Training chatbots to improve their Natural Language Understanding (NLU) skills.

One of the main factors that can influence the effectiveness of a chatbot is its ability to understand natural language (NLU). Our goal was to create a system that would train chatbots to understand and process human language accurately, interpreting user intentions and responding in a relevant manner.
We developed a SaaS that allowed for the creation of models tailored to different business areas or objectives that could be tested and improved on the fly.

The goal.

The training system was meant to help chatbots better understand the meaning behind user queries, grasp local idioms, technical terms, legal jargon, or company-specific language, and based on that, provide more relevant and useful responses to customers using these chats on their websites, apps, and systems.

The beginning.

The programmers, without any help from a designer, had developed a system that fulfilled the task, but it was as difficult to use as it was unpleasant to look at (sorry, guys!). Furthermore, the task flows were lengthy, disconnected from one another, making the process of training the bots rather tedious and truly challenging.

The most important thing.

Analyzing the system of the guys, I understood that one of the most important and underutilized flows was the testing of the chatbots. So, I set to work to propose ways in which the conversation testing could be more natural, simple, practical, and useful; a process that not only yielded a reliability percentage or error data but also dynamically allowed feeding the model while testing the chatbot.

The obvious question.

Despite the advancement of AI, do we still need to train a bot? Yes, because despite the vast amount of knowledge an AI can handle, it cannot keep up-to-date (for now) with all regionalisms, specific services, or policies of all companies. Additionally, in many cases, companies want to train their chatbots to respond in a very particular and specific manner regarding their services or products.

Weeks of work

12

Designed task flows

23

Tasks related to managing workspaces, models, intents, utterances, test, and conversation flows.

Designed screens

42

Screens made from scratch.