LIAAD INESCTEC at SemEval-2026 Task 4: Unsupervised Narrative Similarity via Discourse Representation Structures and Sentence Embeddings

AbstractIn this paper, we describe an unsupervised approach using Discourse Representation Structures (DRS) for the SemEval-2026 Task 4. This task was Narrative Similarity and was formulated in two different tracks. Our team only developed a solution for track A, where the input is composed of a triplet: an anchor story, a story A, and a story B. The output in this formulation is to predict which story, A or B, is more similar to the anchor story. Our approach parsed each story and transformed in a DRS format,then we leveraged its structure and extracted features, performing ablation experiments inthe development dataset. Our strategy achieved 0.5975 accuracy in the official blind test set.