[ad_1]
TimeDial presents a set of challenges in crowdsourced English, for common sense reasoning, formulated as a multiple choice cloze task with approximately 1.5k carefully curated dialogues. The dataset is derived from the DailyDialog (Li et al., 2017), which is a multi-turn dialogue corpus.
For each masked range, there is more than one correct answer in the options in the dialog boxes ~ 1.1k. This makes the task more challenging for models as a more complete understanding of the context is required to recognize all correct choices. In our dataset, we guarantee two wrong answers for each masked range. Some incorrect options are selected to be falsely related to the dialog context. For example, we include time slots in the context of the dialogue as negative options, which will test models that rely primarily on superficial pattern matching only without proper temporal reasoning.
See README for more information.
[ad_2]
picture credit