Skip to main navigation Skip to search Skip to main content

Enhancing Argument Mining with Interpretable Sources and Semantic Information

  • Somaye Moslemnejad

Student thesis: Doctoral Thesis › Doctor of Philosophy

Abstract

This thesis addresses critical challenges in Argument Mining (AM), specifically in Argument Relation Classification (ARC), by introducing interpretable and semantically enriched models that operate effectively in low-resource and domain-diverse settings. Existing ARC approaches often rely on opaque representations within Pre-trained and Large Language Models (PLMs/LLMs), lacking transparency and structured semantic reasoning between Argument Components (ACs). To overcome these limitations, this dissertation introduces three complementary and novel frameworks that integrate interpretable sources and structured semantic knowledge. Identifying Argumentative Relations using Interpretable Sources and Evidence-based Theory under Uncertainty introduces the DST-ARI (DempsterShafer Theory-based Argument Relation Identification) framework, which models uncertainty by fusing multiple interpretable sources—causal effect, causal concept, aspect stance, and opinion stance—to classify relations among ACs as Support, Attack, or No-relation without direct supervision. Prompt Templates for Argument Relation Classification using Frame Semantic Parsing presents the FSP-ARC framework, which introduces a semantic-aware pipeline that leverages frame semantic parsing via prompt templates within a RoBERTa-based Siamese model with multi-head cross attention to perform fine-grained and context-sensitive ARC. Event Knowledge Alignment for Argument Relation Classification proposes the EKA-ARC framework, which enriches argument components by aligning them with structured event knowledge extracted via Event Argument Extraction. Two alignment strategies—substitution-based and prompt-based—are explored for injecting event-aware semantics into ACs through a RoBERTa-based Siamese model with multi-head cross attention, enhancing the model’s capacity to capture implicit and contrastive argument relations. Extensive experiments on diverse AM datasets—including QT30, US2016, Essays, AbstRCT, AMPERE++, CMV, and Zillow—demonstrate that the proposed methods consistently outperform competitive baselines across accuracy, precision, recall, and macro F1 metrics. This thesis highlights the importance of integrating interpretable reasoning and structured semantic knowledge to develop robust, explainable, and adaptable ARC systems. The proposed frameworks advance AM toward real-world applications requiring transparency and semantic awareness.
Date of Award2026
Original languageEnglish
Awarding Institution
  • University of Dundee
SupervisorChris Reed (Supervisor) & John Lawrence (Supervisor)

Cite this

'