Stepwise Noise Elimination for Better Motivational and Advisory Texts Classification

There is little research into designing artificial motivational agents. The end-goal of our studies is therefore to create a dialogue system that would motivate users to do their everyday tasks using natural language. In this paper, we present a method of distinguishing texts containing motivational advice from regular texts to sort out noise in training data for our dialogue system. We implemented a novel method of chaining two shallow networks together by utilizing the output results of the first network to determine the input for the second one. We achieved F-score of 0.94 and 0.97 with our proposed method. The contributions of this paper are threefold: first, we successfully identified 14 hand-crafted features that make a text motivational/advisory. Secondly, we were able to create a classifying algorithm that distinguishes motivational/advisory texts from regular ones. Finally, our proposed method can be applied to other text classification tasks.