Classifier neural network for a chatbot application

Author: David Adamashvili
Co-authors: Davit Adamashvili, Beqa Mamacashvili, Ramaz Bluashvili
Keywords: neural network, classifier, NLP, BERT, transformer
Annotation:

Nowadays, as the rate of online users increases tremendously, one of the challenges is the ability to communicate automatically with them. Most companies use chatbots to communicate with their customers on their websites or social networks. The chatbot is software that finds out what are the user needs and helps them to solve their problems. These chatbots either provide customers with a large menu with lots of choices or connect with company employees. It is not convenient for the user to navigate such large menus and make the desired choice, sometimes it is even impossible: for example for visually impaired people. We believe that the integration of artificial intelligence in modern chatbots will give us much better results, so we created our program-bot, which uses neural networks to identify the "content of conversation". We use Transformer-based machine-learning technology BERT, developed by Google for natural language processing. We have added a new layer so any company can input their data and the program will train the neural network on it. More precisely, the company will be able to adopt a neural network that will help its developer to create a more comfortable environment for the customer. The chatbot will no longer need to list all the options, these options will be combined into different classes. If the company has a sufficient amount of data and all options are given a separate class, the chatbot will no longer need to list anything at all. Note, that such an approach makes the user interfaces cleaner because from the beginning it will not give you a lot of choices, but offer only relevant ones. This way we can simplify navigation for users and make chatbot useful also for those people who want to use Text to speech instead of physically typing text as they will be able to make requests to the chatbot directly.