고전시가 분석에서의 자연어처리 활용 ― 次韻詩詩句의 부자연스러움 증명

Abstract

Rhyme-matching poem is a kind of replying poem. It is written based on the rule that the rhymes in the original poem should be used in order. Loved by great poets, it achieved widespread popularity after the Song Dynasty. However, rhyme-matching poems had been criticized for unnatural expressions due to the use of fixed rhymes, and such works had been pointed out. So this study performed the ‘fill-mask’ task using the embeddings of pre-trained models that learned Chinese classical text in a Masked Language Modeling (MLM) so as to verify whether the expressions of the verse in rhyme-matching poems is unnatural compared to those of original poems. The subject of this experiment is the poems of the literary network led by Su-Shi in the North Song Dynasty. The experimental result shows the probability to predict the rhyme tokens in rhyme-matching poem was lower than that of original poem; the probability was much lower in the long poem. This proves that the expressions in rhyme-matching poems are not general compared to that of original poems, but more strongly represents the need for using rhyme-matching method in the literary networks. In conclusion, this study demonstrates that rhyme-matching poems were advantageous for close connection with the original poems and the authors.