Emotion Classification in Kannada Poetry: Contrasting TF-IDF, MuRIL, and RDF2Vec Approaches Amid Semantic Challenges

Abstract

Emotion classification in Kannada Hanigavana (short poems) is
challenging due to metaphorical language, cultural references, lowresource linguistic tools, and semantic ambiguity. This study
compares traditional, contextual, and ontology-based approaches for
classifying poems into the Navarasa (nine emotions) framework. Six
pipelines were evaluated: TF-IDF with Support Vector Machine
(SVM) and Random Forest (RF), MuRIL embeddings with SVM and
RF, and RDF2Vec semantic embeddings with a feed-forward Neural
Network. A manually annotated corpus of 600 Kannada poems was
used. Feature extraction included TF-IDF, MuRIL’s contextual
embeddings, and RDF2Vec vectors generated from a custom emotion
ontology. While SVM and RF with TF-IDF achieved perfect scores on
conventional metrics, they failed to capture deeper semantic meaning.
MuRIL-based models offered slight improvement in context
understanding but struggled with domain-specific abstraction. The
best performance was observed with RDF2Vec + Neural Network,
achieving 59.00% accuracy and 57.00% F1-score, along with
comparatively high semantic coherence. The classification faces
challenges like metaphorical ambiguity, limited annotated resources,
and the difficulty of capturing semantic nuance in low-resource poetic
corpora. This study demonstrates the advantage of incorporating
structured semantic knowledge via ontologies for emotion
classification in low-resource languages like Kannada and provides a
scalable framework for similar literary NLP tasks The proposed
classification framework can support cultural AI applications, such as
emotion-aware Kannada voice assistants and educational tools for
literary analysis.