argument quality assessment

This paper addresses the pronounced subjectivity and multiplicity of perspectives in argument quality assessment, as well as the lack of annotator metadata and multidimensional quality annotations in existing datasets. We propose a perspectivist framework for argument quality evaluation. Through systematic literature review and hierarchical taxonomy design, we establish the first unified classification scheme and evaluation standard for argument datasets in perspectivist NLP. We conduct the first systematic analysis of annotator metadata—including background and stance—to quantify their influence on quality judgments. Pilot experiments validate the comparability and interoperability of our taxonomy. Key contributions include: (1) a standardized, multidimensional classification framework covering coherence, relevance, evidential support, and rhetorical effectiveness; (2) identification of high-quality datasets amenable to non-aggregated, fine-grained modeling; and (3) practical guidelines for annotator diversity control and transparent reporting of perspective-dependent judgments.

This study addresses argument mining in educational settings—specifically, identifying, classifying, and evaluating arguments in student persuasive essays. We propose a lightweight framework based on small open-source decoder-only large language models (e.g., Phi-3, TinyLlama). Methodologically, it integrates few-shot prompting with supervised fine-tuning, leveraging sequence labeling for argument segmentation and text classification for type identification and quality assessment—enabling localized, low-overhead, privacy-preserving real-time feedback. Our key contribution is the first systematic investigation of multi-task adaptability of compact LLMs for educational argument mining, moving beyond traditional encoder-based architectures while balancing deployability and performance. Experiments show that fine-tuned models significantly outperform baselines on the Feedback Prize dataset (argument segmentation F1 +8.2%, type classification accuracy +6.5%); few-shot prompting achieves baseline-level performance in quality assessment, validating the efficacy of the lightweight paradigm.

Existing argument mining approaches in educational settings predominantly rely on simplistic binary relations, failing to capture the intricate vertical (hierarchical) and horizontal (sequential/cohesive) argumentative structures prevalent in student essays. Method: We propose a fine-grained argumentation analysis framework tailored for education, introducing the first taxonomy of 14 argument relation types—including vertical relations (e.g., claim-support, evidence-explanation) and horizontal relations (e.g., contrast, concession, adversative shift). The framework jointly addresses argument unit identification, argument relation classification, and automated essay scoring via a unified architecture integrating large language model fine-tuning with multi-task learning, thereby co-modeling argument structure, writing quality, and discourse relations. Contribution/Results: Experiments demonstrate substantial improvements in argument unit detection and relation classification accuracy. Fine-grained relational annotation enhances both the interpretability and precision of automated scoring, advancing argument mining from binary modeling toward multidimensional, structured representation.

This study addresses argument mining from online comments on polarized public controversies (e.g., abortion), proposing a three-stage predefined argument mining framework: argument existence detection, span extraction, and logical relation classification. We conduct the first systematic evaluation of four state-of-the-art large language models (LLMs)—including fine-tuned variants—on multi-topic, few-shot, high-affectivity argument understanding. Validated on a dataset comprising 2,000+ comments across six polarized topics, our framework achieves strong overall performance. Concurrently, we uncover systematic LLM deficiencies in long-text reasoning and emotionally charged language processing, and quantify their environmental footprint. Key contributions include: (1) a reproducible, end-to-end argument mining pipeline; (2) a cross-topic benchmark for argument understanding under realistic social media conditions; and (3) identification of critical bottlenecks limiting current LLMs’ applicability to public opinion analysis tasks.

Legal argument mining has long been hindered by structural bottlenecks—such as data standardization, modeling efficacy, and domain adaptability—stemming from the absence of structured representations that balance theoretical expressiveness with computational feasibility. This study systematically examines the current state and inherent tensions in the field across data, technical, and theoretical dimensions, proposing a novel paradigm that integrates legal dogmatics with computational modeling. By synthesizing multidisciplinary approaches—including legal text analysis, formal argumentation theory, traditional machine learning, and large language models—the work articulates design principles for effective structured representations. It clarifies core challenges and outlines pathways for breakthroughs, offering a theoretically grounded and computationally efficient framework to advance legal artificial intelligence through synergistic theoretical reconstruction and technical innovation.

This work addresses the lack of explicit, verifiable reasoning mechanisms in current large language models for debate analysis, which hinders structured representation of support and attack relations among arguments and their collective acceptability. The paper proposes the first unified framework that integrates large language model–driven argument mining, quantitative argumentation semantics, and fuzzy description logic to automatically construct a fuzzy argumentation knowledge base from raw debate texts. By leveraging an efficient query rewriting technique, the framework enables interpretable formal reasoning over this knowledge base. This approach overcomes the black-box limitations of purely statistical models, supports complex semantic queries, and significantly enhances the transparency, verifiability, and logical rigor of computational debate analysis.

Existing argument analysis tools struggle to evaluate the conditional validity of arguments across diverse worldviews. This work proposes a multi-perspective reasoning framework that operationalizes conditional validity for the first time by integrating structured worldview modeling, three-tier natural language inference, and conditional reasoning with large language models. The system automatically identifies value conflicts and assumption gaps, generating perspective-specific explanations. It further supports interactive visualization to effectively reveal differences in logical and normative coherence of the same argument under pluralistic value systems, thereby enabling users to explore multidimensional interpretations of complex arguments.

This work proposes a novel end-to-end approach to argument mining by reframing the task as a text-to-text generation problem, thereby circumventing the complexity of traditional multi-stage pipelines that rely on rule-based post-processing and extensive hyperparameter tuning. Leveraging a pretrained encoder-decoder language model, the method jointly generates argument spans, their component types, and relational structures in a unified framework, eliminating the need for task-specific post-processing steps. This design significantly simplifies the overall pipeline and enhances structural adaptability. The approach achieves state-of-the-art performance across three established benchmark datasets—AAEC, AbstRCT, and CDCP—demonstrating both its effectiveness and generalizability in diverse argumentative contexts.

This study addresses the high labor intensity and poor scalability of human debate coaching by investigating persuasive argument construction differences between GPT-4 and undergraduate debaters during pre-competition preparation. Methodologically, it pioneers the systematic application of Aristotle’s rhetorical triad—ethos (credibility), pathos (emotional appeal), and logos (logical structure)—to human-AI argument comparison, integrating prompt engineering, qualitative coding, quantitative rhetorical dimension analysis, and double-blind comparative experiments. Results indicate that while AI demonstrates robust performance in logos, it exhibits significant deficiencies in ethos and pathos; human arguments consistently achieve superior contextual adaptability and narrative tension. Based on these findings, the study proposes a “rhetorically augmented AI prompting template” that enhances the humanistic quality and persuasive efficacy of generated arguments. This contribution provides both a theoretical framework and actionable pedagogical strategies for integrating AI into critical thinking instruction.

This work proposes an end-to-end approach based on large language models (LLMs) for automatically reconstructing argument structures from natural language texts. The method identifies premises, conclusions, and their logical relations—specifically support, attack, and undercut—and generates corresponding abstract argumentation graphs in the form of directed acyclic graphs. By integrating natural language understanding with structured graph representation through a multi-stage pipeline, this study presents the first fully LLM-driven argument reconstruction framework, which is flexible enough to accommodate diverse annotation schemes. Experimental results demonstrate that the system accurately reproduces argument structures from theoretical textbooks according to human evaluation and achieves performance comparable to existing methods across multiple standard datasets.