Research on Question Answering System Based on BERT Model

The core of natural language processing is the science of how computers understand and respond to the influence of human language. This is also the main research direction in the field of machine intelligence development. Question answering system is one of the development directions. How to make the answer of question answering system more intelligent is the focus of research in recent years. Based on this, this thesis proposes a question answering system based on the BERT model that combines the co-attention mechanism and the self-attention mechanism. In order to better show the representation of questions and articles at their respective semantic levels, this article will choose BERT language to encode them. Attention and fusion are then performed horizontally and vertically at different levels of granularity between questions and paragraphs. Experiments show that the model has an average EM of 46.56 and an average F1 of 58.90 on all datasets. The average F1 for the development dataset is 54.85, and the average F1 for the test dataset is 62.95. The content answered by the question answering system based on the BERT model is basically reasonable, and the intelligence of the question answering system is effectively improved.