KoTAB: Korean Template-Based Arithmetic Solver with BERT
Kyung Seo Ki, Dong Geon Lee, Gahgene Gweon · 2020
Math word problems are a type of mathematical question expressed in the linguistic form which can be mathematically formulated from situations encountered in everyday life. In this study, we propose KoTAB, a Korean template-based arithmetic solver with BERT, which is modified from TAB (Template-based Arithmetic Solver with BERT) framework that is used to solve English math word problems. Using KoTAB, we solved Korean math word problems using two datasets, CC and IL, which were translated from English to Korean. Four different pre-trained BERT models were used to fine-tune KoTAB; BERT-base-cased, BERT-multilingual-cased, Korean_BERT_WordPiece, Korean_BERT_ Morphology. Among these four pre-trained BERT models, the highest performance in accuracy was achieved when the Korean_BERT_Morphology model was used to evaluate CC (99.3%) and IL (85.22%) datasets. This result is comparable to the performance of English models for CC (99.5%), but not for IL (89%).