Scientific discovery depends on connecting claims scattered across a fragmented literature—yet researchers’ tools offer either creative breadth or rigorous traceability, rarely both. We present a framework that unites large language models with Semantic Web knowledge graphs, enabling hypothesis exploration grounded in explicit, navigable chains of evidence. Our approach models a hypothesis as a semantic relation between concepts and introduces evidence paths : sequences of semantic predications, each published as a nanopublication with provenance metadata. These paths transform knowledge graphs into hypertextual spaces where navigation corresponds to traversing semantically meaningful, source-anchored links across the literature. The framework operates through an iterative loop. This iterative back-and-forth keeps exploration anchored to structured evidence whilst exploiting the generative capacity of language models to surface novel connections. We illustrate feasibility through a biomedical case study, showing how evidence paths expose both supporting mechanisms and countervailing risks. The framework contributes to hypertext research by treating scientific knowledge as a linked, traversable structure—advancing a vision of the WWW as a substrate for machine-assisted reasoning.

Science advances by connecting ideas—yet the literature recording those ideas has grown beyond any individual's capacity to survey [12]. Researchers confront not merely a mountain of publications but a fragmented landscape: findings relevant to a hypothesis may lie buried in distant disciplines, expressed in unfamiliar vocabularies, invisible to those best placed to recognise their significance [23, 25]. This problem is sharper today because of scale, but it is not new [3]. The long neglect of Mendel's laws of heredity illustrates how fragmentation can delay recognition of important connections [3].

Computational systems can help by surfacing implicit connections, generating candidate hypotheses, and retrieving evidence that supports or contradicts them [5, 6, 22]. Two kinds of machine capability are especially relevant today. Large language models (LLMs) process vast corpora and can propose associations, reformulations, and speculative hypotheses directly from natural language [7, 29]. Knowledge graphs (KG) and Semantic Web representations, by contrast, encode claims in a logical, machine-processable form that supports explicit reasoning and traceable evidence retrieval [9, 17, 18].

Each capability has characteristic limitations. Language models capture statistical regularities but do not inherently produce explicit evidence structures; their outputs may be plausible yet ungrounded [11, 28]. Knowledge graphs offer traceability but depend on structured data and cannot easily accommodate the open-ended reformulation that exploratory science demands [5, 6, 8]. This complementarity creates a natural opportunity: generative and reasoning-based processes can interact, each compensating for the other's weaknesses.

Iterative Interaction. We introduce a framework realising this interaction. The system combines an LLM-based hypothesis generator with a symbolic reasoner operating over a Semantic Web knowledge graph (SWKG) of scientific claims [8, 18]. A hypothesis is modelled as a semantic relation between two concepts; its evidence path is a sequence of semantic predications connecting them. Because each predication is published as a nanopublication carrying provenance metadata, the resulting chains can be inspected, validated, and reused by both humans and machines [9, 17].

Not all paths in the graph constitute meaningful evidence. To ensure that retrieved connections can be interpreted as genuine support, paths must respect semantic constraints: predicates carry specific domains and ranges, concepts bear semantic types, and admissible predicate sequences must form coherent reasoning chains [6]. Enforcing these constraints prevents the system from retrieving invalid or misleading connections [6, 8].

The hypertextual space can be further enriched by incorporating knowledge from the broader Semantic Web, including Linked Open Data and domain ontologies [1, 13, 19, 26]. Where ontologies are formally specified—for example in OWL—their axioms can be applied during reasoning, enabling the system to interpret richer semantic structures and retrieve more meaningful paths [4, 26].

Contributions. This paper makes three contributions:

We introduce evidence paths as a formal representation for hypothesis support, grounding each path in provenance-bearing nanopublications.\

We present an architecture that integrates LLM-based hypothesis generation with symbolic reasoning over a SWKG, maintaining strict separation between generative and evidential roles.\

We illustrate the framework through a biomedical case study, showing how evidence paths expose both supporting mechanisms and countervailing risks.\

The ambition of enlisting machines in the service of science has a long history. In 1945, Bush identified the need for tools that could help researchers navigate the ‘growing mountain of research’ and trace unexpected connections across fields [3, p. 101]. This body of knowledge is not merely immense; owing to disciplinary specialisation, it is dispersed across literatures and may remain invisible to those best placed to recognise its significance.

Literature-Based Discovery. A foundational idea in literature-based discovery (LBD) is that new hypotheses may emerge by connecting claims reported in separate publications. Swanson's work on undiscovered public knowledge established the classical example: the relation between fish oil and Raynaud's disease, identified by linking intermediate findings not explicitly connected anywhere in the literature [22, 23]. Later systems explored computational support for identifying such connections [20, 24].

Subsequent studies have shown that discovery depends on more than retrieving related documents or co-occurring terms. Graph-based approaches demonstrate that support for a hypothesis can emerge through paths connecting intermediate concepts and relations [5, 6]. Evaluation work has distinguished between exploratory settings, in which a system searches for previously unnoticed relations, and closed settings, in which it examines whether support exists for a specific hypothesis [27]. This distinction frames discovery as the identification and examination of structured connections across distributed claims.

Semantic Predications. A major line of work represents scientific assertions as semantic predications: tuples of the form $({it subject}, {it predicate}, {it object})$. In the biomedical domain, SemRep extracts semantic predications from the literature using domain knowledge and linguistic analysis [15, 21], whilst SemMedDB provides a large repository derived from PubMed abstracts [16]. By making relation types explicit, predications offer a more suitable basis for structured reasoning than approaches relying solely on term co-occurrence [5, 25].

The effectiveness of reasoning over scientific claims depends on semantic normalisation. In biomedical applications, normalisation typically uses the UMLS Metathesaurus and Semantic Network, which provide controlled concepts, semantic types, and relation definitions [2, 14]. These resources enable machines to connect claims expressed with lexical variation and to constrain reasoning through type information [6, 15].

Previously, we introduced the web of machine-readable scientific facts: a Semantic Web infrastructure in which scientific claims are represented as semantic predications and published as nanopublications enriched with provenance [18]. Recently, we showed that symbolic inference over such a network can generate scientific hypotheses through explicit inference chains, evaluated using a temporally-grounded methodology [8]. These results establish semantic predications as a practical unit for reasoning over distributed scientific claims.

Semantic Web Representations. SemWeb technologies provide a framework for representing scientific claims as machine-readable, reusable, and traceable knowledge. RDF (Resource Description Framework) represents claims as structured statements. Nanopublications offer a model for publishing atomic scientific claims together with provenance and publication metadata [9, 17]. This representation is especially relevant in scientific settings, where interpretation of a claim depends on its source, context, and epistemic status.

Because provenance is included in the encoding [18], evidence paths can be inspected for validation and reused across systems. Unlike opaque language model outputs, our evidence paths are intended to be read as traceable connections among published claims.

Ontologies further enrich this space by contributing semantic knowledge exploited during reasoning. They constrain admissible concepts, relations, and combinations, helping the system interpret richer structures and retrieve more meaningful paths. When formally defined in OWL, they support verification of semantic constraints through axioms, restrictions, and property definitions. This combination of machine-readable claims, provenance, and ontological structure makes Semantic Web representations well suited to hypertextual exploration of scientific knowledge.

Large Language Models. Large language models have recently been explored for scientific hypothesis generation and research ideation [7, 29]. Such models can propose candidate hypotheses from natural language—including speculative or novel ones—by capturing statistical regularities and latent associations across large corpora. They can also reformulate research questions and suggest plausible connections. These capabilities make them relevant to exploratory settings where the starting point is a natural language information need rather than a structured query. They do not, however, produce explicit evidence structures to support those hypotheses.

Recent work on retrieval-augmented generation (RAG) has highlighted the importance of grounding generated outputs in external sources whenever provenance and evidential support matter [11, 28]. In related work, language models have been used as constrained mediators between natural language and explicit reasoning outputs, preserving a separation between linguistic interaction and symbolic inference. The present paper extends that direction by considering language models as components in hypothesis generation and exploration, whilst keeping evidence retrieval within the reasoning layer.

Positioning. Two directions appear especially relevant today: language models that process text and suggest associations, and KGs that support explicit reasoning. The present work is situated at their intersection, exploiting the strengths of each whilst mitigating their limitations through structured interaction.

The proposed framework differs from existing approaches by combining LLM-based generation with semantic reasoning over a Web of Facts. Reasoning-oriented Semantic Web systems, such as HyQue [4], evaluate user-defined hypotheses over RDF knowledge bases using SPARQL queries and predefined scoring rules; our framework instead supports exploratory hypothesis formulation and retrieves multi-step evidence paths for assessment. SemMedDB-based LBD systems use semantic predications as connections, often mediated by contextual structures such as MeSH descriptor clusters [5, 6]. They do not, however, reason over paths whose predications can be assessed through semantic types, provenance, qualifying values, and epistemic information.

LLM-only approaches generate hypotheses [7, 29], but rely on generated plausibility and therefore remain exposed to hallucinations or unsupported associations.

In contrast, our framework uses the LLM to propose and refine candidate hypotheses, whilst the reasoning layer retrieves evidence paths whose predications can be constrained and inspected directly according to predicate semantics, semantic types, provenance, and qualifying values.

This section introduces the formal representation of hypotheses, evidence, and assessment. The model is grounded in Semantic Web representations and interprets the KG as a hypertextual structure of scientific claims.

Knowledge Representation. We consider a KG $mathcal {G}$ composed of semantic predications extracted from the literature. Each predication is a tuple: p = (s, r, o, π) where s is the subject concept, r is the predicate (relation), o is the object concept, and π represents provenance information modelled according to the nanopublication framework. Concepts are drawn from a controlled vocabulary or ontology; predicates belong to a predefined set of semantic relations (e.g., CAUSES, TREATS, INHIBITS).

The graph $mathcal {G}$ is a directed labelled multigraph: nodes correspond to concepts, and each predication defines a directed edge from s to o labelled by r. The provenance π links each statement to its supporting sources. This representation follows the Semantic Web paradigm, where each semantic predication can be interpreted as an RDF statement enriched with provenance. The resulting structure forms a hypertextual network of scientific claims, where navigation corresponds to traversing semantically meaningful links.

Hypotheses. A hypothesis is a target relation between two concepts: H = (a, rh, b) where a and b are concepts and rh is the relation of interest. Hypotheses may be derived from natural language statements and mapped to this structured representation.

Evidence Paths. An evidence path connecting a to b is a sequence of semantic predications: P = [p1, p2, …, pn] where each pi = (si, ri, oi, πi) defines a directed edge from si to oi, with: s1 = a,  on = b,  oi = si + 1 for i < n.

From a graph perspective, an evidence path is a traversal from a to b. From a Semantic Web perspective, it is a sequence of RDF statements forming a structured argument. Each path carries provenance elements {π1, …, πn}, allowing the entire chain to be traced to supporting literature.

Semantic Constraints. Not all paths in $mathcal {G}$ are meaningful for assessing hypotheses. We define a constraint set $mathcal {C}$ restricting admissible paths:

compatibility between predicate domains and ranges and concept types.\

admissible sequences of relations forming coherent reasoning chains.\

bounds on the number of edges in a path.\

For example, the composition:

begin{eqnarray} texttt{INHIBITS}(X,Y) wedge texttt{CAUSES}(Y,Z) Rightarrow text{negative influence of } X text{ on } Z end{eqnarray} \

defines a semantically meaningful chain. A path P is valid if all constituent predications satisfy $mathcal {C}$.

Supporting and Contradicting Evidence. Given a hypothesis H = (a, rh, b), we define:

$mathcal {P}^+(H)$: paths whose predicate composition is compatible with rh.\

$mathcal {P}^-(H)$: paths whose predicate composition contradicts orweakens rh.\

Compatibility and contradiction are determined by domain knowledge and predicate semantics.

Assessment Function. We define an assessment function evaluating a hypothesis based on its evidence:

begin{eqnarray} mathrm{score}(H) = fbigl (mathcal {P}^+(H), mathcal {P}^-(H), Pi mid mathcal {C}bigr) end{eqnarray} \

where Π denotes provenance elements associated with the paths. The function f may consider:

the number of supporting and contradicting paths;\

the diversity and coherence of predicate sequences;\

the quantity and reliability of provenance sources;\

structural properties of the paths.\

The score provides a structured measure of support within available knowledge; it does not constitute formal proof.

Iterative Refinement. The system operates iteratively:

A hypothesis H0 is generated from a natural language query.\

Evidence paths $mathcal {P}(H0)$ are retrieved and filtered under $mathcal {C}$.\

The hypothesis is assessed using score(H0).\

The retrieved evidence and its assessment are interpreted by the language model, which may stop, reformulate the hypothesis, or propose a refined hypothesis H1.\

This produces a sequence H0 → H1 → H2 → ⋅⋅⋅ interpretable as navigation within a hypertextual space, where each step traverses semantically grounded, provenance-aware links.

The Discovery Orchestrator sends prompts to the LLM Gateway, which responds with hypotheses. The Orchestrator sends entities to the Normaliser, which sends CUIs to the Hypothesis Compiler, which sends queries to the SAETTA Reasoner. SAETTA interacts bidirectionally with the KG and returns evidence paths to the Orchestrator.

The proposed system implements an iterative discovery loop between a language model and a reasoning component operating over a resident knowledge graph. Its purpose is to let the language model propose, reformulate, and prioritise candidate hypotheses whilst a structured reasoning layer evaluates them against explicit evidence and constraints. The architecture enforces strict separation of roles: the language model drives search and reformulation; the reasoner remains the authoritative layer for evidential support.

Front-end. The user-facing part of the system is depicted on the left-hand side of Figure 1. The front-end uses natural language. The accepts a seed query, and returns an overall summary accompanied by evidence summaries derived from the retrieved paths.

Core Components. The architecture, depicted on the right-hand side of Figure 1, centres on five components:

Manages the interaction loop and maintains session state across iterations. It sends the current query to the language model with a prompt that derives a discovery objective and produces one or more candidate hypotheses in structured form. It determines when to continue exploration, when to refine, and when to terminate.\

Handles prompts for hypothesis generation and for interpreting reasoning outcomes. The gateway abstracts the specific language model, allowing different models to be substituted.\

Maps free-text entities and predicates to controlled graph identifiers—UMLS concept unique identifiers (CUIs) and allowed predicate names. Normalisation failures are reported to the orchestrator, which may prompt the LLM to reformulate.\

Translates normalised hypotheses into executable requests for the reasoning engine. The compiler enforces structural constraints before queries reach the reasoner.\

The reasoning engine is SAETTA (Semantic Analysis Engine for Traceable Text-based Assertions), a simplified Datalog reasoner developed in Rust for this project. It evaluates requests against a preloaded graph built from semantic predications derived from SemMedDB and nanopublications and returns support status, bindings, witnesses, and execution metadata.\

SAETTA. SAETTA operates over a preloaded resident graph and is designed for interactive exploration through very low-latency evaluation of bounded queries. In initial testing on a graph of approximately 1.5 × 107  semantic predications, a two-hop query between fixed endpoints completed in 7.1 × 10− 4 ms, whilst a two-hop query starting from a fixed node and retrieving all connected nodes completed in 22.1 ms.

Latency matters because users should not wait long between iterations if human intervention is to remain practical (see §5).

SAETTA is not the orchestrator and does not interpret natural language. Its role is strictly to evaluate structured requests against the graph and return results. This separation ensures that evidential claims are grounded in explicit graph traversal rather than language model generation.

Iterative Discovery Loop. The runtime flow begins with a seed query in natural language. The orchestrator sends this to the LLM Gateway with a prompt that derives a discovery objective and produces candidate hypotheses in structured form.

These hypotheses are normalised into controlled identifiers and compiled into SAETTA request programs. The reasoner evaluates them against the resident graph and returns whether they are supported, unsupported, or unresolved, together with supporting witnesses. The orchestrator packages these results into a second prompt, and the language model decides whether to stop, reformulate, or expand the search.

The loop continues until a stopping condition is reached:

sufficient support is found;\

repeated normalisation failure indicates the hypothesis cannot be grounded;\

the iteration budget is exhausted;\

the system detects cycling over equivalent hypotheses.\

Constraints. Hypotheses are evaluated only after normalisation and compilation into a constrained reasoning form. Constraints apply at multiple levels:

entities must resolve to acceptable identifiers;\

predicates must belong to the controlled set;\

rules must remain within the SAETTA language subset;\

reasoning is bounded by branching, iteration, and budget limits.\

Additional semantic-type constraints can be injected where useful. These restrictions keep the language-model search space aligned with the graph and control the cost and meaning of reasoning.

Division of Roles. Within this design, the reasoner is the source of truth for support. The language model handles search, reformulation, prioritisation, and evidence interpretation. The reasoning layer handles truth-conditional evaluation against structured evidence. This distinction is a defining property of the architecture: natural-language generation and symbolic reasoning cooperate without being conflated. Implications of this distinction are described in §6.

Prototype. A prototype of this architecture is under development1. The reasoning layer operates over a graph of literature-derived semantic predications with in-memory, resident evaluation; the language-model layer receives natural-language information needs, proposes hypotheses, and interprets returned evidence. The architecture remains modular: alternative language models, normalisation strategies, ontologies, or reasoning engines can be integrated without changing the overall workflow.

To illustrate the framework in practice, we present a concrete example from the biomedical domain. This case study illustrates how an initial hypothesis is evaluated, how evidence paths are interpreted, and how interpretation drives iterative refinement.

Initial Hypothesis. Consider the hypothesis:

begin{eqnarray} H0 = (text{Metformin}, texttt{TREATS}, text{Alzheimer'{}s Disease}) end{eqnarray} \

This reflects a genuine open question. Metformin, a first-line treatment for type 2 diabetes, has attracted interest as a potential neuroprotective agent owing to shared pathological mechanisms—metabolic dysfunction, insulin resistance, and neuroinflammation—between metabolic disease and neurodegeneration.

Evidence Retrieval. The following paths are constructed to illustrate the framework's behaviour. The system queries $mathcal {G}$ and retrieves candidate evidence paths connecting Metformin to Alzheimer's Disease:

begin{align} P1 &: text{Metformin} xrightarrow {texttt{ACTIVATES}} text{AMPK} xrightarrow {texttt{INHIBITS}} text{mTOR} nonumber &quad xrightarrow {texttt{ASSOCIATED_WITH}} text{Alzheimer'{}s Disease} P2 &: text{Metformin} xrightarrow {texttt{REDUCES}} text{Inflammation} nonumber &quad xrightarrow {texttt{ASSOCIATED_WITH}} text{Alzheimer'{}s Disease} P3 &: text{Metformin} xrightarrow {texttt{CAUSES}} text{Vitamin B12 Deficiency} nonumber &quad xrightarrow {texttt{ASSOCIATED_WITH}} text{Cognitive Impairment} end{align} _

Each predication carries provenance linking it to source publications.

Path Filtering. Retrieved paths are filtered according to semantic constraints $mathcal {C}$ encoding admissible predicate compositions. For example:

begin{eqnarray} texttt{ACTIVATES}(X,Y) wedge texttt{INHIBITS}(Y,Z) Rightarrow text{modulation of } Z text{ by } X end{eqnarray} \

Applying these constraints:

P1 is retained as mechanistically plausible: AMPK and mTOR provide a well-characterised molecular basis linking Metformin to neuroprotection.\

P2 is retained with lower confidence: the path passes through neuroinflammation and therefore provides a broader, less specific mechanistic connection to Alzheimer's disease than the predicate sequence in P1.\

P3 is classified as risk-related evidence: it suggests a potential adverse effect on cognitive function and does not support the hypothesis.\

The resulting sets are:

begin{eqnarray} mathcal {P}^+(H0) = lbrace P1, P2rbrace, quad mathcal {P}^-(H0) = lbrace P3rbrace end{eqnarray} \

Evidence Assessment. The hypothesis H0 is assessed by weighing supporting and contradicting evidence. Mechanistic paths such as P1 carry strong evidential weight: their predicate sequences reflect well-characterised biological processes. Indirect associations such as P2 contribute weaker but non-negligible support: the chain passes through intermediate concepts whose connection to the source is less tightly constrained.

The presence of P3 introduces a counter-signal that qualifies the overall assessment. Support indicators associated with the underlying predications could help determine the strength of this counter-signal. Rather than invalidating the hypothesis, it motivates further refinement. The assessment integrates number, structure, and provenance of retrieved paths into a nuanced evidential judgement driving the next iteration.

At this point, a human in the loop [10] could refine the natural language prompt or issue a new one. For this example, we assume batch-mode operation.

Summarisation and Interpretation. The evidence paths are synthesised into a natural language summary:

Metformin may influence Alzheimer's Disease through activation of AMPK and downstream modulation of mTOR signalling. Additional evidence suggests a contributory role for its anti-inflammatory effects. However, a potential adverse pathway —via vitamin B12 deficiency and associated cognitive impairment— qualifies this support and warrants further investigation.

This summary is grounded in structured, provenance-bearing evidence and can be returned to the user as an interpretable account of the current state of the hypothesis. At the same time, the refinement process may continue on the basis of the structured information retrieved from the knowledge graph.

Hypothesis Refinement (retrieval of evidence path). Based on the assessed evidence, the system proposes a refined hypothesis replacing the direct treatment claim with a mechanistic account:

begin{eqnarray} begin{aligned} H1 =;& (text{Metformin}, texttt{MODULATES}, text{AMPK/mTOR Pathway}) &wedge (text{Pathway}, texttt{ASSOCIATED_WITH}, text{Alzheimer'{}s Disease}) end{aligned} end{eqnarray} _

Path P1 provides explicit evidence for this more specific mechanistic formulation. This formulation is also amenable to further assessment, since each conjunct can be assessed independently.

Iterative Exploration. The refined hypothesis H1 can serve as the starting point for subsequent iteration. Alternatively, the counter-signal in P3 motivates a parallel line of enquiry:

begin{eqnarray} H2 = (text{Metformin}, texttt{AFFECTS}, text{Neuroinflammation}) end{eqnarray} \

Each hypothesis generates new evidence paths, which are filtered, assessed, and interpreted—driving continued exploration of the hypertextual space. All edges carry provenance, linking them to concrete sources.

Summary. This example demonstrates how the framework supports structured hypothesis assessment. The knowledge graph is interpreted as a hypertextual space where navigation is guided by semantic constraints and grounded in provenance-aware evidence. The process highlights the coexistence of supporting and contradicting evidence and supports traceable hypothesis refinement—essential for practical use in scientific workflows where hypotheses must be evaluated amid heterogeneous and sometimes conflicting evidence.

The proposed framework combines two capabilities often developed separately: language models that formulate hypotheses from natural language, and symbolic reasoning that retrieves explicit evidence paths from a knowledge graph. Their interaction defines an iterative process in which hypothesis exploration remains grounded in structured, traceable evidence.

The Role of Evidence Paths. A central aspect of the framework is the role assigned to evidence paths. Support is returned as paths composed of semantic predications, each associated with provenance, making results inspectable at the level of intermediate concepts, predicate sequences, and source claims. In scientific settings, this matters because the value of a hypothesis depends on how support is obtained and how clearly it can be inspected.

Evidence paths also serve an explanatory function. Rather than presenting a binary supported/unsupported verdict, the system exposes the reasoning chain, allowing researchers to evaluate the strength and relevance of each link. This transparency distinguishes the framework from approaches that treat the knowledge graph as a black-box retrieval system.

The major contribution is to enlarge the portion of objective knowledge that becomes effectively available as working knowledge, by exposing explicit, traceable connections that would otherwise remain difficult to identify.

Hypertext Perspective. The framework is also relevant from a hypertext perspective. Scientific claims are treated as linked units of knowledge, explored through typed relations across a connected space of assertions. A candidate hypothesis's meaning depends on the paths connecting it to other claims, the intermediate mechanisms those paths reveal, and any contradictory paths that emerge—supporting navigation through a hypertextual space where exploration and assessment remain closely related.

The case study illustrates this: the initial hypothesis is examined through supporting and risk-related paths, interpreted in light of provenance and predicate structure, and reformulated into a more precise mechanistic hypothesis.

Human in the Loop. Researchers can inspect evidence paths, evaluate provenance, and guide refinement by modifying natural language prompts. This human-in-the-loop capability is essential for scientific workflows where automated systems should augment rather than replace human judgement.

The low latency of SAETTA is designed with this interaction in mind: By keeping response times short, the system supports the kind of iterative, reflective exploration that characterises scientific reasoning.

Limitations. The quality of exploration depends on coverage and reliability of the underlying graph, quality of normalisation, and availability of meaningful semantic constraints. True relations may be missed if relevant evidence is: absent, weakly represented, or not reachable by the reasoning patterns.

On the language-model side, quality depends on discipline imposed by the orchestration layer. Weak or repetitive hypotheses reduce the usefulness of the iterative loop. Also, the system operates only on knowledge that has been structured, normalised, and made available for reasoning—a partial body of knowledge, as researchers themselves work with.

Scope of Applicability. The application requires claims in machine-readable form, and for claims to be inspectable, the provenance must be encoded. The biomedical domain, with resources such as SemMedDB and UMLS Metathesaurus, provides a natural testbed.

The vision motivating this work is of the WWW as a substrate for machine-assisted scientific reasoning—a space where knowledge is not merely stored but actively traversed, where hypotheses are not merely asserted but grounded in navigable chains of evidence. LLMs contribute hypothesis generation, reformulation, and interpretation; the reasoning layer retrieves explicit evidence paths connecting machine-readable scientific claims.

The contribution concerns both scientific hypothesis exploration and the broader hypertext perspective of traversing linked, provenance-aware knowledge structures.

Evidence paths are the central innovation: they transform the KG from a static repository into a navigable space where each link carries meaning and each chain of links constitutes an argument. By grounding LLM outputs in explicit, inspectable evidence, the framework addresses a fundamental limitation of generative AI in scientific contexts.

1 We plan to demonstrate the prototype at the Conference.

ACM full-text HTML

ACM Digital Library

Imported from ACM’s structured HTML source. ACM Reference Format: Nicola Raffaele Di Matteo and James Blustein. 2026. Evidence Paths: Grounding LLM-Generated Hypotheses in Semantic Web Knowledge Graphs. In 37th ACM Conference on Hypertext (HT '26), September 14--18, 2026, London, United Kingdom. ACM, New York, NY, USA 7 Pages. https://doi.org/10.1145/3800935.3830863

[1] François Belleau, Marc-Alexandre Nolin, Nicole Tourigny, Patrice Rigault, and Jean Morissette. 2008. Bio2RDF: Towards a mashup to build bioinformatics knowledge systems. Journal of Biomedical Informatics 41, 5 (2008), 706–716. <10.1016/j.jbi.2008.03.004>

[2] Olivier Bodenreider. 2004. The Unified Medical Language System (UMLS): Integrating biomedical terminology. Nucleic Acids Research 32, Database issue (Jan. 2004), D267–D270. <10.1093/nar/gkh061>

[3] Vannevar Bush. 1945. As We May Think. The Atlantic 176, 1 (July 1945), 101–108. Retrieved Dec. 01, 2020 from https://www.theatlantic.com/magazine/archive/1945/07/as-we-may-think/303881/

[4] Alison Callahan, Michel Dumontier, and Nigam H. Shah. 2011. HyQue: evaluating hypotheses using Semantic Web technologies. Journal of Biomedical Semantics Suppl 2, S3 (May 2011), 11 pages. <10.1186/2041-1480-2-s2-s3>

[5] Delroy Cameron, Olivier Bodenreider, Hima Yalamanchili, Tu Danh, Sreeram Vallabhaneni, Krishnaprasad Thirunarayan, Amit P. Sheth, and Thomas C. Rindflesch. 2013. A graph-based recovery and decomposition of Swanson's hypothesis using semantic predications. Journal of Biomedical Informatics 46, 2 (April 2013), 238–251. <10.1016/j.jbi.2012.09.004>

[6] Delroy Cameron, Ramakanth Kavuluru, Thomas C. Rindflesch, Amit P. Sheth, Krishnaprasad Thirunarayan, and Olivier Bodenreider. 2015. Context-driven automatic subgraph creation for literature-based discovery. Journal of Biomedical Informatics 54 (April 2015), 141–157. <10.1016/j.jbi.2015.01.014>

[7] Qiguang Chen, Mingda Yang, Libo Qin, Jinhao Liu, Zheng Yan, Jiannan Guan, Dengyun Peng, Yiyan Ji, Hanjing Li, Mengkang Hu, Yimeng Zhang, Yihao Liang, Yuhang Zhou, Jiaqi Wang, Zhi Chen, and Wanxiang Che. 2025. AI4Research: A Survey of Artificial Intelligence for Scientific Research. <10.48550/arXiv.2507.01903>

[8] Nicola Raffaele Di Matteo. 2025. Automatic Scientific Discoveries Using a Public Collection of Characterized Semantic Predications. In 2025 IEEE International Conference on Information Reuse and Integration and Data Science(IRI 2025). IEEE, San Jose, CA, USA, 104–109. <10.1109/IRI66576.2025.00026>

[9] Paul Groth, Andrew Gibson, and Jan Velterop. 2010. The anatomy of a nanopublication. Information Services & Use 30, 1–2 (Sept. 2010), 51–56. <10.3233/ISU-2010-0613>

[10] Yeming Hu, Evangelos E. Milios, and James Blustein. 2014. Interactive document clustering with feature supervision through reweighting. Intelligent Data Analysis 18, 4 (2014), 561–581. <10.3233/IDA-140658>

[11] Yizheng Huang and Jimmy Xiangji Huang. 2026. A Survey on Retrieval-Augmented Text Generation for Large Language Models. Comput. Surveys 58, 12 (Sept. 2026), 1–38. <10.1145/3805774>

[12] Arif E. Jinha. 2010. Article 50 million: An estimate of the number of scholarly articles in existence. Learned Publishing 23, 3 (2010), 258–263. <10.1087/20100308>

[13] Saminda Kanza and Jeremy G. Frey. 2019. A new wave of innovation in Semantic Web tools for drug discovery. Expert Opinion on Drug Discovery 14, 5 (2019), 433–444. <10.1080/17460441.2019.1586880>

[14] Halil Kilicoglu, Graciela Rosemblat, Marcelo Fiszman, and Thomas C. Rindflesch. 2011. Constructing a semantic predication gold standard from the biomedical literature. BMC Bioinformatics 12, 486 (Dec. 2011), 17 pages. <10.1186/1471-2105-12-486>

[15] Halil Kilicoglu, Graciela Rosemblat, Marcelo Fiszman, and Dongwook Shin. 2020. Broad-coverage biomedical relation extraction with SemRep. BMC Bioinformatics 21, 188 (May 2020), 28 pages. <10.1186/s12859-020-3517-7>

[16] Halil Kilicoglu, Dongwook Shin, Marcelo Fiszman, Graciela Rosemblat, and Thomas C. Rindflesch. 2012. SemMedDB: a PubMed-scale repository of biomedical semantic predications. Bioinformatics 28, 23 (Dec. 2012), 3158–3160. <10.1093/bioinformatics/bts591>

[17] Tobias Kuhn, Paolo Emilio Barbano, Mate Levente Nagy, and Michael Krauthammer. 2013. Broadening the Scope of Nanopublications. In The Semantic Web: Semantics and Big Data(Lecture Notes in Computer Science), Philipp Cimiano, Oscar Corcho, Valentina Presutti, Laura Hollink, and Sebastian Rudolph (Eds.). Springer, Berlin, Heidelberg, 487–501. <10.1007/978-3-642-38288-833>

[18] Nicola Raffaele Di Matteo, Andrea Schimmenti, Fabio Vitali, and James Blustein. 2024. Building and Exploiting a Web of Machine-Readable Scientific Facts to Make Discoveries. In Information and Research science Connecting to Digital and Library science(IRCDL 2024). CEUR Workshop proceedings, Bressanone, Brixen, Italy, 15 pages. https://ceur-ws.org/Vol-3643/paper1.pdf

[19] Eric K. Neumann, Eric Miller, and John Wilbanks. 2004. What the Semantic Web could do for the life sciences. Drug Discovery Today: BIOSILICO 2, 6 (2004), 228–236. <10.1016/S1741-8364(04)02420-5>

[20] Wanda Pratt and Meliha Yetisgen-Yildiz. 2003. LitLinker: capturing connections across the biomedical literature. In Proceedings of the 2nd international conference on Knowledge capture(K-CAP ’03). Association for Computing Machinery, New York, NY, USA, 105–112. <10.1145/945645.945662>

[21] Thomas C Rindflesch and Marcelo Fiszman. 2003. The interaction of domain knowledge and linguistic structure in natural language processing: interpreting hypernymic propositions in biomedical text. Journal of Biomedical Informatics 36, 6 (Dec. 2003), 462–477. <10.1016/j.jbi.2003.11.003>

[22] Don R Swanson. 1986. Fish Oil, Raynaud's Syndrome, and Undiscovered Public Knowledge. Perspectives in Biology and Medicine 30, 1 (Autumn 1986), 7–18. <10.1353/pbm.1986.0087>

[23] Don R. Swanson. 1986. Undiscovered Public Knowledge. The Library Quarterly 56, 2 (April 1986), 103–118. <10.1086/601720>

[24] Don R Swanson and Neil R Smalheiser. 1997. An interactive system for finding complementary literatures: a stimulus to scientific discovery. Artificial Intelligence 91, 2 (April 1997), 183–203. <10.1016/S0004-3702(97)00008-8>

[25] Menasha Thilakaratne, Katrina Falkner, and Thushari Atapattu. 2020. A Systematic Review on Literature-based Discovery: General Overview, Methodology, & Statistical Analysis. Comput. Surveys 52, 6 (Jan. 2020), 1–34. <10.1145/3365756>

[26] World Wide Web Consortium. 2015. Semantic Web. https://www.w3.org/standards/semanticweb/

[27] M. Yetisgen-Yildiz and W. Pratt. 2008. Evaluation of Literature-Based Discovery Systems. In Literature-based Discovery, Peter Bruza and Marc Weeber (Eds.). Springer, Berlin, Heidelberg, 101–113. <10.1007/978-3-540-68690-37>

[28] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. 2025. Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models. Computational Linguistics 51, 4 (Dec. 2025), 1373–1418. <10.1162/COLI.a.16>Retrieved April 23, 2026 from https://direct.mit.edu/coli/article/51/4/1373/131631/Siren-s-Song-in-the-AI-Ocean-A-Survey-on

[29] Yangqiaoyu Zhou, Haokun Liu, Tejes Srivastava, Hongyuan Mei, and Chenhao Tan. 2024. Hypothesis Generation with Large Language Models. In Proceedings of the 1st Workshop on NLP for Science (NLP4Science), Lotem Peled-Cohen, Nitay Calderon, Shir Lissak, and Roi Reichart (Eds.). Association for Computational Linguistics, Miami, FL, USA, 117–139. <10.18653/v1/2024.nlp4science-1.10>