Improving Emotion Recognition Accuracy with Combination of Bidirectional and Long Short-Term Memory Models

Abstract

Emotions play a vital role in shaping human behavior and mental health, making accurate emotion recognition essential for mitigating potential negative impacts. This study explores the application of Bidirectional Long Short-Term Memory (Bi-LSTM) for recognizing emotions from text-based data. Bi-LSTM extends the standard LSTM by enabling the model to process input sequences in both forward and backward directions, thereby capturing contextual dependencies more effectively.  The research methodology consists of data collection, manual emotion labeling, and pre-processing techniques, including stemming, tokenization, and one-hot encoding. Visualization of the dataset and the distribution of labeled emotions was conducted to gain deeper insights into the data. The Bi-LSTM model was trained for 25 epochs, achieving a training accuracy of 0.9954 and validation accuracy of 0.8790, along with a training loss of 0.0133 and validation loss of 0.658. A confusion matrix was used to further evaluate model performance and classification accuracy across various emotion categories.  The experimental results confirm that the Bi-LSTM model is highly effective in recognizing emotions from textual input. Its ability to capture long-term dependencies in both directions contribute to improved learning and prediction. However, opportunities for enhancement remain, particularly in refining the model architecture, expanding the dataset, and exploring additional feature extraction techniques.  This research demonstrates the potential of Bi-LSTM in building intelligent emotion-aware systems for applications in mental health monitoring, customer feedback analysis, and human-computer interaction.