A Multi-Level Semantic Fusion Algorithm Based on Historical Data in Question Answering

Abstract

To solve the problem that existing machine reading comprehension models can't understand the semantics of articles and questions well, this paper proposes a multi-level semantic fusion neural network based on natural language understanding and deep learning. The model combines calculations of attentions at different levels with cross-layer fusions of historical data. Experiments on the SQuAD dataset prove that our proposed model is superior to traditional models in terms of the exact match rate and F1 value. Our model can better transmit and interpret data and enhance the performance of question answering systems.