A combined multi-margin contrastive learning with granulated data for warrant identification in computational argumentation

Information Sciences. 2025. Vol. 699. P. 1–17.

Behzadidoost R.

Argumentation reflects the cognitive processes humans use to justify and persuade through natural language. Computational argumentation, the attempt to model this reasoning process, is a challenging task in the field of natural language processing. Among the main components of argument reasoning, warrants play a central role in connecting premises to claims by explaining the implicit reasoning that justifies how a claim logically follows from its premises. Accurate identification of warrants is important for uncovering the underlying reasoning structure and enhancing argumentation models, particularly in applications like stance detection. To deal with such inherent complexities of extracting these logical structures, we propose a deep learning model that combines a pre-trained language model and a novel multi-margin contrastive learning method for warrant identification. The pre-trained language model learns complex linguistic patterns and contextual information; on the other hand, the multi-margin contrastive loss function fine-tunes feature learning, thereby distinguishing between positive and negative classes effectively. To that end, a novel decomposition granulation algorithm was given by the contrastive learning approach, enriching the textual data set by generating many meaningless textual data. This algorithm decomposes the data into fine-grained components, which increase the training set and thereon enhances the identification of warrants by the model. Experimental results demonstrate that while the existing models have accuracy rates from 47.30% to 58.56%, the proposed one holds a much higher accuracy of 76.80%, outperforming state-of-the-art approaches to warrant identification.

Keywords: Granular computing, Contrastive learning, Pre-trained language models, Computational argumentation, Natural language processing