Evaluating the Effect of Model Inductive Bias and Training Data in Predicting Human Reading Times