Stance detection represents a relatively recent addition to the family of automatic information extraction tasks in natural language processing. It encompasses two primary types: targeted stance detection and text-based stance detection. This study develops a text-based stance detector capable of analyzing headlines and articles to predict, for a given headline-article pair, whether their relationship should be labelled “agree”, “disagree”, “unrelated” or “discuss”. The Fake News Challenge dataset was utilized to train the model. Previous research has primarily focused on enhancing overall model accuracy or improving metrics defined by the dataset creators on their website. In this research, the F1 measure is the principal metric used to evaluate performance of the model for each class. The primary objective is to ensure the model comprehends the contextual relationship between headlines and article bodies. A cascaded models approach is presented, training two models using an attention-based Siamese network: one for relevance detection and the other for category classification. Additionally, a pretrained BERT model is used for tokenization and embedding representations of both article bodies and headlines. Results on the test set provided by dataset creators indicate significant improvement in class-wise F1 values compared to a baseline model.