This section presents the first set of 20 tasks for testing text understanding and reasoning in the bAbI project. The tasks are described in detail in the paper:
Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M. Rush, Bart van Merriënboer, Armand Joulin and Tomas Mikolov. Towards AI Complete Question Answering: A Set of Prerequisite Toy Tasks, arXiv:1502.05698.
Please also see the following slides:
Antoine Bordes Artificial Tasks for Artificial Intelligence, ICLR keynote, 2015.
The aim is that each task tests a unique aspect of text and reasoning, and hence test different capabilities of learning models. More tasks are planned in the future to capture more aspects.
Training Set Size: For each task, there are 1000 questions for training, and 1000 for testing. However, we emphasize that the goal is to use as little data as possible to do well on the tasks (i.e. if you can use less than 1000 that’s even better) — and without resorting to engineering task-specific tricks that will not generalize to other tasks, as they may not be of much use subsequently. Note that the aim during evaluation is to use the same learner across all tasks to evaluate its skills and capabilities.
Supervision Signal: Further while the MemNN results in the paper use full supervision (including of the supporting facts) results with weak supervision would also be ultimately preferable as this kind of data is easier to collect. Hence results of that form are very welcome. E.g. this paper does include weakly supervised results.
For the reasons above there are currently several directories:
- en/ — the tasks in English, readable by humans.
- hn/ — the tasks in Hindi, readable by humans.
- shuffled/ — the same tasks with shuffled letters so they are not readable by humans, and for existing parsers and taggers cannot be used in a straight-forward fashion to leverage extra resources– in this case the learner is more forced to rely on the given training data. This mimics a learner being first presented a language and having to learn from scratch.
- en-10k/ shuffled-10k/ and hn-10k/ — the same tasks in the three formats, but with 10,000 training examples, rather than 1000 training examples. Note the results in the paper use 1000 training examples.
The file format for each task is as follows:
ID question[tab]answer[tab]supporting fact IDS.
The IDs for a given “story” start at 1 and increase. When the IDs in a file reset back to 1 you can consider the following sentences as a new “story”. Supporting fact IDs only ever reference the sentences within a “story”.
1 Mary moved to the bathroom.
2 John went to the hallway. 3 Where is Mary? bathroom 1
4 Daniel went back to the hallway.
5 Sandra moved to the garden.
6 Where is Daniel? hallway 4
7 John moved to the office.
8 Sandra journeyed to the bathroom.
9 Where is Daniel? hallway 4
10 Mary moved to the hallway.
11 Daniel travelled to the office.
12 Where is Daniel? office 11
13 John went back to the garden.
14 John moved to the bedroom.
15 Where is Sandra? bathroom 8
1 Sandra travelled to the office.
2 Sandra went to the bathroom.
3 åWhere is Sandra? bathroom 2
Some data statistics including overlap between train and test (which is minimal) can be found here. Code Code to generate tasks is available on github. We hope this will encourage the machine learning community to work on, and develop more, of these tasks.