Sentiment-based masked language modeling for improving sentence-level valence–arousal prediction

Sentiment word masking bert for valence and arousal

Introduction

This repository contains the code for replicating results from

• Sentiment-based masked language modeling for improving sentence-level valence–arousal prediction
• Jheng-Long Wu; Wei-Yi Chung
• Applied Intelligence, 2022

Flow Chart

Getting Started

• Build a new virtual environment
• Install python3 requirements: pip install -r requirements.txt
• Run cd ./Models to the model folder
• Choose the parameter you want to run in the sh files to train different models
• Train your own models

Training Insturctions

• Experiment configurations are found in ./Models/*.sh
• Results model and logs are stored in the corresponding output directory under VA_BERT_mask_sentiment*.

Adding Sentiment Words

• Under the sense_data folder, that is negative sentiment word and positive sentiment word, you can add the sentiment word to mask it

Other Insturctions

• result_statistic can grep all the prediction result and indicator to the csv file
• Graph stores the visul of experimental result from the paper
• data contains the experimental data which already split to five fold, the split code and the statistic of masking coverage code is here too
• notebooks stores the multilabel regression model and the visulization code