Design and implementation of a multilingual chatbot for global fashion E-commerce CX using the RASA machine learning framework

The increasing globalization of markets necessitates robust multilingual customer service solutions to meet diverse consumer linguistic preferences. Traditional customer service systems, often constrained by language barriers, limited availability, and scalability issues, are insufficient to meet the demands of a globalized customer base. This paper presents the design and implementation of a multilingual chatbot leveraging the open-source RASA machine learning framework to address these challenges. The proposed architecture integrates a language detection module, a multilingual Natural Language Understanding (NLU) pipeline, a sophisticated Core Dialogue Manager, and an Action Server capable of dynamic response translation. The chatbot supports three languages: English, Devanagari Hindi, and Hinglish, using a custom-built multilingual dataset comprising 1350 utterances manually annotated, 450 utterances for each of the three supported languages. The dataset comprises 11 customer-service intents , with the three language-specific intents represented by a larger number of training examples than the supporting conversational intents, reflecting the primary objective of multilingual language identification and routing. The dataset was divided into training and testing sets using an 80:20 split. Evaluation metrics, including Accuracy, Precision, Recall, and F1-score for intent classification, demonstrate strong performance across all three languages. The proposed RASA-based system outperforms a non-multilingual RASA baseline and a Translate-to-English pipeline, while achieving competitive performance relative to the standalone fine-tuned XLM-R model, with additional advantages in end-to-end dialogue management, deployment flexibility, and data privacy. However, the Hinglish component requires further development. User testing results, encompassing Customer Satisfaction (CSAT), average resolution time, and overall satisfaction rates, further validate the system’s practical utility. Key outcomes highlight RASA’s flexibility in adapting to diverse linguistic requirements and its potential to significantly enhance global customer support by providing instant, culturally sensitive, and cost-effective AI assistance. The findings underscore the importance of robust NLU and a hybrid approach to translation for maintaining semantic consistency across languages, while also identifying critical research gaps in domain-specific multilingual corpora, particularly for code-switched varieties such as Hinglish.