Few-Shot Learning with Self-supervised Classifier for Complex Knowledge Base Question Answering

Complex Question Answering (CQA) over Knowledge Base (KB) involves transferring natural language questions to a sequence of actions, which are utilized to fetch entities and relations for final answer. Typically, meta-learning based models regard question types as standards to divide dataset for pseudo-tasks. However, question type, manually labeled in CQA data set, is indispensable as a filter in the support set retrieving phase, which raises two main problems. First, preset question types could mislead the model to be confined to a non-optimal search space for meta-learning. Second, the annotation dependency makes it difficult to migrate to other datasets. This paper introduces a novel architecture to alleviate above issues by using a co-training scheme featured with self-supervised mechanism for model initialization. Our method utilizes a meta-learning classifier instead of pre-labeled tags to find the optimized search space. Experiments in this paper show that our model achieves state-of-the-art performance on CQA dataset without encoding question type.