PaperlayerView exact revision on arXiv

Source-linked AI summary

SciAtlas: A Large-Scale Knowledge Graph for Automated Scientific Research

Shuofei Qiao, Yunxiang Wei, Jiazheng Fan, Bin Wu, Busheng Zhang, Mengru Wang, Yuqi Zhu, Ningyu Zhang, Keyan Ding, Qiang Zhang, Huajun Chen

arXiv:2605.22878v1cs.AIcs.CLcs.IRcs.LG

TL;DR

Fragmented academic knowledge limits interdisciplinary retrieval and topological reasoning. SciAtlas builds a large multidisciplinary knowledge graph with neuro-symbolic retrieval, and reports a resource spanning 109.70 million authors, 3.76 million keywords, and 0.12 million institutions.

  • Problem

    Fragmented, unstructured academic knowledge obscures entity relationships and limits interdisciplinary integration and global topological understanding.

  • Method

    SciAtlas combines a heterogeneous academic knowledge graph with tri-path neuro-symbolic recall and graph reranking for deterministic topological retrieval.

  • Results

    109.70 million authors, 3.76 million keywords, and 0.12 million institutions are connected across 11 relationship types in SciAtlas.

  • Takeaways & Limitations

    SciAtlas supports automated research trend synthesis, idea positioning, and academic trajectory exploration as structured scientific-search applications.

  • Takeaways & Limitations

    SciAtlas primarily represents paper-centric entities and presents downstream tasks as running examples because many scientific tasks lack high-quality benchmarks and definitive verifiers.

Abstract

from arXiv · show

The exponential growth of global academic output has confronted researchers and AI agents with an unprecedented ``information explosion,'' where fragmented and unstructured knowledge organization impedes deep interdisciplinary integration. Current academic retrieval tools predominantly rely on superficial keyword matching or vector-space semantic retrieval, which lack the topological reasoning capabilities required to navigate complex logical connections. Agentic deep-research-based frameworks are often prone to logical hallucinations and consuming high inference costs. To bridge this gap, in this report, we introduce SciAtlas, a large-scale, multi-disciplinary, heterogeneous academic resource knowledge graph designed as a panoramic scientific evolution network. By integrating over 43M papers from 26 disciplines, and a total of 157M entities and 3B triplets, SciAtlas provides a structured topological cognitive substrate that dismantles disciplinary barriers and furnishes AI agents with a global perspective. Furthermore, we develop a neuro-symbolic retrieval algorithm featuring tri-path collaborative recall and graph reranking, achieving a seamless transition from simple semantic matching to deterministic association discovery. We also present key application directions of SciAtlas, including literature review, automated research trend synthesis, idea positioning, and academic trajectory exploration, to demonstrate that SciAtlas can serve as an effective ``cognitive map'' to empower the full loop of automated scientific research while significantly reducing reasoning costs. We have released the interfaces for KG retrieval and various downstream tasks in our GitHub repo.

1. Introduction

SciAtlas addresses fragmented academic knowledge and retrieval methods that lack topological reasoning by organizing research resources into a structured knowledge graph. Its neuro-symbolic retrieval and proposed applications aim to support automated scientific research with a global cognitive perspective.

  • Motivation: Academic resources remain scattered in unstructured text, creating knowledge islands that impede interdisciplinary integration and hide logical connections between entities.These barriers affect both novice researchers and AI agents.
  • Motivation: Existing retrieval tools rely mainly on keyword matching or vector-space semantic retrieval, while agentic deep-research frameworks remain vulnerable to logical hallucinations and high inference costs.These approaches use flattened feature comparisons rather than genuine topological reasoning.
  • SciAtlas: SciAtlas organizes fragmented academic resources into a structured logical topology and provides AI agents with a panoramic scientific network and global cognitive perspective.The graph is designed as a topological cognitive substrate for accelerating scientific discovery.
  • SciAtlas: 9 entity-node categories and 12 relational-edge categories give SciAtlas a heterogeneous schema with attributes and links spanning papers, authors, institutions, keywords, and research fields.Examples include paper abstracts and PDF URLs, author citations, and citation relations.
  • Retrieval and applications: SciAtlas’s neuro-symbolic retrieval combines lexical matching, vector retrieval, and graph propagation through tri-path collaborative recall and graph reranking.The mechanism fuses paper semantic relevance, graph topological support, and global-citation-based importance metrics.
  • Retrieval and applications: Proposed applications include research trend synthesis, idea positioning, and academic trajectory exploration, positioning SciAtlas as a cognitive map for automated scientific research.The applications also target highly relevant academic-author retrieval and support the full research loop.

2. SciAtlas

SciAtlas is a heterogeneous academic knowledge graph centered on papers and organized across semantic, conceptual, disciplinary, and social levels. It spans 26 disciplines and 43.30 million papers, with large-scale entity coverage and billions of relational edges supporting retrieval and reasoning.

  • Schema: SciAtlas centers academic literature on Paper entities and organizes related authors, institutions, keywords, sources, topics, fields, subfields, and domains.These hybrid entities organize papers across four levels.
  • Schema: Citation and relevance edges connect papers semantically, while keyword co-occurrence links them conceptually.The semantic level uses CITES and RELATED_TO; the conceptual level uses salient keywords and COOCCUR relationships.
  • Schema: Disciplinary hierarchies organize papers by domains, fields, subfields, and topics, while coauthor, authorship, and affiliation edges create social links.Together, these structures provide a foundation for deep retrieval and reasoning.
  • Statistics: 26 disciplines and 43.30 million papers comprise SciAtlas, with Medicine contributing 18.56% and the five largest disciplines together contributing 51.43%.SciAtlas also contains 109.70 million authors, 3.76 million keywords, and 0.12 million institutions across 11 relationship types.
  • Construction: OpenAlex supplies the scholarly data, which SciAtlas restructures, normalizes, deduplicates, filters, and connects through entity identifiers.The pipeline filters non-English papers and papers with very short abstracts to ensure high-quality data.
  • Construction: SciAtlas adds dense paper-level keywords and pre-computed semantic vectors from titles, abstracts, and keywords to support hybrid, efficient retrieval.Embeddings use bge-large-en-v1.5, and the completed graph is deployed with Neo4j.

3. Neuro-Symbolic Retrieval

SciAtlas’s neuro-symbolic retrieval maps diverse query formats to candidate KG nodes through keyword, semantic, and title pathways, then propagates and reranks them using graph topology, citation importance, and transparent path-based explanations. The system returns the top-20 papers and completes retrieval within 2 minutes.

  • Query-to-node recall: The system accepts keywords, scientific questions, abstracts, idea texts, and complete papers, mapping each query into KG nodes through three distinct pathways.The pathways are keyword matching, semantic matching, and title matching.
  • Query-to-node recall: Keyword matching combines LLM-extracted importance scores with exact and vector matching, retaining vector matches above the default 0.7 threshold and at most top-3 nodes per keyword.When multiple matches occur, each node receives the maximum score across matching routes.
  • Query-to-node recall: Semantic matching retrieves top-60 papers by title and abstract embeddings, reranks them with bge-reranker-large, and retains top-15 results for each representation.For whole-paper queries, only the abstract is embedded.
  • Query-to-node recall: Title matching extracts and normalizes the top-10 LLM-ranked titles, assigns exact matches a score of 1.0, and discards fuzzy matches below the default 0.88 threshold.Fuzzy similarity weights longest-common-subsequence similarity at 0.65 and token overlap at 0.35.
  • Graph propagation and reranking: 2-hop graph propagation treats edges as undirected, limits each entity type to 500 nodes per hop, and uses weighted random walks to deepen topological reasoning from seed nodes.Paper importance can use citation counts or be set to 1 for relevance-only tasks; propagation uses a restart probability and stops at ε = 10^-6 or T_max = 50.
  • Graph propagation and reranking: The final reranking combines initial relevance, graph support, citation importance, and title bonuses, returns top-20 papers with path explanations, and completes within 2 minutes.The default weights are λ_pre = 0.35, λ_graph = 0.45, and λ_imp = 0.20; graph expansion can add papers absent from the initial seed set.

4. Downstream Application of SciAtlas

SciAtlas supports downstream scientific research through customizable literature retrieval, idea grounding and evaluation, exploratory idea generation, trend prediction, author retrieval, and academic trajectory summarization. These applications use graph-based evidence collection and LLM analysis to organize literature, assess ideas, synthesize interdisciplinary concepts, identify influential developments, and profile researchers.

  • Literature Review: Literature review uses customizable retrieval to collect papers for a research direction and support automated review synthesis.The retrieved paper collection can be adapted to various LLM-based automated literature review methods.
  • Idea Grounding and Evaluation: Idea grounding retrieves relevant papers and paragraphs, extracts claims across motivation, methodology, and experimental design, and compares similarities and differences.Grounded evidence can support novelty, feasibility, and soundness evaluation, with criteria adjustable to downstream needs.
  • Idea Generation: Exploratory KG search can collect literature, identify gaps, synthesize concepts across domains, and generate interdisciplinary ideas by relaxing distant-node constraints.The example proposes federated and privacy-preserving knowledge editing by combining federated-learning constraints with knowledge editing.
  • Research Trend Prediction: Research trend prediction emphasizes paper influence by increasing citation weights during search to characterize the development of a research direction.The example organizes biologically plausible learning in spiking neural networks into successive developmental stages from 2006 onward.
  • Academic Trajectory Exploration: Author retrieval can rank and filter relevant researchers using citation counts and author-order-weighted AUTHORED edges, while trajectory summaries cluster work by research direction.The generated profile describes a trajectory spanning Natural Language Processing, Artificial Intelligence, Large Language Models, reasoning, planning, agentic AI, interpretability, and safety.

5. Limitations and Future Work

SciAtlas is under continuous maintenance, with future work focused on improving accessibility, expanding the knowledge represented, strengthening evaluation, and enabling more timely updates. These directions aim to facilitate automated scientific discovery.

  • CLI and Skills: CLI and Skills will encapsulate knowledge-graph retrieval and invocation functions to reduce the need for users to write Neo4j queries.The KG is currently accessed primarily through Neo4j, despite available usage guidelines.
  • Integrating More Knowledge Forms: More knowledge forms will extend the paper-centric graph to atomic knowledge, theorems and standards, experimental experiences, datasets, and code.Future work will acquire these forms and establish their associations with papers.
  • Benchmark and Evaluation: Benchmark and evaluation work will address the lack of high-quality benchmarks, faithful real-world research simulations, and definitive verifiers for long-form outputs.Many scientific tasks involve long-form responses whose evaluation is currently ambiguous.
  • Dynamic Update: Dynamic-update work will systematize real-time strategies to support daily knowledge-graph updates instead of relying primarily on periodic manual execution of fixed scripts.User-initiated updates are supported, but automated real-time updates are considered essential for the evolving knowledge landscape.

6. Related Work

Related work frames LLMs as central to automated scientific discovery, whose workflow spans literature reviewing, method implementation, and manuscript writing. Scientific retrieval has traditionally used academic platforms and keyword- or vector-based search over local paper collections.

  • Automated Scientific Discovery: LLMs now occupy a central role in automated scientific discovery, whose workflow includes literature reviewing, method implementation, and manuscript writing.Literature reviewing consolidates papers, method implementation turns hypotheses into code and evaluates them experimentally, and manuscript writing documents research rationale, methods, and results.
  • Scientific Retrieval: Human scientists retrieve research through general-purpose academic search engines, domain-specific preprint servers, and official publisher platforms.Examples include Google Scholar, Semantic Scholar, arXiv, ChemRxiv, PubMed, and publisher platforms for journals and conferences.
  • Scientific Retrieval: Early automated scientific research systems primarily relied on keyword- or vector-based retrieval within local paper collections.This approach is described as the initial retrieval paradigm in automated scientific research.

7. Conclusion

SciAtlas is introduced as a large-scale, multidisciplinary, heterogeneous academic knowledge graph designed as a panoramic scientific evolution network. It integrates diverse entities and relations across more than 43M papers and develops a neuro-symbolic retrieval algorithm featuring tri-path collaboration.

  • Conclusion: SciAtlas is a large-scale, multidisciplinary, heterogeneous academic knowledge graph designed as a panoramic scientific evolution network.The report presents SciAtlas as its central resource for representing scientific knowledge.
  • Conclusion: Over 43M papers are integrated with 9 categories of entity nodes and 12 categories of relational edges.This structure is described as a topological cognitive substrate that dismantles disciplinary barriers and gives AI agents a global perspective.
  • Conclusion: SciAtlas includes a neuro-symbolic retrieval algorithm featuring tri-path collaboration.The passage identifies this retrieval algorithm as a further contribution of the report.

A. Full Schema of SciAtlas

SciAtlas defines a Neo4j schema covering node types, their attributes, relationship types, and both standard and vector indexes.

  • Node types and attributes: The Neo4j schema specifies node types and their attributes, including fields for keywords, papers, sources, and topics.Listed attributes include normalized keyword text, frequency, embeddings, paper metadata, source identifiers, and topic update dates.
  • Relationship types: The schema defines relationship types connecting entities in the knowledge graph.These relationship definitions are documented separately from node and attribute specifications.
  • Indexes: The schema includes standard indexes and vector indexes to support Neo4j data access and embedding-based retrieval.The two index categories are presented in separate schema tables.

B. Prompts Used in this Report

The report defines prompts for generating retrieval queries, grounding evidence against research ideas, proposing novel ideas, predicting trends, and profiling authors. Each task specifies structured inputs, strict JSON outputs, and task-specific fields or judgments.

  • B.2.1. Idea Grounding – Query Generation: Idea Grounding query generation creates dense-retrieval queries from structured scientific extraction results and returns strict JSON.The system prompt frames the task as scientific grounding and retrieval-oriented paragraph search query generation.
  • B.2.1. Idea Grounding – Query Generation: Queries use only motivation and method sections, select retrieval-useful items, and produce at most max_queries across both sections.Outputs identify each source section and sentence alongside a retrieval-oriented rewrite optimized for paragraph retrieval.
  • B.2.2. Idea Grounding – Grounding: Idea Grounding evaluates how retrieved evidence aligns with a research idea unit using the paragraph as primary evidence and context only for clarification.The prompt requires evidence-grounded judgments without invented claims and explicit acknowledgment of weak relationships.
  • B.2.2. Idea Grounding – Grounding: The grounding output records status, focus aspect, evidence spans, shared and different points, coverage label, and match rationale.Coverage labels range from high and partial to limited and none, corresponding to increasingly weak or absent alignment.
  • B.2.3. Idea Generation: Idea Generation proposes novel research ideas from papers by extending, combining, or contrasting existing work in strict JSON.Each idea includes a title, description, novelty, significance, and key references, with exactly {idea_count} ideas requested.
  • B.2.4. Research Trend Predicting: Research Trend Predicting summarizes chronologically ordered papers using trend, stage, methodological-shift, emerging-topic, gap, direction, and representative-paper fields.Representative papers include titles, years, and explanations of why they represent the topic’s evolution.
  • B.2.5. Author Research Profile: Author Research Profile summarizes a researcher’s publication-based trajectory through an overall profile, directions, technical arsenal, and representative papers.Direction entries include themes, active years, and descriptions; representative papers include titles, years, and selection rationales.
Loading 2605.22878v1…