Designing a Transactional Smart Assistant in Indonesian using Rasa Framework

The use of chatbots in the industrial world is increasing along with the increasing number of smartphone users in Indonesia. Many companies take this opportunity to develop their chatbot services with the help of Natural Language Processing (NLP) and Neural Network (NN) technologies. In this study, the creation of an Indonesian-based chatbot is more emphasized on the questions and answer domain, complaint handling, and transaction support at “pulsa” distributor company. Chatbots must have the ability to understand natural language, provide accurate predictions of subsequent actions and responses, especially for Indonesian chatbots. We compared several pipeline configurations in building the Natural Language Understanding (NLU) model, namely the default Rasa NLU with DIETClassifier, pretrained FastText with DIETClassifier and Spacy Multi Language Model using FastAi’s ULMFit for intent classification. The formation of the intent classification based on the complaint data log, customer service chat history and company FAQ resulted in 216 intents. The evaluation results on the pipeline configuration show that the intent classification and entity extraction tasks are better using the pretrained FastText model with an F1-Score intent evaluation reaching 73.4% and entity extraction 83%.