This research work presents a novel approach for detecting and classifying text related to smoking, leveraging Natural Language Processing (NLP) and Graph Neural Networks (GNNs). In this method, we innovatively differentiate between texts that promote smoking and those that either advice against smoking or merely contain related keywords without intent to discuss the act or consequences of smoking. Our methodology integrates various text preprocessing methods, Latent Dirichlet Allocation (LDA) for topic modeling, and graph theory to construct a relational representation of textual data. The LDA model identifies thematic structures within a dataset of 480 texts, which is further encapsulated into a graph structure, with nodes representing individual documents and edges weighted by TF-IDF-based cosine similarity scores. This graph serves as the backbone for a GNN that learns to classify the textual data effectively. Our system demonstrates a high degree of precision in identifying smoking-related content. By bridging the gap between traditional text classification and contemporary graph-based learning, our approach paves the way for novel applications in thematic text analysis.