Pretrained Encyclopedia: Weakly Supervised Knowledge-Pretrained Language Model

Recent breakthroughs of pretrained language models have shown the of self-supervised learning for a wide range of natural language (NLP) tasks. In addition to standard syntactic and semantic NLP, pretrained models achieve strong improvements on tasks that involve-world knowledge, suggesting that large-scale language modeling could be an method to capture knowledge. In this work, we further investigate the to which pretrained models such as BERT capture knowledge using a-shot fact completion task. Moreover, we propose a simple yet effective supervised pretraining objective, which explicitly forces the model to knowledge about real-world entities. Models trained with our new yield significant improvements on the fact completion task. When to downstream tasks, our model consistently outperforms BERT on four-related question answering datasets (i.e., WebQuestions, TriviaQA, and Quasar-T) with an average 2.7 F1 improvements and a standard-grained entity typing dataset (i.e., FIGER) with 5.7 accuracy gains.