Language Translation for Low Resource Languages

Jassiris Nunez

Co-Presenters: Individual Presentation

College: Hennings College of Science Mathematics and Technology

Major: BS.INFO/TECH

Faculty Research Mentor: Navya Martin Kollapally

Abstract:

This project explores the potential of artificial intelligence (AI) to recognize human emotions through social media posts. With millions of users sharing their thoughts and feelings online, social media provides a rich dataset for understanding human emotions. Given recent societal concerns such as school shootings and mental health crises, developing an AI capable of analyzing emotional expression could help identify warning signs and support early intervention. Using the BERT transformer model as its foundation, this project investigates whether AI can accurately discern emotions like love, sadness, anger, and fear from real-world social media text.
The research builds upon Google’s GoEmotions dataset and involves training and testing an emotion detection system based on BERT and Large Language Models (LLMs). The model analyzes each word in a sentence to infer the emotional tone. Early experiments demonstrated that BERT effectively recognized clear, singular emotions but struggled with mixed or complex emotional expressions—especially in shorter posts. For instance, when processing posts that combined sadness and hope, the model often defaulted to a “neutral” classification. However, accuracy improved significantly with longer posts exceeding 100 words, which provided more context for emotional interpretation.
The team collected over 2,000 posts from Twitter and Reddit, focusing on content related to depression and anxiety. After manual cleaning, 600 posts per platform were selected, ranging from 100 to 500 words. The hypothesis was that longer posts would provide richer contextual cues, enhancing emotion recognition performance. Preliminary results suggest that AI could achieve 75–80% accuracy, approaching human-level understanding, particularly when analyzing nuanced emotions in extended text.
Future directions include fine-tuning the model to detect emotions in shorter posts and developing applications such as mental health chatbots or browser extensions that flag potentially harmful content for authorities. Challenges encountered include API limitations, the difficulty of handling posts with multiple emotions, and the model’s bias toward neutral predictions. Overall, the project demonstrates both the promise and current limitations of transformer-based models like BERT in emotion detection, highlighting their potential role in promoting digital mental health awareness and human well-being.