A Non-transfer Learning Approach in Exploring the Variations of Extractive Summarization Components for a Low-Resource Language

Creating corpora for low-resource languages like Filipino poses challenges in NLP, particularly in text summarization. Researchers often rely on pre-existing annotated corpora for high-quality summaries. However, a corpus for Filipino text summarization is currently unavailable. Hence, this study aimed to explore different variations of tokenizer, embedding model, and algorithm for extractive summarization in the context of a low-resource language, with a specific focus on a non-transfer learning approach. Eight combinations of summaries in total were generated and evaluated based on their accuracy and coherence using user feedback from thirty undergraduate students. The different component combinations were tested for generating short and medium-length summaries. The results showed that Summary #1, which used the Filipino Sentencepiece tokenizer, RoBERTa-TL embedding model, and TextRank algorithm, produced the most accurate and coherent summary overall. Comparing the components used in summary #1, the tokenizer, embedding model, and algorithm consistently performed better. However, Word2Vec models may face challenges if certain words are missing from their vocabulary, resulting in incomplete summaries. Additionally, a summary corpus may be created to fine-tune a pre-trained transformer model and evaluate the model performance using ROUGE Scores and BLEU.