Attending to Emotional Narratives

2019 8th International Conference on Affective Computing and Intelligent Interaction (ACII)

cite contains a model implementation for our paper. If you find this code useful in your research, please consider citing:

@inproceedings{wu2019attending,
  title={Attending to Emotional Narratives},
  author={Wu, Zhengxuan and Zhang, Xiyu and Zhi-Xuan, Tan and Zaki, Jamil and Ong, Desmond C.},
  journal={IEEE Affective Computing and Intelligent Interaction (ACII)},
  year={2019}
}

What it is?

This repo consist 5 different deep neural networks predicting emotion valence with multi-modal inputs. This repo contains evaluation scripts and pre-trained models.

Description

Attention mechanisms in deep neural networks have achieved excellent performance on sequence-prediction tasks. Here, we show that these recently-proposed attention-based mechanisms---in particular, the Transformer with its parallelizable self-attention layers, and the Memory Fusion Network with attention across modalities and time---also generalize well to multimodal time-series emotion recognition. Using a recently-introduced dataset of emotional autobiographical narratives, we adapt and apply these two attention mechanisms to predict emotional valence over time. Our models perform extremely well, in some cases reaching a performance comparable with human raters. We end with a discussion of the implications of attention mechanisms to affective computing.

Models

Memory Fusion Transformer

Simple Fusion Transformer

Baseline 1 - LSTM Model For Multi-modal Inputs

Baseline 2 - Transformer With A Linear Header For Multi-modal Inputs

Baseline 3 - Memory Fusion Network

How To Run?

Requirements:

VM with image, Data science for linux, on Microsoft Azure

Command:

python train.py (with all default settings)