This Author published in this journals
All Journal TEKNIK INFORMATIKA
Kiki Adnan Muzaki
Religious Studies Department, University of California, Riverside, USA

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

Interpretable Machine Learning for Narrative Framing Classification: A TF-IDF-Based Study of Frankenstein Luthfiana Dwi Rakhmawati; Rusdi Aulia Romadhon; Kiki Adnan Muzaki; Anna Nurawalia; Dewi Aprilia Ningrum
JURNAL TEKNIK INFORMATIKA Vol. 19 No. 2: JURNAL TEKNIK INFORMATIKA
Publisher : Department of Informatics, Universitas Islam Negeri Syarif Hidayatullah

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15408/jti.v19i2.54477

Abstract

This study evaluates interpretable supervised machine learning for three-class narrative framing classification of literary text, using Mary Shelley's Frankenstein (1818) as a case study; narrative framing is the evaluative position a passage assigns to the Creature: Threat, Victim, or Ambivalent. The methodology comprised five stages: (1) extraction of passages narrated by Victor Frankenstein, the Creature, and Robert Walton from the 1818 Project Gutenberg text; (2) annotation of 195 passages, with a blinded second annotator re-labeling 68 passages (Cohen's κ = 0.6926); (3) sublinear TF-IDF unigram and bigram representation; (4) classification with Multinomial Naive Bayes, Linear SVM, and Logistic Regression against Majority and VADER baselines under stratified 5-fold cross-validation; and (5) a designation-masking ablation and coefficient analysis. The results show that Logistic Regression obtained the highest mean scores (66.67% accuracy; 47.75% Macro-F1), although its margin over Linear SVM (46.18%) and VADER (44.85%) lies within fold-level variability. Pooled F1 reached 69.0% for Threat and 74.3% for Victim, whereas none of the 29 Ambivalent passages was correctly identified. Masking explicit designations reduced Macro-F1 only to 45.20%, indicating partial robustness to this procedure rather than discourse-level modeling. It is concluded that interpretable classical classifiers can separate Threat from Victim framing using transparent lexical evidence but cannot yet detect ambiguous framing; given the single-novel, small, and imbalanced dataset, the findings are a preliminary proof of concept. Future work should enlarge and balance the corpus, apply group-based and repeated cross-validation, and compare context-aware models, including large language models, on the Ambivalent class.