Zero-Shot Commonsense Question Answering with Cloze Translation and Consistency Optimization

Commonsense question answering (CQA) aims to test if models can answer questions regarding commonsense knowledge that everyone knows. Prior works incorporate external bases have shown promising results, but are expensive construct and often limited a fixed set of relations. In this paper, we instead focus on better utilizing the implicit stored in pre-trained language models. While researchers found embedded be extracted by having them fill blanks carefully designed prompts for relation extraction text classification, it remains unclear adopt paradigm CQA where inputs outputs take much more flexible forms. To end, investigate four translation methods translate natural into cloze-style sentences solicit from models, including syntactic-based model, an unsupervised neural two supervised addition, combine different methods, propose encourage consistency among model predictions translated with unlabeled data. We demonstrate effectiveness our three datasets zero-shot settings. show complementary base improved combining lead state-of-the-art performance. Analyses also reveal distinct characteristics cloze provide insights why great improvements. Code/dataset is available at https://github.com/PlusLabNLP/zero_shot_cqa.