Ekstraksi Topik dan Deteksi Keberpihakan Portal Berita: Pendekatan Inverted Pyramid Prompting Menggunakan DeepSeek
An Inverted Pyramid Prompting Approach Using DeepSeek
DOI:
https://doi.org/10.62712/juktisi.v5i1.1218Keywords:
large language model, topic extraction, stance detection, prompt engineering, media analysisAbstract
Media daring turut membentuk persepsi publik melalui pilihan isu dan cara pembingkaian pemberitaan, sehingga pemetaan kecenderungan keberpihakannya menjadi penting namun sulit dilakukan secara manual pada volume berita yang besar. Metode pembelajaran mesin klasik terkendala dalam menemukan topik yang belum diketahui dan mendeteksi sikap tersirat tanpa pelabelan manual yang masif. Penelitian ini bertujuan menemukan isu dominan dan mengklasifikasikan keberpihakan dua portal berita, yaitu Detik.com dan Kompas.com, terhadap pemerintah menggunakan model bahasa besar DeepSeek. Berita dikumpulkan melalui scraping tanpa kata kunci pada periode 1–31 Mei 2026, menghasilkan 27.789 artikel (2.200 dari Detik.com dan 25.589 dari Kompas.com). Untuk mengatasi batas token dan biaya, diterapkan ekstraksi teras berita berbasis prinsip piramida terbalik (inverted pyramid) dan two-pass prompting yang didukung context caching, sementara penemuan topik dan deteksi sikap dijalankan melalui prompting bertahap. Pemrosesan seluruh korpus menggunakan model deepseek-v4-flash hanya menelan biaya sebesar USD 1,56 atau diperkirakan sekitar 82% lebih hemat dibanding pemrosesan teks utuh. Pada evaluasi terhadap gold standard hasil anotasi manual (n = 12), klasifikasi sikap memperoleh akurasi 91,67% dan F1-Score makro 0,930. Pemetaan keberpihakan mengungkap bahwa mayoritas pemberitaan bersifat netral, dengan proporsi Pro-Pemerintah lebih tinggi pada Detik.com (32,9%) sementara proporsi Non-Pro-Pemerintah lebih tinggi pada Kompas.com (8,4%). Penelitian ini menunjukkan bahwa kombinasi DeepSeek, inverted pyramid, dan two-pass prompting merupakan pendekatan yang efektif dan efisien untuk analisis wacana media berskala besar tanpa pelabelan manual.
Downloads
References
[1] R. M. Entman, "Framing: Toward clarification of a fractured paradigm," Journal of Communication, vol. 43, no. 4, pp. 51–58, 1993, doi: 10.1111/j.1460-2466.1993.tb01304.x.
[2] S. M. Mohammad, S. Kiritchenko, P. Sobhani, X. Zhu, and C. Cherry, "SemEval-2016 Task 6: Detecting stance in tweets," in Proc. 10th Int. Workshop on Semantic Evaluation (SemEval-2016), San Diego, CA, USA, 2016, pp. 31–41, doi: 10.18653/v1/S16-1003.
[3] W. Yin, J. Hay, and D. Roth, "Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach," in Proc. 2019 Conf. Empirical Methods in Natural Language Processing and 9th Int. Joint Conf. Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, 2019, pp. 3914–3923, doi: 10.18653/v1/D19-1404.
[4] D. M. Blei, A. Y. Ng, and M. I. Jordan, "Latent Dirichlet allocation," Journal of Machine Learning Research, vol. 3, pp. 993–1022, Jan. 2003. [Online]. Available: https://jmlr.org/papers/v3/blei03a.html
[5] D. Küçük and F. Can, "Stance detection: A survey," ACM Computing Surveys, vol. 53, no. 1, pp. 1–37, Feb. 2020, doi: 10.1145/3369026.
[6] T. B. Brown et al., "Language models are few-shot learners," in Proc. Advances in Neural Information Processing Systems (NeurIPS), vol. 33, 2020, pp. 1877–1901, arXiv: 2005.14165.
[7] J. Wei et al., "Chain-of-thought prompting elicits reasoning in large language models," in Proc. Advances in Neural Information Processing Systems (NeurIPS), vol. 35, 2022, pp. 24824–24837, arXiv: 2201.11903.
[8] DeepSeek-AI, "DeepSeek API documentation," DeepSeek, 2026. [Online]. Available: https://api-docs.deepseek.com [Accessed: Jun. 15, 2026].
[9] H. Pöttker, "News and its communicative quality: The inverted pyramid—when and why did it appear?" Journalism Studies, vol. 4, no. 4, pp. 501–511, 2003, doi: 10.1080/1461670032000136596.
[10] DeepSeek-AI, "DeepSeek-V3 technical report," arXiv:2412.19437, 2024, doi: 10.48550/arXiv.2412.19437.
[11] F. Gilardi, M. Alizadeh, and M. Kubli, "ChatGPT outperforms crowd workers for text-annotation tasks," Proc. Nat. Acad. Sci. U.S.A., vol. 120, no. 30, p. e2305016120, 2023, doi: 10.1073/pnas.2305016120.
[12] B. Wilie et al., "IndoNLU: Benchmark and resources for evaluating Indonesian natural language understanding," in Proc. 1st Conf. Asia-Pacific Chapter Assoc. Comput. Linguistics and 10th Int. Joint Conf. Natural Language Processing (AACL-IJCNLP), Suzhou, China, 2020, pp. 843–857, doi: 10.18653/v1/2020.aacl-main.85.
[13] M. Grootendorst, "BERTopic: Neural topic modeling with a class-based TF-IDF procedure," arXiv:2203.05794, 2022, doi: 10.48550/arXiv.2203.05794.
[14] M. Mets, A. Karjus, I. Ibrus, and M. Schich, "Automated stance detection in difficult topics and small languages: The challenging case of immigration in polarizing news media," PLoS ONE, vol. 19, no. 4, p. e0302380, 2024, doi: 10.1371/journal.pone.0302380.
[15] T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwasawa, "Large language models are zero-shot reasoners," in Proc. Advances in Neural Information Processing Systems (NeurIPS), vol. 35, 2022, pp. 22199–22213, arXiv: 2205.11916.
Downloads
Published
How to Cite
Issue
Section
License
Copyright (c) 2026 Hullio Kaisar Leisina, Faisal Farobi Ahmad, Suprianto, Hendra Supendar, Riza Fahlapi
This work is licensed under a Creative Commons Attribution 4.0 International License.
