Performance of the Latent Ranking Neural Network models. Models with CR include the candidate retrieval scores as input, models with TAG use the features from the best performing TAG model (1G_CT+2G_CT), and models with embeddings include an average embedding for each of the questions, the answers, and the text from which the justification graphlet was derived. Significance tests were performed using bootstrap resampling with 10,000 iterations, but none of the differences between the neural network models and the CR baseline were significant.