Few-shot Question Answering with Entity-Aware Prompt

Providing simple task descriptions or prompts in natural language for large pre-trained language models yields impressive few-shot learning results in different tasks, such as text classification, knowledge probing, machine translation, and named entity recognition. In this paper, we apply this idea to question-answering task to fine-tune pre-trained language models by constructing entity-type prompts. Specifically, we augment the context sequences with semantic labels to enhance the understanding of pre-trained models, and dynamically adjust the prompts via intention recognition of the questions. Our proposition is simple yet powerful over traditional fine-tune training strategies and robust under few-shot conditions. The contributions of our work are as follows: 1. We proposed a few-shot learning method with entity-aware prompts for question-answering tasks to fine-tune the pre-trained language model. 2. Based on the SQuAD dataset, we extract a subset with 1,131 samples containing different categories of answer type, in which the answers to all questions are entities. 3. Experiments on multiple pre-trained language models validate that our method can effectively improve the performance of few-shot learning of question-answering tasks over the promptless ones.