…structured explanations…describe explanations that are not entirely free-form, e.g., there are constraints placed on the explanation-creating process, such as the use of specific inference rules.

English

Dataset Task Explanation Type Collection Method # Instances # Explanations per Instance Total # Annotators
WorldTree v1 science exam QA explanation graphs authors 1,680 1 4
OpenBookQA open-book science QA 1 fact from WorldTree crowd 5,957 1 n/a
WorldTree v2 science exam QA explanation graphs experts 5,100 1 n/a
QED reading comprehension QA inference rules authors 8,991 1 3
QASC science exam QA 2-fact chain authors + crowd 9,980 1 62
eQASC science exam QA 2-fact chain automatic + crowd 9,980 ~10 n/a
eQASC + perturbed science exam QA 2-fact chain automatic + crowd n/a n/a n/a
eOBQA open-book science QA 2-fact chain automatic + crowd n/a n/a n/a
Ye et al. (2020) SQuAD QA semi-structured text crowd + authors 164 1 n/a
Ye et al. (2020) NaturalQuestions QA semi-structured text crowd + authors 109 1 n/a
R4C reading comprehenesion QA chains of facts crowd 4,588 3 45
StrategyQA implicit reasoning QA reasoning steps with highlights crowd 2,780 3 54
TriggerNER Named Entity Recognition group of tokens with highlights crowd ~7K 2 3
COPA-SSE (Semi-Structured Explanations for COPA) Balanced COPA (commonsense QA, causal reasoning) ConceptNet-like triples with free-form head and tail concepts* crowd 1,500 4-9 (9747 total) N/A

* The author classed this as structured but says it’s not very rigid and can also be used as free text.