Özet:

Abstract In 2021 there was an increase in the number of people who have internet access, and the number of users increased from 175 million users to 202 million users. News classification in general still uses traditional techniques such as word embedding with TF-IDF and machine learning.  The latest development for technology that can classify NLP news using the BERT model is a state-of-the-art pre-trained model. However, the pre-train model of BERT is only limited to use in English. So in this study, IndoBERT will be used in making news recommendations based on the category. This dataset uses an Indonesian news dataset that has 5 categories, including football, news, business, technology, and automotive. The IndoBERT method will be compared with other pre-train models, such as XLNET, BERT multilingual, XLMRoberta. Meanwhile, the machine learning method with TF-IDF word embedding was compared using the XGBoost method, LGB, and random forest. In this study, we see that the classification method using IndoBERT gives the best results with an accuracy value of 94% and also provides the smallest computation time compared to other methods with a time of one minute 56 seconds and a validation time of 10 seconds. BERT can give the best results because BERT is a type of pre-trained model that is trained from various kinds of Indonesian words such as news and several website sources to add to the corpus of vocabulary sources in the model. In the future research will be carried out to implement the dual IndoBERT model and the Siamese IndoBERT.

Anahtar Kelimeler: