Punjabi Poetry Genre Classification Using Graph Neural Networks

Punjabi poetry, famous for its multi-faceted culture and language, is unique in posing a challenge to computational linguistics and machine learning. A wide range of factors including writing styles, authors, emotions, and historical periods involved in the classification of Punjabi poems requires an innovative approach that takes full advantage of GNNs'(Graph Neural Networks) power. In this research, a dataset of 400 Punjabi poems curated from various online sources, was divided into four different genres: Festival; Humorous; Spirituality; and Patriotism. Our methodology involves thorough preprocessing, constructing an innovative text graph based on word probability(PMI - Pointwise mutual exclusion) and intra-poem weightings (TF IDF). It reflects intricately all those complex links between words, their cooccurring patterns with other words as well as their relationship to any given poem. It is then reconstructed as a graph-based representation in which words and their relations are expressed explicitly through the nodes of the graph. Next, text GCN(Graph Convolutional Networks) is used to analyze and classify the poems using graph structures. Unlike traditional classification methods, our GNN model gets much better results and has achieved an accuracy of 82.50%.