Abstract
Despite extensive research efforts in the recent years, computational modeling of argumentation remains one of the most challenging areas of natural language processing (NLP). This is primarily due to inherent complexity of the cognitive processes behind human argumentation, which commonly combine and integrate plethora of different types of knowledge, requiring from computational models capabilities that are far beyond what is needed for most other (i.e., simpler) natural language understanding tasks. The existing large body of work on mining, assessing, generating, and reasoning over arguments largely acknowledges that much more common sense and world knowledge needs to be integrated into computational models that would accurately model argumentation. A systematic overview and organization of the types of knowledge introduced in existing models of computational argumentation (CA) is, however, missing and this hinders targeted progress in the field. In this survey paper, we fill this gap by (1) proposing a pyramid of types of knowledge required in CA tasks, (2) analysing the state of the art with respect to the reliance and exploitation of these types of knowledge, for each of the for main research areas in CA, and (3) outlining and discussing directions for future research efforts in CA.
Introduction: The Complexity of Argumentation
Computational Argumentation (CA) represents one of the most sophisticated challenges in Natural Language Processing. While tasks like sentiment analysis focus on identifying what an opinion is, CA seeks to understand the underlying why and how—modeling how premises support conclusions and how stances are logically constructed. This process is inherently "knowledge-hungry," requiring an integration of common sense, logic, and an understanding of human values that goes far beyond simple pattern matching.
Despite a decade of rapid progress, the field has lacked a systematic framework to categorize the types of knowledge being utilized. Researchers frequently call for "more knowledge," but without a clear taxonomy, it is difficult to measure progress or identify specific gaps. This work addresses this by providing the first comprehensive survey of knowledge usage across 162 CA publications.
Mapping the Computational Argumentation Landscape
The authors categorize the field into four primary subareas, each representing a different stage of argumentative processing:
- Argument Mining: Extracting argumentative structures from raw text.
- Argument Assessment: Evaluating the quality, stance, or framing of arguments.
- Argument Reasoning: Identifying implicit gaps or warrants between claims.
- Argument Generation: Synthesizing new claims or full arguments.
The Knowledge Pyramid
To analyze how these subareas leverage information beyond the input text, the authors propose a hierarchical "Knowledge Pyramid." This taxonomy ranks knowledge by its specificity to the domain of argumentation:
- Linguistic Knowledge: The foundation, consisting of syntax, word embeddings, and general language models.
- World and Topic Knowledge: Factual data, common sense, and logical axioms (often sourced from Wikipedia or Wikidata).
- Argumentation-Specific Knowledge: Rhetorical strategies, sentiment, and theoretical models of argument quality.
- Task-Specific Knowledge: Highly narrow information, such as user-specific beliefs or specific discourse markers.
Analysis of the State of the Art
By auditing 162 papers and conducting an in-depth manual annotation of 40 representative works, several key trends emerge:
The Dominance of Linguistic Patterns
Despite the inherent complexity of argumentation, the survey reveals a significant reliance on lower-level knowledge. Approximately 37% of the surveyed papers rely purely on linguistic features. While 102 papers utilize some form of "advanced" knowledge, this is often limited to argumentation-specific lexicons rather than deep world knowledge.
The Reasoning Bottleneck
Argument Reasoning—the task of identifying implicit premises or "warrants"—is the most reliant on high-level world and topic knowledge. Conversely, Argument Mining remains heavily driven by linguistic cues. The analysis suggests that the biggest hurdle in the field is not a lack of algorithms, but a scarcity of structured knowledge resources. While raw text is abundant, "knowledge-rich" argumentative corpora that link claims to logical axioms or real-world facts are rare.
Implicit vs. Explicit Knowledge
In modern neural architectures, knowledge is often "latent," baked into the weights of pre-trained Transformers. However, the authors find that explicit modeling of world knowledge (e.g., through knowledge graphs) consistently leads to performance gains, yet remains underutilized.
Conclusion and Future Directions
The study validates the Knowledge Pyramid as a robust framework (achieving a Cohen’s κ of 0.74). To move the field forward, the authors propose several strategic shifts:
- Unified Representations: Developing a space to consolidate fragmented knowledge sources like lexicons and knowledge graphs.
- Knowledge-Rich Corpora: Building datasets that explicitly map argumentative text to the underlying world knowledge required for its comprehension.
- Cross-Pollination: Adapting techniques from argument generation (which uses significant task-specific knowledge) to improve the identification of complex structures in argument mining.
Ultimately, while "knowledge is power," the current state of CA shows that we have only begun to tap into the higher levels of the pyramid. Transitioning from linguistic pattern recognition to true argumentative reasoning will require a concerted effort to integrate structured, explicit world knowledge into our models.
Summary updated