TL;DR
This paper provides the first systematic survey of how "knowledge"—information beyond the immediate input text—is used in Computational Argumentation (CA). The authors propose a four-level "knowledge pyramid" and reveal that while CA is inherently complex, most current models still rely heavily on basic linguistic patterns rather than structured world facts or rhetorical theory.
What Problem Does This Paper Address?
Computational Argumentation (CA) is one of the "final bosses" of Natural Language Processing. Unlike sentiment analysis, which asks what an opinion is, CA tries to model why a stance is held and how premises lead to conclusions. This requires more than just understanding words; it requires "knowledge"—common sense, logic, and an understanding of human values.
Despite a decade of progress in sub-tasks like argument mining and generation, the field lacked a clear map of what kinds of knowledge are actually being used. Researchers knew they needed "more knowledge," but there was no taxonomy to categorize it or a systematic study to show where the gaps were. This paper fills that void by defining what "knowledge" means in CA and auditing the state of the art.
Methods
The authors conducted a two-stage study:
- The Survey: They collected and categorized 162 CA publications across four main subareas: Mining (extracting structures), Assessment (judging quality), Reasoning (filling in implicit gaps), and Generation (creating new arguments).
- The Taxonomy (The Knowledge Pyramid): They developed a hierarchical model of knowledge specificity:
- Linguistic Knowledge: Basic syntax, word embeddings, and language models.
- World and Topic Knowledge: Common sense, factual data (e.g., from Wikipedia), and logic.
- Argumentation-Specific Knowledge: Rhetorical strategies, sentiment, and models of what makes an argument "good."
- Task-Specific Knowledge: Information relevant only to a narrow task, like user-specific beliefs or discourse markers.
- In-Depth Analysis: Three experts manually annotated 40 representative papers (10 per subarea) to identify exactly which types of knowledge were used and how they were integrated (e.g., via manual features, multi-task learning, or external lexicons).
Datasets
As a survey paper, this work does not introduce a new task dataset. Instead, it systematizes the use of existing resources mentioned across 162 papers, including:
- Structured Knowledge Bases: WordNet, Wikipedia, FrameNet, and Wikidata.
- Specialized Lexicons: Sentiment and subjectivity lexicons used for argument assessment.
- NLI Models: Pre-trained Natural Language Inference models used as a proxy for reasoning.
Key Insights and Ideas
- The "Knowledge-Hungry" Subarea: Argument Reasoning (identifying implicit premises or "warrants") relies most heavily on high-level world and topic knowledge. In contrast, Argument Mining is still largely driven by linguistic cues.
- The Distribution Gap: While 102 out of 162 papers used some form of "advanced" knowledge, 60 papers (roughly 37%) relied purely on linguistic features.
- The Bottleneck: The biggest hurdle isn't a lack of algorithms, but a lack of structured knowledge resources. We have plenty of raw text, but very few "knowledge-rich" argumentative corpora that map claims to real-world facts or logical axioms.
- Implicit vs. Explicit: Knowledge is often baked into models implicitly (e.g., through the weights of a pre-trained Transformer) rather than explicitly (e.g., a knowledge graph). The authors argue that explicit modeling of world knowledge consistently leads to performance gains but remains underutilized.
Key Results
- Taxonomy Validation: The authors achieved a "substantial" inter-annotator agreement (Cohen’s κ = 0.74) for their knowledge pyramid, suggesting it is a robust framework for future research.
- Argumentation-Specific Knowledge is King: Among "advanced" knowledge types, argumentation-specific knowledge (like rhetorical structures) is the most frequently used across all subareas.
- World Knowledge is Rare: Only 7 out of the 40 papers analyzed in-depth utilized world or common-sense knowledge, despite these being crucial for human-like reasoning.
Limitations and Future Work
The authors acknowledge that their in-depth study was limited to 40 papers, which, while representative, cannot capture every nuance of the field. Furthermore, they note that "knowledge" in modern neural models is often "latent" and hard to disentangle from basic linguistic patterns.
Future Directions: * Unified Representation: The authors call for a "unified knowledge representation space" that can consolidate fragmented sources (lexicons, KGs, and corpora) for CA. * Knowledge-Rich Corpora: There is an urgent need to build datasets that explicitly link argumentative text to the underlying world knowledge required to understand it. * Cross-Pollination: Techniques used in argument generation (which uses a lot of task-specific knowledge) should be adapted for argument mining to help identify more complex, implicit structures.
Summary updated