Balinese story texts dataset for narrative text analyses.

Area of Science:

• Computational Linguistics
• Natural Language Processing (NLP)
• Artificial Intelligence (AI)

Background:

• Computational linguistic tools for narrative text analysis are advancing, but are primarily English-focused due to data scarcity in other languages.
• Character identification is a critical first step for deeper narrative analysis, yet remains challenging in diverse linguistic contexts.
• Low-resource languages like Balinese lack sufficient annotated datasets for developing such analytical tools.

Purpose of the Study:

• To present the first annotated Balinese story texts dataset for narrative text analysis.
• To facilitate character identification, alias clustering (named entity linking), and character classification in Balinese narratives.
• To support the development of computational linguistic tools and machine learning models for low-resource languages.

Main Methods:

• Manual annotation of 120 Balinese stories by native speakers, including a sociolinguistics expert.
• Creation of four sub-datasets for character identification (word and sentence level), alias clustering, and protagonist/antagonist classification.
• Calculation of inter-annotator agreement using Cohen's Kappa, Jaccard Similarity, and Mean F1-score to ensure dataset reliability.

Main Results:

• A comprehensive dataset comprising 89,917 annotated words, 6,634 annotated sentences, and 930 character groups.
• Successful annotation for character identification, alias clustering, and classification of 848 character groups into protagonists (66.16%) and antagonists (33.84%).
• Demonstrated reliability and consistency of the dataset through high inter-annotator agreement scores.

Conclusions:

• The developed Balinese narrative dataset is a valuable resource for advancing computational linguistics and AI in low-resource languages.
• This dataset enables enhanced research in character identification, network development, and relationship extraction within narrative texts.
• It provides a foundation for building sophisticated machine learning and deep learning models for analyzing non-English narrative structures.