Rabbani, Dzakki Damar (2026) Analisis Sentimen dan Stance Detection Berbasis Aspek pada Persepsi Publik terhadap Film "Jumbo 2025" di Media Sosial. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 2043221024-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
2043221024-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (9MB) | Request a copy |
Abstract
Industri perfilman Indonesia mengalami pergeseran signifikan pada tahun 2025 ketika film animasi lokal Jumbo 2025 berhasil menarik lebih dari 10 juta penonton dan menjadi salah satu film Indonesia terlaris sepanjang masa. Fenomena ini menunjukkan tingginya antusiasme publik terhadap film animasi lokal yang sebelumnya kurang mendominasi pasar domestik. Penelitian ini bertujuan untuk memahami persepsi publik terhadap film Jumbo 2025 melalui pendekatan Aspect-Based Sentiment Analysis (ABSA) dan stance detection pada komentar media sosial. Identifikasi aspek dilakukan menggunakan metode Latent Dirichlet Allocation (LDA), analisis sentimen menggunakan model IndoBERT, sedangkan stance detection dilakukan menggunakan pendekatan berbasis NRC Emotion Lexicon yang diterjemahkan ke dalam Bahasa Indonesia dan dikombinasikan dengan hasil klasifikasi sentimen. Data penelitian terdiri dari 5.000 komentar yang diperoleh dari YouTube, Instagram, Facebook, TikTok, dan Twitter. Tahap preprocessing meliputi case folding, normalisasi, cleansing, dan tokenisasi, dengan tambahan stopword removal serta stemming untuk kebutuhan pemodelan topik. Hasil pemodelan LDA pada 4.974 komentar menghasilkan empat topik utama dengan nilai coherence score sebesar 0,4264, yaitu Peran dan Karakter Pemain (21,1%), Karakter dan Alur Cerita (36,8%), Kualitas Animasi dan Kebanggaan Nasional (22,0%), serta Genre dan Tren Film (20,2%). Analisis sentimen menunjukkan bahwa tiga topik didominasi sentimen positif, sedangkan topik Karakter dan Alur Cerita didominasi sentimen negatif sebesar 49,5%, yang mengindikasikan tingginya ekspektasi publik terhadap pengembangan karakter dan narasi film. Hasil stance detection menunjukkan bahwa seluruh topik didominasi oleh stance favor dengan proporsi keseluruhan sebesar 72,9%, diikuti oleh against sebesar 14,7% dan neutral sebesar 12,5%. Temuan utama penelitian ini adalah banyaknya komentar bernada negatif yang justru tetap menunjukkan dukungan terhadap film, yakni sebanyak 1.008 komentar (20,2%), yang memperlihatkan bahwa sentimen dan stance tidak selalu searah. Topik Karakter dan Alur Cerita menjadi contoh paling menonjol karena meskipun memiliki sentimen negatif tertinggi (49,5%), topik ini tetap didominasi oleh stance favor sebesar 74,0%. Evaluasi awal model IndoBERT dan LSTM menunjukkan akurasi sebesar 65% untuk klasifikasi sentimen dan 73% untuk klasifikasi stance, namun kedua model menunjukkan performa rendah pada kelas minoritas (neutral dan against) akibat ketidakseimbangan distribusi kelas (class imbalance). Penanganan masalah ini melalui penerapan class weighting berhasil meningkatkan F1-score kelas neutral pada model sentimen dari 0,26 menjadi 0,41 dan F1-score kelas against pada model stance dari 0,15 menjadi 0,32, meskipun akurasi keseluruhan menurun menjadi 60% dan 57%. Hasil penelitian ini menunjukkan bahwa stance detection dapat melengkapi analisis sentimen dalam memberikan pemahaman yang lebih komprehensif mengenai sikap publik terhadap film Jumbo 2025 sebagai representasi film animasi lokal Indonesia.
========================================================================================================================================
The Indonesian film industry underwent a significant shift in 2025 when the local animated film Jumbo 2025 attracted more than 10 million viewers and became one of the highest-grossing Indonesian films of all time. This phenomenon reflects the strong public enthusiasm toward local animated films, which had previously been less dominant in the domestic market. This study aims to examine public perception of Jumbo 2025 through an Aspect-Based Sentiment Analysis (ABSA) and stance detection approach applied to social media comments. Aspect identification was carried out using Latent Dirichlet Allocation (LDA), sentiment analysis was performed using the IndoBERT model, while stance detection was conducted through an NRC Emotion Lexicon-based approach translated into Indonesian and combined with the sentiment classification results. The research data comprised 5,000 comments collected from YouTube, Instagram, Facebook, TikTok, and Twitter. The preprocessing stage included case folding, normalization, cleansing, and tokenization, with additional stopword removal and stemming for topic modeling purposes. The LDA modeling on 4,974 comments produced four main topics with a coherence score of 0.4264, namely Roles and Character Voice Actors (21.1%), Characters and Storyline (36.8%), Animation Quality and National Pride (22.0%), and Film Genre and Trends (20.2%). The sentiment analysis showed that three topics were dominated by positive sentiment, whereas the Characters and Storyline topic was dominated by negative sentiment at 49.5%, indicating high public expectations for character development and film narrative. The stance detection results showed that all topics were dominated by a favor stance, with an overall proportion of 72.9%, followed by against at 14.7% and neutral at 12.5%. The main finding of this study is the considerable number of negatively toned comments that nonetheless expressed support for the film, namely 1,008 comments (20.2%), demonstrating that sentiment and stance are not always aligned. The Characters and Storyline topic represents the most prominent example, as despite having the highest negative sentiment (49.5%), it remained dominated by a favor stance at 74.0%. The initial evaluation of the IndoBERT and LSTM models yielded an accuracy of 65% for sentiment classification and 73% for stance classification; however, both models exhibited low performance on the minority classes (neutral and against) due to class imbalance in the dataset. Addressing this issue through the application of class weighting increased the F1-score of the neutral class in the sentiment model from 0.26 to 0.41 and the F1-score of the against class in the stance model from 0.15 to 0.32, although the overall accuracy decreased to 60% and 57%, respectively. These findings indicate that stance detection can complement sentiment analysis in providing a more comprehensive understanding of public attitudes toward Jumbo 2025 as a representation of Indonesian local animated films.
Actions (login required)
 |
View Item |