Over the past two decades, conversational Artificial Intelligence has become an increasingly prevalent part of our daily lives. With companies relying heavily on the use of chatbots for e-commerce, customer service, and education, it is safe to say that these technologies are not going away any time soon. While machine learning based chatbots provide revolutionary advances in the way these companies conduct business online, they are often vulnerable to conversational bias emanating from toxic training data. If left unchecked, these chatbots have the potential of reflecting offensive elements of biased conversation. In this paper, we develop a novel approach to eliminating bias from training data, including user input. More specifically, we create a filtering algorithm that assesses the toxicity level of a chatbotÕs response and eliminates statements from the training data that surpass a predetermined threshold of conversational bias. Our model includes a toxicity assessment framework that evaluates such a bias based on the content of a given statement, as well as a toxicity scoring system that evaluates the level of bias present based on this framework. Our chatbot implements this technique by evaluating each statement in its initial training dataset, as well as new user input, and filtering out statements that contain high levels of toxicity so that harmful outcomes are successfully mitigated.

PDF