Abstract:
The rapid expansion of e-wallet services in Indonesia necessitates robust customer service to maintain competitiveness
and user satisfaction. Current customer service chatbots in Indonesian e-wallet services primarily rely on rule-based approaches,
limiting their adaptability to user needs and resulting in communication issues and negative feedback. The transition to AI-based
chatbots is challenging, particularly in accurately classifying user intents in Indonesian-language due to the language's complexities
and the absence of specialized models and datasets. This research proposes a customized intent classification model for AI-based
customer service chatbots in e-wallet services, employing transformer-based embedding methods, specifically IndoBERT and
Sentence-BERT (SBERT), with TextConvoNet classification model. Comparative analysis is conducted against conventional
transformer models and the original TextConvoNet framework. The findings consistently showcase the superiority of the proposed
models across various metrics, demonstrating significant advancements compared to baseline approaches. Notably, SBERT
embeddings with TextConvoNet classification achieved the highest accuracy (86.90%), precision (84.81%), recall (86.90%), and F1-
score (85.11%) with a learning rate of 0.001, indicating its potential to enhance customer service chatbots in e-wallet platforms.
These findings not only advance AI-driven customer service within the financial sector but also offer valuable insights into the
broader application of natural language processing technologies for addressing real-world challenges
