Representation of the Logical Structure (RLS)

• Representation of the Logical Structure (RLS) is a method for explicitly encoding logical content via structured forms, enabling robust and inspectable reasoning.
• RLS applies across neural networks, natural language processing, and abstract logic by externalizing atoms, rules, and dependencies into clear, task-oriented structures.
• It bridges practical implementations with theoretical insights by unifying symbolic, graph-based, and probabilistic approaches to capture and preserve inferential organization.

Searching arXiv for the cited RLS-related papers to ground the article in current records. Representation of the Logical Structure (RLS) denotes a family of approaches in which logical content is made explicit in a structured form rather than being left implicit in numeric parameters or unconstrained text. Across the literature, the represented object varies: it may be a neural topology whose connectivity mirrors rules, a task-dependent symbolic skeleton extracted from natural-language arguments, a logical structure tree over discourse connectives, an indexed semantic graph encoding scope and quantification, or an abstract consequence structure represented by valuations or scheme-theoretic objects. Despite these differences, the common aim is stable: to preserve the inferentially relevant organization of atoms, rules, connectives, contexts, and dependencies so that reasoning, inspection, or theorematic characterization becomes possible in a direct way (Wang, 2021, Shah et al., 20 Aug 2025, Bos, 2019, Roy et al., 2021, Breiner, 2014).

1. Conceptual definition and general scope

In explicit task-oriented form, RLS is defined as a structured representation of a natural-language statement that encodes all the relevant information for a given reasoning task. One formulation states that for a statement $S$ in the context of a reasoning task $R$, the corresponding RLS must be such that, if $S$ is replaced by its RLS in any inference or reasoning chain in which $S$ participates, the chain remains valid and evaluates to the same result as before (Shah et al., 20 Aug 2025). This definition is deliberately task dependent and is not constrained to a single formal syntax.

A broader philosophical formulation treats logic itself as the abstract theory of representation. On that view, a representational system is a set of entities $S$ for a universe $U$ such that $U$ can be reconstituted from $S$, at least virtually, according to a representational function associating contents in $U$ with representations in $S$. The same account distinguishes analogical components, which directly present their content, from symbolic components, whose content is not direct and therefore requires analogical mediation. Logical laws are then interpreted as laws governing how representations present contents and combine into larger representational structures (Plagnol, 2023).

This conceptual range explains why the term appears in disparate technical settings. In neural models, the emphasis is on storing if-then rules in graph topology rather than in opaque weights. In natural-language reasoning, the emphasis is on extracting logical atoms and rules from text before running symbolic inference. In semantic graph formalisms such as AMR, the emphasis is on separating predicate-argument structure from a distinct logical layer for scope, negation, and quantification. In universal logic and categorical logic, the emphasis is on representation theorems showing that abstract consequence relations or first-order theories can be recovered from suitably defined semantic structures (Wang, 2021, Wang, 2019, Bos, 2019, Roy et al., 2021, Breiner, 2014).

A useful unifying interpretation is that RLS externalizes inferential structure. This suggests that the central criterion is not any single notation, but preservation of the structure required for valid consequence, faithful reconstruction, and explicit readout.

2. Structural primitives and representational constraints

The most explicit symbolic accounts of RLS decompose it into atoms, rules, connectives, and contexts. In the natural-language argument setting, the atomic units are logical literals represented as tuples of natural-language words or phrases plus polarity, with rules taking the form $R$0. Facts are conjunctive sets of atoms, conclusions are consequent atoms, and polarity enables contradiction detection by conflicts between positive and negative literals (Shah et al., 20 Aug 2025).

A different but complementary decomposition appears in the theory of representation. There, the key primitives are the window of presence, the elementary analogical fragment (EAF), the analogical basis, the symbolic web, the semantic function, links, paths, unfolds, and analogical extensions. An analogical extension is defined as a set of fragments of universe reconstituted from some EAFs and some fragments of the web. On this view, conjunction expresses co-presence de jure, disjunction expresses co-display of alternatives, negation expresses incompatibility, and implication is interpreted primarily as a particular implication: “Given $R$1, $R$2” or “From $R$3, $R$4” (Plagnol, 2023).

That same framework states three logical properties that any adequate representational system should have: completeness, faithfulness, and coherence. Completeness requires that any fragment of the represented universe can be fully given within an analogical extension; symbolic completeness requires that any fragment can be coded. Faithfulness requires that reconstituted content $R$5 from a fragment $R$6 satisfy $R$7; otherwise the system reconstitutes a different universe. Coherence requires that explicitation by iterative projection gives rise to an analogical unified display, except for frontiers between different worlds. The account further distinguishes intrinsic from extrinsic incoherence (Plagnol, 2023).

In AMR-based logical structure, the primitives are indexed sub-AMRs and contextual constraints. AMR$R$8 assigns an index to each sub-AMR and adds structural constraints using four relations: identity =, negated context \lnot, conditional context \Rightarrow, and presuppositional context <. Here logical structure is not fused with ordinary predicate nodes; it is represented by how indexed contexts relate (Bos, 2019).

The following brief comparison summarizes several recurrent representational media.

Representation medium	Core units	Main purpose
Task-dependent RLS	atoms, polarity, conjunction, implication	preserve reasoning-relevant structure of arguments
Theory of representation	analogical basis, symbolic web, links, paths, extensions	reconstitute universes beyond immediate presence
AMR$R$9 logical layer	indexed sub-AMRs, contextual constraints	encode scope, negation, quantification, presupposition
Logical neural structure	neurons for things, links for logical relations	store if-then rules in connectivity
Logical structure tree	connectives as non-terminals, arguments as terminals	track hierarchical logic flow in statements

A recurring misconception is that RLS must be equivalent to full semantics. The task-dependent account explicitly rejects this: RLS is not a full semantic parse, but a relevant logical skeleton sufficient for the reasoning task (Shah et al., 20 Aug 2025). A related misconception is that symbolic representation is self-sufficient. The theory of representation denies this by insisting on the principle of analogical mediation: symbolic content, to be accessed, needs to be analogically presented (Plagnol, 2023).

3. Neural and graph-topological realizations

A prominent line of work realizes RLS directly in neural architecture. In “A Logical Neural Network Structure With More Direct Mapping From Logical Relations” (Wang, 2021), the design principle is explicit: neurons represent things only, and links represent logical relations only. The model targets if-then rules such as $S$0, $S$1, and standard operators including AND, OR, XOR, NAND, NOR, and XNOR. Positive Link (PL) represents $S$2, Negative Link (NL) represents $S$3, Composite Excitatory Link (CEL) fires only when all pre-end neurons are activated, and Composite Inhibitory Link (CIL) jointly inhibits an excitatory link. The paper contrasts this with earlier KBNN and continuous-logic LNN approaches, arguing that the newer mapping is more direct, more readable from topology, and uses fewer neurons; for XOR it contrasts its structure with a continuous-logic LNN requiring 7 neurons and 3 layers (Wang, 2021).

“A Novel Neural Network Structure Constructed according to Logical Relations” (Wang, 2019) extends the same separation of roles in the form of PLDNN, a Probabilistic Logical Dynamical Neural Network. Here a neuron $S$4 represents a thing or proposition and takes states $S$5 (resting), $S$6 (positively activated), and $S$7 (negatively activated). Links represent logical relations: EL for excitation and IL for inhibition, with composite variants CEL and CIL. The logical relation form is

$S$8

The network is built synchronically as events occur, with a four-stage cycle: perceive the event $S$9, associate to infer a reasoning set $S$0, perceive the next event $S$1 as the actual set $S$2, and learn or adjust by comparing $S$3 and $S$4. Wrong cases trigger the addition of ELs or ILs to refine representation. A frequency-counting mechanism assigns each link $S$5 and $S$6, with weight

$S$7

interpreted as a conditional probability for ELs or ILs. The paper therefore shifts RLS from purely structural mapping to structural-plus-probabilistic representation (Wang, 2019).

The sulfuric acid production example in PLDNN illustrates why inhibition is structurally central. The event sequence

$S$8

shows that naive all-to-all excitation can over-generate outputs; inhibitory links are inserted to block unwanted pathways such as incorrect excitation of $S$9 (Wang, 2019). The same structural logic underlies the treatment of competing rule bodies in the more direct logical ANN, where inhibitory links prevent interference between rules such as $S$0 and $S$1 (Wang, 2021).

Experimental demonstrations in PLDNN are representation-oriented rather than purely predictive. On the Zoo dataset, with incremental expansions from 10 attributes / 20 animals to 15 attributes / 40 animals and 20 attributes / 60 animals, the reported recalling degree is “all” in each case. On Breast Cancer Wisconsin (Original), feature values are used as pre-event and diagnosis class as post-event, allowing diagnosis to be recalled from the constructed links (Wang, 2019).

A related but non-neural graph setting appears in knowledge-graph rule learning. “Logical Entity RePresentation in Knowledge-Graphs for Differentiable Rule Learning” (Han et al., 2023) introduces LERP, defined for an entity $S$2 as

$S$3

where each component is a logical function over the entity’s surrounding sub-graph. The recursive function family includes true, chaining, negation, and merging by $S$4 or $S$5. This allows branched, conjunctive, disjunctive, and negated contextual constraints rather than only chain-like Horn clauses. The result is a shift from chain-only reasoning to chain-plus-context reasoning. On WN18RR, the reported MRR 0.622 exceeds NeuralLP at 0.435, DRUM at 0.486, and RNNLogic+ at 0.513; on WN18, the paper reports MRR 0.958 (Han et al., 2023).

These models share a strong thesis: logical relations should be stored in structure that can be inspected, extended, or read back into symbolic form. A plausible implication is that RLS in neural systems is best understood not as “neural reasoning” in the conventional black-box sense, but as topology-level knowledge representation with differentiated roles for nodes, edges, and inhibitory controls.

4. Natural-language and discourse-level representations

In natural-language reasoning, RLS commonly appears as an intermediate structure extracted from text and then consumed by a symbolic or structure-aware reasoner. “Reasoning is about giving reasons” (Shah et al., 20 Aug 2025) makes this separation explicit: first extract the core logical structure of the argument from text, then run deterministic symbolic reasoning over that structure. The paper uses a sequence-to-sequence transformer, specifically T5, to map natural-language sentences into encoded RLS strings. It uses T5-small for D3 and T5-base for the other experiments. No separate alignment algorithm is used; instead, target RLS strings are generated automatically from dataset metadata. Once extracted, reasoning on RuleTakers is handled by ProbLog. Reported extraction results are 95.9% exact match on CLUTRR, 99.6% on Leap-Of-Thought Hypernyms, and 99.8% on Leap-Of-Thought Counting. For reasoning, trained on D3 and tested on D5, the system reports 100% across all depths 0–5; on Birds-Electricity it reports 99.8% overall, compared with ProofWriter at 97.0% and PRover at 86.5% (Shah et al., 20 Aug 2025).

A distinct discourse-oriented construction appears in “Boosting Logical Fallacy Reasoning in LLMs via Logical Structure Tree” (Lei et al., 2024). There the logical structure tree is defined with relation connectives as non-terminal nodes and textual arguments as terminal nodes, mostly elementary discourse units. The tree is constructed in an unsupervised manner guided by a constituency tree and a taxonomy of ten common logical relations: conjunction, alternative, restatement, instantiation, contrast, concession, analogy, temporal, condition, and causal. Construction proceeds top-down and left-to-right through the constituency tree, recursively matching connective phrases and extracting their left and right arguments. The representation is designed to “explicitly represent and track the hierarchical logic flow among relation connectives and their arguments,” especially because logical fallacies often involve a mismatch between the relation suggested by a connective and the actual semantic support between arguments (Lei et al., 2024).

That tree is injected into LLMs in two ways. First, it is textualized as bottom-up triplets of left argument, relation connective, and right argument, and supplied as a hard text prompt: $S$6 Second, a relation-aware tree embedding is composed bottom-up and inserted as a soft prompt. For a simple tree with leaf children,

$S$7

and for nested trees,

$S$8

The tree embedding is then projected to the LLM space. On fallacy detection, the full model reports Argotario: Precision 86.02, Recall 88.40, F1 87.19; Reddit: Precision 70.05, Recall 84.80, F1 76.72; Climate: Precision 69.17, Recall 100.00, F1 81.78, with F1 increased by up to 3.45%. On fallacy classification, the F1 gain reaches up to 6.75% (Lei et al., 2024).

AMR-based approaches externalize logical structure at the graph level. “Abstract Meaning Representation-Based Logic-Driven Data Augmentation for Logical Reasoning” (Bao et al., 2023) treats AMR as “a structural representation of the semantic meaning and logical structure of a text via a rooted directed acyclic graph (DAG).” AMR-LDA uses the loop

$S$9

The parser is parse_xfm_bart_large and the generator is T5Wtense. Four logical equivalence laws are applied directly as graph transformations: contraposition, implication, commutative law, and double negation. For example,

$U$0

On ReClor, RoBERTa + AMR-LDA improves dev accuracy to 65.26 from 59.73 baseline, and DeBERTaV2 + AMR-LDA improves test accuracy to 77.63 from 70.46 baseline. For GPT-4 on LogiQA test, the paper reports 58.06 with AMR-LDA versus 53.88 baseline (Bao et al., 2023).

A more semantically exact use of AMR for logical structure appears in AMR$U$1. “Separating Argument Structure from Logical Structure in AMR” (Bos, 2019) argues that standard AMR is adequate for predicate-argument structure but not for scope, quantifiers, bound variables, valid inference behavior with negation, and presuppositional phenomena. AMR$U$2 preserves the original AMR content, minus polarity attributes, and adds a separate logical layer of indexed contexts and structural constraints. Existentials can share context; negation introduces a new negated context; universals introduce conditional context; proper names and definites introduce presuppositions; and bound variables are handled through indexed equality and context relations. The resulting framework is described as similar to Discourse Representation Theory (Bos, 2019).

Taken together, these approaches show that discourse-level RLS can be tree-shaped, graph-shaped, or literal-and-rule-shaped. What they share is explicit decomposition of text into structure that supports reasoning, augmentation, or fallacy diagnosis more directly than raw surface form.

5. Abstract logical structures and representation theorems

RLS also has a theorematic meaning in universal logic. “Lindenbaum-type Logical Structures” (Roy et al., 2021) studies logical structures as pairs

$U$3

with associated consequence operator

$U$4

The paper distinguishes Tarski-type structures, defined by reflexivity, monotonicity, and transitivity, from four Lindenbaum-types based on variants of extension principles using saturated or relatively maximal sets. It proves that the five classes—Tarski-type plus the four Lindenbaum-types—are distinct, establishes the inclusions

$U$5

and then studies the joint classes $U$6 (Roy et al., 2021).

Its central representational result concerns $U$7. Defining

$U$8

the paper proves the full theorem

$U$9

This gives an exact bivalent semantics for $U$0-type structures via characteristic functions of strongly closed saturated sets, together with adequacy and minimality results (Roy et al., 2021). Here RLS is not a data structure for computation but a representation theorem connecting proof-theoretic behavior to a semantic family of bivaluations.

A scheme-theoretic version appears in “Scheme representation for first-order logic” (Breiner, 2014). There the guiding analogy is between commutative rings and first-order theories or pretoposes, and between affine schemes $U$1 and logical affine schemes $U$2. For every pretopos $U$3, the paper constructs a topological groupoid $U$4 and an equivariant sheaf of pretoposes $U$5 such that

$U$6

For a coherent theory $U$7, the affine logical scheme is

$U$8

where points are labeled models and arrows are isomorphisms. Basic opens are determined by satisfaction of parameterized formulas, and the stalk at a point is the local diagram of the model. Definability is characterized geometrically: equivariant compact subsheaves are exactly definable ones (Breiner, 2014).

These abstract representations differ sharply from neural or NLP-oriented RLS, but they answer the same structural question at another level: what mathematical object exactly carries the logical organization of a theory or consequence relation? In one case the answer is a family of valuations over strongly closed saturated sets; in the other it is a space-plus-sheaf object over a spectrum of models.

6. Capabilities, interpretations, and limitations

The capabilities attributed to RLS are broad but not uniform across domains. In task-driven natural-language reasoning, once RLS is extracted, the same symbolic machinery is said to support deduction, abduction, contradiction detection, mistake rectification, and interactive discussion. The reason given is that these are all operations over the same explicit logical structure rather than over hidden neural states (Shah et al., 20 Aug 2025). In logical fallacy reasoning, explicit logical structure helps distinguish texts where connective phrases align with semantic support from those where they do not (Lei et al., 2024). In logical neural networks, direct mapping is intended to make rules readable from topology and to allow dynamic addition or adjustment of rules (Wang, 2021, Wang, 2019). In knowledge graphs, contextual logical representations allow more expressive rule families than pure chains (Han et al., 2023).

Several limitations are also explicit. The task-dependent RLS approach focuses on propositional structure and does not handle quantifiers; it extracts only the core logical structure, not full semantics; and exact string matching can fail where fuzzy unification would be needed (Shah et al., 20 Aug 2025). AMR-LDA states that AMR is a proxy and may not fully capture natural language variation, complex logical constructs, idioms, nuanced context, or varied logical frameworks; it mainly considers individual sentences and propositional logic statements in natural language (Bao et al., 2023). AMR$U$9 is more expressive with respect to scope and binding, but still less expressive than full DRS and may require inferred labels that are not always intuitive for annotators (Bos, 2019). In the theory of representation, completeness is described as often utopian for natural systems, even if possible for artificially fixed universes such as a website or a robot’s limited environment (Plagnol, 2023).

An important controversy concerns whether logical structure should be embedded in a fixed generic architecture or explicitly externalized into a representation that can be read, edited, or formally characterized. The neural papers reject fixed feedforward structures when different logical relations are forced into the same topology, because the corresponding logical relations cannot be read out clearly (Wang, 2021). The argument-centered RLS paper similarly rejects end-to-end systems that conflate structure extraction and reasoning (Shah et al., 20 Aug 2025). By contrast, the more abstract logical literature shows that explicit representation can itself become the object of theorems, not merely of engineering convenience (Roy et al., 2021, Breiner, 2014).

The overall significance of RLS lies in this convergence: whether implemented as neural links, discourse trees, AMR contexts, logical atoms and rules, bivaluation families, or logical schemes, the represented target is the inferential organization that makes reasons explicit. This suggests that RLS is best understood not as a single formalism, but as a methodological commitment to make logical dependencies structurally manifest rather than latent.