Data for the paper "Contextual Out-of-Domain Utterance Handling with Counterfeit Data Augmentation" by Sungjin Lee and Igor Shalyminov [Paper] [Slides]
babi_task6
- clean version of bAbI Dialog Task 6 for Hybrid Code Network trainingbabi_task6_ood_0.2_0.4
- bAbI Dialog Task 6, version with OOD augmentations. OOD turns distributed as follows: OOD turn sequence starts with a probabilityp_start=0.2
and keeps going withp_cont=0.4
. Every OOD sequence ends up with a segment-level OOD turn. For more detail on data augmentation, check out our papers: 1 and 2- Google datasets - coming soon
Data augmentation code can be found in this repo