The framework, developed by Yuhang Du of Boda College of Jilin Normal University, rests on a deceptively simple observation: literary texts contain recurring imagery, typed symbolic relations, and long-range emotional shifts that sequence models do not represent explicitly. Convolutional networks, LSTMs, and pretrained Transformers can identify lexical polarity and local semantic patterns, but literary passages often reuse imagery across distant discourse units and assign different functions to the same expression. A sentiment classifier may register that a passage is melancholic, yet it cannot explain how a predicted emotion connects to a recurring symbol through a traceable chain of relations. SymbolFusion was built to close that gap by predicting a typed symbolic graph, window-level emotion distributions, an emotion trajectory, transition points, and ranked evidence paths directly from raw text.

The technical core of the method is what the author calls a literary compatibility tensor. The system first extracts candidate nodes from a passage using deterministic rules: maximal noun, adjective, and verb phrases are identified with spaCy, and a span is retained as an imagery candidate when its head word has a concreteness rating of at least 3.0 in published English norms or belongs to a frozen sensory lexicon. It becomes a symbolic candidate when a training-fitted metaphor classifier returns a probability of at least 0.55, or when it recurs in at least two sentence groups with thematic similarity of at least 0.45. Three node types populate the resulting graph: imagery units, conceptual-symbolic units, and predicted affective states. Edges are restricted to exactly three relation types: symbolic entailment, which links a concrete image to a concept or affective function; metaphorical recurrence, which links repeated or semantically equivalent images; and narrative echo, which connects non-adjacent units whose thematic function reappears after at least three sentence groups.

What distinguishes SymbolFusion from relational graph neural networks such as R-GCN and HGT, which parameterize edge types but do not encode literary validity rules, is that a single binary compatibility tensor is used twice. During message passing, it produces a hard attention mask that sets invalid node-relation combinations to minus infinity, guaranteeing zero invalid attention at inference. During training, the same tensor supplies a pre-mask regularizer that penalizes large scores on incompatible candidates, teaching the unmasked scorer to separate valid from invalid relations. Relation-specific transformations and narrative distance are therefore learned only within an auditable validity space, and every accepted edge carries a traceable audit trail back to its type and threshold decision. A gated module then aligns literal, contextual, and narrative emotion representations with the constrained graph, so that emotion propagates only along symbolically meaningful pathways rather than simple token adjacency.

The evaluation is unusually rigorous for this field. The corpus contains 3,120 English literary passages drawn from 104 public-domain works by 47 authors on Project Gutenberg and the Gutenberg Poetry Corpus: 760 poetry samples, 920 narrative-fiction samples, 780 prose-essay samples, and 660 dramatic excerpts. Two trained annotators independently marked imagery spans, node categories, typed relations, window-level emotion distributions, and transition points, achieving Cohen’s kappa of 0.86 for symbolic-unit identification, 0.82 for relation typing, and 0.83 for dominant emotion before adjudication. Crucially, the passages are divided by literary work into 2,184 training, 468 validation, and 468 test passages, so that no source work appears in more than one partition, and a stricter author-disjoint evaluation further prevents stylistic overlap from inflating results. No gold symbolic or affective annotation is supplied during validation or testing; the model sees only raw text and frozen external lexical resources.

Across ten random seeds, SymbolFusion achieved 91.3 percent symbolic graph agreement, an emotion consistency score of 0.880, a symbolic-emotion alignment score of 0.820, and a transition-point F1 of 0.842. Against the strongest constraint-based baseline, a BERT variant augmented with the constraint-based learning loss of Diligenti and colleagues, the symbolic graph agreement gain was 2.2 percentage points, with a 95 percent bootstrap confidence interval of 1.5 to 2.9 and a p-value below 0.001. The comparison suite included BiLSTM, BERT, RoBERTa, DeBERTa-v3, Longformer, GATv2, R-GCN, HGT, and GraphGPS, all trained under identical splits, task heads, and seed lists. The latency cost was modest: end-to-end decoding rose from 60.1 to 63.2 milliseconds, well within the study’s prespecified five-millisecond bound relative to the constraint baseline.

Controlled ablations isolate exactly where the improvement comes from. Comparing unconstrained relational attention, a soft finite penalty, the proposed hard mask, and randomly permuted relation labels, performance improves monotonically from unconstrained attention through the soft penalty to the hard compatibility mask, and permuting relation types degrades all three principal outcomes below even the unconstrained model. This demonstrates that the gains arise from meaningful literary type semantics rather than from additional relation parameters. Removing the trajectory loss most strongly hurts emotion consistency, which falls from 0.880 to 0.810, and transition F1, which drops from 0.842 to 0.766, while removing gated fusion most strongly affects symbolic-emotion alignment, which falls from 0.820 to 0.720. Sensitivity analysis identified three graph layers, four attention heads, a 256-dimensional representation, and a 0.55 relation threshold as the validation-optimal configuration.

A worked example illustrates what the system actually produces. For a held-out public-domain poetry passage containing the candidates midnight, chamber, curtain, raven, and shadow, SymbolFusion restored four of five annotated relations and predicted a fear trajectory of 0.43, 0.56, 0.71, 0.82, and 0.78 across five discourse windows, closely tracking the annotated values of 0.41, 0.58, 0.73, 0.84, and 0.79. Both annotated transition points were recovered within the one-window tolerance, and the three highest-weight propagation paths, midnight-shadow, raven-shadow, and chamber-raven, expose precisely the symbolic relations associated with the predicted emotional escalation. The output presents these paths as inspectable evidence rather than as the only defensible interpretation, a deliberate epistemological restraint.

The error analysis is equally candid. Among the 58 of 468 work-disjoint test passages containing at least one incorrect typed relation or unmatched transition decision, irony or unreliable narration accounted for 18 errors, culture-specific symbolism for 15, long-range recurrence beyond the retained graph for 14, and ambiguous dramatic speakers for 11. Hard constraints can eliminate invalid type combinations, but they cannot resolve pragmatic inversion, missing cultural knowledge, or dependencies pruned outside the eight-window limit. Scalability measurements show approximately linear growth in latency, from 29.1 milliseconds for 200-word passages to 125.2 milliseconds for 1,000-word passages, with peak GPU memory reaching 10.9 gigabytes, supporting passage-level use while indicating that full novels would require hierarchical retrieval or graph compression.

The study is explicit about its boundaries: SymbolFusion is positioned as an inspectable computational aid rather than a replacement for human literary interpretation. Its outputs should be treated as testable hypotheses about symbolic structure and emotional dynamics, not as definitive reconstructions of authorial intention or aesthetic value. Yet the framework’s contribution is substantial for digital humanities: it operationalizes imagery and symbolic units, separates automatically computed test-time features from training annotations, and demonstrates that explicit literary compatibility rules, when coupled with hard masking and constraint regularization, measurably improve both graph restoration and emotion trajectory modeling. As computational literary studies continues to grapple with interpretability and data scarcity, SymbolFusion offers a template for how symbolic AI and neural learning can be fused without sacrificing the auditability that scholarly analysis demands.

Subject of Research: Neuro-symbolic graph learning with relation constraints for modeling literary symbolic structure and emotion trajectories

Article Title: SymbolFusion uses relation constraints in neurosymbolic graph learning to model literary structure and emotion trajectories

Article References: Du, Y. (2026). SymbolFusion uses relation constraints in neurosymbolic graph learning to model literary structure and emotion trajectories. Discover Artificial Intelligence, 6(1), Article 1540. https://doi.org/10.1007/s44163-026-02546-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02546-9

Keywords: SymbolFusion, neuro-symbolic AI, graph neural networks, computational literary studies, emotion trajectory modeling, relation constraints, symbolic graph learning, sentiment analysis, digital humanities, metaphor detection, attention masking, natural language processing