Recent advances in machine learning and deep learning have expanded Natural Language Processing (NLP) applications in stance detection, stance distribution analysis, and claim verification. Yet many high-performing models remain opaque, producing accurate predictions without revealing how decisions were reached. Building systems that provide predictions with clear, faithful explanations is essential for trust, accountability, and model refinement. Existing explainability techniques, such as attention visualizations, surface only shallow signals, while chain-of-thought prompting can produce convincing but unfaithful post-hoc justifications. This thesis develops targeted frameworks that produce structured, evidence-grounded explanations for stance detection, stance distribution, and claim verification. Stance detection determines whether an article supports, opposes, or remains neutral toward a claim, critical for fact-checking and understanding media coverage. We introduce Stance Tree, an unsupervised framework rooted in Rhetorical Structure Theory. Stance Tree organizes articles into hierarchical discourse structures, evaluates Elementary Discourse Unit stances, and combines them using Dempster-Shafer Theory. This produces accurate predictions and structured explanations highlighting the most relevant argumentative units. Understanding opinion distribution across article collections reveals patterns in public discourse. Our graph-based framework organizes arguments into communities where nodes represent arguments and edges capture relationships. The framework explains homogeneity (single-stance communities), controversy (competing-stance communities), and how subtopics connect arguments from multiple authors. This can help readers navigate complex discourse and identify where alignment or conflict arises. We address claim verification from tabular data, essential for scientific publishing, financial auditing, and sustainability reporting. We propose the Multi-Agentic Claim vErification (MACE) framework, a prompt-based system using specialized agents that collaboratively verify claims and generate step-by-step explanations indicating veracity labels, reasoning strategies, mathematical operations, and assumptions. These works advance explainability in three interconnected tasks. Stance Tree provides discourse-structured explanations at the article level. Stance distribution analysis extends insights to article collections, revealing how perspectives cluster or conflict. MACE enables transparent verification of claims against tabular evidence. By grounding explanations in discourse structures, argument communities, and agent-based reasoning, this thesis contributes frameworks that make stance reasoning and evidence use more interpretable and practically useful.
Rudra Saha (2026) studied this question.