A Chinese Question Answering System based on GPT
Shuai Liu, Xiaojun Huang · 2019
The Chinese question-answering system needs to select the most appropriate answer from the answer library for user according to the given question on the natural language form. Previous question-answering systems required modeling for specific task characteristics and designing multiple modules. This paper first proposes to use the Generative Pre-trained Transformer (GPT) to implement the Chinese question-answering system. To optimize and improve the model, this Chinese model pays more attention to the contextual content and semantic characteristics, and we designed a new method to train this model. This model reduces the number of modules in the question-answering system. This paper evaluates the model on the Document-Based Chinese Question and Answer (DBQA) dataset and achieves a 2.5% improvement in MRR/MAP over the latest lattice convolutional neural networks (Lattice CNNs).