In our work, we acquire a big QA dataset of about 300 movies with difficult questions that require semantic reasoning over an extended temporal domain. Participants were not informed which video was which, to prevent any bias when responding to the questions. Within the below talked about example, there is no (implicit/explicit) bias in direction of any id, but model misclassifies due to the presence of phrases related to an id. For instance, "tesseract" and "Hellicarrier bridge" which are main key phrases tend to be connected with "Nick Fury" a major character within the film.