A Comparative Analysis of Quantized and Non-Quantized BERT Model Performance for the Low-Resource Tagalog Language through Binary Text Classification

The development of artificial intelligence and its integration to human society is an unstoppable force. Hence, many subsets of this phenomenon such as natural language processing are expected to continuously evolve alongside other advancements made in the field. One of the challenges under such subset are low-resource languages which impact the growth of artificial intelligence in countries whose language or languages are not globally popular and therefore lack training data and/or tailored NLP technologies for machine and deep learning use. The fairly recent, state-of-the-art transformer-based language model Bidirectional Encoder Representations from Transformers or BERT has given NLP research for low-resource languages such as Filipino/Tagalog the opportunity to advance, but is held back for many by its high hardware requirements. Therefore, the researchers wish to contribute to the accessibility of Filipino natural language processing research by determining if quantization would be beneficial for low-resource BERT models such as for Tagalog.