Discovering Emotion and Reasoning its Flip in Multi-Party Conversations using Masked Memory Networks and Transformer

This folder contains the code required to reproduce the results of the paper - "Discovering Emotion and Reasoning its Flip in Multi-Party Conversations using Masked Memory Networks and Transformer".

The paper introduces a novel problem, called Emotion-Flip Reasoning aka EFR in conversations. The goal is to find all utterances that trigger a flip in emotion of a speaker within a dialog. An example scenario is depicted in the figure below:

There is a single emotion-flip, i.e., u_6 (neutral) -> u_8 (fear). The responsible trigger utterances are marked with arrows in the figure.

We augment MELD, a benchmark ERC dataset with ground-truth EFR labels. The resultant dataset, called MELD-FR, contains ~8,500 trigger utterances for ~5,400 emotion-flips.

Contribution

In this work, our major contributions are four-fold:

1. We propose a novel task, called emotion-flip reasoning (EFR), in the conversation dialog.
2. We develop a new ground-truth dataset for EFR, called MELD-FR.
3. We benchmark MELD-FR through a Transformer-based model and present a strong baseline for the EFR task.
4. We develop a masked memory network based architecture for ERC, which outperforms several recent baselines.

Dataset

We provide the dataset along with the code ("Data/MELD_<set>_efr.csv"). You can use your own dataset if you follow the expected format for the data. The models in this repo expects the data to follow the following structure:

Dialogue Id	Speaker	Emotion	Utterance	EFR Label

Models

Emotion Recognition in Conversations (ERC)

We employ a memory network to supplement the global and local emotion dynamics in dialogues captured through a series of recurrent layers. In the repo we provide with 3 models that perform the ERC task.

• ERC-MMN
• EFR-ERCmulti
• EFR->ERCcas (Best Performance)

Emotion Flip Reasoning (EFR)

We employ a transformer-based model to model the EFR task. We provide 4 models to perform the ERC task in this repo.

• EFR-TX
• EFR-ERCmulti
• ERC->EFRcas
• ERCTrue->EFRcas (Best Performance)

Code Reproducibility

The results reported in the paper are achieved on the full dataset and hence would not be achieved by the dataset provided in this folder. However you can run the codes by following these steps:

1. Run the Dataloaders for ERC and EFR tasks respectively to get the necessary pickle files.
2. Run the train_model file where model can be one of "erc_mmn", "efr_tx", "multitask", "cascade_efr_erc", "cascade_erc_efr" and "cascade_ercT_efr".

Running the code for custom data

Since our model uses BERT embeddings as utterance reprsentation, first you'll need to generate these BERT embeddings for your data. After you have a dictionary mapping for your utterances to embeddings, you'll need to change the paths in the ERC and EFR dataloaders. Specifically, the paths for training file, testing file and embedding files needs to be changed. After executing the teh dataloaders you'll get the necessary pickle files to run the models.