Neural Networks based Dialogue System for Customer Support

Abstract

Customer service is one of the most important components of online services. However, as natural language processing methods are on the rise, the market is looking at automated conversational models based solutions to deliver high-quality services to a user base that is always expanding. The Deep Learning based con versational agent (CA) is a challenging Natural Language Processing (NLP) task in a language with poor resources like Urdu, as well as the scalability and gen eralisation capacities of the neural conversational models that were lacking in previously employed manually annotated and rule-based systems. Although con versational agents have been developed for other languages, recent state-of-the-art neural network-based techniques have not yet been investigated for conversational agents in Urdu. We have compiled a dataset of about 12000 question-answer pairs and implemented two basic deep learning architectures: Long Short Term Mem ory (LSTM) with and without Attention mechanism. These have been used in our work to examine the powerful deep learning techniques for an Urdu conversa tional agent in the customer support area. In this study, we developed an Urdu conversational agent model based on Transformer that entirely follows the atten tion mechanism. The suggested and baseline methodologies were implemented on Urdu and English customer care datasets from Amazon, where the suggested model outperformed all other deep learning techniques when the results of these techniques were examined. The Transformer attained a BLEU score of 38.13, 40.2, and 31.7 on the small, large, and English data sets, respectively, outperforming the basic deep learning models.