Year: 2025
Type: conference-paper
Abstract: Online support groups for smoking cessation provide an accessible, cost-effective method for assisting individuals looking to quit. However, these groups often face challenges related to low user engagement and the stigma associated with smoking cessation. Non-native English speakers encounter additional barriers in accessing effective support, as language differences increase feelings of isolation and discourage participation. To address these issues, this study focuses on developing a multilingual conversational agent designed to offer real-time, culturally sensitive support in multiple languages, including Mandarin, Arabic, Hindi, Spanish, and Farsi. A critical aspect of the system is intent classification, enabling the agent to understand user inputs and generate contextually appropriate responses. Due to the limited availability of high-quality training data in non-English languages, we introduce a synthetic data augmentation strategy that combines both synthetic data generation and real data scraping. Initially, we fine-tune a pre-trained large language model on multilingual smoking cessation group data. To enhance intent detection accuracy, we augment the training dataset by generating synthetic posts using advanced prompt engineering. We compare the performance of the fine-tuned LLM with its non-fine-tuned counterpart, demonstrating that synthetic data augmentation results in a 35.5% increase in F1 scores. This improvement validates the effectiveness of our augmentation approach in boosting the accuracy of non-English conversational agents. By integrating this conversational agent into widely used social media platforms, such as GroupMe or X, we aim to offer a scalable, inclusive solution that reduces stigma, enhances engagement, and improves the effectiveness of smoking cessation programs for non-native English speakers.
Authors: Salar Hashemitaheri, Ian Harris
Institution: University of California, Irvine
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W4416144406