Identifikasi Kuantitas Relevan pada Soal Cerita Matematika Menggunakan Model BiLSTM dan CRF

Abstract

Mathematical word problems often contain information that is not entirely relevant, causing students to make mistakes in selecting important information. This study aims to identify relevant quantities in Indonesian mathematical word problem texts using a sequence labeling approach. This research applies a quantitative approach through computational experiments. The dataset consists of 700 problems collected from elementary school mathematics e-books and annotated using the BIO scheme with B-REL, I-REL, and O labels. The model used is a combination of BiLSTM and CRF, with an 80:20 split for training and testing data. Evaluation was conducted using precision, recall, F1-score, and accuracy, with a focus on the F1-score of the REL class. The results show that the model achieved an F1-score of 0.9726. The dominant error occurred in the classification of REL as O, indicating that quantity spans were still often truncated. However, no BIO rule violations were found, indicating that label sequence consistency was maintained. These results show that the model still has difficulty recognizing some parts of relevant quantities completely. This study is limited to elementary school arithmetic word problems involving basic operations. It also focuses only on identifying relevant quantities, not on solving the final answers to the problems. Future research may add more varied data and compare the BiLSTM-CRF model with other models.