Automated Customer Support With Conversational Agents Employing Text Mining: A Case of online University Application
The use of online information systems that can be and are accessed on a 24 hour basis has continued to grow both locally and globally, these systems require available customer support at all times. Customer support is services that assist customers to make cost effective and correct use of products or services; it is divided into voice-based and non-voice based. Requests received by support personnel are routine in nature and their numbers surpass the staff employed to handle them hence organizations need to device mechanisms of automating customer support to ensure available staff deal with requests that have been escalated due to their uniqueness as opposed the routine ones; thereby enhancing productivity and reducing staff related costs and improving the response times. This study demonstrates automation of customer support, through design and development of an automated customer support system that uses the online university application as a case study. The system employs conversational agents (CA) which are computer systems that are intended to converse with human beings. The demonstrated CA use classification algorithms specifically Naive Bayes (NB) algorithm and Latent Semantic Indexing (LSI) to categorize the emails received. The developed application can process and respond to 500 emails in 3 minutes, while support personnel can handle the same in approximately 5000 minutes, hence demonstrating its applicability in possibly reducing staff costs and improving response times. A comparative analysis of responses generated was performed on a subset of the sample support emails to determine the accuracy of the application as compared to the customer support personnel. NB algorithm was found to have an error rate of 88 percent while LSI was found to have an accuracy of 74%. Further evaluation conducted on the algorithms resulted in LSI accuracy 0.89 and FScore 0.94 while NB accuracy 0.94 and F-Score 0.96. The study recommends that any future studies to study the improvement of accuracy of the NB classifier and system enhancement to be able to handle multi- part emails.