Aspect Sentiment Triplet Extraction Menggunakan Transformer pada Data Teks Berita dan X

Pane, Rafli Syahputra (2026) Aspect Sentiment Triplet Extraction Menggunakan Transformer pada Data Teks Berita dan X. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221038_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Analisis sentimen telah berkembang dari klasifikasi polaritas sentimen biasa menjadi pendekatan yang lebih fine-grained seperti Aspect Sentiment Triplet Extraction (ASTE) yang mengekstraksi aspek, opini, dan polaritas sentimen sebagai kesatuan triplet dari suatu kalimat. Penelitian ASTE pada Bahasa Indonesia dan pada teks dengan karakteristik beragam masih terbatas, terutama menggunakan pendekatan generatif. Penelitian ini mengkaji efektivitas model generatif Transformer untuk task ASTE pada dua karakteristik dataset, yaitu teks media sosial X dan teks berita, dalam dua versi bahasa, yaitu Bahasa Indonesia dan Bahasa Inggris hasil translasi. Tiga arsitektur generatif dibandingkan, yaitu T5, BART dengan arsitektur encoder-decoder, dan GPT-2 dengan arsitektur decoder-only. Pseudolabeling menggunakan LLM diterapkan untuk menghasilkan triplet sebagai training data, dengan validasi manual untuk memverifikasi kualitas pseudolabel. Evaluasi dilakukan menggunakan metrik Precision, Recall, dan F1-score pada mode exact match dan fuzzy match. Hasil eksperimen menunjukkan bahwa T5 mencapai F1-score tertinggi 0,5091 pada konfigurasi X Bahasa Inggris, diikuti BART dengan 0,4194, dan GPT-2 dengan 0,2154. Performa pada dataset X yang berkarakteristik teks pendek konsisten lebih tinggi dibandingkan Teks Berita yang berkarakteristik teks panjang di seluruh konfigurasi model dan bahasa. Penggunaan strategi translasi ke Bahasa Inggris meningkatkan F1-score terutama pada model T5 dan BART. Penelitian ini memberikan insight bahwa panjang teks dan kompleksitas struktural lebih berdaampak terhadap performa ASTE dibandingkan formalitas teks, serta mengonfirmasi efektivitas arsitektur encoder-decoder untuk task ekstraksi struktural generatif.
======================================================================================================================================
Sentiment analysis has evolved from conventional sentiment polarity classification into more fine-grained approaches such as Aspect Sentiment Triplet Extraction (ASTE), which extracts aspects, opinions, and sentiment polarities as a unified triplet from a sentence. ASTE research in Indonesian and on texts with diverse characteristics remains limited, particularly those employing generative approaches. This study examines the effectiveness of generative Transformer models for the ASTE task across two distinct dataset characteristics, social media texts from X and news texts, in two language versions: Indonesian and translated English. Three generative architectures are compared: T5 and BART with encoder-decoder architectures, and GPT-2 with decoder-only architecture. LLM-based pseudolabeling is applied to generate triplets as training data, accompanied by manual validation to verify the quality of the pseudolabels. Evaluation is conducted using Precision, Recall, and F1-score metrics under both exact match and fuzzy match modes. Experimental results show that T5 achieves the highest F1-score of 0.5091 in the English X configuration, followed by BART at 0.4194, and GPT-2 at 0.2154. Performance on the X dataset short texts is consistently higher than that on the News Text dataset long texts across all model and language configurations. Furthermore, utilizing an English translation strategy improves the F1-score, particularly for the T5 and BART models. This study provides insight that text length and structural complexity have a greater impact on ASTE performance than text formality, and confirms the effectiveness of the encoder-decoder architecture for generative structural extraction tasks.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Aspect Sentiment Triplet Extraction (ASTE), Transformer Generatif, T5, BART, GPT-2, Media Sosial dan Teks Berita, Pseudolabeling. Aspect Sentiment Triplet Extraction (ASTE), Generative Transformer, T5, BART, GPT-2, Social Media and News Text, Pseudolabeling.
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA76.9.D343 Data mining. Querying (Computer science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Rafli Syahputra Pane
Date Deposited:	21 Jul 2026 06:21
Last Modified:	21 Jul 2026 06:21
URI:	http://repository.its.ac.id/id/eprint/136048

Actions (login required)

View Item