Machine Reading Comprehension(MRC) is to analyzing and inferring a paragraph received as input by a machine. Ushing machine reading comphrehension to understand given questions and paragraphs and outputting appropriate answers is question and answer u...
Machine Reading Comprehension(MRC) is to analyzing and inferring a paragraph received as input by a machine. Ushing machine reading comphrehension to understand given questions and paragraphs and outputting appropriate answers is question and answer using machine reading comprehension. Building machine reading learning data is difficult task, and you have to manually create the correct answers and questions that can derive the correct answers that appear in the document. In order to solve this problem research on automatic question generation has been actively studied. As opposed to machine reading comprehension, question generation is a task of generating question that can derive correct answers by looking at documents and correct answers. BERT is a language model showing excellent performance in various natural language processing tasks in recent years, and it learns a language model with a transformer with bedirectionality for a large-scale corpus. The pre-trained BERT can be applied to natural language processing tasks by adding an output layer. In this paper, we use KorQuAD 1.0 and KorQuAD 2.0 which are Korean question and answer datasets for machine reading comprehension learning. We propose machine reading comprehension model that adds a SRU(Simaple Recurrent Unit) on pre-trained BERT model and features suitable each dataset and BERT-based Sequence-to-sequence model that adds copying mechanism to then model that automatically generates a question from the document to which the correct answer belongs. As a result of the experiment, when the proposed in this paper was applied to KorQuAD 1.0 data, EM 85.35%, F1 93.24% were shown in the development set, and when was applied to KorQuAD 2.0 data, EM 49.2%, F1 71.21% were shown. In addition, the performance of the BERT-based Transformer decoder model was better than that of the exising model and the BERT + GRU decoder model.