for classification. Note that we used the whole dataset but reduced the number of neutral utterances to balance it.

2. Fine-tune the gpt-3.5-turbo-1106 model using the OpenAI fine-tuning UI. Please refer to their documentation.

Here is our model:

Unfortunately, we cannot share our model due to the OpenAI policy.

To use the model, you should run:

python src/emotion_annotation.py

This will annotate utterances from data/Subtask_1_test.json and save them to data/Subtask_1_test_gpt.json. Do not forget to specify api_key and model (in get_completion).

Cause extraction

If you want to train a CauseExtractor neural network by yourself, run:

python src/training.py

It will save the model to model/new_best_model.pt (now our model is model/best_model.pt; it is also used for prediction).

To run prediction, you need to have data/Subtask_1_test_gpt.json annotated with the emotion classifier in advance, then:

python src/prediction.py

It will save predictions to data/Subtask_1_pred.json.

Note that all the files mentioned in this README.md are already saved in the data directory.