Automating the Classification of Economic Activities in Official Statistics: A Comparative Study of Neural Networks and Transformers

This paper explores the process of automation of the classification of open-ended questions regarding the economic activities of enterprises, in official statistics. Neural networks (NN) and transformer-based models such as BERT are compared. This study's aim is the improvement of the accuracy and efficiency of the economic activities classification in alignment with the NACE classification system. The textual data from official statistics are processed beforehand. NN and BERT models are utilized to classify at the 2-digit and 4-digit levels. To assess and compare the effectiveness of these models, performance metrics, methods such as accuracy and F1-scores, are used. The results show the potential of transformer-based models to improve the process of automation codification of economic activity, by the reduction of manual work and increasing consistency in classifications. This research makes an essential contribution by exploiting the potential of the application of transformer models to domain-specific data such as the ones in official statistics, advancing the field of automatic text classification.