Topic model based on co-occurrence word networks for unbalanced short text datasets