Detecting dynamics of action in text with a recurrent neural network

According to the dynamics of action (DoA)-theory, action is an interplay of instigating and consummatory forces over time. The TAT/PSE—a psychological test instrument—should measure this dynamics. Therefore, people get presented different pictures with the instruction to invent stories. In those stories, the periodical tendencies should be visible, but this could not be shown yet. I reanalyzed two datasets regarding category IS: They were coded by a human expert, a recurrent neural network (RNN), and a convolutional neural network (CNN). It is visible that in Heckhausen's origin data category IS produces saw-tooth related curves in the stories across the pictures and that this could potentially better be detected by the RNN than by the CNN or the human coder. Second, I reanalyzed a study that experimentally assessed the DoA with a picture x position effect and rejected it. Here again, only the RNN coded IS-score produces a statistical significant picture x position effect. This shows that because of its sequential structure the RNN detects different phrases in the text that are barely capable by human coder or other neural networks but are related to motive theories.