15:45 - 16:30
The opportunity of empowering business by replacing human labor with machines without affecting customer satisfaction is attracting the community more and more nowadays. Even conversational agent concept is not new, the conversational agent (or specifically chatbot) development became a very hot topic in both the research and software development areas in the last decade with the advances in artificial intelligence. In this paper, we introduce the development and deployment of a Turkish chatbot in the telecom domain. We started with the general architecture and explained every component including Dialog Engine, Middleware, and the NLP/NLU unit. We explain the main architecture of the chatbot, the data collection and training steps for intent classification, and our evaluation strategy for calculating the performance. We propose two different approaches for intent classification, using conventional methodology and Transformers. We employed both ML algorithm Fasttext and the deep learning methods BERT and ELECTRA for the intent matching. We introduced novel evaluation parameters which may shed light on other commercial chatbot evaluations. WOur chatbot is deployed for a telecom company with more than $25M$ customers and $15M$ active mobile application users.