PaperlayerView exact revision on arXiv

Source-linked AI summary

SUDO: From Scientific Publications to Global Scholarly Discourse

Nipun D. Pathirage, Henrique Santos, Deborah L. McGuinness

arXiv:2610.33018v1cs.CLcs.AIcs.DL

TL;DR

Existing scholarly knowledge graphs preselect entities or statements, capturing only parts of how papers build and communicate knowledge. SUDO models that discourse through linked, source-grounded structures, and evaluations show broader representational coverage and improved scientific question answering. Its direct comparisons with some existing systems are limited by evaluation and corpus coverage.

  • Problem

    Existing scholarly knowledge graphs preselect concepts or statements, leaving publication discourse only partially represented and source connections difficult to preserve across scholarly knowledge-modeling tasks.

    Source: §1 · pp. 1–2, §2.1 · p. 3

  • Method

    SUDO models scholarly knowledge through linked information and propositions, retaining source-grounded relations while connecting recurring concepts across publications.

    Source: §1 · p. 2, §4.3 · p. 6, §4.4 · p. 8

  • Results

    SUDO achieved broader, more balanced use-case coverage than the examined concept- and statement-centric graphs, and query-guided retrieval reached an F1 score of 82.44.

    Source: §5.1.2 · p. 10, §5.2.1 · p. 11

  • Takeaways & Limitations

    SUDO provides a discourse-aware representation that preserves how knowledge is communicated within publications and connects recurring concepts across the literature.

    Source: §7 · p. 12

    Directly comparable scores were unavailable for CS-KG, which yielded too few reliable alignments, and ORKG, which covered only 34 of 413 QASPER papers.

    Source: §5.2.1 · p. 11

Abstract

from arXiv · show

Scholarly Knowledge Graphs (SKGs) support literature search, scientific discovery, and agent-assisted research. Yet, most approaches to SKG construction pre-specify the knowledge unit to be represented, typically named scholarly entities or contribution-level statements, and retain only the publication content expressible through that unit. This choice produces two dominant yet distinct paradigms: concept-centric SKGs that preserve entities and local relations, and statement-centric SKGs that capture high-level discourse in sentences. In both cases, the graph reflects a chosen extraction target rather than how a paper constructs and communicates knowledge. We introduce SUDO, the Scholarly Upper Discourse Ontology, which defines how knowledge elements are organized and related to form a paper's research narrative, namely its scholarly discourse. Using linguistic principles that generalize across fields and tasks, SUDO enables each publication's scholarly discourse to be constructed from the bottom up as a coherent, source-grounded representation. Applied consistently across papers, the same structure connects their research narratives into a global scholarly discourse without discarding local context or provenance. We demonstrate SUDO's capabilities across a comprehensive suite of scholarly knowledge-modeling use cases and performance improvements on a downstream Large Language Model (LLM)-based scientific QA task.

1. Introduction

Existing scholarly knowledge graphs capture selected entities or high-level statements, rather than how a paper builds and communicates knowledge. SUDO models that discourse from the bottom up, preserving local context and source grounding while connecting papers.

1. Introduction

  • Concept-centric graphs can represent a term such as “attention” uniformly, leaving unclear whether a paper uses it as an approach, motivation, or baseline.This obscures how concepts are applied, adapted, or interpreted in particular papers.Source: §1 · p. 1
  • Statement-centric graphs capture research assertions but may omit methodological detail and cross-paper discourse, while requiring abstractions not directly stated in a single passage.The resulting descriptions can be weakly grounded in identifiable source text.Source: §1 · pp. 1–2
  • SUDO models scholarly discourse from linguistic structure, connecting concepts and propositions within a paper and recurring concepts across papers.Its principles are intended to apply across domains and tasks.Source: §1 · p. 2
  • The paper evaluates SUDO on scholarly knowledge-modeling use cases and a downstream LLM-based scientific question-answering task.The evaluation uses an ISWC main-track graph covering publications from 2012 to 2025.Source: §1 · p. 2

2. Related Work

Scholarly graphs have shifted from large-scale metadata and concept networks toward statement-level representations, but each retains only some publication knowledge. SUDO draws on linguistic theories to model papers as structured contributions to a broader scholarly dialogue.

2.1. The Data-Source Paradigm and Information Loss

  • Web-scale scholarly infrastructures support discovery and analysis, while concept-centric graphs link recurring concepts but leave their roles within individual papers underspecified.That limits searches and analyses requiring publication-specific meaning and argumentative context.Source: §2.1 · pp. 2–3
  • Statement-centric graphs retain more source meaning but often cover selected dimensions, isolate statements, and face challenges in faithful, traceable construction.These limitations constrain their use in question answering, knowledge discovery, and literature-review workflows.Source: §2.1 · p. 3

2.2. Speech Acts and the Positioning of SUDO

  • SUDO applies linguistic foundations to represent whole publications and treats the literature as a longitudinal, multi-party dialogue in which papers respond to, build on, or challenge prior work.Citations and shared conceptual vocabularies connect papers within that dialogue.Source: §2.2 · p. 3

3. Conceptual semantic view of scholarly knowledge

The paper treats linguistic structure as the basis for understanding a publication’s meaning and its relationships to other works. Concepts take their roles from their participation in clauses and the broader discourse, while recurring concepts connect papers.

3. Conceptual semantic view of scholarly knowledge

  • In the Transformer example, one sentence proposes a new architecture based on attention, while a preceding sentence gives background about attention.Readers distinguish these roles from linguistic cues in phrases, clauses, word choices, and relations between sentences.Source: §3 · p. 3, Figure 1 · p. 4
  • A concept’s role depends on its clause and that clause’s contribution to the paper; recurring concepts such as Transformer connect local discourse across papers.In the BERT example, Transformer describes BERT’s architectural basis rather than serving as the principal contribution being introduced.Source: §3 · p. 4
  • Conceptual semantics separates surface language from structured meaning, while distinguishing propositional content from communicative information such as topic and focus.SUDO adapts this architecture by recovering scholarly objects and propositions through lexical and clausal analysis, then assigning scholarly functions and relations through discourse analysis.Source: §3 · p. 4

4. SUDO

SUDO represents a publication through linked information and propositional structures, alongside metadata, document structure, and provenance. This design retains source-specific context while connecting recurring concepts and their surrounding discourse across publications.

4.1. Overview

  • SUDO combines information and propositional structures with metadata and document structure; Arguments capture research arguments, while Descriptors elaborate or contextualize concepts.The structures are intended to form a complete conceptual-semantic representation of a publication.Source: §4.1 · pp. 5–6, Figure 2 · p. 5

4.2. Information structure

  • The information structure represents noun-phrase expressions as source-located Artifacts and maps their canonical identities to Concepts for alignment across publications.SUDO uses broad categories—Task, Method, Metric, Material, and Term—and general relations such as used-for, part-of, and compare.Source: §4.2 · p. 6
  • Separate Artifact occurrences preserve lexical and publication context even when closeMatch aligns them to a shared canonical Concept.In the example, Transformer occurrences in the Transformer and BERT papers retain distinct roles in their respective Approach propositions.Source: §4.2 · p. 6

4.3. Propositional Structure

  • Source-clause-anchored Propositions link to participating Artifacts, separating argumentative acts from propositions that provide semantic detail.This preserves both the research argument and the source-grounded information from which it is built.Source: §4.3 · p. 6
  • The argumentative backbone links Issues, Ideas or Approaches, Evidence, Warrants, and Claims through relations including motivates, supports, and responds-to.The roles draw on Toulmin’s model for Evidence, Warrant, and Claim, complemented by Issue, Idea, and Approach roles.Source: §4.3 · pp. 6–7
  • Descriptors attach semantic detail to Artifacts or Propositions using eleven binary relations drawn from Rhetorical Structure Theory.Later Transformer-method clauses can elaborate the Approach without becoming separate contributions, while remaining linked to their source clauses.Source: §4.3 · p. 7

4.4. Conceptual Semantic View of Global Scholarly Discourse

  • SUDO’s global discourse aligns recurring Artifacts through canonical Concepts while retaining distinct source-specific Artifacts and their proposition structures.Source Propositions also preserve nuanced relations that a fixed concept-predicate vocabulary might omit or generalize.Source: §4.4 · pp. 7–8

4.5. Implementation

  • SUDO version 3.5.2 has undergone three major revisions across 19 update cycles and is published in multiple RDF and ontology formats.The ISWC graph covers 898 main-track papers from 2012–2025; a separate graph covers 413 QASPER publications.Source: §4.5 · p. 8, §4.6 · p. 8

5. Evaluation

The evaluation tests SUDO’s structural coverage across scholarly knowledge-modeling use cases and its utility for LLM-based scientific question answering. It reports broader coverage than the evaluated concept- and statement-centric models and improved QASPER retrieval performance.

5. Evaluation

  • The evaluation combines use-case-based coverage analysis with downstream LLM-assisted scientific question answering to assess SUDO’s representation and practical utility.It asks whether SUDO represents scholarly knowledge throughout a publication and whether its discourse-aware representation improves information access.Source: §5 · pp. 8–9

5.1.2. Results and Discussion

  • SUDO achieves the highest, most balanced coverage, with most use cases at ACL 2 or 3 and substantial support across all five families.CS-KG has direct coverage for some concept relations but limited coverage of document structure, argument, and pragmatic descriptions; ORKG and NLP-AKG cover some statement-oriented cases but decline on other requirements.Source: Figure 4 · p. 9, §5.1.2 · p. 10

5.1. RQ1: Structural Representation Coverage

  • The ontology-independent collection contains 30 use cases across five recurring families, enabling comparable information requirements to be operationalized as model-specific competency questions.Each family contains six use cases; use-case satisfiability averages Answer Capability Levels across its competency questions, and family coverage macro-averages the six use-case scores.Source: §5.1 · pp. 9–10

5.1.1. Experimental Setup

  • The RQ1 comparison evaluates CS-KG, ORKG, and NLP-AKG using model-specific competency-question sets and the ACL procedure.NLP-AKG’s graph was generated by applying its implementation to the ISWC publication collection because no comparable dump was available; CS-KG and ORKG questions used public graph dumps.Source: §5.1.1 · p. 10

5.2.1. Results and Discussion

  • Query-guided retrieval reaches an F1 score of 82.44, with BERTScore precision 82.57 and recall 82.39, improving by about 10 percentage points over NLP-AKG.Embedding-based retrieval achieves an F1 score of 78.16, exceeding the optimized NLP-AKG agent by approximately six percentage points.Source: Table 2 · p. 11, §5.2.1 · p. 11

5.2. RQ2: Utility for Scholarly Downstream Tasks

  • The RQ2 experiment evaluates BERTScore precision, recall, and F1 on 600 QASPER question–answer pairs spanning 413 publications.The questions are document-level and vary in complexity; SUDO graphs for these publications use the same neuro-symbolic pipeline as the ISWC collection.Source: §5.2 · p. 11

5.2.1. Results and Discussion

  • Comparable scores could not be established for CS-KG or ORKG under the same QASPER protocol.CS-KG yielded too few reliable question-to-concept alignments, while the available ORKG graph covered only 34 of 413 QASPER publications.Source: §5.2.1 · p. 11

6. Future Work

Future work proposes SUDO as an intermediary between publication text and specialized SKGs, preserving source grounding through derived representations. The authors also plan to investigate downstream applications and evaluate extensions on larger scholarly collections.

6. Future Work

  • A construction pipeline could represent publications through SUDO first, then derive domain- or task-specific representations from that shared layer.The derived knowledge would remain connected to SUDO Artifacts, Propositions, source locations, and provenance.Source: §6 · p. 12
  • Future work will implement and evaluate ontology alignments, identity mappings, and transformation pipelines for concept- and statement-centric views.Concept-centric models could align entities with SUDO Concepts and inherit publication-local roles; statement-centric models could build claims over propositions with traceable source connections.Source: §6 · p. 12
  • The authors plan to investigate domain- and task-specific extensions for evidence tracing, claim verification, cross-publication analysis, and LLM-assisted scholarly workflows.They also plan to evaluate construction and use across larger scholarly collections.Source: §6 · p. 12

7. Conclusion

The conclusion presents SUDO as a domain-agnostic, discourse-aware ontology that represents scholarly knowledge while preserving how it is communicated within publications and connected across literature. Evaluations report broader coverage than examined concept- and statement-centric SKGs and an advantage in focused information retrieval for LLM-based scientific QA.

7. Conclusion

  • SUDO represents scholarly knowledge through complementary information and propositional structures grounded in conceptual semantics.It connects concepts, source-grounded propositions, research arguments, descriptive context, document structure, and provenance.Source: §7 · p. 12
  • SUDO preserves how knowledge is communicated within a publication while connecting recurring concepts across literature into global scholarly discourse structure.Source: §7 · p. 12
  • The evaluation reports broader representational coverage than examined concept- and statement-centric SKGs and a measurable retrieval advantage in an LLM-based scientific QA workflow.The conclusion identifies a shared foundation for specialized SKG extensions, cross-publication analysis, and future web-scale scholarly AI workflows.Source: §7 · p. 12

Declaration on Generative AI

The authors report using GPT-5.5 and Gemini for grammar and spelling checks, then reviewing and editing the content and taking full responsibility for the publication.

Declaration on Generative AI

A.1. Scholarly Publication Knowledge-Representation Use Cases

The use cases span cross-publication concept dynamics and paper-local pragmatic descriptions. They ask how concepts recur and change across papers, and how individual publications define, deploy, and evaluate them.

A.1. Scholarly Publication Knowledge-Representation Use Cases

  • Cross-paper cooccurrence, role shifts, and definitional variance ask how canonical concepts recur and differ across publications.The use cases also examine dialectical claims, inter-paper influence, and concepts’ evolutionary trajectories.Source: §A.1 · p. 17
  • Paper-local questions cover a concept’s intensional definition, functional intent, and operationalization parameters within a publication.Operationalization asks how a concept is configured or structured in the publication’s workflow.Source: §A.1 · p. 17
  • Other paper-local questions examine attributed affordances, boundary conditions, and pragmatic contrasts with alternatives.These concern stated capabilities, operating constraints or assumptions, and comparative utility or suitability.Source: §A.1 · p. 17

A.2. Online Resources

The appendix lists SUDO resources, its concept-structure classes and relations, and vocabularies used across representational layers.

A.2. Online Resources

  • SUDO’s ontology, documentation, ISWC corpus and graph, and QASPER graph are listed as online resources.The passages provide links for each resource.Source: §A.2 · p. 17

A.3. Concept-Structure Classes and Relations

  • SUDO uses five domain-agnostic Artifact subclasses—Task, Method, Metric, Material, and Term—to organize scholarly nomenclature.These broad semantic categories adapt the cross-domain SciERC NER schema.Source: §A.3 · p. 17
  • Artifact relations capture functional, taxonomic, and compositional associations, while SKOS properties align canonical concepts with instances and lexical variants.Table 5 summarizes the relations used in the concept structure.Source: §A.3 · p. 17, Table 5 · p. 18

A.4. Ontology Reuse and Vocabulary

  • SUDO reuses vocabularies for publication metadata, document structure, knowledge organization, and argumentation across its representational layers.Tables 6 and 7 list the reused ontologies and the classes, datatypes, and properties used by layer.Source: §A.4 · p. 17, Table 6 · p. 18, Table 7 · p. 19
Loading 2610.33018v1…