Periodic Table of Arguments

• Periodic Table of Arguments is a taxonomy that categorizes argument forms by organizing facts, stance, inference patterns, and linguistic realizations.
• It demonstrates that identical facts can yield distinct arguments when different reasoning schemes and stances are explicitly varied.
• The framework enables controllable generation and analysis through annotated corpora, graph structures, and computational models that map facts to structured argument texts.

The Periodic Table of Arguments is a proposed structured inventory of argument forms that organizes arguments along dimensions such as topic/content, stance/polarity, argument scheme / inference pattern, and surface realization. In computational argumentation, the idea is operationalized most directly by systems that separate what an argument is about from how it reasons: the same factual inputs can yield different arguments with different reasoning structures when stance and argument scheme are varied explicitly, making the table a taxonomy of controllable argumentative forms rather than a list of topics or conclusions (Saha et al., 2023).

1. Conceptual basis

The central premise of a periodic table of arguments is that argument types are not reducible to topic, conclusion, or lexical surface alone. The same set of facts may support multiple arguments whose inferential mechanisms differ, and those differences matter because arguments using the same facts can still differ in perlocutionary effect depending on their structure. In this view, a useful taxonomy must distinguish what facts or concepts an argument uses from how it reasons from premises to conclusion (Saha et al., 2023).

This orientation treats arguments as structured forms rather than mere text instances. A periodic table, in this sense, seeks recurrent, reusable patterns of inference that are stable across domains. The underlying idea is that the same inferential skeleton can recur in abortion, minimum wage, nuclear energy, gun control, the death penalty, or school uniform, even when the constituent facts differ. The table is therefore best understood as a structured inventory of reasoning archetypes plus the controls needed to instantiate them in text (Saha et al., 2023).

A common misconception is that a periodic table of arguments would simply classify claims as pro or con. The computational work most directly associated with the concept rejects that reduction. It separates stance from argument scheme, so that arguments can be controlled for polarity without collapsing distinct forms such as consequence-based, authority-based, or principle-based reasoning into a single class (Saha et al., 2023).

2. Organizing dimensions

The most explicit formulation organizes arguments along four dimensions: topic/content, stance/polarity, argument scheme / inference pattern, and surface realization. Topic specifies what facts or concepts are being used; stance specifies whether the argument supports or opposes a position; scheme specifies the inferential form; and surface realization specifies the linguistic instantiation. This suggests that a periodic table of arguments is not merely classificatory but generative: it can serve as a schema for constructing multiple textual realizations from the same underlying argumentative coordinates (Saha et al., 2023).

In computational form, this organization appears as a controlled mapping from topic, fact variables, stance, and scheme to argument text. The paper does not present the equation formally, but the implemented functional structure is:

$$(\text{topic}, \text{facts}, \text{stance}, \text{scheme}) \rightarrow \text{argument text}$$

This structure is significant because it turns a theoretical taxonomy into an executable representation: same “elements” (facts), different “compounds” (argument forms), depending on the inferential pattern selected (Saha et al., 2023).

The scheme dimension is especially important because it isolates recurring patterns of inference that are invariant to the constituent facts. A periodic table built on that principle treats scheme labels as stable building blocks. The stance dimension then cuts across those blocks, while surface realization captures the fact that one and the same inferential form can be rendered by multiple templates or phrasings. This suggests a layered architecture in which inferential form is primary, stance is orthogonal, and wording is secondary.

3. Scheme families

The most developed computational inventory currently linked to the periodic-table idea uses Walton’s argument schemes as explicit generation controls. The relevant operative inventory contains five substantive schemes for generation, with Others retained as a residual class in annotation and classification (Saha et al., 2023).

Scheme	Characterization in the data	Typical role
Means for Goal / Goal from Means	links an action or means to a desired end	instrumental reasoning
From Consequence	“X is/are (not) favourable as it Y”	consequence-based evaluation
From Source Knowledge	appeals to what is known, shown, or evidenced	evidential warrant
From Source Authority	appeals to an expert, advocate, supporter, or authoritative source	authority-backed support
Rule or Principle	reasons from normative standards, laws, rights, or principles	deontic or normative reasoning
Others	residual class in annotation/classification	non-productive residual class

Means for Goal / Goal from Means links an action or means to a desired end, or infers from a desired goal to adopting the means. Operationally, it expresses reasoning like “We want goal $G$, therefore we should adopt means $M$” or “$M$ is valuable because it advances $G$.”

From Consequence is the clearest evaluative-causal form. It reasons that an action or policy should or should not be accepted because it leads to a consequence. The paper’s canonical paraphrase is “$X$ is/are (not) favourable as it $Y$.”

From Source Knowledge appeals to what is known, shown, or proved, rather than to who says it. It introduces an epistemic warrant: evidence or knowledge establishes that a policy or action has some property or effect.

From Source Authority appeals to an expert, advocate, supporter, or authoritative source. The distinction from consequence-based reasoning is explicit: “Access to abortion is important towards reproductive rights” differs from “Reproductive rights advocates say enabling access to abortion is important towards reproductive rights.”

Rule or Principle reasons from conformity to, or violation of, a principle, law, right, or normative standard. It isolates deontic or normative reasoning from empirical consequence reasoning.

Others is a residual class used in annotation and classification. In the final generation corpus, samples predicted as “Others” are removed, so the class is not a productive building block in the final generation setup.

A common misunderstanding is that these schemes are merely topical paraphrases. The data state the opposite: “Argument schemes represent structures of inference and are invariant to the constituent facts.” A plausible implication is that a mature periodic table of arguments would require more than this starter set, but the existing inventory already spans causal evaluation, instrumental reasoning, authority, evidence, and normativity (Saha et al., 2023).

4. Corpora, annotation, and controllable generation

A periodic table of arguments requires more than a taxonomy; it requires data that align text with facts and inference types. The most substantial current corpus contribution reports 69,428 arguments spanning six topics and six argument schemes, with the six topics being abortion, minimum wage, nuclear energy, gun control, the death penalty, and school uniform. The corpus is described as the largest publicly available corpus for identifying argument schemes (Saha et al., 2023).

The construction proceeded in two phases. Phase 1 used the BASN dataset and produced a fully annotated 2,990-example corpus. Each example contained an argument and between 2 and 5 fact variables from a knowledge base of 205 facts (variables) across the six topics. The annotation tasks were Span Detection, which identifies non-overlapping factual spans in argument text, and Span Grounding, which grounds each factual span to one of the available knowledge-base variables or “Others.” 1,153 randomly sampled examples were manually annotated in Doccano by 2 expert annotators with backgrounds in computational linguistics and computer science; 66 shared samples yielded Cohen’s Kappa = 0.79, interpreted by the authors as substantially high agreement (Saha et al., 2023).

To label the remainder of BASN, the authors trained ArgSpan, a RoBERTa-based model that concatenates argument text and fact variables, uses a BOS token for each variable, applies a feedforward layer, and uses a biaffine layer to model interaction between argument tokens and variables while predicting BIO labels. Against one human evaluator on 300 random remaining examples, ArgSpan achieved Span Detection F1: 91.1% average across 300 examples and Span Grounding Accuracy: 89.2%. Those 300 examples were then added to the training set, giving 1,453 manually checked examples before retraining and inference over the rest of BASN (Saha et al., 2023).

Phase 2 expanded the corpus using 66,180 examples from the Aspect-Controlled Reddit and CommonCrawl corpus and 733 examples combined from Sentential Argument Mining, Arguments to Key Points, and debate portal-based Webis datasets. The initial parallel corpus had 1,272,548 examples, but only a filtered subset survived to the final 69,428-example corpus. Automatic labeling in this phase used ArgSpanScheme, which jointly handled BIO span labels and a multi-label (six labels including ‘Others’) classification for schemes. The stance labels used in generation were Pro and Con, with control tokens <pro> and <con>; scheme control codes included <from_consequence>, <from_source_authority>, <from_source_knowledge>, <goal_from_means/means_for_goal>, and <rule_or_principle> (Saha et al., 2023).

The generation system itself takes as input a topic, a set of knowledge-base fact variables, a desired stance, and a desired argument scheme, and outputs a short factual argument sentence that uses the supplied facts, expresses the requested stance, and instantiates the requested inference pattern. It also experiments with an “argument template” stage that establishes an inference strategy before actual argument generation. The importance of this workflow for a periodic table of arguments is direct: it treats the table not only as a descriptive taxonomy but as a control interface for factual argument generation (Saha et al., 2023).

5. Internal structure, relation types, and constructive methods

A periodic table of arguments cannot be limited to scheme labels alone. In formal argumentation, abstract graph structure is often too coarse, because arguments treated as atomic nodes lose distinctions in premises, claims, and inferential dependencies. Work on instantiating bipolar argument graphs with deductive arguments addresses exactly this issue by connecting the graph level—support, attack, or ambiguous influence—to the internal logical level of arguments with explicit supports and conclusions (Hunter, 2023).

In this framework, a deductive argument is a pair $(\Phi,\alpha)$ where $\Phi$ is minimal such that it is consistent, $\Phi \vdash \alpha$, and no strict subset of $G$0 entails $G$1. This makes internal structure explicit through support $G$2 and claim $G$3. Once such arguments instantiate graph nodes, relation types can be differentiated much more finely than by polarity alone. Attack can appear as defeater, undercut, direct undercut, canonical undercut, or defeating rebuttal. Support can range from consistency-only relations to premise support, subset-of-premises support, and full-premise support. The framework also includes an explicit * label for ambiguous influence, capturing cases where a source-target relation could reasonably be interpreted as support or attack depending on the logical or semantic context (Hunter, 2023).

This makes an important corrective to another common misconception: a periodic table of arguments is not exhausted by named schemes. It must also classify relation types, target loci, and graph construction policies. The data explicitly distinguish whether support concerns one premise, a subset of premises, or the full support of another argument; whether attack targets the claim or the premises; whether ambiguity is intrinsic; and whether graphs merely use a knowledge base, display it, or exhaust it. A plausible implication is that a mature periodic table would need at least two interacting layers: a layer of inferential forms and a layer of inter-argument relations.

Constructive proof methods add a second formal layer. In the argumentation tableau approach, a tableau node is not a set of formulas but a set of arguments, and the basic object is $G$4, where $G$5 is a conclusion and $G$6 is its support. Closure is itself argument-generating: from $G$7 and $G$8, the method derives $G$9. The framework extends from standard propositional and predicate logic to defeasible rules and reasoning by cases, thereby distinguishing deductive, defeasible, inconsistency, undercutting, and case-based arguments within one proof-theoretic machinery (Roos, 2022).

For a periodic table of arguments, this suggests that classification can also be constructive. Arguments may be typed not only by scheme but by how they are generated: atomic premise arguments, deductive composite arguments, defeasible rule applications, inconsistency arguments, undercutting arguments against rules or premises, and case-based arguments. The tableau perspective adds subargument structure, minimality, and branching pattern to the taxonomy (Roos, 2022).

6. Representation, limitations, and prospective extensions

The current state of the periodic-table idea is that of a computationally grounded but incomplete framework. The available scheme inventory is a coherent starter set, not an exhaustive universal classification; the field of computational argument generation is described as relatively new and as lacking datasets conducive to training; and the formal frameworks that enrich schemes with internal structure still remain tied to specific logical settings, especially classical deductive arguments (Saha et al., 2023).

Several limitations follow directly. First, classifying arguments only by graph polarity is too coarse, because two supporters or two attackers may differ radically in logical role. Second, a single set of instantiation constraints is not universally appropriate: some applications require dense analytical graphs, others sparse communicative graphs. Third, the deductive-graph framework does not yet provide full machinery for comparing instantiations, identifying similarity between instantiated graphs, or handling approximate arguments and enthymemes. Fourth, the tableau method is foundational rather than optimized: efficient implementation and adaptation to other logics are left as future work (Hunter, 2023).

A possible extension comes from work on using a scientific table itself as a structured input representation for machine learning. There, the general lesson is that if symbolic entities can be placed in a principled table with meaningful adjacency and topology, then a neural model can “read” the table through convolution. The paper explicitly proposes the abstraction that arguments, claims, fallacies, rhetorical moves, or inferential forms are placed in a structured 2D table, nearby positions encode similarity, opposition, transformability, or shared inferential role, and an instance is represented by activating the corresponding cells (Konno, 2019). This suggests a future computational interpretation of the periodic table of arguments as a learned or hand-designed coordinate system for argumentative forms.

Taken together, these strands define the topic with some precision. The periodic table of arguments is not a finished chart but a research program: a structured inventory of argument forms grounded in controllable scheme labels, factual inputs, and stance controls; enriched by logical internal structure, support and attack subtypes, and constructive derivation procedures; and potentially representable as a structured symbolic space for downstream computational models. Its distinctive claim is that argumentative regularities can be made explicit at the level of inferential architecture, not only at the levels of subject matter, sentiment, or discourse surface.