EVALUASI MAXIMUM SEQUENCE LENGTH PADA INDOBERT-BILSTM-CRF UNTUK NAMED ENTITY RECOGNITION BAHASA INDONESIA

EVALUATING THE IMPACT OF MAXIMUM SEQUENCE LENGTH ON INDOBERT-BILSTM-CRF FOR INDONESIAN NAMED ENTITY RECOGNITION

Abstract

Maximum sequence length merupakan salah satu parameter penting pada model berbasis Transformer karena memengaruhi cakupan input sekaligus kebutuhan komputasi. Penelitian ini menganalisis pengaruh maximum sequence length terhadap kinerja dan efisiensi komputasi model IndoBERT-BiLSTM-CRF pada tugas Named Entity Recognition Bahasa Indonesia menggunakan dataset IDNer-News-2K. Empat konfigurasi diuji, yaitu 64, 128, 256, dan 512 token, dengan parameter model lainnya dipertahankan tetap. Setiap konfigurasi dijalankan menggunakan tiga random seed. Kinerja dievaluasi menggunakan Token Macro-F1 dan Entity F1, sedangkan efisiensi dianalisis berdasarkan training time, penggunaan GPU memory, training throughput, dan inference time. Hasil menunjukkan bahwa MAX_LEN=64 masih menyebabkan truncation sekitar 0,55% dan menghasilkan Entity F1 sebesar 0,9119. MAX_LEN=128 menghilangkan seluruh truncation serta mencapai 100% word coverage dan token coverage, dengan Entity F1 sebesar 0,9171. Nilai Entity F1 tertinggi diperoleh pada MAX_LEN=256 sebesar 0,9177, tetapi peningkatannya terhadap MAX_LEN=128 sangat kecil dan disertai kenaikan biaya komputasi. Pada MAX_LEN=512, waktu pelatihan dan penggunaan GPU memory meningkat secara substansial, sementara Entity F1 menurun menjadi 0,9159. Temuan ini menunjukkan bahwa peningkatan maximum sequence length setelah seluruh input tercakup memberikan diminishing return. Berdasarkan keseimbangan antara kinerja dan efisiensi, MAX_LEN=128 merupakan konfigurasi yang paling rasional untuk dataset yang digunakan.

Maximum sequence length is an important parameter in Transformer-based models because it affects both input coverage and computational requirements. This study analyzes the effect of maximum sequence length on the performance and computational efficiency of an IndoBERT-BiLSTM-CRF model for Indonesian Named Entity Recognition using the IDNer-News-2K dataset. Four configurations were evaluated, namely 64, 128, 256, and 512 tokens, while keeping the remaining model parameters unchanged. Each configuration was run using three random seeds. Model performance was evaluated using Token Macro-F1 and Entity F1, while computational efficiency was assessed based on training time, GPU memory usage, training throughput, and inference time. The results show that MAX_LEN=64 still caused approximately 0.55% truncation and achieved an Entity F1 of 0.9119. MAX_LEN=128 eliminated truncation completely and achieved 100% word coverage and token coverage, with an Entity F1 of 0.9171. The highest Entity F1 was obtained at MAX_LEN=256, reaching 0.9177, but the improvement over MAX_LEN=128 was marginal and accompanied by higher computational cost. At MAX_LEN=512, training time and GPU memory usage increased substantially, while Entity F1 decreased to 0.9159. These findings indicate that increasing maximum sequence length after full input coverage has been achieved results in diminishing returns. Considering the trade-off between performance and computational efficiency, MAX_LEN=128 provides the most practical configuration for the dataset used in this study.