기계 번역 구조를 활용한 문장형 수학 문제풀이에 관한 연구

Abstract

Deep learning has been developed and applied to natural language processing and applied to a variety of fields. However, even when large scale language models are utilized, they lead to unstable results for questions that require mathematical reasoning. To complement these weaknesses and expand the application areas of language models, language understanding-based mathematical inference learning is needed. Research on MWP (Math Word Problem) is a task that began in the 1960s to infer mathematical expressions through language understanding. MWP refers to problems consisting of mathematical logic, numbers, and natural language. MWP solvers, which are based on language models, must accept the problem text as input and make mathematical inferences after understanding the context. However, designing an MWP solver has been rated challenging due to differences in the inherent meanings of words and prior knowledge understood by humans and machines, and deep Research on learning models is necessary. Among the research on MWP solvers is not only the presentation of model designs and methodologies, but also the disclosure of datasets for different types of learning. However, English and Chinese-based datasets are mainly published, and Korean datasets are lacking. In this study, I explores the design of an equation generation model to narrow down the semantic gap by leveraging the Conformer model proposed in the speech recognition domain, and implement a Korean-based MWP generator for collecting large amounts of Korean data. The data generator proposed in this study is based on the elementary mathematics subject level and consists of four types: arithmetic operations, ordering, finding unknown, and geometry. Each type is made up of sub types, for a total of 42 subtypes. Additionally, four data variations are applied during data generation to design a robust model. A total of 210,311 pieces of data were used in the experiment. Of this, 210,000 pieces of data are generated data. The training dataset consists of 150,000 pieces of data, and the validation and evaluation datasets each consist of 30,000 pieces. The 311 pieces of data are real datasets collected from commercially available math problem books. The real dataset was for additional experiments to evaluate the effectiveness of the proposed data generator and MWP solver with Conformer structure. Experiments confirmed that the Conformer model had an accuracy of 91.19% on the produced dataset and 42.76% on the real dataset, outperforming the Sequence-to-Sequence and Transformer models. Although the performance declined on the real dataset due to problem types and words that the model was not trained on, the performance of the Conformer model suggests that the proposed data generator partially reflects real math problems. In this study, I designed an MWP solver by applying other domain models and generated the data necessary for learning the Korean MWP solver. When applied to math education and the service industry, these models can help improve convenience and productivity. However, the performance decreased due to the difference between the proposed data generator and the actual math problem. In addition to complementing the problem types of data generators, future research will need to design MWP solvers with enhanced context understanding capabilities by leveraging large-scale pre-trained models.