SciAtlas: A Cognitive Map of 43 Million Papers
Background: The Knowledge-Island Problem
Global academic output exceeds 480 million publications in OpenAlex and grows by millions per year. Existing retrieval tools have two main limitations: keyword matching misses semantic links such as "protein structure prediction" → AlphaFold, while vector similarity search cannot answer topological questions like "which paper's method inspired which improvement." Subject silos further block cross-disciplinary discovery. Agentic LLM-based research frameworks help but incur huge API costs and hallucination risk from long reasoning chains. SciAtlas proposes a deterministic, graph-based alternative: structured knowledge topology over fuzzy semantic similarity, and graph propagation over probabilistic model inference.
Core Design: A Four-Layer Cognitive Architecture
SciAtlas integrates 43 million papers into 157 million entities and 3 billion edges, currently the largest open academic knowledge graph in coverage and breadth.
Nine entity types:
| Entity | Count | Role | |--------|-------|------| | Paper | 43.3M | Core node with title, abstract, DOI, citations | | Author | 109.7M | Social network layer | | Keyword | 3.76M | Concept layer, LLM-extracted reusable phrases | | Topic / Subfield / Field / Domain | 4,520 / 252 / 26 / 4 | Direction layer, four-level hierarchy | | Institution | 120K | Social network layer | | Source | 280K | Publication outlets |
Twelve relation types span five layers: semantic (CITES, RELATED_TO), concept (HAS_KEYWORD, COOCCUR), direction (DOMAIN_OF, FIELD_OF, SUBFIELD_OF), social (AUTHORED, COAUTHOR, AFFILIATED_WITH), and publication (PUBLISHED_IN). CITES totals 214M edges; COAUTHOR reaches 2.06B edges, the densest relation type.
The four-layer design is task-configurable: literature review emphasizes CITES and RELATED_TO; collaboration mining relies on COAUTHOR and AFFILIATED_WITH; trend prediction combines FIELD_OF and COOCCUR.
Data Construction: From 480M Papers to 3B Edges
Source data is OpenAlex with 480M papers. Four filtering decisions shape the final corpus:
- Language filter: English-only papers, reducing multilingual complexity but excluding non-English scholarship, with potential bias in medicine, social sciences, and humanities.
- Abstract-length filter: Papers with too-short abstracts are dropped, because abstract quality directly affects LLM keyword extraction.
- PDF-availability filter: Papers without PDF URLs are removed, ensuring traceability to full text.
- No author deduplication: Name ambiguity is prevalent; merge errors (combining distinct people) are considered worse than retaining duplicates. This is acute for East Asian names like "Wang Wei."
- Literature review: Configurable weights on venue, AUTHORED, and AFFILIATED_WITH let the same algorithm serve quick scouting or systematic review.
- Idea Grounding: A research idea is used as a query, retrieved papers are segmented, the LLM extracts fine-grained sub-queries (motivation, methodology, experiments), segments are retrieved, and the LLM compares the idea to prior work to assess novelty, supporting evidence, or true innovation.
- Other applications: Trend prediction, collaboration mining, and academic trajectory exploration all rely on transforming textual similarity into relational proximity.
- No quantitative benchmarks: Full evaluation chapters with baselines, hit-rate, and NDCG metrics were not accessible, leaving retrieval quality unverified externally.
- English-only bias: Coverage gaps in medicine, social sciences, and humanities.
- Author duplication: 109.7M author nodes with unresolved name collisions.
- Limited propagation interpretability: Random Walk with Restart is opaque inside the 50 iterations; final scores are transparent but intermediate paths are not.
- LLM dependency for keywords: Quality cascades downstream; weak LLM domain understanding degrades retrieval.
- SciAtlas technical report: arXiv:2605.22878
- Project page: http://scigraph.openkg.cn/
- OpenAlex open academic data platform
- Neo4j graph database
- bge-large-en-v1.5 embedding model
- Qwen3-30B-A3B-Instruct-2507 keyword extraction LLM
The central data innovation is keyword extraction with Qwen3-30B-A3B-Instruct-2507. OpenAlex's original Concept layer has only 65K entries—too sparse for fine-grained inference. Each paper yields 3–8 reusable keywords, avoiding paper-specific jargon, system names, or marketing language, and preferring cross-paper phrases (e.g., "protein structure prediction" stays; "hierarchical dual-path adaptive learning framework" is filtered). Each keyword carries an importance score, used as an edge weight for HAS_KEYWORD. Co-occurrence within a paper creates COOCCUR edges weighted by frequency.
Semantic embeddings are precomputed for titles, abstracts, and keywords using bge-large-en-v1.5 and stored as node attributes. The graph is hosted on Neo4j with Cypher query support. Updates are daily via OpenAlex API and bimonthly via change files; missing papers can be supplemented with GROBID-extracted metadata.
Neural-Symbolic Three-Path Retrieval
The core technical contribution combines three parallel recall paths, graph propagation reranking, and a fused final ranking.
Path 1 — Keyword Matching: An LLM extracts query keywords with normalized importance scores. Two parallel channels run: exact text match (full score) and vector match (threshold θ_kw=0.7, top-3, decayed score). Output is a weighted seed set K_seed.
Path 2 — Semantic Matching: The query is encoded as embedding e_q (abstract used if full paper). Title-embedding channel retrieves top-60, abstract-embedding channel retrieves top-60; both are reranked by bge-reranker-large to top-15. Final score combines title (0.4) and abstract (0.6) via weighted average.
Path 3 — Title Matching: Activated when the query contains titles (e.g., a reference paper and its reference list). GROBID extracts titles, an LLM assigns confidence, top-10 are kept, then exact (1.0) or fuzzy (0.65×LCS + 0.35×Jaccard, threshold 0.88) matching is applied. This enables citation tracing and variant detection.
Node merging applies a title-match bonus: +0.35 for exact, +0.10 for fuzzy. Paper seed weights incorporate citation importance: w_p^seed = s_p^pre × (1 + γ × imp(p)), where imp(p) is a log-scaled citation count.
Graph propagation: A 2-hop subgraph is constrained (≤500 nodes per entity type per hop), then Random Walk with Restart runs. Transition probabilities normalize by edge weights reflecting signal strength: CITES 1.00, RELATED_TO 0.90, AUTHORED 0.80, COAUTHOR and COOCCUR 0.60. The update rule is r_v^(t+1) = α·s_v + (1−α)·Σ_u r_u^(t)·P(v|u), with α as restart probability. Convergence uses L1 norm < ε=10⁻⁶ or max 50 iterations.
This is the key topological reasoning mechanism: a paper sharing no keywords or direct similarity with a query can still surface because it is cited by multiple high-weight seed papers.
Final ranking: A graph-support gating factor g_p = max(0.25, s̃_p^pre) prevents graph scores from overriding initial relevance. Final score: s_p^final = min(1, 0.35·s̃_p^pre + 0.45·s̃_p^graph·g_p + 0.20·imp^final(p)). The weights encode the design philosophy that relational topology is more reliable than isolated semantic features. Each result includes a full score breakdown across all paths and a path-based explanation, supporting transparency. End-to-end latency is reported as significantly under two minutes.
Downstream Applications
Technical Assessment
Strengths: Scale (157M entities, 3B edges) leads open academic knowledge graphs. The OpenAlex-to-43M filtering pipeline balances quality and coverage. Qwen3-30B-A3B offers a pragmatic cost-quality tradeoff for keyword extraction.
Open tensions:
Strategic Significance
SciAtlas signals a shift from academic metadata management to cognitive topology construction. It provides a structured, traceable substrate for AI-for-Science workflows: LLMs handle high-level reasoning and hypothesis generation, while SciAtlas handles factual retrieval and relational verification. Coverage of 26 first-level domains plus keyword co-occurrence enables cross-disciplinary discovery—for example, a materials scientist querying "graph neural network" can find GNN applications in molecular representation through COOCCUR and citation chains, which domain-specific vector stores cannot match.
SciAtlas does not deliver fully autonomous science; it solves the information retrieval and association discovery stages, while hypothesis generation, experiment design, and analysis still require humans or stronger AI. Its role is a navigable cognitive map in the information explosion, not an autonomous driver.