Maqasid: A Hybrid CNN-BiLSTM Framework for Nuanced Thematic Classification of Arabic Poetry

The computational analysis of Arabic poetry is hindered by its thematic richness, overlapping meanings, and a critical lack of multi-label annotated corpora—rendering traditional single-label approaches insufficient. This study presents Maqasid (Arabic: maqāṣid , denoting literary intentions or thematic purposes), the first end-to-end framework for multi-label thematic classification of Arabic poetry, and introduces the first large-scale corpus annotated for this task. Our contributions address these challenges through three main components. First, we develop a hierarchically structured, expert-validated multi-label taxonomy, used to annotate a large-scale Arabic poetry corpus. Second, we propose a hybrid CNN-BiLSTM model that processes custom-trained FastText embeddings through convolutional and sequential layers, with hyperparameters optimized via Bayesian search. Third, the model is deployed at scale to annotate the full Diwan corpus, and an interactive web-based platform is developed to support thematic exploration across genres, periods, and poets. Expert-based validation confirms the model’s high accuracy and interpretive reliability. These contributions provide a robust and extensible foundation for data-driven research in Arabic literary studies and digital humanities.