Results of the full model on narration test data. We show the mean and standard deviation of bootstrapped scores, pooled over four training runs (chance recall@10 = 10%; chance triplet accuracy = 50%).