Textual emotion detection utilizing a transfer learning approach.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Traditional deep learning models (LSTM, GRU, BiLSTM) face challenges in automated textual emotion detection.
• These models require extensive datasets, significant computational resources, and long training durations.
• They exhibit limitations with smaller datasets and are prone to catastrophic forgetting.

Purpose of the Study:

• To demonstrate the efficacy of transfer learning techniques for enhanced textual emotion detection.
• To showcase the ability of pre-trained models to capture contextual meaning with reduced data and training time.
• To compare the performance of transfer learning models against traditional recurrent neural network (RNN)-based models.

Main Methods:

• Utilized EmotionalBERT, a pre-trained model based on bidirectional encoder representations from transformers (BERT).
• Conducted experiments on two benchmark datasets to evaluate model performance.
• Focused on analyzing the impact of varying training data amounts on model effectiveness.

Main Results:

• Transfer learning with EmotionalBERT achieved superior performance in textual emotion detection compared to RNN-based models.
• The study highlighted the effectiveness of EmotionalBERT even with limited training data.
• Reduced training time was observed when using the transfer learning approach.

Conclusions:

• Transfer learning, specifically using EmotionalBERT, offers a more efficient and effective solution for automated textual emotion detection.
• This approach mitigates the need for large datasets and extensive computational resources.
• EmotionalBERT demonstrates robust performance, particularly in scenarios with limited data availability.